Biomedical Engineering and Science [1 ed.] 9781683921868

Citation preview

PROCEEDINGS OF THE 2017 INTERNATIONAL CONFERENCE ON BIOMEDICAL ENGINEERING AND SCIENCES

Editors Hamid R. Arabnia Leonidas Deligiannidis Fernando G. Tinetti

CSCE’17 July 17-20, 2017 Las Vegas Nevada, USA americancse.org ©

CSREA Press

This volume contains papers presented at The 2017 International Conference on Biomedical Engineering and Sciences (BIOENG'17). Their inclusion in this publication does not necessarily constitute endorsements by editors or by the publisher.

Copyright and Reprint Permission Copying without a fee is permitted provided that the copies are not made or distributed for direct commercial advantage, and credit to source is given. Abstracting is permitted with credit to the source. Please contact the publisher for other copying, reprint, or republication permission.

© Copyright 2017 CSREA Press ISBN: 1-60132-451-0 Printed in the United States of America

Foreword It gives us great pleasure to introduce this collection of papers to be presented at the 2017 International Conference on Biomedical Engineering and Sciences (BIOENG’17), July 17-20, 2017, at Monte Carlo Resort, Las Vegas, USA. An important mission of the World Congress in Computer Science, Computer Engineering, and Applied Computing, CSCE (a federated congress to which this conference is affiliated with) includes "Providing a unique platform for a diverse community of constituents composed of scholars, researchers, developers, educators, and practitioners. The Congress makes concerted effort to reach out to participants affiliated with diverse entities (such as: universities, institutions, corporations, government agencies, and research centers/labs) from all over the world. The congress also attempts to connect participants from institutions that have teaching as their main mission with those who are affiliated with institutions that have research as their main mission. The congress uses a quota system to achieve its institution and geography diversity objectives." By any definition of diversity, this congress is among the most diverse scientific meeting in USA. We are proud to report that this federated congress has authors and participants from 64 different nations representing variety of personal and scientific experiences that arise from differences in culture and values. As can be seen (see below), the program committee of this conference as well as the program committee of all other tracks of the federated congress are as diverse as its authors and participants. The program committee would like to thank all those who submitted papers for consideration. About 65% of the submissions were from outside the United States. Each submitted paper was peer-reviewed by two experts in the field for originality, significance, clarity, impact, and soundness. In cases of contradictory recommendations, a member of the conference program committee was charged to make the final decision; often, this involved seeking help from additional referees. In addition, papers whose authors included a member of the conference program committee were evaluated using the double-blinded review process. One exception to the above evaluation process was for papers that were submitted directly to chairs/organizers of pre-approved sessions/workshops; in these cases, the chairs/organizers were responsible for the evaluation of such submissions. The overall paper acceptance rate for regular papers was 27%; 18% of the remaining papers were accepted as poster papers (at the time of this writing, we had not yet received the acceptance rate for a couple of individual tracks.) We are very grateful to the many colleagues who offered their services in organizing the conference. In particular, we would like to thank the members of Program Committee of BIOENG’17, members of the congress Steering Committee, and members of the committees of federated congress tracks that have topics within the scope of BIOENG. Many individuals listed below, will be requested after the conference to provide their expertise and services for selecting papers for publication (extended versions) in journal special issues as well as for publication in a set of research books (to be prepared for publishers including: Springer, Elsevier, BMC journals, and others). • • • • • •

Prof. Abbas M. Al-Bakry (Congress Steering Committee); University President, University of IT and Communications, Baghdad, Iraq Prof. Nizar Al-Holou (Congress Steering Committee); Professor and Chair, ECE Department; Vice Chair, IEEE/SEM-Computer Chapter; University of Detroit Mercy, Detroit, Michigan, USA Prof. Hamid R. Arabnia (Congress Steering Committee); Graduate Program Director (PhD, MS, MAMS); The University of Georgia, USA; Editor-in-Chief, Journal of Supercomputing (Springer); Fellow, Center of Excellence in Terrorism, Resilience, Intelligence & Organized Crime Research (CENTRIC). Prof. Dr. Juan-Vicente Capella-Hernandez; Universitat Politecnica de Valencia (UPV), Department of Computer Engineering (DISCA), Valencia, Spain Prof. Kevin Daimi (Congress Steering Committee); Director, Computer Science and Software Engineering Programs, Department of Mathematics, Computer Science and Software Engineering, University of Detroit Mercy, Detroit, Michigan, USA Prof. Leonidas Deligiannidis (Congress Steering Committee); Department of Computer Information Systems, Wentworth Institute of Technology, Boston, Massachusetts, USA; Visiting Professor, MIT, USA

• • • • • • • • • • • •

• • •

Prof. Mary Mehrnoosh Eshaghian-Wilner (Congress Steering Committee); Professor of Engineering Practice, University of Southern California, California, USA; Adjunct Professor, Electrical Engineering, University of California Los Angeles, Los Angeles (UCLA), California, USA Prof. Byung-Gyu Kim (Congress Steering Committee); Multimedia Processing Communications Lab.(MPCL), Department of Computer Science and Engineering, College of Engineering, SunMoon University, South Korea Prof. Tai-hoon Kim; School of Information and Computing Science, University of Tasmania, Australia Prof. Dr. Guoming Lai; Computer Science and Technology, Sun Yat-Sen University, Guangzhou, P. R. China Dr. Muhammad Naufal Bin Mansor; Faculty of Engineering Technology, Department of Electrical, Universiti Malaysia Perlis (UniMAP), Perlis, Malaysia Dr. Andrew Marsh (Congress Steering Committee); CEO, HoIP Telecom Ltd (Healthcare over Internet Protocol), UK; Secretary General of World Academy of BioMedical Sciences and Technologies (WABT) a UNESCO NGO, The United Nations Prof. Dr., Eng. Robert Ehimen Okonigene (Congress Steering Committee); Department of Electrical & Electronics Engineering, Faculty of Eng. and Technology, Ambrose Alli University, Edo State, Nigeria Prof. James J. (Jong Hyuk) Park (Congress Steering Committee); Department of Computer Science and Engineering (DCSE), SeoulTech, Korea; President, FTRA, EiC, HCIS Springer, JoC, IJITCC; Head of DCSE, SeoulTech, Korea Dr. Akash Singh (Congress Steering Committee); IBM Corporation, Sacramento, California, USA; Chartered Scientist, Science Council, UK; Fellow, British Computer Society; Member, Senior IEEE, AACR, AAAS, and AAAI; IBM Corporation, USA Ashu M. G. Solo (Publicity), Fellow of British Computer Society, Principal/R&D Engineer, Maverick Technologies America Inc. Prof. Fernando G. Tinetti (Congress Steering Committee); School of CS, Universidad Nacional de La Plata, La Plata, Argentina; Co-editor, Journal of Computer Science and Technology (JCS&T). Prof. Shiuh-Jeng Wang (Congress Steering Committee); Director of Information Cryptology and Construction Laboratory (ICCL) and Director of Chinese Cryptology and Information Security Association (CCISA); Department of Information Management, Central Police University, Taoyuan, Taiwan; Guest Ed., IEEE Journal on Selected Areas in Communications. Prof. Layne T. Watson (Congress Steering Committee); Fellow of IEEE; Fellow of The National Institute of Aerospace; Professor of Computer Science, Mathematics, and Aerospace and Ocean Engineering, Virginia Polytechnic Institute & State University, Blacksburg, Virginia, USA Prof. Jane You (Congress Steering Committee); Associate Head, Department of Computing, The Hong Kong Polytechnic University, Kowloon, Hong Kong Dr. Wen Zhang; Icahn School of Medicine at Mount Sinai, New York City, Manhattan, New York, USA; Board member, Journal of Bioinformatics and Genomics; Board member, Science Research Association

We would like to extend our appreciation to the referees, the members of the program committees of individual sessions, tracks, and workshops; their names do not appear in this document; they are listed on the web sites of individual tracks. As Sponsors-at-large, partners, and/or organizers each of the followings (separated by semicolons) provided help for at least one track of the Congress: Computer Science Research, Education, and Applications Press (CSREA); US Chapter of World Academy of Science; American Council on Science & Education & Federated Research Council (http://www.americancse.org/); HoIP, Health Without Boundaries, Healthcare over Internet Protocol, UK (http://www.hoip.eu); HoIP Telecom, UK (http://www.hoip-telecom.co.uk); and WABT, Human Health Medicine, UNESCO NGOs, Paris, France (http://www.thewabt.com/ ). In addition, a number of university faculty members and their staff (names appear on the cover of the set of proceedings), several publishers of computer science and computer engineering books and journals, chapters and/or task forces of computer science associations/organizations from 3 regions, and developers of high-performance machines and systems provided significant help in organizing the conference as well as providing some resources. We are grateful to them all. We express our gratitude to keynote, invited, and individual conference/tracks and tutorial speakers - the list of speakers appears on the conference web site. We would also like to thank the followings: UCMSS (Universal Conference Management Systems & Support, California, USA) for managing all aspects of the conference; Dr. Tim Field of APC for coordinating and managing the printing of the proceedings; and the staff of Monte Carlo Resort (Convention department) at Las Vegas for the professional service they

provided. Last but not least, we would like to thank the Co-Editors of BIOENG’17: Prof. Hamid R. Arabnia, Prof. Leonidas Deligiannidis, and Prof. Fernando G. Tinetti. We present the proceedings of BIOENG’17.

Steering Committee, 2017 http://americancse.org/

Contents SESSION: BIOMEDICAL ENGINEERING, METHODOLOGIES, AND NOVEL APPLICATIONS Analysis of 3D Cone-Beam CT Image Reconstruction Performance on a FPGA Devin Held, Michael Bauer

3

A Review of Brain Signal Processing Methods Selena Snyder, Xiaoping Shen

10

An Ultra-Low Power modified SERF Adder in Subthreshold region for Bio-medical Applications Pradeep Sarva, Santosh Koppa, Eugene John

17

Digital Compensation Method for Enhancing the Blood Pressure Estimation Based on the Arterial Pulse Waveform Boyeon Kim, Yunseok Chang

23

Using a Non-invasive Approach to Evaluate the Effects of Smoking on Swallowing and Respiration Coordination Wann-Yun Shieh, Chin-Man Wang, Hsin-Yi Kathy Cheng

30

Discovery of a Naturally Occurring Temperature Scale Steve Richfield

36

SESSION: POSTER PAPERS A Simulation Study on the Safety of Brain during Ocular Iontophoresis Sangjun Lee, Chany Lee, Chang-Hwan Lim

45

Conductive Cottons for Patch type Wearable Bio Potential Monitoring Seong-A Lee, Ha-Chul Jung, Sanghun Lee, Dahye Kwon, A-Hee Kim, Jin-Hee Moon

46

Drug Delivery for Colon Cancer Therapy by Doxorubicin with Oligonucleotide Modified Gold-Nanoparticles Moo-Jun Baek, Dongjun Jeong, Hyung Ju Kim

48

Remote Measurement of Infant Emotion via Heart Rate Variability Guang-Feng Deng, Yu-Shiang Hung, Wei-Kuang Ho, Hsiao-Hung Lin

50

Digital Arterial Pulse Waveform Measurement System with the PPG Sensor Yunseok Chang, Munseong Jeong, Boyeon Kim

52

A Study on Brain Activation by Addition Task with Driving Woo-Ram Kim, Ji-Hun Jo, Mi-Hyun Choi, Hyung-Sik Kim, Soon-Cheol Chung

54

Classifying Clinical Citations based on Semantic Polarity Hassan Alam, Aman Kumar, Tina Werner, Manan Vyas, Rachmat Hartono

56

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

SESSION BIOMEDICAL ENGINEERING, METHODOLOGIES, AND NOVEL APPLICATIONS Chair(s) TBA

ISBN: 1-60132-451-0, CSREA Press ©

1

2

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

3

Analysis of 3D Cone-Beam CT Image Reconstruction Performance on a FPGA Devin Held and Michael Bauer Department of Computer Science The University of Western Ontario London, Ontario, N6A 5B7, Canada {dheld2,bauer}@uwo.ca

Abstract Computed tomography (CT) scans are used to analyze internal structures through capture of x-ray images. CT scans are prone to multiple artifacts, including motion blur, streaks, and pixel irregularities, and therefore must be run through image reconstruction software to reduce visual artifacts. Efficient and accurate tomographic image reconstruction algorithms and software have been an intensive topic of research. The most common algorithm used is the Feldkamp, Davis, and Kress back projection algorithm. The algorithm is computationally intensive due to the O(n4) back projection step. Processing large CT data files on GPUs has been shown to be effective. An emerging alternative for implementation of this algorithm is via Field Programmable Gate Arrays (FPGAs). With companies, like Intel and IBM, starting to bring FPGAs into more mainstream computing systems, there are opportunities to leverage the potential performance of FPGAs and their reduced power consumption. In this paper we present an analysis of the performance of a 3D cone-beam CT image reconstruction implemented in OpenCL on a FPGA and compare its performance and power consumption to a GPU. keywords: Image reconstruction, image processing, computed tomography, cone-beam CT images, FPGAs, performance.

operations are time consuming on the central processing unit (CPU) of a computer, so many researchers have looked to implementations that can leverage accelerators, such as a graphics processing unit (GPU); they essentially look to offload a lot of the computations from the CPU to the accelerator themselves. In this paper we look at the use of an alternative accelerator – a field-programmable gate array (FPGA). FPGAs are an interesting component for the efficient execution of algorithms due to the low energy consumption by these devices [1]. This makes them a viable alternative to consider as they can run with low power usage in remote locations. We compare the implementation of a common reconstruction algorithm, the Feldkamp algorithm for cone-beam image reconstruction [2], on a GPU and FPGA and look at performance and power consumption. The remainder of the paper is organized as follows. In the next section we provide some background on CT Scans and the FDK reconstruction algorithm. Section 3 provides an overview of other efforts to improve the performance of the reconstruction algorithm, these typically involve parallelization or use of GPUs or both. We present our results in Section 4 and in Section 5 we provide conclusions and future directions.

2. CT Scans and Reconstruction 2.1. Overview of CT Scans

1. Introduction Efficient and accurate tomographic image reconstruction has been an intensive topic of research due to its increasing everyday use in areas such as radiology, biology, and materials science. The capture of tomographic images allows specialists to feed two dimensional images through a software application, run computationally intensive image processing algorithms on the images, and reconstruct a clear three dimensional image. CT scanners record x-ray image slices through a variety of approaches, including fan-beam and cone-beam computed tomography, providing different techniques to section image slices. By far, the most computationally intensive task is reconstructing the image slices. Such

After the first computed tomography (CT) scanner was released in 1971, it was quickly described as the “greatest diagnostic discovery since x-rays themselves in 1895” [3]. At the time of the scan, the patient will lay on a flat table attached to the CT scanner. Patients must lay very still within the CT scanner to avoid noise and error within the images taken. Each image takes half a second to acquire [4] as the scanner rotates and takes tomographic slice images through the use of x-rays. In CT scans, x-rays are not stationary. The x-rays are rotated about an axis, continuously gathering information about a target through a focal point. This data is run through a software program in order to create two dimensional image slices, which can later be combined to

ISBN: 1-60132-451-0, CSREA Press ©

4

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

create the three dimensional image; the software enables certain filtering operations to be done on the images. X-ray images are taken from a single source, diverging in either a fan-beam or cone-beam shape as they rotate about a fixed point and record data from different angles. In x-ray CT scans, fan-beam imagery is a quite common approach to the projection geometry [5]. The xray “fans out” from a focal point, angling out to create a fan shape before hitting the sensors on the opposite side. It is also important to note that the fan-beam rotates about the center pixel, and data is gathered using this narrow fan-beam, slightly changing the angle of the source to collect data from many slices. After data is collected, the image slices are then compiled to create two dimensional representations of parts of the object. A disadvantage of fan-beam imaging is that part of the image data may be truncated due to the limited field of view from the magnification factor, causing spikes and other artifacts around the edges of the image [6]. Unlike the thin fan-beam geometry, cone-beam geometry immediately captures a two dimensional image through a three dimensional x-ray beam, which scans in the shape of a cone. The x-ray tube and detector rotate about the object of interest, as displayed in Figure 1. This three dimensional beam proves efficient as at most one rotation around the object provides enough information to reconstruct a three dimensional image 7].

Figure 1: Cone-Beam CT System [7] The rapid acquisition of information causes conebeam projections to be more favorable than fan-beam projections due to the shorter procedure time and therefore less possibility of quality loss due to patient movement. [8] The shorter procedure time also limits the patient’s exposure to radiation. The ability to capture two dimensional images in a single attempt provides higher quality images than the fan-beam x-ray geometry. X-ray images taken through either fan-beam or conebeam tomography risk discrepancies called artifacts. These artifacts are errors between the numbers evaluated during the CT scan and the exact attenuation coefficients. These discrepancies are undesirable as they can make certain internal features appear brighter or darker

depending on the variation. There are many ways artifacts can appear in generated x-ray images, including errors due to physics, patients, and scanner hardware inconsistencies [9.10]. Many of these types of errors can be corrected through filtering and deblurring the images through image processing algorithms.

2.2.

3D CT Image Reconstruction

Among the most popular algorithms used for 3D CT image reconstruction is Feldkamp Davis and Kress (FDK) weighted filtered back projection algorithm [2]. Due to the circular trajectory required for calculating this algorithm, only an approximate reconstruction can be acquired. Yet, this algorithm remains popular due to its simplicity, the fact that is amenable to parallel implementations, and the good enough reconstructed image. Weighting and filtering the two dimensional image slices takes little time when compared to the time consuming back projection step, which is the bottleneck of the algorithm itself. The core of the FDK algorithm makes use of Fourier Transforms, which are the basis for the Ramp Filter used in the reconstruction algorithm. At the core of filtered back projection algorithm there exist both the Fourier and inverse Fourier Transforms. The Fourier Transform essentially reduces any waveform in the real world into sinusoids, providing another way to interpret waveforms [11]. Typically, the function transforms a function of time into corresponding frequencies that make up the function. The Ramp Filter is a high-pass filter and is used to create a clearer image without changing projection data before the back projection step. This filter assists in reducing blur and other visual artifacts found in the projection images. The Ramp Filter, defined by the inverse Fourier Transform, targets image noise and imperfections and smooths them out through filtering techniques. Cone-Beam CT (CBCT) [8] reconstructs three dimensional images rather than two dimensional images. Since CBCT captures two dimensional images rather than data to be reconstructed, it is prone to many more artifacts than traditional Fan-Beam CT technology. Because of this, a more intensive mathematical filtering process is applied to each CBCT projection to effectively reduce appearance of artifacts in the images. The main troublesome artifacts, noise and contrast imperfections, are more apparent in CBCT projections due to the higher amount of radiation present [8]. These artifacts must be smoothed through image processing applications or image clarity will be impaired. The process of CT image reconstruction begins with acquiring the data from the Cone-Beam CT (CBCT) scanner, and then reconstructing the two dimensional images. This process not only creates the three dimensional representation of an object, but also

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

essentially reduces the artifacts in the images. Figure 2 illustrates the reconstruction elements.

5

approach. It essentially converts the CBCT image reconstruction problem into the fan-beam reconstruction problem, using a circular focal point trajectory [5] but uses a cone-beam coordinate system (see Figure 3). However, an additional dimension - z - is added to represent the third dimension of the cone-beam projection. An important variable to note is the length from the detector to the focal point, D, which is the main determinant of a weighting function for this algorithm. CBCT uses the Fourier transform and ramp filtering to reconstruct the CT images. Ramp-filtering is applied to the images row-by-row and then back projection is calculated on the resulting data. Zeng et al. [5] provide an exceptional straightforward diagram of the cone-beam coordinate systems and a clear mathematical equation to assist in understanding and expressing important aspects. The mathematical equation is expressed as follows:

Figure 2: Steps in the reconstruction process [8]. The first stage in CBCT image reconstruction is the acquisition stage, beginning with image data collection. Throughout a single rotation of the CT scanner, about 150 to 600 images are taken. Each individual image consists of greater than a million pixels, with each pixel holding 12 to 16 bits of data [8]. The second stage in the CBCT process is the reconstruction stage (see Figure 2). This stage involves processing image relations and recombining slices to create a three dimensional volume. Three popular algorithms to reconstruct cone-beam projections are Grangreat’s [12], Katsevich’s[13], and the most used one – Feldkamp, Davis and Kress (FDK) [2]. All three of the popular algorithms result in reconstructed images with little data loss, and are able to be effectively parallelized to an extent. Algorithms developed by Katsevich and Greangreat provide exact image reconstructions, but

, ,

sin

, and

1 2





√







, ̂,





The mathematical equation may appear complex, but it is simply broken up into different steps. This equation is performed by weighting, ramp filtering (including the inverse Fourier Transform), and then performing back projection to reconstruct the three dimensional image [14]. This algorithm has the potential to be highly parallelized due to data independence and so running such an algorithm on an accelerator is promising. Design wise, the projections are first multiplied by the weighting function. The ramp filter is then applied to the data and finally back projection is performed; this step is computationally intensive that runs in O(N4) time [15]. With the move to GPUs, the back projection step has conclusively sped up the overall run time. Yet, in using an accelerator, such as the GPU, a new bottleneck surfaces. Projection data must be loaded onto the accelerator in order to run computations on the device, thus creating the memory transfer bottleneck. The question then becomes how do we implement efficient memory transfer? This bottleneck can be reduced through transferring data all at once and doing as many computations as possible on the GPU before inducing another single transfer back to the CPU. But, with the exceptionally large size of the CT scan data files, this still results in considerable overhead.

3. Related Work Feldkamp et al. provide approximate image reconstructions. Figure 3: Feldkamp Algorithm Visual [5]. The FDK algorithm [2] is the most widely used algorithm today, due to the practicality and power of the

The FDK [2] has been the focus for many of the research studies on the performance of CT reconstruction due to its efficiency and its suitability to parallel processing. The research studies reviewed are all focused on the FDK algorithm. The majority of studies of this

ISBN: 1-60132-451-0, CSREA Press ©

6

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

algorithm are implemented on GPUs in either CUDA or OpenCL; only one study was found that discussed an implementation on an FPGA, although it uses a simulated memory bus. Wang et al. [14] benchmark the Feldkamp et al. algorithm through a parallel implementation for the GPU. Unlike many implementations of this algorithm, they decided to use OpenCL to write the device code for the GPU. The parallel capabilities of the OpenCL language make it a useful choice for optimizing implementations for accelerator hardware as the code can also be run on CPUs, GPUs, and FPGAs. With OpenCL, The ability to abstract sections of the algorithm into device code contained in kernels ensures that the developer has a clear sense of what is running where and how to write code to run on both the CPU and GPU at the same time. To minimize data transfer between the CPU and GPU, Wang et al. introduced pinned (also called pagelocked) memory to store the image slices rather than nonpinned memory [14]. Both pinned and non-pinned memory are ways the GPU accesses memory from the host device. Pinned memory is quicker due to the ability to execute writes and application instructions concurrently, and large sequential reading [16]. Nonpinned memory invokes multiple levels of caching, involving full loads of data [16]. Non-pinned memory seems less useful in this case due to the change of data throughout the reconstruction process. Kernels were designed to run each of the steps in the FDK algorithm, as described in the previous section. Image-by-image, data is loaded onto the GPU, with kernels executing preweighting, filtering, reconstruction, and weighting before transferring the data back to the GPU. Running all GPU computations at the same time reduces the need to transfer data multiple times which would inhibit performance. In comparing the OpenCL implementation of the FDK algorithm on the CPU compared to the GPU, Wang et al. found an overall speedup of over 57 times. Wang et al. used a generated head phantom with 1283 voxels. They also noticed that as the volume of the data increased, the speedup numbers also increased. With a head phantom of 2563 voxels, the GPU implementation performed 250 times better than the CPU implementation for the weighting step! Noel et al. approached the 3D cone-beam reconstruction for the FKD algorithm on a GPU as well. A CUDA implementation is used to parallelize the intensive back projection step on the GPU. The paper evaluates the run time of the CUDA implementation on the CPU vs. the GPU. Multiple volume sizes are tested on both the CPU and GPU to compare speedup numbers. Specific tests were run on an NVIDIA GeForce GTX 280, making the results relevant to real world applications due to the low-cost of the device. The approach consists of harnessing all shared memory, loading all of the image data in the memory on the GPU device, and parallelizing computations on individual

voxels. Because the GPU cannot handle both the volume and projection data at the same time, they decided the best approach was to load all of the projection data onto the device, and only add parts of the volume data at a time; each sub-row of the volume data are then back projected through multithreading on the GPU device. Noel et al. found that the execution took about 3.2 seconds on the GPU as compared to 25 minutes on the CPU for a low-quality image [17]. For a high-quality image with 5123 volumes, the GPU reconstruction implementation ran in about 8.5 seconds. They note that there still exists a bottleneck in the GPU implementations of this algorithm due to the data transfer of large-scale images from CPU to GPU memory. Okitsu et al. [18] focused on harnessing the processing power of the GPU through multiple optimization techniques, expanding on the work by Noel et al. Such optimization techniques included loopunrolling, multithreading for multiple GPUs, and reducing off-chip memory access. Due to the independent data computations in the FDK algorithm, Okitsu et al. observed that there is potential to run computations on different voxels, as well as different projections between multiple GPUs. This breaks up the large amount of data and runs it on multiple accelerators at the same time, allowing further parallelization of data. To limit data exchange, only volume data is stored in the device memory due to the limited device memory and the ability to discard projection data after running the algorithm. Okitsu et al. concluded that both parallelizing projection and voxel data resulted in the runtime of the back projection step on a single GPU to be almost cut in half for both low and high quality data. Spreading data access multiple GPUs reduces the number of computations on a single GPU, and further parallelizes the FDK algorithm, thus achieving a 41% better performance than alternative implementations in CUDA. This multiple GPU approach also results in 24 times quicker reconstruction than on a native CPU. This approach also solves the problem of running out of memory for computation on a single GPU by effectively splitting up the data, as large-scale datasets can easily diminish resources in a single GPU device memory. Gac et al. [19] experimented with implementations of 3D CT image reconstruction algorithms on the CPUs, GPUs, and FPGAs. This paper explores optimization techniques for implementing the back projection on the FPGA hardware. They approach the memory access bottleneck, designing a pipeline that improves efficiency. This is among a few studies done with the 3D image reconstruction problem on FPGAs, comparing speedups in comparison with both the CPU and GPU.

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

7

4. Algorithm and Implementation

4.1. Test Systems

The Feldkamp, Davis, and Kress algorithm for 3D cone-beam CT image reconstruction is implemented in a parallel manner across many threads. Each thread can work simultaneously on separate image slices as they are data independent. In the OpenCL program, the slices can be divided so every available thread in the machine can process a set of data concurrently. Our overall implementation is represented in the rough pseudo code of Tables 1 through 4. In Table 2, an overview of the entire algorithm is presented; in Table3 2 through 4, the basic pseudo-code for the main processing kernels is presented.

To test the effectiveness of the OpenCL FPGA implementation, benchmarking was done across two test systems: the built in CPU/GPU on an Apple iMac (late 2015) and a Nallatech FPGA card [20] attached to IBM's S822L power processor. We opted for an “off-the-shelf” GPU as previous research had already demonstrated the improved performance of the FDK algorithm on GPUs, and this GPU performed well and was “reasonable” in power consumption. The specifications of the systems are as follows:

load all image data from file onto CPU   initialize OpenCL device  load OpenCL kernels onto device  copy image data to device  invoke Weighting kernel  invoke Ramp Filtering kernel  invoke Back Projection kernel  copy image data to medical image file  end OpenCL device program

 Apple iMac (late 2015). CPU: 3.2 GHz Intel Core i5, 1867 MHz Memory: 16 GB 1867 MHz; GPU: AMD Radeon R9 M380 2048 MB, Memory: 4096MB, Memory Bandwidth 96 GB/s, Memory clock speed: 1500 MHz.  IBM Power System S822L. CPU: 2 8-core 4.15 GHz Power8 Processing Cards, 16 GB/512GB min/max memory, process to memory bandwidth: 192 GB/s per socket, I/O Bandwidth: 96 GB/s; ,FPGA: Altera Stratix V [21].

Table 1. Pseudo Code for Reconstruction Algorithm. Weighting kernel       for each image in projections // parallelized over threads            for each x row in image                 for each y col in image                       image[x][y] *= weightingDistanceFactor            copy image to output variable 

4.2. Test Data The image used to test the performance of the implemented FDK algorithm is a generated Shepp Logan Phantom. See the image displayed in Figure 4.

Table 2. Pseudo Code for Weighting. Ramp Filtering kernel       for each image in projections // parallelized over threads            for each x row in image                 for each y col in image                      perform Fourier transform on x,y                      multiply by ramp filter value in frequency domain                      perform inverse Fourier transform on x,y            copy image to output variable 

Figure 4: Shepp Logan Phantom.

Table 3. Pseudo Code for Filtering. Back Projection kernel       for each image in projections // parallelized over threads            for each x row in image                 for each y col in image                      for every angle                           calculate x,y angles from focal point                           calculate upper, lower, left, right weighting                           perform back projection arithmetic                 scale image 

Table 4. Pseudo Code for Back Projection.

The Shepp Logan Phantom is frequently used to test Fourier reconstruction problems modelling the human head. It is easily generated in MatLab with built-in function calls. Each individual projection is a sinogram of the phantom, representing two dimensional slices of the three dimensional image. Two variations of the image were used to test different data volumes: 

A 2563 3D image from 400 projections;



A 5123 3D image from 600 projections.

ISBN: 1-60132-451-0, CSREA Press ©

8

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

5. Experimental Results Tables 5 and 6 present the resulting performance of the OpenCL implementation of the FDK algorithm on the GPU and FPGA. The three kernels, developed in OpenCL, are run on the GPU and FPGA. The kernels are timed individually to also compare the back projection step relative to the weighting and filtering steps. The total reconstruction timing also includes reading and writing the image; the results are averaged over 10 testing sessions. A summary of the computation times for each phase and for each system is presented in Table 5; standard deviations (in parentheses) are also included. It is clear to see that the back projection computation is the most time consuming. In comparing the performance of the GPU and FPGA, the GPU outperformed the FPGA by a factor of 1.7 and for the back projection step it outperformed the FPGA by 1.8. It should be noted, however, that this is an initial implementation, and there are likely additional improvements that can be done to improve the FPGA implementation. Shepp Logan Phantom with 2563 Voxels GPU

FPGA

Weighting

0.124 (0.005)

0.207 (0.008)

Ramp Filtering

1.914 (0.058)

3.775 (0.032)

Back Projection

1.72 (0.007)

3.215 (0.023)

Total Time for Reconstruction

4.244 (0.122)

7.646 (0.071)

Shepp Logan Phantom with 5123 Voxels Weighting

0.398 (0.011)

0.782 (0.019)

Ramp Filtering

8.557 (0.025)

16.168 (0.236)

Back Projection

12.392 (0.112)

22.909 (0.235)

Total Time for Reconstruction

26.021 (0.495)

43.666 (0.522)

Table 5. Summary of Timing Results. One of the key advantages of the FPGA is its power consumption. With a move to create green, energyefficient technology, energy usage per implementation should be a factor in choosing the optimal device. The power consumption of the FPGA is estimated due to its power usage being directly correlated with the number of

logic elements implemented. We use an average figure for the Stratix V FPGA; Altera estimates that it typically runs with 4W of power (altera.com). Table 6 summarizes the average power consumption of each device, the total runtime and estimate energy consumption (in joules) for each different sized image. Based on these numbers, the FPGA runs the most energy-efficient version of the FDK algorithm. The GPU implementation runs with about 10.6x more power with the 2563 image and 11.2x more power with the 5123 image than the FPGA. Time for 2563 Voxels

Power (joules)

Time for 5123 Voxels

(75W)

4.244

318.3

26.021

1951.6

FPGA (4W)

7.646

30.6

43.666

174.7

  

Power (joules)

GPU

Table 6. Power Consumed for Computations by Different Systems.

5. Conclusion and Directions Cone-beam tomography is the most common CT imagery technology today. Unlike fan-beam tomography, which collects data to reconstruct a 2D image, cone-beam technology captures 2D images and reconstructs a 3D image. The x-ray projects in a cone shape from a focal point, capturing 2D image data of the area of interest. The x-ray focal point and detector rotate around the patient, slightly varying angles between image capture. The CT scans are prone to many imperfections, called artifacts, such as irregularities due to patient movement, hardware calibrations, and foreign objects in the scan field. Most of these artifacts and removed post scan through a software program that runs 3D CT image reconstruction. This program is essentially image processing software, reducing blur and applying filters to clean the images. The final step of the software actually reconstructs the 3D image, so radiologists can view the image in more detail. The most common algorithm used for 3D CT image reconstruction is the Feldkamp, Davis, and Kress [2] algorithm. It is preferred due to the highly parallel nature of the algorithm, making it compatible to run on accelerators. In general, this algorithm applies a weighting function, then ramp filtering, and then back projection. The back projection step is the bottleneck of the algorithm due to its computationally intensive process, resulting in O(N4) run time. With the high volumes of data from CT scanners, this deems the CPU inadequate to run the reconstruction in a short amount of time. Therefore, a push to use accelerators, such as graphics processing units

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

(GPUs) and field programmable gate arrays (FPGAs), has been a focus of much research. When compared with the run times on the GPU, the GPU outperforms the FPGA by a factor of 1.7. The back projection step runs about 10 seconds faster on the GPU. These results indicate that the FPGA is a viable solution to speeding up this algorithm. However, it is also interesting to note that the FPGA while the implementation is slower, but it is 10-11x more energy efficient than the GPU. Future work could proceed in several directions:  Implementation in the OpenCL language abstracts out the circuit design level in the development process. A careful review of the OpenCL code and knowledge of circuits could result in improvements in both the processing time and the even lower the power consumption. We are currently looking into this.  The Feldkamp, Davis, and Kress (FDK) algorithm for cone-beam image reconstruction is not the only algorithm designed to reconstruct 3D images. It actually provides an approximate image reconstruction in comparison to exact reconstructions such as those resulting from Grangreat's or Katsevich's algorithms. It would be interesting to explore how these algorithms perform on the FPGA.  Increasingly, use of computed tomography is being used not just to generate a single image, but to utilize multiple scans, thus requiring reconstruction of many images. This may mean hundreds of scans, each with hundreds of images. Reconstructions of such sequences can take many minutes. Further work to optimize FPGA implementation may be a real advantage. Finally, with Intel’s acquisition of Altera and with IBM’s promotion of FPGAs as accelerators, one can expect FPGAs as accelerators to become more common and the cost to continue to drop. More algorithms implemented in OpenCL to run of FPGAs and GPUs are likely to become more common.

References [1] Jamieson, Peter, et al. An energy and power consumption analysis of FPGA routing architectures: Field Programmable Technology, IEEE International Conference on Field Programmable Technology, 2009. [2] Feldkamp, L. A., L. C. Davis, and J. W. Kress. Practical cone-beam algorithm. Journal of the Optical Society of America A, Vol. 1, No. 6, 1984, pp. 612-619. [3] Webb, Steve. A brief history of tomography and CT, Proceedings of the 7th Asian and Oceanian Congress of Radiology. 1998, pp. 429-430.

9

[4] Physics Central, CT Scans, www.physicscentral.com. [5] Zeng, Gengsheng Lawrence. Medical Image Reconstruction. Springer, 2010. [6] Gengsheng L, Zeng et al. Fan-Beam Reconstruction Algorithm for a Spatially Varying Focal Length Collimator, IEEE Transactions on Medical Imaging, Vol. 12, No. 3, 1993, pp. 575-582. [7] Sedentexct, Technical Description of CBCT, www.sedentexct.eu. [8] Scarfe, William C., and Allan G. Farman. What is cone-beam CT and how does it work?. Dental Clinics of North America. Vol. 52, No.4, 2008, pp. 707-730. [9] Barrett, Julia F., and Nicholas Keat. Artifacts in CT: recognition and avoidance". Radiographics Vol. 24, No. 6, (2004), 1679-1691. [10] Ohnesorge, B., et al. Efficient correction for CT image artifacts caused by objects extending outside the scan field of view. Medical Physics, Vol. 27, No. 1, 2000, 39-46. [11] The Fourier Transform, www.thefouriertransform.com. [12] P. Grangeat, Mathematical framework of cone-beam 3D reconstruction via the first derivative of the Radon transform, Proc. Mathematical Models in Tomography, Springer Lecture Notes in Mathematics, Vol. 1497, 1990, pp. 66-97. [13] Katsevich A. A general schedule for constructing inversion algorithm for cone beam CT. Int. Journal Math. Sciences, Vol. 21, 2003, pp. 1305–1321. [14] Wang, Bo, et al. Accelerated cone beam CT reconstruction based on OpenCL. 2010 IEEE International Conference on Image Analysis and Signal Processing. 2010, pp. 291-295. [15] Zhao, Xing, Jing-Jing Hu, and Peng Zhang. GPUbased 3D cone-beam CT image reconstruction for large data volume. Journal of Biomedical Imaging, Vol. 8, 2009. [16] Pinned and Non-Pinned Memory, IBM, http://www.ibm.com/support/knowledgecenter/ SSFKCN_3.5.0/com.ibm.cluster.gpfs.v3r5.gpfs300.d oc/bl1ins_pnonpmem.htm [17] Noël, Peter B., et al. GPU-based cone beam computed tomography. Comp. Methods and Pgms. in Biomedicine.Vol. 98, No. 3. 2010, 271-277. [18] Okitsu, Yusuke, Fumihiko Ino,, Kenichi Hagihara. High-performance cone beam reconstruction using CUDA compatible GPU. Parallel Computing, Vol. 36, No. 2, 2010, pp. 129-141. [19] Gac, Nicolas, et al. High speed 3D tomography on CPU, GPU, and FPGA. EURASIP Journal on Embedded systems, Vol. 5. 2008. pp. 1-2. [20] Nallatech,, 385 Hardware Reference Guide. [21] Stratix V, Altera, https://www.altera.com/ content/dam/alterawww/global/en_US/pdfs/ literature/hb/stratix-v/stx5_51001.pdf.

ISBN: 1-60132-451-0, CSREA Press ©

10

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

A Review of Brain Signal Processing Methods Selena Snyder and Xiaoping A. Shen Department of Mathematics Ohio University, Athens, OH, USA [email protected]

Abstract— Brain signals can be obtained and analyzed using a variety of methods as described in this literature review. Understanding the possibilities of analytical methods expands researchers’ horizons for developing technological approaches to detecting biological events. Specifically, EEG signals can be analyzed using a variety of methods, suggesting a combination of methods could be optimal for ease of computerized analysis and diagnosis of epileptic seizures.

acquisition techniques can be found in Table 1. A summary of published literature on these techniques is in Figure 1.

Keywords—Brain-Computer Interface, EEG signals, signal processing, feature extraction, epileptic seizures

I. INTRODUCTION The human brain has been studied since the time of the ancient Egyptian mummifications to 18th century scientific research on neurons. Today, neuroscience has advanced to exploring the technological possibilities of the human brain. The human brain is composed of glial cells, neural stem cells, blood vessels, and neurons. Neurons are the source of electrical activity in the brain, communicating through action potentials (AP), or spikes, as seen in electrophysiological recordings. The individual APs translate into thoughts and actions that require large groups of neurons to communicate at spatial and temporal resolution of interactions [1]. These large groups of AP can be recorded using a grid of electrical conductors (e.g. wire ‘tetrode’ arrays or silicon probes). The recorded electrical signals of the brain can be combined with technology to create a Brain-Computer Interface (BCI). A BCI processes the brain’s output pathway and uses the activity to control the external environment [2, 3]. There are five stages to developing a BCI: signal acquisition, preprocessing, feature extraction, classification, and application interface. This review paper builds on concepts presented in [2]. II. SIGNAL ACQUISITION A. Invasive v. Noninvasive Electrical brain signals are obtained through invasive or noninvasive measures. Invasive procedures, such as electrocorticography or local field potential, run the risk of tissue rejection or infection due to electrodes being placed on the surface or penetrating the cortex of the brain [4]. Noninvasive procedures are, hence, preferred due to their affordability and ease in recording electrical activity in the brain [2]. However, noninvasive procedures tend to have high contamination of the signal by noise and artifacts. Signal acquisition methods are mainly employed to observe spontaneous brain activity through two methods: electrical and hemodynamic [5]. A summary of the following signal

Fig. 1. Publications from 2014 – 2017 as grouped by signal acquisition technique. Data retrieved from Web of Science on 3/3/2017.

B. Electrically-based Techniques Electroencephalography (EEG) was first employed around 1970 to mechanize the detection of epileptic seizures [6]. Epileptic seizures are the result of excessive neural electrical activity in the brain and can be characterized in EEG through the occurrence of spikes and sharp waves. EEG data is collected using the International 10-20 system placement of electrodes (Fig. 2) [30]. Since then, algorithms, such as wavelet transform, Fourier transform, multi-wavelet transform, smoothed pseudo-Wigner-Vile distribution, and multifractional analysis, have been developed to interpret the electrical signals obtained [7, 8]. Epileptic seizures are the result of excessive neural electrical activity in the brain and can be characterized in EEG through the occurrence of spikes and sharp waves [8]. EEG signals are ideal due to the low signal-to-noise ratio (SNR) and minimal artifacts in the signal; however, EEG have poor spatial resolution [2, 7, 8]. Magnetoencephalography (MEG) is a noninvasive brain imaging technique based on magnetic fields in the brain induced by synchronized neural currents [29]. MEG and EEG present complimentary information. MEG measures magnetic signals generated by electrical activities in the brain [2]. MEG maintains excellent spatiotemporal resolution.

Research for this paper was supported by Ohio University Provost Undergraduate Research Fund (PURF) and the Air Force Office of Scientific Research (AFOSR).

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

Fig. 2. The interanational 10-20 electrode placement system seen from (A) left and (B) above the head. A = ear lobe, C = central, Pg = nasopharyngeal, P = parietal, F = frontal, Fp = frontal polar, O = occipital.

Since publication of Lakshmi et. al. in 2014, a new technology, NeuroGrid, has emerged to record electrical signals on the surface of the brain. NeuroGrid was developed based on the hypothesis that AP could be recorded from the surface of the cortex without penetrating the brain based on previous observations of AP of hippocampal pyramidal neurons. NeuroGrid was designed because current electrode arrays do not conform to the curvilinear surface of the brain, consequently decreasing the stability and efficiency of electrical and mechanical contacts. It was observed that NeuroGrid could record the average spike waveform amplitude consistently over a 10-day duration, indicating the promise of stability to monitor individual neurons over time with minimal physiological disruption. Specifically, NeuroGrid records local field potentials (LFP) and AP from superficial cortical neurons. Since NeuroGrid is an invasive option, the data provides better spatiotemporal resolution, as seen by the high SNR capable of spike detection, than other noninvasive techniques [1].

11

observe oxygenation changes due to abnormal brain activity. BOLD fluctuations have been investigated since identification of patterns in oxygenation in blood flow in 1995 [9]. fMRI presents data in a series of 3D images composed of volumetric pixels, or voxels [5]. The voxels are analyzed using a variety of techniques, such as principal component analysis (PCA), independent component analysis (ICA), canonical correlation analysis, and fuzzy clustering [5, 10]. These techniques utilize the high temporal and spatial resolution based on the time required to acquire signal from the voxel. In particular, fMRI ICA studies use relatively long temporal resolution (2 – 3 seconds) to increase the BOLD weighting under scan durations of 5 – 10 minutes [2, 10 – 12]. Hence, ICA techniques are used to identify consistent patterns in fMRI in a more exploratory manner. fMRI has two limitations: (i) most fMRI data consists of incomplete images of the brain due to signal dropout over the inferior temporal, orbitofrontal, or lateral midtemporal cortex; (ii) the signal obtained reflects many factors, but only some are related to neural activity. Functional Near-Infrared Spectroscopy (fNIRS) is a functional neuroimaging technique that measures brain activity through hemodynamic responses associated with neuron behavior. Like fMRI, fNIRS uses BOLD to obtain signal data, primarily from the primary motor cortex and prefrontal cortex [2, 3]. Specifically, fNIRS uses 650-1000nm wavelength to measure concentration changes of oxygenated hemoglobin and deoxygenated hemoglobin. These changes occur in the local capillary network of the brain because of neuron firings [3]. While fNIRS is not susceptible to electrical activity, fNIRS can be combined with EEG to form a hybrid BCI that is low cost, portable, safe, produces low noise in signals, and easy to use [2, 3]. The first occurrence of this hybrid development was in 2004 when studies were published supporting the feasibility of fNIRS-EEG BCI [3].

C. Hemodynamically-based Techniques Functional Magnetic Resonance Imaging (fMRI) uses blood oxygenation level dependent (BOLD) methodology to TABLE I.

Technique

COMPARISON OF SIGNAL ACQUISITION TECHNIQUES

Type of Signal Obtained

Advantages

Disadvantages

EEG

Electrical via scalp

• •

High Temporal resolution Noninvasive

• • •

Poor spatial resolution Non stationary signal Susceptible to motion artifacts

MEG

Magnetic generated by electrical activities

• •

Wider frequency range Excellent spatio-temporal resolution

• •

Bulky setup Expensive

Electrical obtained from superficial cortical neurons

•

NeuroGrid

•

Stable with minimal physiological disruption High spatiotemporal resolution

•

Invasive

fMRI

Hemodynamic using BOLD

•

High spatiotemporal resolution

• •

Expensive Delay in data acquisition process

fNIRS

Hemodynamic using BOLD

• • • •

High spatial resolution Inexpensive Portable Safe and easy to use

• •

Low temporal resolution Less performance

ISBN: 1-60132-451-0, CSREA Press ©

12

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

adjustments and is suitable for online applications, such as positioning an arrow tip within a circle on a screen and maintaining the position for 300ms [18].

III. SIGNAL PREPROCESSING TECHNIQUES Signal preprocessing occurs after the signal is acquired. When signals are acquired, generally, they are contaminated by noise and artifacts. To remove the noise and artifacts, several techniques can be employed to identify the true signal. •

Wavelet Transform (WT) is the wavelet decomposition of a noisy signal that concentrates on identifying the intrinsic signal information in a few wavelet coefficients that have large absolute values without modifying the random distribution of noise. This is also known as denoising achieved through means of thresholding wavelet coefficients. If a timevariant composition is given, this technique holds advantage over Wiener filtering, allowing the user to select different filter settings to account for different time ranges [7].

•

Independent Component Analysis (ICA) was first applied to signal preprocessing in 1996 to separate artifacts from EEG signals based on characteristics of the data without relying on reference channels [2, 13]. Separation of these artifacts generated from physiological noises allows the restoration of the hemodynamic signals [3]. The defining characteristics for ICA can change depending on the type of ICA used: temporal or spatial. Temporal ICA (tICA) corrects overlapping spatial distributions of different signal sources by identifying the temporally independent signals. tICA has good interpretability but is rarely used due to the complexity of the computations required and the limited amount of data points available to perform the calculations [10]. Spatial ICA (sICA) cannot accomplish the same task, but its calculations are simpler [14].

•

Principal component analysis (PCA) was first proposed by Karl Pearson in 1901 and supported by Harold Hotelling in 1930 [2]. PCA removes physiological noises and motion artifacts and performs dimensionality reduction, allowing for singular value decomposition to be applied, creating the principal components to be used in ICA [3, 10].

•

•

Common spatial patterns (CSP) is the transformation of EEG signals into covariance matrices that maximally discriminates between different classes [2]. CSP extracts features relevant to motor imagery tasks, such as grasping an object per Cutkosky’s grasping taxonomy, and is used for abnormality detection [15]. However, CSP is not optimal for multichannel data that has different degrees of correlation. To deal with data exhibiting this noncirculatity, an extension of CSP is introduced, the augmented complex CSP [16]. CSP models are trained on a subject by subject basis to prevent neglect of inter-subject information [17]. Common Average Referencing (CAR) removes noise by subtracting common activity from the position of interest. CAR enhances SNR to reference the position of interest [2]. CAR does not need training or

•

Surface Laplacian (SL) estimates the electrical current density leaving or entering the scalp through the skull. In the center scalp, SL is insensitive to motion artifacts because there are no underlying muscles [19]. It also solves the electrode reference problem – identifying the noise present on all channels to subtract it out. [2, 20].

Additional signal preprocessing techniques include single value decomposition (SVD), common spatio-spatial patterns (CSSP), frequency normalization, local averaging technique (LAT), Robust Kalman filtering, common spatial subspace decomposition (CSSD), adaptive filtering. Figure 3 summarizes publications from 2014 – 2017 based on signal preprocessing and feature extraction techniques.

Fig. 3. Publications from 2014 – 2017 as grouped by signal preprocessing and feature extraction techniques. Data retrieved from Web of Science on 3/3/2017.

IV. FEATURE EXTRACTION TECHNIQUES Feature extraction occurs after noise is removed from the raw signal. Feature extraction techniques emphasize essential characteristics of the signal for easy detection of biomedical events. Some feature extraction methods, PCA, ICA, and WT, are also applicable as signal preprocessing methods. PCA extracts information from all-time series multichannels as principal components [2]. This eliminates artifacts to reduce dimensionality of the signal. Then SVD can be used to have data compatible with ICA. ICA separates independent signals from noise in a Blind Source Separation (BSS). BSS estimates a set of unknown source signals from an observed mixture of data. BSS has been employed in ICA since the 1980s [21]. This methodology is often seen with fMRI data [10].

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

WT was applied to feature extraction in 1984 by Goupillard, Grossman, and Morlet using B-spline parameters to act as low and high pass filters [22]. WT was designed with the intention of improving upon the shortcomings of other techniques by capturing details about the sudden changes in EEG signals. Essentially, WT acts as a mathematical microscope (Fig. 4). WT is applied when the main wavelet is shifted by a small interval on the x-axis and the correlation coefficients are then computed [7]. This can be done through two types of WT: .

•

Continuous time wavelet analysis (CWT) evaluates the coefficients for continuous variation in the x- and y-axes.

•

Discrete Wavelet Transform (DWT) processes input signals with finite impulse response filters.

13

V. CLASSIFICATION METHODS Classification methods are used to organize the processed signal and its features into understandable data sets - to split the desirable from the undesirable. Classification methods are based on various principles of mathematics: linearity, calculus, discrete networks, etc. Each method has its benefits and its drawbacks. There are two types of linear classifiers: linear discriminant analysis (LDA) and support vector machine (SVM). LDA models the probability density function by separating the obtained hyperplane to maximize the distance between two classes’ means and minimize interclass variances, assuming a normal distribution of data [2, 23]. Ideally, LDA recovers a vector v in the feature space such that two projected clusters, yes-decision (Y) and no-decision (N), can be separated from each other. LDA is the most common classification method used for fNIRS-BCI studies [3]. However, SVM is the most popular method in BCI applications in general [2]. SVM was developed by Vapnik and Cortes in 1995 [24]. SVM follows statistical learning theory, following the principle of structural risk minimalization with the goal to provide good generalization while maximizing the distance between the separated hyperplane and the nearest training points [2, 23]. Nonlinear classifiers include artificial neural networks (ANN) and nonlinear Bayesian classifiers (NBC). ANN are composed of many interconnected elements. The structure is so like that of the human brain that these elements are termed neurons. ANN come in several different designs: multilayer perceptron neural network (MLPNN), Gaussian classifier, learning vector quantization, RBF neural network, and others [2, 3, 25]. ANN mimic the pattern recognition technique of the human mind to solve problems. Specifically, in MLPNN, three layers of neurons – input, hidden, and output – allow for fast operation, ease of implementation, and small training sets. When training MLPNN, errorback propagation method is used so the network can learn to map a set of inputs to outputs [26]. NBC use nonlinear decision boundaries to perform more efficient rejections of uncertain samples as compared to discriminative classifiers such as LDA. NBC do so by assigning a feature vector to the class with the highest probabilities [2]. An example of NBC is the Hidden Markov method, a nonlinear probabilistic classifier that provides the probability of observing a given set of features suitable for classification of time series [3]. Other examples of NBC include partial least squares discriminant analysis and quadratic discriminant analysis.

Fig. 4. Examples of different types of wavelets that have been used to analyze data in practice.

Additional feature extraction techniques include adaptive auto regressive parameters (AAR), fast Fourier transform (FFT), genetic algorithms (GA), and wavelet packet decomposition (WPD).

Nearest neighbor classifiers (NNC) predict an object’s value or class membership based on the k-closest training examples in the feature space [2]. In other words, NNC assigns a feature vector to a class based on its nearest neighbors. For example, NNC can be applied to EEG signals in the form of a weighted adaptive nearest neighbor (WDNN). WDNN assigns a weight to each training set to control its influence over classifying test samples. Influence increases the larger the weight. WDNN is applied in this example to

ISBN: 1-60132-451-0, CSREA Press ©

14

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

determine the quality of the EEG signal segment by assigning different weights for the classification task [27]. Empirical Mode Decomposition (EMD) is a classification method based on fractional order calculus. The signal is passed through the fractional linear prediction filter (FLP) and the coefficients of the filter are calculated by least squares analysis. Then, the prediction error, the difference between the modeled signal and the prediction signal, is calculated and used to train a SVM which then classifies the data sets into ictal and seizure-free categories [8]. EMD was proposed based on the effectiveness of fractional order modeling for speech signals [28]. VI. FUTURE WORK Based on the information presented, the intended path of this research is to identify the best methods for analyzing biological signal data. Focusing on EEG signals, they can be analyzed a variety of ways using WT, ICA, PCA, CSP, CAR, and SL. A wide range of literature has been developed in the subject of EEG analysis since Lakshmi, Prasad, and Prakash published their review on the subject in 2014. A keyword search in the citation index, Web of Science, reveals over 5,000 individual research papers on EEG analysis (Fig. 5). The number of publications per year has steadily increased in the past 20 years. The peak year for publication was 2015. Even with 2017 barely a quarter through, over 50 publications on the subject have occurred. The trend of increased publication may be expected to continue as the significance of mathematical analysis of EEG signals continues to be realized. A further analysis of publications per year can be seen illustrating the amount of publications per year in the field of Mathematical Computational Biology (Fig. 6). The general trend in this field is still upwards.

Fig. 5. Publications per year in the area of EEG signal analysis. generated via Web of Science on 3/10/2017.

Fig. 6. Publications by year for EEG analysis in Mathematical Computational Biology field. Data taken from Web of Science 3/10/2017.

Another look at the literature shows the significant areas in which EEG signal analysis research is being completed. The percentage of papers per research field clearly shows neuroscience as the leading field (Fig. 7). However, engineering, computer science, and mathematical computational biology are still strong contributors. This graph does not sum to 100% as many of the papers are published under several research fields and are not constrained to single fields. Future research intends to evaluate all current methods used to analyze EEG signals and identify the best characteristics of all to develop a new methodology. The intention is to replicate results from the references and then evaluate better approaches. Ideally, future work will lead to an automated way of classifying seizures based on EEG signals without the necessity of a person to monitor the process. This would speed up diagnosis and decrease the wait for treatment.

Image

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

15

Fig. 7. Percentage of EEG research in each field. Chart created using data from Web of Science accessed 3/10/2017.

Office of Scientific Research (AFOSR) for funding and support in this research. VII. CONCLUSIONS This review presented ideas from [2] and built upon them to reflect increased research between 2014 and 2017 on signal processing methods in addition to development of new signal acquisition methods. It is important to note that this is not an exhaustive review and lesser known methodologies may have been overlooked. A comprehensive review, covering signal acquisition, signal preprocessing, feature extraction, and signal classification methods, was presented. Suggestions of applications of multiple methods at once in a hybrid project were given. It is important to remain updated on the new methods surrounding the field of BCI research provides new avenues for research to pursue. Specifically, the possibility of developing the ultimate analysis methodology for EEG was suggested and the extent of such an application was described. ACKNOWLEDGMENT

REFERENCES [1]

[2]

[3] [4] [5]

[6]

[7]

D. Khodagholy et al., “NeuroGrid: recording action potentials from the surface of the brain,” Nat Neurosci, vol. 18, no. 2, pp. 310-315, Feb. 2015. M. R. Lakshmi, D. T. Prasad, and D. V. C. Prakash, “Survey on EEG signal processing methods,” International Journal of Advanced Research in Computer Science and Software Engineering, vol. 4, no. 1, 2014. N. Naseer and K.-S. Hong, “fNIRS-based brain-computer interfaces: a review,” Front. Hum. Neurosci., vol. 9, 2015. A. Burns, H. Adeli, and J. A. Buford, “Brain-Computer Interface after Nervous System Injury,” The Neuroscientist, vol. 20, no. 6, 2014. J. D. Power, B. L. Schlaggar, and S. E. Petersen, “Studying Brain Organization via Spontaneous fMRI Signal,” Neuron, vol. 84, no. 4, pp. 681–696, Nov. 2014. S. S. Viglione and G. O. Walsh, “Proceedings: Epileptic seizure prediction,” in Electroencephalography and Clincal Neurophysiology, vol. 39, no. 4, pp. 435-436, 1975. O.Faust, U. R. Acharya, H Adeli, and A. Adeli, “Wavelet-based EEG processing for computer-aided seizure detection and epillepsy diagnosis,” Seizure, vol. 26, pp. 56-64, Mar. 2015.

The authors would like to thank Ohio University Provost Undergraduate Research Fund (PURF) and the Air Force

ISBN: 1-60132-451-0, CSREA Press ©

16

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 | V. Joshi, R. B. Pachori, and A. Vijesh, “Classification of ictal and seizure-free EEG signals using fractional linear prediction,” Biomedical Signal Processing and Control, vol. 9, pp. 1–5, Jan. 2014. B. Biswal, F. Z. Yetkin, V.M. Haughton, and J. S. Hyde, “Functional connectivity in the motor cortex of resting human brain using echoplanar MRI,” Magn. Reson. Med., vol. 34, ppl 537-541, 1995. R. N. Boubela, K. Kalcher, W. Huf, C. Kronnerwetter, P. Filzmoser, and E. Moser, “Beyond Noise: Using Temporal ICA to Extract Meaningful Information from High-Frequency fMRI Signal Fluctuations during Rest,” Front. Hum. Neurosci., vol. 7, 2013. S.-G. Kim and S. Ogawa, “Biophysical and physiological origins or lood oxygenation level-dependent fMRI signals,” J. Cerb. Blood Flow Metab., vol. 32, pp. 1188-1206, 2012. B. B. Biswal, M. Mennes, X.-N. Zuo, S. Gohel, C. Kelly, S.M. Smith, et al., “Toward discovery science of human brain function,” Proc. Natl. Acad. Sci. U.S.A., vol. 107, pp. 4734-4739, 2010. A. Delorme, S. Makeig, “EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis,” J. of Neuroscience Methods, vol. 134, 2004. E. B. Beall and M. J. Lowe, “The non-separability of physiologic noise in functional connectivity MRI with spatial ICA at 3T,” J. Neurosci. Methods, vol. 191, pp. 263-276, 2010. E. Uchiyama, W. Takano, and Y. Nakamura, “Multi-class grasping classifiers using EEG data and a common spatial pattern filter,” Advanced Robotics, vol. 0, no. 0, pp. 1–14, Jan. 2017. C. Park, C. C. Took, and D. P. Mandic, “Augmented Complex Common Spatial Patterns for Classification of Noncircular EEG From Motor Imagery Tasks,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 22, no. 1, pp. 1–10, Jan. 2014. H. Kang and S. Choi, “Bayesian common spatial patterns for multisubject EEG classification,” Neural Networks, vol. 57, pp. 39–50, Sep. 2014. H. Rehbaum and D. Farina, “Adaptive common average filtering for myocontrol applications,” Med Biol Eng Comput, vol. 53, no. 2, pp. 179–186, Feb. 2015.

[19] P. L. Nunez and R. Srinivasan, Electric Fields of the Brain: The Neurophysics of EEG. Oxford University Press, pp. 282-287, 2006. [20] S. P. Fitzgibbon et al., “Surface Laplacian of scalp electrical signals and independent component analysis resolve EMG contamination of electroencephalogram,” International Journal of Psychophysiology, vol. 97, no. 3, pp. 277–284, Sep. 2015. [21] G.R. Naik and W. Wang, “Blind source separation,” Berlin, Heidelberg, Springer, pp. 151-193, 2014. [22] P. Goupillaud, A. Grossmann, and J. Morlet, “Cycle-octave and related transforms in seismic signal analysis,” Geoexploration, vol. 23, no. 1, pp. 85-102, 1984. [23] N. Naseer, M. J. Hong, and K.-S. Hong, “Online binary decision decoding using functional near-infrared spectroscopy for the development of brain–computer interface,” Exp Brain Res, vol. 232, no. 2, pp. 555–564, Feb. 2014. [24] C. Cortes and V. Vapnik, “Support-vector networks,” Mach Learn, vol. 20, no. 3, pp. 273–297, Sep. 1995. [25] M. Anthony and P. L. Bartlett, “Neural Network Learning: Theoretical Foundations,” [26] M. B. I. Reaz, M. S. Hussain, and F. Mohd-Yasin, “Techniques of EMG signal analysis: detection, processing, classification and applications,” Biological Procedures Online, vol. 8, no. 1, pp. 11–35, Dec. 2006. [27] E. Parvinnia, M. Sabeti, M. Zolghadri Jahromi, and R. Boostani, “Classification of EEG Signals using adaptive weighted distance nearest neighbor algorithm,” Journal of King Saud University - Computer and Information Sciences, vol. 26, no. 1, pp. 1–6, Jan. 2014. [28] K. Assaleh and W. M. Ahmad, “Modeling of speech signals using fractional calculus,” Proc. 9th International Symposium on Signal Processing and its Applications, Sharjah, UAE, pp. 1-4, Feb. 2007. [29] W. Wu, S. Nagarajan, and Z. Chen, “Bayesian Machine Learning: EEG/MEG signal processing measurements,” IEEE Signal Processing Magazine, vol. 33, no. 1, pp. 14–36, Jan. 2016. [30] “13.Electroencephalography.” [Online]. Available: http://www.bem.fi/book/13/13.htm. [Accessed: 22-Mar-2017].

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

17

An Ultra-Low Power modified SERF Adder in Subthreshold region for Bio-medical Applications Pradeep Sarva

Santosh Koppa

Eugene John

Department of Electrical and Computer Engineering, UTSA San Antonio, TX 78249, USA [email protected]

Department of Electrical and Computer Engineering, UTSA San Antonio, TX 78249, USA [email protected]

Department of Electrical and Computer Engineering, UTSA San Antonio, TX 78249, USA [email protected]

Abstract— In this paper, we propose an ultra-low power modified static energy recovery full adder operating in sub-threshold region with full range output voltage swing. The full adder proposed consumes less energy per addition compared to the other existing full adders. Since the speed requirements are low for implantable medical devices, only the power needs are considered while designing this full adder for this application. At supply voltage of 0.25 V, this adder consumed 2.184 fJ per addition with restored output voltages. The proposed full adder consumed less than 50% of energy required per cycle compared to the conventional static energy recovery full adder. A comparative power-performance analysis is performed with other adders by using a 32-bit ripple carry adder and it consumes only 19.67 nW power per cycle with rail-rail output swing operating in sub-threshold region. Keywords – Full Adder, SERF, Sub-Threshold, Energy recovery, Ripple carry adder.

I.

INTRODUCTION

The growing demand for energy-efficient bio-medical devices such as pacemakers, implantable cardioverter defibrillator, hearing aids etc. have increased the need for the design of new circuits and systems to improve their battery life and hence their operating life [1]. The speed requirements for bio-medical devices are low and hence these devices are designed keeping only power budget as the main constraint. One of the effective ways to improve the operating life of these devices without compromising on the battery size is to design new low-power circuits and systems. Designing systems aiming for low power is not a straight forward task, as it involves integrating low-power techniques in all the IC design stages from system behavioral description to the fabrication and packaging processes [2]. There are three major components of power dissipation in CMOS circuits, given by Equation (1) which is the summation of switching power, short circuit power and static power. 𝑃𝑡𝑜𝑡𝑎𝑙 = 𝛼𝐶𝑉𝑑𝑑 2f + 𝐼𝑠𝑐 𝑉𝑑𝑑 + 𝐼𝑠𝑡𝑎𝑡𝑖𝑐 𝑉𝑑𝑑 𝐸𝑡𝑜𝑡𝑎𝑙 = 𝛼𝑁𝐶𝑉𝑑𝑑 2 +

𝑁𝐼𝑠𝑡𝑎𝑡𝑖𝑐 𝑉𝑑𝑑 𝑓

(1) (2)

where 𝐶 is load capacitance ; 𝛼 is switching activity factor; f is system clock frequency; 𝑉𝑑𝑑 is supply voltage; 𝐼𝑠𝑐 is short

circuit current, 𝐼𝑠𝑡𝑎𝑡𝑖𝑐 is the leakage current, and N is gate count. Supply voltage scaling has been the most adopted approach to power optimization due to its quadratic relationship to the power dissipation [3]. Reduced transistor count also contributes to significant power savings. Complex blocks built using low transistor count gates (N) operating at lower supply voltages have an edge over other designs in terms of smaller silicon area and lower energy consumption, modeled in Equation (2). Combining these techniques result in maximum energy savings for low frequency bio-medical devices. Current generation bio-medical devices consist of computational and power intensive units such as signal processing units, ALU, FFT, multipliers etc. The basic building block of these units are full adders and hence focus on reducing the power consumption of these adders impact the whole system performance and power consumption. Over the years, several full-adder circuits were proposed with a primary focus on transistor count. Given the fact that the addition process is the basis of subtraction, multiplication and division operations, designing these adder circuits with low transistor count is essential. Pass transistor logic (PTL) enables one to perform this logic operation with lower transistor count than CMOS logic. However, the threshold-voltage loss problems in PTL circuits needs level restoration at the gate outputs which is necessary to avoid static currents at the subsequent logic gates [4]. An extensive analysis of the existing low power 10transistor static energy recovery full-adder (SERF) [10] revealed threshold-voltage loss problems. As a solution, a new full adder cell is designed by modifying the existing SERF circuit. This paper presents the new modified SERF adder cell for ultra-low power sub-threshold operation in implantable bio-medical devices. The efficiency and scalability is tested by implementing a 32-bit ripple carry adder (RCA) block. A comparative analysis of both single-bit and 32-bit addition processes with other adder cells as a building block is also performed. All the simulations are performed using NCSU 45nm SPICE models. The rest of this paper is organized as follows. Section II describes and analyzes the SERF adder. Section III describes and analyzes the circuit design of the new modified SERF

ISBN: 1-60132-451-0, CSREA Press ©

18

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

adder. In Section IV, the proposed circuit is compared with other adder designs in terms of power, delay and energy consumption. In Section V, a 32-bit adder operating in subthreshold region is implemented using existing full adder cells and the simulation results are compared. Finally, Section VI concludes the paper. II.

THE SERF ADDER

The conventional SERF adder block diagram and transistor schematic is shown in Fig.1. Reportedly it has the lowest power consumption among all the existing full adder circuits. It was inspired by the 4T XNOR gate full adder design. The 4T XNOR gate with no direct path to the ground forms the heart of the SERF adder for generating the intermediate signal ‘F’ and the output signal ‘Sum’. PTL based multiplexer circuit consisting of single PMOS and NMOS devices was used for generating ‘Cout’ signal.

reason, the output voltage level rises only to Vdd-Vtn [4]. For the input vectors ABCin=110 and ABCin=111, the voltage loss problem in SERF deteriorates due to multiple nodal voltage drop. The voltage loss at intermediate node F causes further degradation of output signals. Vectors ABCin=110, ABCin=111 are identified as the worst-case vectors. Table I shows the intensity of this problem for possible input combinations. TABLE I: TRUTH TABLE OF SERF ADDER WITH LOGIC PROPAGATION.

ABCin 000 001 010 011 100 101 110 111

Node(F) GOOD 1 GOOD 1 GOOD 0 GOOD 0 GOOD 0 GOOD 0 BAD 1 BAD 1

Sum

Cout

GOOD 0 BAD 1 GOOD 1 GOOD 0 GOOD 1 GOOD 0 GOOD 0 BAD 1

GOOD 0 GOOD 0 BAD 0 GOOD 1 BAD 0 GOOD 1 BAD 1 BAD 1

Conventional Level Restoration Circuits: The output signals need to be level restored for the worst-case input combinations which even worsens during the sub-threshold operations for biomedical applications. Block diagram of SERF adder with level restorers is shown in Fig.2. For implementing designs using SERF adder as a building block, highly robust form of modified SERF adder is desired. Conventional techniques in the form of external level restoring circuits such as buffers or NMOS-only level restoring circuits can be adopted [3].

(a)

Figure 2. Block diagram of SERF adder with conventional level restorers.

(b) Figure 1. SERF adder (a) Gate level Block Diagram, (b) Transistor Schematic.

In non-energy recovery designs, the charge applied to the load capacitance during logic level high is drained to ground during the logic level low [10]. The elimination of a direct path to the ground removes the short circuit component of total power 𝑃𝑡𝑜𝑡𝑎𝑙 , making this combination an energy efficient design. However, the elimination of direct path to the ground distorts the outputs and hence the outputs requires level restoration. Threshold-Voltage Losses in SERF: In SERF, the output signal strength has a proximity relationship with the MOS device that drives the logic to charge or dis-charge the load capacitance. The output signal suffers from a significant voltage drop when an NMOS device passes logic 1. Due to this

The back to back connected inverters called buffers restore the output voltage swing and it is used in the design to maintain the signal integrity. The availability of different range of inverter sizes makes a commonly adopted practice of level restoration. For NMOS only logic, PMOS-only level resorting circuit helps to restore the degraded output logic 1. Because the PMOS transmits only good logic 1’s, the actual output load capacitance is charged back to Vdd . It offers full logic swing. However, it is more complex due to large capacitances. Sizing of the PMOS transistor is also a big challenge. III.

THE MODIFIED SERF ADDER

As a solution to the threshold voltage loss problems in SERF, a new circuit design is proposed by modifying the existing transistor schematic. The modified circuit design of SERF consists of XNOR and a 2:1 multiplexer based on transmission gates. The modified SERF adder block diagram is shown in Fig.3. XNOR Circuit using Transmission Gates: Each NMOS transistor connected to output node “Sum” in XNOR module

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

19

is replaced with a transmission gate. The inverted form of internal signal “F” or input signal “Cin” is connected to the gate terminal of each PMOS device which passes good high voltage. So, on applying the worst-case input combinations, the voltage level of the output signal “Sum” is restored to full swing.

Figure 3. Block diagram of proposed modified SERF adder.

Figure 5. Transistor schematic of modified SERF adder.

2:1 Multiplexer using Transmission Gates: The 2:1 multiplexer circuit driving the output signal net “Cout” is replaced with transmission gates connected in parallel. For a common control input connected to both the switches, either of the input signals propagates to the output. The control input “F” and its inversion "𝐹̅ " are the gate inputs of MOS devices in each transmission gate. When F = 0, Cout = Cin, else Cout = A. Transmission gate based XNOR and 2:1 multiplexer is shown in Fig.4.

(a)

(b)

Figure 4. Transmission gate based (a) XNOR, (b) 2:1 multiplexer.

Analysis of Modified SERF: The circuit schematic of the proposed new SERF adder is shown in Fig.5. Robustness requirement in SERF is evaluated by measuring its output voltages representing high logic VOH and low logic VOL while driving equivalent loads. Simulations are performed at lower supply voltages applying worst case input combinations. The output waveforms shown in Fig.6 prove the resilience of the output signals in the proposed design to the thresholdvoltage losses while the conventional SERF had distorted outputs when operating in sub-threshold region. At Vdd =0.3 V, the SERF adder exhibits 70% transition rates. The proposed circuit design shows 99% transition rates with significantly higher noise margins.

Figure 6. Input/output waveforms of SERF and Modified SERF.

The adder was simulated and analyzed in SS, TT & FF corners to overcome the challenges involved in the process variations. Simulations are run at corner dependent temperatures (0-125) oC at multiple supply voltages. The product of average power and worst case delay was used to estimate the energy consumption or the power-delay product (PDP). The circuit consumes maximum energy in SS corner operating at 125o C temperature. The circuit when operating in FF corner and at 0o C shows the best energy savings. Analysis in TT corner at 25o C shows ideal energy savings. The obtained PDP curves in Fig.7 prove the energy efficiency of the proposed circuit design.

ISBN: 1-60132-451-0, CSREA Press ©

20

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

the product of the consumed power and latency, is expected to be minimal at a point [11]. In other words, the least possible amount of energy incurred is identified for ideal Vdd [12], [13]. The PDP or energy consumption curves of adders are shown in Fig.9. Unlike SERF, the other circuit designs have full output voltage swing at Vdd =0.3 V. Average Power: It is worth mentioning that the 14-T Transistor, TF and SERF adders have the advantage of lower power dissipation. At Vdd=0.3 V, DVL has the lowest power dissipation and CPL adder has the higher power dissipation compared to all others designs. The proposed circuit has equivalent power dissipation to that of 14-T and TF adders. Figure 7. PDP curves of Modified SERF.

The proposed design was compared with other forms of level restoration techniques as described earlier. Using a similar setup, simulations were performed at lower supplyvoltages. Fig.8 shows the simulated results. The SERF adder with NMOS-only level restoration circuit shows the highest energy consumption. The proposed work outperforms the other techniques by ensuring lowest energy consumption, significant performance improvement and lowest power consumption than the other two restoration techniques. This energy savings proves the usefulness of the proposed design for implementing complex arithmetic units.

Delay: The proposed modified SERF adder shows comparable performance at near-threshold voltages than at sub-threshold voltage. The DVL adder has a very high latency at lower power supply-voltages. At Vdd = 0.3 V, DVL circuit has an average propagation delay of 45.6 ns. CPL adder is close to DVL in terms of delay. The proposed adder circuit had better performance attributes when compared to 14-T and TF cells. TG-CMOS and CMOS circuits take longer time to charge and discharge the load capacitances than proposed SERF. Simulation results of all the adders are tabulated in increasing order of PDP in Table II.

Figure 8. Comparative Analysis of Modified with other Level Restorers in terms of Energy Consumption.

IV.

COMPARATIVE ANALYSIS OF 1-BIT ADDERS

For comparative analysis, the other full adders considered are 32T complimentary pass-transistor logic (CPL) [2], 28T [2] CMOS full adder, 20T dual voltage logic (DVL) adder [7], 20T transmission gate (TG-CMOS) [2] adder, 16T transmission function (TF) [8] adder and 14-T [9] full adder. A comparative analysis of these designs is performed at lower supply voltages targeting the near-threshold and subthreshold regions. However, operating in such region requires detailed analysis for obtaining the minimum operating point. As the operating voltage is reduced, the benefits of power savings exceed the cost of delay incurred, resulting in overall energy savings which is highly required for biomedical applications. However, this trend reverses when the operating voltage reduces beyond a threshold point. Energy dissipation, which is

Figure 9. PDP curves of 1-bit full adders

Energy Consumption: At lower supply-voltages, the energy consumptions of these designs seem to increase due to increased delay dependency. At Vdd = 0.3 V, CPL design incur excessive power dissipation than DVL, resulting in higher energy consumption. The proposed modified SERF adder cell offers significant energy savings than TG-CMOS, CMOS, DVL, CPL and SERF adder designs. It has its ideal operating point between VDD = 0.4~0.3 V. The lower transistor count 14-T and TF designs offer more energy savings. However, from Fig.9 these low power designs have higher energy consumption as the operating voltage is reduced. They have minimum operating points at higher voltages.

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

TABLE II: SIMULATION RESULTS AT VDD=0.3V, SORTED BY PDP

21

modified SERF adder based 32-bit RCA consumed only 4.80 fJ at 0.19 V. All the results are tabulated in Table IV.

Figure 10. PDP curves of 32-bit RCA using different full adder cells. TABLE III: COMPARATIVE ANALYSIS OF 32-BIT RCA.

V.

COMPARATIVE ANALYSIS OF 32-BIT ADDERS

With improved signal strengths and ultra-lower power operation, the new adder circuit proves to be highly suitable for energy constrained bio-medical devices. A 32-bit RCA using the modified SERF adder is implemented to test the adder cell’s power efficiency. The designed 32-bit adder is also compared with other 32 bit RCA realized using various other adder cells used for comparison in this research as a building blocks. All the analysis is performed at sub-threshold voltages. The 14-T, TF, TG-CMOS and CMOS adders which exhibited better energy savings when used as a single bit adder as described in previous section are chosen for comparison. The PDP curves of the implemented designs under similar setup are shown in Fig.10. The RCA using modified SERF had decreasing energy trend as the supply voltage moves towards sub-threshold operation. This signifies that the modified SERF adder has lower energy consumption in sub-threshold region suitable for bio-medical applications. The 14-Transistor, TF and CMOS based RCA exhibited increase in energy consumption as the supply voltage was reduced. Table III tabulates the simulation results at Vdd = 0.25V. The 32-bit adder realized using the modified SERF has better energy savings when compared to TG-CMOS and CMOS based adders. Comparison with the Previous Work: Fin-FET based 32bit adder design operating at sub-threshold region [14] consumed 10 fJ per computation at 0.4 V while the proposed modified SERF adder consumes only 2.184 fJ per computation at 0.4 V. The energy of a 32-bit adder designed in [11] consumed 34 fJ per computation at 0.37 V while the proposed modified SERF adder cell operating at 0.37 V consumed only 2.096 fJ per computation. A 32-bit TG full adder implemented in sub-threshold region [5] consumed 0.4 pJ at 0.19 V with a maximum operating frequency of 10MHz while the proposed

TABLE IV: COMPARISON WITH PREVIOUS WORK SORTED BY ENERGY CONSUMPTION

Technology

Vdd (V)

Energy Consumption

45nm CMOS

0.4

2.18 fJ

32-bit Adder [14]

45nm Fin-FET

0.4

10 fJ

32-bit Adder [11]

45nm CMOS

0.37

34 fJ

32-bit TG Adder [5]

90nm CMOS

0.19

0.40 pJ

Design Proposed 32-bit SERF RCA

ISBN: 1-60132-451-0, CSREA Press ©

%Energy savings

-----

78.16%

93.83%

98.8%

22

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

VI.

CONCLUSIONS

In this paper, an ultra-low power and energy efficient modified SERF adder cell is presented. Unlike the other PTL based adder cells, this design has rail-rail output voltages even at extremely low operating voltages. The ultra-low power consumption of the proposed modified SERF adder results in significant energy savings. Analysis proved the robustness and tolerance of the proposed circuit design to process variations at extremely low supply voltages. Since the adder circuit is energy efficient and highly robust, it can be integrated as a useful component in the low power design for biomedical devices. The implementation of a 32-bit adder based on modified SERF adder cell showed up to 5 times energy savings when compared to other approaches. The energy consumption is estimated to range between 2 fJ to 10 fJ for 45nm process variations. Finally, the presented work highlights the practical usefulness of a modified SERF adder and delivers a viable option for saving both power and energy in complex arithmetic blocks in IC design for bio-medical applications.

[11] Anh T. Tran and Bevan M. Baas, “Design of an Energy-Efficient 32-bit Adder Operating at Subthreshold Voltages in 45-nm CMOS”, International Conference on Communications and Electronics (ICCE), 2010. [12] B. H. Calhoun and A. Chandrakasan, “Characterizing and modeling minimum energy operation for subthreshold circuits,” in Proc. IEEE Int. Symp. Low Power Electronics and Design (ISLPED), 2004, pp. 90– 95. [13] B.H. Calhoun, A. Wang, A. Chandrakasan, “Modeling and sizing for minimum energy operation in subthreshold circuits”, Proceedings of the IEEE, Sep. 2005, pp. 1778-1786. [14] Mohsen Jafari, Mohsen Imani, Mohammad Ansari, Morteza Fathipour and Nader Sehatbakhsh, “Design of an Ultra-Low Power 32-bit Adder Operating at Subthreshold Voltages in 45-nm FinFET”, International Conference on Design & Technology of Integrated Systems in Nanoscale Era, 2013. [15] Eng Sue Chew, Myint Wai Phyu, Wang Ling Guo,” Ultra Low –power full adder for biomedical applications”, IEEE International Conference on Electronic Devices and Solid-State Circuits , 25-27 December , 2009 ,pp.115-117.

Acknowledgement Research reported in this paper was supported in part by National Institute of General Medical Sciences of the National Institutes of Health under award number 1SC3GM09693701A1. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health REFERENCES [1]

Xin Liu, Yuanjin Zheng, Myint Wai Phyu, Bin Zhao, Minkyu Je and Xiaojun Yuan, “Multiple Functional ECG Signal is Processing for Wearable Applications of Long-Term Cardiac Monitoring”, IEEE Transactions on Biomedical Engineering, vol. 58, no. 2, February 2011. [2] A. M. Shams, T.K. Darwish, and M. A. Bayoumi, “Performance analysis of low-power 1-bit CMOS full adder cells”, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 10, no. 1, pp. 20– 29, 2002. [3] J.M. Rabaey, A. Chandrakasan, B. Nikolic, "Digital integrated circuits: a design perspective”, 2nd.Ed” Prentice-Hall, 2002. [4] Reto Zimmermann and Wolfgang Fichtner, “Low-Power Logic Styles: CMOS Versus Pass-Transistor Logic”, IEEE Journal of Solid-State Circuits, vol. 32, no. 7, July 1997. [5] Nele Reynders and Wim Dehaene, “Variation-Resilient Building Blocks for Ultra-Low-Energy Sub-Threshold Design”, IEEE Transactions on Circuits and Systems—II: Express Briefs, vol. 59, no. 12, December 2012. [6] H.Soeleman and K. Roy, “Ultra-low-power digital subthreshold logic circuits,” Proc. Int. Symp. Low-Power Electronics Design, 1999. [7] V. G. Oklobdzija, M Soderstrand and B. Duchene “Development and Synthesis Method for Pass-Transistor Logic Family for High-Speed and Low Power CMOS” Proceedings of the 1995 IEEE 38th Midwest Symposium on Circuits and Systems, Rio de Janeiro, 1995. [8] Nan Zhuang and Haomin Wu, “A New Design of the CMOS Full Adder”, IEEE Journal of Solid-State Circuits, Vol. 27, no. 5, May 1992. [9] Ahmed M. Shams and Magdy A. Bayoumi, “A New Full Adder Cell for Low-power Applications”, Proceedings of the IEEE Great Lakes Symposium on VLSI, 1998, pp. 45-49. [10] R. Shalem, E. John and L. K. John, “A Novel Low-Power Energy Recovery Full Adder Cell”, in Proc. IEEE Great Lakes VLSI Sym., pp.380-383, Feb 1999.

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

Digital Compensation Method for Enhancing the Blood Pressure Estimation based on the Arterial Pulse Waveform Boyeon Kim1 and Yunseok Chang2 1

Division of Computer Science & Engineering, Hanyang University, Seoul, Korea, 04763 2 Department of Computer Engineering, Daejin University, Pocheon, Korea, 11159

Abstract - Blood pressure measurement method by using the PPG sensor and arterial pulse waveform has a prior condition that the sensor has to maintain constant pressure on the artery all over the measurement process, and it is very hard under the 24-hours continual measurement environment. To get the reliable blood pressure measurement result, we have to find a way to compensate the sensor pressure variation over the artery. In this work, we proposed the new digital compensation method that can measure the arterial pulse waveform and sensor pressure simultaneously and can compensate the blood pressure estimation equation along with the sensor pressure variation to find out more precise blood pressure. we also designed and implemented the arterial pulse waveform measurement system includes digital PPG sensor, ADC, Raspberry PI 3 and smartphone app that has new blood pressure estimation equation that can compensate the sensor pressure variation on the differential values of the arterial pulse waveform and BP-relation equation. The android smartphone app processes the arterial pulse waveform from the Raspberry PI 3 through the Bluetooth link and estimates more accurate blood pressure with the proposed estimation equation. To evaluate the estimation accuracy, we implemented the experimental environment and compared the experimental results to the commercial tonometer. The experimental results showed the proposed estimation equation can improve the estimation accuracy up to 5.5% on systolic and 30% on diastolic for hypertension and normal volunteers, and can make continuous estimation within 5% error rate for hypertension. Keywords: Digital compensation, Blood pressure estimation, Arterial pulse waveform, BP-relation equation, Digital tonometer

1

Introduction

A traditional Kortokoff-type tonometer oppresses the artery to block and release the blood current that would hurt the vessel, it cannot apply to the continuous blood pressure measurement.[1] Since non-Kortokoff-type blood pressure

measurement device using PPG sensor such as a smart watch has lower pressure on the vessel or artery than oscillometric tonometer, it can be applied on the 24-hours continuous blood pressure measurement area but affected by sensor movement and sensor pressure variation that can increase the measurement error rate.[2] Recently some kind of smart watch or smart products uses PPG sensor on heart rate or blood pressure measurement apps based on the arterial pulse waveform analysis. Since the arterial pulse waveform can directly reflect the blood pressure variation in the vessel, PPG sensor can provide an easy way to measure the arterial pulse waveform with simple circuit and device. In continuous blood pressure measurement method by using PPG sensor and arterial pulse waveform has a prior condition that the sensor has to maintain constant sensor pressure all over the measurement process. But in clinical or real life environment, there are many sensor pressure variations during 24-hours continuous measurement due to the patient’s movement that can make irregular sensor pressure and abnormal arterial pulse waveform. To solve this problem, we have to measure arterial pulse waveform and the sensor pressure variation simultaneously. On the contrary of cuffs type tonometer, PPG sensor oppresses the vessel with very weak pressure to sustain the continuous measurement that makes it difficult to maintain a constant sensor pressure. Therefore, digital compensation method against the sensor pressure variation is more effective than maintaining constant sensor pressure continuously. The previous works on the continuous arterial pulse waveforms measurement used the wristband structured cuff that has air-pressure sensors, CPU, Bluetooth device and battery inside as shown in Fig. 1[3]. But this cuff has expensive and complicate structure, and the air-pressure sensor could be easily affected by the wrist movement. But the PPG sensor can be installed at the end of the index finger and has less movement and noise than wristband structure. Therefore we use arterial pulse waveform measurement method based on the PPG sensor and smart device instead of previous wristband structured cuff in this work, and propose new digital compensation method against the sensor pressure variation that can enhance the blood pressure estimation method more stable and precisely than previous works.

ISBN: 1-60132-451-0, CSREA Press ©

23

24

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

improve the measurement method with only air pressure parameter without any other options. In this paper, we focused on the arterial pulse waveform analysis and digital compensation method to estimate the blood pressure continuously instead of blocking blood current inside the vessel.

3

Blood pressure estimation method

Fig. 1 Air-pressure cuffs sensor vs. PPG sensor The remaining part of this paper is organized as follows: Section 2 describes the related works and studies concerned with the proposed method; in Section 3 we proposed digital compensation method and reliability enhancement method on BP-relation equation between the arterial pulse waveform and blood pressure and in Section 4 we provide the experimental environment that carries the proposed method; Section 5 contains analysis of the proposed method including accuracy enhancement of the blood pressure estimation equation and estimation results; the paper is concluded with some summarizing remarks and further work in Section 6.

2

Related works

For a long time, the traditional Kortokoff-type tonometer has been used in the clinical field. Recently, many works are focused on the non-Kortokoff-type blood pressure measurement method or non-invasive way from the outside of blood vessel. Some Japanese researchers had studied a kind of non-invasive method that can measure the blood pressure from the blood vessel inside [4]. Another research studied on the remote technique to monitor the patient’s blood pressure and health status through the wrist-banded type sensor device [5]. Although the existing wrist-type tonometer based on the oscillometric method still oppresses the artery of the wrist by using the air-cuff at every blood pressure measurement trial, the electronics sensor based non-invasive methods measure the blood pressure with the pressure value from the sensor unit [6]. Since non-invasive method could not guarantee the accuracy within the acceptable error range, PPG and ECG signal based electronics techniques are upcoming in the recent medical device field.[7][8] Another digital technique by using PTT(Pulse Transit Time) can also be adopted to estimate the blood pressure without cuffs or air pressure. Since the PTT can reflect the blood pressure variation inside vessel directly, it can also apply to the blood pressure estimation with the same as the arterial pulse waveform. But PTT has a drawback that is very noise-sensitive compare to the arterial pulse waveform and needs many noise filtering techniques.[9][10] The u-Healthcare area is one the most important digital technique areas. Many electronics and computer techniques are connected to the u-Healthcare devices and products. For example, recent researches on the digital mobile tonometer have focused on Bluetooth link or Wi-Fi connection with other device or computer system.[11] But these researches has the approach to take the air-pressured digital sensor for blood pressure measurement and there is a severe limitation to

3.1

Differential value

The differential value can be defined as a gap ΔPi between the maximum and minimum value of waveform Wi(W1, W2, W3, W4) in the continuous arterial pulse waveform W showed in Fig. 2. For example, ΔP1 and ΔP4 show the differential values corresponding to W1 and W4. In continuous arterial pulse measurement, there are lots of arterial pulses in the waveform that can get so many differential values as arterial pulses.[12] If there are many differential values from the arterial pulse waveform and lots of blood pressures corresponding to the arterial pulse measurement, they can be used to calculate the real blood pressure including systolic and diastolic with the differential value from the arterial pulse waveform through some kind of relation equation.

Fig. 2 Differential Values ΔP1 and ΔP4 from Arterial Pulse Waveform Wi in W

3.2

BP-relation equation

The BP(Blood Pressure)-relation equation can show the correlation between the differential values from the arterial pulse waveform and blood pressures from the commercial tonometer. It can be deduced through the regression with lots of blood pressures and arterial pulse waveforms from many volunteers.[13] As a result of regression, we can define the relation between the differential value and blood pressure, and can complete the BP–relation equation to the differential values ΔP as follows: Systolic(ΔP) = Regression(Systolicavg, ΔPavg)

(1)

Diastolic(ΔP) = Regression(Diastolicavg, ΔPavg)

(2)

ΔPavg: Average of ΔPi, i = 0, .., n (n: number of peaks Wk in W) Systolicavg: Average systolic value of a volunteer Diastolicavg: Average diastolic value of a volunteer

To get the ΔPavg, Systolicavg, and Diastolicavg, we have to measure the blood pressure and arterial pulse waveform

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

simultaneously from many volunteers, and analyze them to get the average differential value ΔPavg from the arterial pulse waveform and average blood pressures Systolicavg, Diastolicavg from the commercial tonometer. The regression process can find the two BP-relation equations (3) and (4) with ΔPavg, Systolicavg, and Diastolicavg.as a linear function and can complete the angle S_angle, D_angle and offset S-offset, D_offset. Systolic(ΔPavg) = S_angle * ΔPavg +S_offset

(3)

Diastolic(ΔPavg) = D_angle * ΔPavg +D_offset

(4)

These two BP-relation equations compose the blood pressure estimation equation. Once we complete the blood pressure estimation equation, we just only need one’s ΔPavg to estimate the systolic Systolic(ΔPavg) and diastolic Diastolic(ΔPavg) by using the blood pressure estimation equation. If we get the ΔP1 and ΔP2 from one’s arterial pulse waveform, we can directly estimate the Systolic1 and Systolic2 from the BP-relation equation for systolic from the blood pressure estimation equation as shown in Fig. 3.

Fig. 3 Blood pressure estimation with regression graphs of systolic vs ΔPavg

3.3

Digital compensation method

Blood pressure measurement method by using the PPG sensor and arterial pulse waveform has prior condition that the sensor has to maintain constant pressure on the artery all over the measurement process. When the PPG sensor pressure is maintaining almost the same initial sensor pressure to complete the BP-relation equation, the blood pressure estimation equation can estimate near the accurate blood pressure through the BP-relation equation. But under the real situation, the PPG sensor pressure would be varied according to the patient’s movement or activities during the 24-hours continuous measurement. The PPG sensor pressure variation could change the angle and offset of the BP-relation equation that produces inaccurate results. Therefore, we have to compensate the variation of the initial PPG sensor pressure to the BP-relation equation and blood pressure estimation equation within the allowable range.

25

If we assume the initial PPG sensor pressure as Pinit, and let the actual sensor pressure during the arterial pulse measurement as Pair, the real pressure gap between Pinit and Pair can be calculated as ΔPdiff as like the equation (5). Therefore the differential value ΔPi of the arterial pulse Pi has to be compensated as ΔPk_diff in equation (6). ΔPdiff = Pinit - Pair

(5)

ΔPk_diff = ΔPk - ΔPdiff

(6)

ΔPavg_diff = avg(∑ΔPk_diff)

(7)

Since the average differential value ΔPavg can be compensated as the ΔPavg_diff as the equation (7), the final blood pressure estimation equation can also be modified as equation (8) and (9). Systolic(ΔPavg_diff) = S_angle * ΔPavg_diff +S_offset

(8)

Diastolic(ΔPavg_diff) = D_angle * ΔPavg_diff +D_offset (9)

3.4

Digital waveform filtering

The heart rate is an important checkpoint factor in the continuous arterial pulse measurement with digital PPG sensor.[14] To get the high blood pressure estimation accuracy by using arterial pulse waveform, the waveform has to be measured under the stable environment. If there are sorts of movement or activity to affecting on the PPG sensor pressure during the measurement process, the sensor pressure would be varied and produce the abnormal waveform. At the moment of the abnormal waveform processing, the smartphone app can find the rapid irregular heart rates. If we use some commercial smart device that can check the heart rate along with the arterial pulse waveform measurement, the abnormal arterial pulse waveforms with the irregular heart rates can be easily filtered by comparing to the heart rate from the smart device. In this work, we use two smart devices to filter the abnormal waveform during continuous arterial pulse waveform measurement by checking the irregular heart rates. We applied built-in heart rate check application on Apple Watch and Samsung Galaxy Gear S2 as shown in Fig. 4. At every arterial pulse waveform measurement process, one of these smart devices is applied at the same time.

Fig. 4 Smart devices for heart rate measurement: in case of heart rate check app on the Apple Watch

ISBN: 1-60132-451-0, CSREA Press ©

26

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

The waveform filtering process with smart devices is also applied on the BP-relation equation regressing process. If there are some interval that has irregular heart rates over the allowable error range during the arterial pulse measurement, corresponding arterial pulse waveform interval could be considered as abnormal and cut or abandoned from the effective waveform data. As an effect of the waveform filtering, the differential values within the allowable error range could be applied to BP-relation equation construction. Fig. 4 and Table 1 show the result of the waveform filtering on the BP-relation equation according to the allowable error range. Since the smaller the allowable error range produces the more accurate BP-relation equation, blood pressure estimation results are also more accurate through the enhanced equations. Fig. 5a shows the experimental results of the BP-relation equation with allowable error ranges 5% and 10% for systolic. Regression angle and offset of two results are almost the same in these cases and also have less standard deviation than 0.05, the blood pressure estimation equation can predict the systolic within 95% reliability. Otherwise, in the diastolic cases in Fig. 5b, angle and offset of two results are somewhat different, and in the case of 10% allowable error range has bigger standard deviation than in the case of 5%. Therefore, less allowable error range could have the acceptable the diastolic accuracy through the reliable diastolic BP-relation equation. Table 1 Results of regression for BP relation equation with 10% and 5% allowable error range Value Factors

St. deviation

10%

5%

10%

5%

Systolic_angle Systolic_offset

0.649 98.21

0.672 98.34

0.049

0.036

Diastolic_angle Diastolic_offset

0.376 57.38

0.281 59.73

0.116

0.067

(b) BP relation regression for Diastolic Fig. 5 Regression graphs of blood pressure vs. ΔPavg_diff with allowable error range within 10% and 5%

4 4.1

Experimental environment Arterial pulse waveform measurement

The arterial pulse waveform measurement system composes the PPG sensor, ADC, and Raspberry PI 3 module. Raspberry PI 3 module sends the arterial pulse waveform data to the Android measuring app via the Bluetooth link. The measuring app processes the arterial pulse waveform data and converts to the blood pressure estimation value including systolic, diastolic and heart rate. Fig. 6 shows the experimental environment in this work.

Fig. 6 Implemented Measurement System (Sensor, ADC, Raspberry PI3 Module)

(a) BP relation regression for Systolic

The PPG sensor detects the blood current inside the vessel and outputs the integral type analog signal. In this work, we use the RP520 sensor unit from Laxtha Co. as the PPG sensor.[15] Since the RP520 sensor is similar to the SPO2 sensor unit and finger-ring type structure that can put on the index finger, it does not oppress the artery or vessel so much

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

as the Kortokoff-type tonometer and can measure the arterial pulse waveform continuously without any harm on the vessel. Although the analog output from the RP520 sensor unit cannot connect to the Raspberry PI 3 directly, it can be connected to the Raspberry PI 3 via external ADC module such as ADC0809. Raspberry PI 3 module processes the digitized arterial pulse waveform, filters noise and sends the waveform data to the smartphone app through the Bluetooth link. The measuring app on the Android smartphone displays the waveform on the screen and checks the number of peaks continuously. When peak count reaches the dedicated number, the measuring app estimates the systolic, diastolic and heart rate, and displays the results by the blood pressure estimation equation. Java programming language and Android Studio environment are applied to build the measuring app.[14] Table 2 shows the default values of the experimental environment parameters. Since human’s heart rate has about 60 ~ 80 bpm, we preset the analog sampling rate of ADC0809 as 150Hz, around double of average heart rate by Raspberry PI 3 pin setting and Linux data sampling program. Table 2 Environment Parameters for Arterial Pulse Measurement System Parameters

Value

ADC Sampling Rate

120Hz

ADC Data Resolution

8bit

Sensor Output Type

Analog Integral Output

DSP Processing Module

RaspberryPI 3

Bluetooth Protocol

BT4.0 (SPP)

4.2

Group setting and measuring procedure

To get an objective arterial pulse waveform measurement and blood pressure estimation, we separated all volunteers into two groups. Volunteers in Group 1 participated in completing the BP-relation equation, and volunteers in Group 2 participated to the blood pressure estimation and its reliability verification. Table 3 shows the number of volunteers consists of each group. Volunteers in Group 1 are selected from one of the social education program classes randomly. Volunteers in Group 2 are selected from the 4 classes of the undergraduate course. Before arterial pulse waveform measuring, each volunteer checked his/her own blood pressure with OMRON101, one of the typical wristband type commercial tonometer. According to the pre-checked blood pressure, volunteers are classified as Hypertension group, normal group and hypotension group according to the nursing expert’s advice. The hypertension group has systolic over 135mmHG, the hypotension group has systolic under 105mmHG and the normal group has systolic 105 ~ 135mmHG. For appropriate experiment and validation, we tried to balance the group sizes in fair. But almost volunteers in Group 2 are students affiliated to the university, there are fewer hypertension

27

volunteers than normal or hypotension volunteers. Otherwise, the majority of the Group 1 is ordinary people and can include all cases in fair. Table 3 Samples in volunteer groups Hypertension

Normal

Hypotension

Total Volunteers

Group 1

26

39

23

88

Group 2

21

71

44

136

Arterial pulse measurement and blood pressure measurement are performed 3 trials for individuals of each group. For each trial, arterial pulse waveform measurement had checked for 1 minute just after taking the arterial pulse signal on the smartphone app. Each arterial pulse waveform measurement and blood pressure measurement by using tonometer has 5 minutes interval at every trial. When we took the arterial pulse waveform, the heart rate by smart watch also checked simultaneously.

5

Results and Analysis

Arterial pulse waveforms from the raspberry PI 3 are transformed to the differential values the ΔPk for individual pulse Wk in the measuring app and finally transformed to the average differential value the ΔPavg. Average differential values of the arterial pulse waveform after compensating the sensor pressure variation ΔPavg_diff from each volunteer is to input for the BP-relation equation deduction or for blood pressure estimation equation to estimate the systolic Systolic(ΔPavg_diff) and diastolic Diastolic(ΔPavg_diff). Only the volunteers in Group 2 participated in the blood pressure estimation experiment. For all measured arterial pulse waveforms, heart rates were compared to the smart watch during differential value transform process. If there is any interval that has irregular heart rate, the corresponding arterial pulse waveform is excluded from the waveform data automatically.

Fig. 7 Compensation results for Systolic with 10% vs. 5% allowable heart rate error range

ISBN: 1-60132-451-0, CSREA Press ©

28

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

BP-relation equation along with the sensor pressure variation.

Fig. 8 Compensation results for Diastolic with 10% vs. 5% allowable heart rate error range Fig. 7 and Fig. 8 show the accuracy analysis results on the blood pressure estimation with digital sensor pressure variation compensation compare to the commercial tonometer. In cases of 10% and 5% allowable heart rate error range, the case of low allowable error range can have the more accurate systolic estimation. The analysis results show that blood pressure estimation equation with digital sensor pressure compensation can estimate the systolic over 95% of accuracy for hypertension and normal volunteers but less than 90% of accuracy for hypotension volunteers. Diastolic estimation cases show less estimation accuracy than systolic estimation cases, and show 80% ~ 90% of accuracy for hypertension and normal volunteers, and even less than 70% of accuracy for hypotension volunteers. Therefore, only the systolic estimation has efficient accuracy compares to the commercial tonometer, and diastolic estimation does not show enough accuracy for all cases of blood pressure even though the digital compensation has applied on the estimation result at all. Fig. 9 and Fig. 10 show the comparison results between before and after the digital sensor pressure variation compensation on the BP-relation equation and blood pressure estimation equation. Fig. 9 shows that systolic estimation with compensation has over 5.5% accuracy enhancement compares to no compensation. The normal volunteers have more estimation accuracy than hypertension and hypotension volunteers. It shows the fact that normal volunteers are more sensitive to the sensor pressure variation than hypertension and hypotension volunteers. Therefore, the digital compensation is very important accuracy enhancement method of the blood pressure estimation for normal patients. In the case of diastolic estimation, there is a larger gap between before and after the digital compensation than the systolic case. In the case of digital compensation, diastolic estimation has 17% ~ 30% accuracy enhancement compares to the case of no compensation. Especially in the case of hypertension and normal volunteers, digital compensation shows over 80% accuracy enhancement. Therefore, Fig. 10 shows the fact that the diastolic estimation can achieve the enough estimation accuracy when digital compensation method applied on the blood pressure estimation equation and

Fig. 9 Blood pressure estimation accuracy for Systolic with 5% allowable heart rate error range

Fig. 10 Blood pressure estimation accuracy for Diastolic with 5% allowable heart rate error range In the clinical case, systolic has more important than diastolic for diagnosis and treatment. Therefore, blood pressure estimation method with arterial pulse waveform data also weights on the accuracy of systolic estimation than diastolic estimation. These analysis results show the blood pressure estimation equation with digital compensation can help enhancing the estimation accuracy of systolic and diastolic. They also show that digital compensation can enhance the systolic estimation accuracy over 95% for hypertension and normal patients. These results can be applied to the real digital tonometer for 24-hours continuous blood pressure measurement instead of the traditional Kortokoff-type tonometer.

6

Conclusion and Further works

In this work, we proposed an enhanced blood pressure estimation method with digital compensation against the sensor pressure variation. It protects to increase BP-relation

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

equation error and decrease the estimation error rate enough to the continuous blood pressure measurement. To make enough compensation on the blood pressure estimation equation, we improved the BP-relation equation and blood pressure estimation equation by using compensated differential values against the sensor pressure variation. Based on the previous work results, the proposed digital compensation can enhance the systolic estimation accuracy, and diastolic estimation accuracy also can greatly enhance for hypertension, normal patients in clinical cases. For the cases of 10% allowable heart rate error range, there is up to 24% systolic estimation accuracy enhancement and under 5% error rates during continuous arterial pulse measurement process. The basic and final goal of the digital blood pressure estimation method based on the arterial pulse waveform analysis is to make the digital tonometer into 24-hours continuous blood pressure measurement area. Although most of the products do not adopt the proposed method in the commercial field yet, it has enough possibility and can lead a new design principle of the continuous blood pressure device. In the next work, we will improve the proposed estimation method that has measurement stability and precision to fit the practical medical device requirement.

7

Acknowledgement

[6] D. Nair, S-Y Tan, H-W Gan, S-F Lim, J. Tan, M. Zhu, H. Gao, N-H Chua, W-L Peh and K-H Mak. “The Use of Ambulatory Tonometric Radial Arterial Wave Capture to Measure Ambulatory Blood Pressure: The Validation of a Novel Wrist-Bound Device in Adults”; Journal of Human Hypertension, Vol. 22 Issue No. 3, 220222, Mar 2008. [7] Youngzoon Yoon, Jung Ho Choi, and Gilwon Yoon. “NonConstrained Blood Pressure Monitoring using ECG and PPG for Personal Healthcare”; Journal of Medical Systems, Vol. 33 Issue No. 4, 261-266, Aug 2009. [8] Woosik Shin, Young Cha and Gilwon Yoon. “ECG/PPG Integer Signal Processing for a Ubiquitous Health Monitoring System”; Journal of Medical Systems, Vol. 34 Issue No. 5, 891-898, Oct 2010. [9] XR. Ding, YT. Zhang, J. Liu, WX. Dai and HK. Tsang, “Continuous Cuffless Blood Pressure Estimation Using Pulse Transit Time and Photoplethysmogram Intensity Ratio”; IEEE Transactions on Biomedical Engineering, Vol. 63 No. 5, 964-972, Sep 2015. [10] Younhee Choi, Qiao Zhang and Seokbum Ko, “Noninvasive cuffless blood pressure estimation using pulse transit time and Hilbert-Huang transform”; Computer & Electrical Engineering, Vol. 39 Issue 1, 103-111, Jan 2013.

This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MEST)(No. NRF-2015R1D1A1A01057703)

[11] Wookjae Ryu, Euntae Kim, Kyungho An, Sunghoon Woo and Yunseok Chang. “A Bluetooth based 5-HD Measurement System for u-Healthcare”; International Journal of Control and Automation, Vol. 6 No. 1, 141-150, Feb. 2013.

8

[12] Boyeon Kim and Yunseok Chang, “Digital Blood Pressure Estimation with the Differential Value from the Arterial Pulse Waveform”; In Proceedings of the International Conference of the Biomedical Engineering and Sciences(BIOENG’16), 7-12, Jul 2016.

References

[1] Bryan. Williams. "Achieving Blood Pressure Control in Clinical Practice"; Journal of the American Society of Hypertension, Vol. 2 No. 4, 10-15, Aug 2008. [2] Gen Yasuda, Nariaki Ogawa, Gaku Shimura, Daisaku Ando, Kazuhiko Shibata, Satoshi Umemura, Osamu Tochikubo. "Effects of Perindopril on 24-H Blood Pressure in Hypertensive Patients with Diabetic Nephropathy"; American Journal of Hypertension, Vol. 17 No. 5, May 2004. [3] Jae Min Kang, Taiwoo Yo and Hee Chan Kim. “A Wrist-Worn Integrated Health Monitoring Instrument with a Tele-Reporting Device for Telemedicine and Telecare”; IEEE Transactions on Instrument and Measurement, Vol. 55, No. 5, 2006. [4] Osamu Tochikubo, Yoshihiro Kaneko, Youji Yukinari, Ikuo Takeda. "Noninvasive Measurement of Baroreflex Sensitivity Index using an Indirect and Continuous Blood-Pressure Recorder"; Japanese Heart Journal, Vol. 27 No. 6, 849-857, Dec 2008.

29

[13] Gretchen Ray, James Nawarskas, Joe Anderson. “Blood Pressure Monitoring Technique Impacts Hypertension Treatment”; Journal of General Internal Medicine, Vol. 27 No. 6, 623-629, Jun 2012. [14] Boyeon Kim, Wenhai Jin, Sung Hoon Woo and Yunseok Chang. “A New Approach on Digital Blood Pressure Measurement Method for u-Healthcare Systems”; International Journal of BioScience and Bio-Technology, Vol. 7, No. 1, 169-178, Jan 2015. [15] Ukjin Lee, Hyeongyeol Park and Hyunchol Shin, “Implementation of a Bluetooth-LE Based Wireless ECG/EMG/PPG Monitoring Circuit and System”; Journal of the Institute of Electronics and Information Engineers, Vol. 51 Issue 6, 261-268, 2014.

[5] Myung Cheon. Ahn, Jong Gu Choi, Il Ho Son, Sang Seok Lee and Keun Ho Kim. “Estimated Blood Pressure Algorithm of Wrist Wearable Pulsimeter using by Hall Device”; Journal of the Korean Magnetics Society, Vol. 20 No. 3, 106-113, Jun 2010.

ISBN: 1-60132-451-0, CSREA Press ©

30

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

Using a Non-invasive Approach to Evaluate the Effects of Smoking on Swallowing and Respiration Coordination Wann-Yun Shieh1,2, Chin-Man Wang2,3, Hsin-Yi Kathy Cheng2,4 1

Department of Computer Science and Information Engineering, Chang Gung University, Taiwan 2 Department of Physical Medicine and Rehabilitation, Chang Gung Memorial Hospital, Taiwan 3 Department of Medicine, College of Medicine, Chang Gung University, Taiwan 4 Graduate Institute of Early Intervention, Taiwan

Abstract: Cigarette smoking leads to numerous diseases and other effects on the human body. Most of the ten leading causes of death reported by the World Health Organization in 2008 were related to intensive or long-term smoking. Numerous studies have focused on proving the relationships between smoking behaviors and lung or cardiovascular diseases, but few have addressed the effect of smoking on swallowing and respiration coordination. In this study, we investigated these effects by using a noninvasive, thin-filmed pressure sensor to detect the movements of the thyroid cartilage. By correlating with surface electromyography on the submental muscle and the nasal airflow cannula in the nasal cavity, we measured respiration and swallowing disruptions. The results show that smoking behavior is a major cause of swallowing disorders. Keywords: smoking; swallowing; respiration; pressure sensor; surface electromyography; nasal airflow

1. Introduction A cigarette contains nicotine, tar, and many other toxic substances. When burning, a cigarette produces thousands of compounds that can be absorbed through the nasal cavity, trachea, and finally the lungs [1]. It has been proved that longterm cigarette smoking engenders a remarkably high risk of contracting hepatic carcinoma, lung and other cancers, cardiovascular disease, and even death. When a person inhales cigarette smoke, approximately 70% of the tar attaches to the lungs, which can block and stimulate and thus damage the lung tissue. Other effects include compromising the immune system, memory degradation, decreased reproductive capacity, and deformed babies. When smoke enters the nasal cavity, the hightemperature air with harmful substances also hurt the oral and oropharyngeal tissue, which inflames the respiratory mucosa. Moreover, numerous studies have shown that the smoke can seriously compromise the pharyngeal functions of nerve endings, oral cavity mucous membranes, pharyngo-upper oesophageal sphincter contractile reflex, reflexive pharyngeal swallow, and pharynxgoglottal closure reflex [2-4], consequently leading to an impairment of swallowing

functions such that food cannot smoothly be pushed from the mouth to the stomach. In the normal swallowing process, the bolus is pushed into the esophagus by the tongue and thyroid cartilage to be conveyed to the stomach. Conventionally, this process is divided into four stages: Oral preparation stage, Oral stage, Pharyngeal stage, and Esophagus stage [5-6]. Each stage requires good coordination between the nasopharyngeal and oropharyngeal organs [7-10]. Cigarette smoking breaks this coordination because the brain stem receives faulty messages from injured sensory receptors at the oral and pharyngeal stages. Consequently, swallowing cannot follow the normal order, leading to symptoms of dysphagia. If these symptoms are not treated properly, it can lead to many complications, such as dehydration, malnutrition, choking injuries, aspiration pneumonia, and even death [11]. Dysphagia can happen at any of the four stages, but the general assessment and therapy for neurogenic dysphagia focuses on the oral and pharyngeal stages. This is because the entrance of the esophagus is in close proximity to the larynx, and both air and the swallowed bolus share a common pathway through the pharynx. More specifically, when the bolus goes through the pharynx, swallowing and respiration cannot happen simultaneously. This physiological mechanism ensures that the bolus can be swallowed through the esophagus smoothly without getting into the trachea and lungs. It also hinders suffocation, aspiration pneumonia, and severe respiratory failure. In fact, most oropharyngeal dysphagia cases are caused by an uncoordinated sequence between respiration and swallowing in the oral and pharyngeal stages. Therefore, in this paper, we examined the effect of cigarette smoking on the coordination between respiration and swallowing. We adopted a noninvasive approach, where we developed a throat belt with an inserted force-sensing resistor (FSR) sensor (Figure 1) to detect the movements of the thyroid cartilage in the pharyngeal stage. The FSR sensor is a type of piezoresistive sensor where the resistance changes proportionally to the pressure on the surface. This enabled us to develop an easy-to-use, lightweight, wearable belt to monitor the participant’s swallowing behavior. A pair of surface electromyography (sEMG) electrode pads were also pasted on the submental muscle to detect the contraction of the submental muscle during swallowing, and a nasal airflow cannula was placed in front of the nose to detect respiration.

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

The testing protocol was approved by the Ethics Committee of Chang Gung Medical Foundation (Number: 102-5615B).

Figure 1. FSR sensor

2. Methods 2.1. Participants We recruited 45 male participants, aged 30–50 years, for the test. They were divided into two groups: the first contained 26 nonsmoking participants and the second 19 smoking participants (Table 1). All participants in the second group had been smoking more than 10 cigarettes daily for over 10 years. Exclusion criteria included any history or symptoms of dysphagia, cardiopulmonary disease, neurological disease, hiatal hernia, chronic indigestion disorder, gastroesophageal reflux disease, cancer or other diseases of the head and neck, or current use of medications with known effects on swallowing or breathing. All participants signed informed consent forms before participation.

2.2 FSR throat belt and other devices We used the FSR sensor to measure the movements of the thyroid cartilage. We designed a belt with Velcro straps, for good elasticity, to fix the FSR sensor in place around participants’ necks (Figure 2). The belt had a maximal width of 5 cm so as not to obstruct a participant’s original swallowing motions. We also inserted a small airbag between the belt and FSR such that the FSR could be fixed on the center of the thyroid cartilage without slipping during the test. When the belt was affixed around the neck, the airbag provided a steady initial pressure on the FSR, which was considered the baseline in each measurement. During swallowing, the larynx and thyroid cartilage retracted, releasing the pressure on the FSR. Our objective was to find the coordination status between swallowing and respiration. Therefore, in addition to the FSR belt around the neck, we pasted a pair of sEMG (surface electromyography) electrode

31

pads on the submental muscle and placed the nasal airflow cannula in front of the nose (Figure 2). Figure 3 shows the synchronized signals of submental sEMG (Figure 3 (a)), nasal airflow (Figure 3(b)), and FSR (Figure 3(c)) from a nonsmoking participant swallowing 10 mL of room-temperature water. The sEMG waveform shows the onset duration of the submental muscle, which can be used to identify the beginning and end of the pharyngeal stage. A pressure transducer was connected to the nasal cannula to measure the nasal airflow in order to observe the respiration pause during swallowing. A typical respiratory response during swallowing is an apnea episode (the period between two red lines in Figure 3(b)). This is a necessary protective respiratory phenomenon to allow for safe swallowing without aspiration. The onsets and offsets among the sEMG, nasal airflow, and FSR signals enabled the evaluation of the coordination between respiration and swallowing. The following definitions are derived from Figure 3: (1) sEMG duration (sEMGD): This is the duration of the submental muscle response, which can be obtained by measuring the interval between the onset and the offset of the submental sEMG wave (Figure 3(a)). (2) sEMG amplitude (sEMGA): This is the peak amplitude of the submental sEMG, which represents the maximal force exerted by the submental muscle during swallowing (Figure 3(a)). (3) Swallowing apnea duration (SAD): This is the apnea period, which can be measured from the onset of the nasal airflow wave to the offset of the same wave (Figure 3(b)). (4) Onset latency (OL): This is the duration from the onset of submental muscle contraction to the onset of laryngeal excursion (the onset of FSR) (Figure 3(a) and (c)). (5) Excursion time (ET): This is the duration from the onset of upward movement of the thyroid cartilage to the thyroid cartilage closing the trachea. (Figure 3(c)). (6) Duration of second deflection (DEFD): This is the duration from the thyroid cartilage closing the trachea, to the offset of the thyroid cartilage returning to the original position. (7) Total excursion time (TET): This is the whole duration of the thyroid cartilage response. It is calculated from the sum of ET and DEFD in the FSR wave (Figure 3(c)). (8) FSR Amplitude (FSRA): This is the peak strength of the thyroid cartilage, which can be detected by FSR (Figure 3(c)).

Table 1. Participants in the nonsmoking and smoking groups. Group

Non-smoking

Smoking

Number of participants

26

19

Average age

38.12±6.45 (years)

37.68±7.13 (years)

Average smoking period

0

16.21±6.04 (years)

Number of cigarettes daily

0

17.95±7.26

* The mean difference is significant at p < 0.05

ISBN: 1-60132-451-0, CSREA Press ©

p-value 0.833*

32

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

Figure 2. FSR throat belt.

Figure 3. Waveforms of (a) submental sEMG, (b) nasal airflow, and (c) FSR.

Figure 4 Average results of (a) sEMGD and (b) sEMGA for the nonsmoking and smoking groups.

ISBN: 1-60132-451-0, CSREA Press ©

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

33

shorter or equal OL for all volumes, except in the 5 mL test, compared with the nonsmoking group. This shows that the smoking participants had a delayed pharyngeal phase in these cases. Figure 6(b) shows that the smoking group also needed longer TET, with greater variations, to complete swallowing compared with the nonsmoking group in all tests. To investigate which part affected the difference in TET between the two groups, we compared ET and DEFD for both groups (Figure 6(c) and (d), respectively). We found that the differences in TET between the two groups mainly came from the fact that the smoking participants performed much longer DEFD than the nonsmoking participants. This means that the smoking participants, on average, needed a longer thyroid cartilage excursion (i.e., longer TET) to finish swallowing than the nonsmoking participants, and also took longer for the thyroid cartilage to return to the original position (i.e., longer DEFD). Finally, FSRA for the two groups is compared in Figure 6(e); no obvious differences can be seen.

3. Results 3.1 Submental muscle results: sEMG Figure 4 shows the average sEMGD and sEMGA between the nonsmoking and smoking groups. From 1 to 20 mL, everyone in the smoking group showed a longer sEMGD and larger sEMGA than the nonsmoking group. The results reveal that the smoking participants on average needed to use greater strength over a longer time to push the water into the esophagus through submental muscle contraction. In addition, although 1 mL was the smallest water volume, it was not the one that the participants required the least time and strength to swallow. This is reasonable for a minute volume of water. Both groups took longer in the 1 and 3 mL tests, and the smoking group took even longer.

3.2 Nasal airflow results: SAD Figure 5 shows the results of SAD for the two groups. The results show that both groups required a longer apnea time when the volume of water increased, but the smoking group had a higher rate of increase. Smoking hurts oral and pharyngeal mucous as well as their sensory receptors, and our study reveals that smoking also affects respiration by prolonged SAD for safe swallowing.

3.4 Piecemeal swallowing When the water volume exceeds a limit (typically 20 mL), numerous participants resort to piecemeal swallowing. This involves dividing the bolus into smaller pieces and swallowing in several gulps. A swallowing limit below 20 mL could be considered inconspicuous dysphagia in neurogenic disorders. Table 2 shows the percentages of piecemeal swallowing over all the trials for the nonsmoking and smoking groups in the 5 mL to 20 mL tests.

3.3 FSR results: OL, TET, ET, DEFD, FSRA Figure 6 shows the results from the FSR measurements. Most participants had an earlier onset of thyroid cartilage than submental sEMG in the 3 to 20 mL tests (OL < 0) (Figure 6(a)). In particular, the smoking group had a

Time (s)

SAD

2.00

Non-smoking

Smoking

1.60 1.20 0.80 0.40 0.00 1mL

3mL

5mL Volume

10mL

20mL

Figure 5. SAD results compared between nonsmoking and smoking groups.

ISBN: 1-60132-451-0, CSREA Press ©

34

Int'l Conf. Biomedical Engineering and Sciences | BIOENG'17 |

Time (s)

3.20

0.4 0.2

Non-smoking

Non-smoking

2.80

Smoking

Time (s) 2.0

(b) TET

Time (s)

(a) OL

Smoking

Non-smoking 1.2

2.00 1.60

-0.2

1.20

-0.4

0.80

0.8 0.4

0.40 -0.6

1mL

3mL

Time (s)

0.0

0.00

5mL 10mL 20mL Volume

1mL

3mL

(d) DEFD

2.00

Non-smoking

Amp (Volts) 2.0

Smoking

1.20

1.2

0.80

0.8

0.40

0.4 3mL

5mL 10mL 20mL

3mL

5mL 10mL 20mL Volume

(e) FSRA Non-smoking

1.6

0.00

1mL

5mL 10mL 20mL Volume

1.60

1mL

Smoking

1.6

2.40

0

(c) ET

Smoking

0.0 1mL

Volume

3mL

5mL 10mL 20mL Volume

Figure 6. Comparison of FSR signals between nonsmoking and smoking groups: (a) OL; (b) TET; (c) ET; (d) DEFD; (e) FSRA.

Table 2. Frequency of piecemeal swallowing. Non-smoking

Smoking

p-value

Number of the subject

26

19

Total trials in each volume of water

78

57

Piecemeal swallowing in 5 mL

2(3%)

7(12%)

0.036*

Piecemeal swallowing in 10 mL

5(6%)

11(19%)

0.030*

Piecemeal swallowing in 20 mL

23(29%)

26(46%)

0.070

* : Represent statistically significant differences p-value 0.90 cosine score) 62 67 64

4. Conclusions Fig. 1. Computing Similarity score

This cosine similarity score computes a value, adjusted for article length, to depict the similarity for each sentence pair, based on the values of shared terms. Mathematically, it is represented as follow. Cosine (D1, D2) = ∑(wD1(j)*wD2(j)/norm(D1)*norm(D2))

3. Experiments We extracted 50 journal articles from PubMed ((https://www.ncbi.nlm.nih.gov/pubmed/batchcitmatch). For each citation in an article the tool extracted the corresponding paper. The tool then extracted two sentences before and two sentences after the citation in the original document and tried to match the words in those sentences with the target document using the cosine similarity metric. This process generated a cosine similarity index for each citation in the original document. Once we had the cosine similarity measurements, we picked up the pairs (citation in the original sentence and relevant parts in the cited document) that scored higher than 0.90. Three subject matter experts (SME - clinical experts in this case) then manually evaluated the citations sentences and the cited documents, and decided which of the correlated documents matched the most. The human experts based their judgement mainly on semantic matching of the sentences in the two documents and not just on matched strings.

In this proof-of-concept study we analyzed the textual similarity between citation text in an original research paper from PubMed and the corresponding text in the cited document. We tried to understand how close the author was in citing the cited paper. We first used cosine similarity measure to come up with a paired list of citation text and cited text. We then looked at 100 such pairs with a cosine similarity score of over 0.90. The system recorded an average accuracy of 64.33% based on the evaluations of the three SMEs. For future work, we plan to extend the similarity metrics using the WordNet synset hierarchy and distributional similarity and Latent Semantic Analysis (LSA) index. Siri.

5. References 1. Bonzi, S.: Characteristics of a literature as predictors of relatedness between cited and citing works. Journal of the American Society for Information Science, 33(4):208-216 (1982). 2. Boyack, K. W., Small, H., and Klavans, R.: Improving the Accuracy of Co-citation Clustering Using Full Text. In Proceedings of 17th International Conference on Science and Technology Indicators. (2012) 3. Corley, C., and Mihalcea, R.: Measuring the Semantic Similarity of Text, in Proceedings of the ACL workshop on Empirical Modeling of Semantic Equivalence and Entailment, pp. 13−18. (2005) 4. Klavans, R., Boyack, K.W.: Identifying a Better Measure of Relatedness for Mapping Science. Journal of the American Society for Information Science and Technology. 57 (2) pp. 251-263 (2006)

3.1 Results

For manual evaluations the three SMEs looked at 100 matched set that scored higher than 0.90 cosine similarity score. SMEs rated their assessment on a scale of 1-100, 100 being the best match. For example, SME-1 found that out of the 100 paired texts, 62 talked about the same concept. In this preliminary study we did not analyze the disagreements between the SMEs. Table 1 gives a summary of this evaluation process.

ISBN: 1-60132-451-0, CSREA Press ©

Author Index Alam, Hassan - 56 Baek, Moo-Jun - 48 Bauer, Michael - 3 Chang, Yunseok - 23 , 52 Cheng, Hsin-Yi Kathy - 30 Choi, Mi-Hyun - 54 Chung, Soon-Cheol - 54 Deng, Guang-Feng - 50 Hartono, Rachmat - 56 Held, Devin - 3 Ho, Wei-Kuang - 50 Hung, Yu-Shiang - 50 Jeong, Dongjun - 48 Jeong, Munseong - 52 Jo, Ji-Hun - 54 John, Eugene - 17 Jung, Ha-Chul - 46 Kim, A-Hee - 46 Kim, Boyeon - 23 , 52 Kim, Hyung Ju - 48 Kim, Hyung-Sik - 54 Kim, Woo-Ram - 54 Koppa, Santosh - 17 Kumar, Aman - 56 Kwon, Dahye - 46 Lee, Chany - 45 Lee, Sanghun - 46 Lee, Sangjun - 45 Lee, Seong-A - 46 Lim, Chang-Hwan - 45 Lin, Hsiao-Hung - 50 Moon, Jin-Hee - 46 Richfield, Steve - 36 Sarva, Pradeep - 17 Shen, Xiaoping - 10 Shieh, Wann-Yun - 30 Snyder, Selena - 10 Vyas, Manan - 56 Wang, Chin-Man - 30 Werner, Tina - 56