International Journal of Electronics

demand or costs or damages whatsoever or howsoever ... methods have been developed for heart ... contrast to the conventional ECG signal measurement...

0 downloads 81 Views 3MB Size
This article was downloaded by:[2007-2008 Inha University] On: 1 August 2008 Access Details: [subscription number 790523467] Publisher: Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Electronics Publication details, including instructions for authors and subscription information: http://www.informaworld.com/smpp/title~content=t713599654

Ubiquitous evolvable hardware system for heart disease diagnosis applications Tae Seon Kim a; Hanho Lee ab; Jaehyun Park b; Chong-Ho Lee b; Yong-Min Lee b ; Chang-Seok Choi b; Seung-Gon Hwang b; Hyun Dong Kim a; Chul Hong Min a a School of Information, Communications and Electronics Engineering, Catholic University of Korea, Bucheon, Korea b School of Information & Communication Engineering, Inha University, Nam-Gu, Incheon, Korea First Published: January 2008 To cite this Article: Kim, Tae Seon, Lee, Hanho, Park, Jaehyun, Lee, Chong-Ho, Lee, Yong-Min, Choi, Chang-Seok, Hwang, Seung-Gon, Kim, Hyun Dong and Min, Chul Hong (2008) 'Ubiquitous evolvable hardware system for heart disease diagnosis applications', International Journal of Electronics, 95:7, 637 — 651 To link to this article: DOI: 10.1080/00207210801923992 URL: http://dx.doi.org/10.1080/00207210801923992

PLEASE SCROLL DOWN FOR ARTICLE Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf This article maybe used for research, teaching and private study purposes. Any substantial or systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material.

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

International Journal of Electronics Vol. 95, No. 7, July 2008, 637–651

Ubiquitous evolvable hardware system for heart disease diagnosis applications Tae Seon Kimb, Hanho Leeab*, Jaehyun Parka, Chong-Ho Leea, Yong-Min Leea, Chang-Seok Choia, Seung-Gon Hwanga, Hyun Dong Kimb and Chul Hong Minb a School of Information & Communication Engineering, Inha University, 253 Yonghyun-Dong, Nam-Gu, Incheon, Korea; bSchool of Information, Communications and Electronics Engineering, Catholic University of Korea, Bucheon, Korea

(Received 28 May 2007; final version received 20 September 2007) This paper presents a stand-alone ubiquitous evolvable hardware (u-EHW) system that is effective for automated heart disease diagnosis applications. The proposed u-EHW system consists of a novel reconfigurable evolvable hardware (rEHW) chip, an evolvable embedded processor, and a hand-held terminal. Through adaptable reconfiguration of the filter components, the proposed u-EHW system can effectively remove various types of noise from ECG signals. Filtered signals are sent to a PDA for automated heart disease diagnosis, and diagnosis results with filtered signals are sent to the medical doctor’s computer for final decision. The rEHW chip features FIR filter evolution capability, which is realised using a genetic algorithm. A parallel genetic algorithm evolves FIR filters to find the optimal filter combination configuration, associated parameters, and the structure of the feature space adaptively to noisy environments for adaptive signal processing. The embedded processor implements feature extraction and a classifier for each group of signal types. The proposed u-EHW system is a promising solution for various applications such as DSPs, communications, and ubiquitous healthcare systems. Keywords: evolvable hardware; reconfigurable; DSP; ubiquitous; healthcare

1. Introduction The electrocardiogram (ECG) signal is a tracing of an electrical activity signal generated by rhythmic contracting of the heart, and affords crucial information for detecting heart disease (Kim and Kim 2005). Since the ECG signal varies from patient to patient and according to measured time and environmental conditions, an adaptable heart disease diagnosis algorithm based on context-aware computing has been proposed in which a genetic algorithm (GA) finds the optimal set of preprocessing feature extraction, and classifier for each group of signal types (Kim and Kim 2005). There have been steady efforts by researchers to develop automated heart disease diagnosis systems with various approaches. Some researchers have focused on signal processing technology for various types of noise reduction. Generally, measured ECG signal data contain muscle noise, baseline noise, in-coupling noise, and 60 Hz power line noise. Generally, 1  50 Hz IIR band pass filter was widely used to remove noise factors.

*Corresponding author. Email: [email protected] ISSN 0020–7217 print/ISSN 1362–3060 online ß 2008 Taylor & Francis DOI: 10.1080/00207210801923992 http://www.informaworld.com

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

638

T.S. Kim et al.

Also, high pass and low pass filters are used separately in many other algorithms (Kohlerm, Hennig and Orgimeister 2002). Feature extraction and selection techniques are also considered very significant steps for accurate diagnosis. Ramli and Ahmed (2003) used cross-correlation analysis to extract features from abnormal signal. A maximum slope detection based detection method known as ‘‘So and Chan’s method’’ and digital analysis of the slope, amplitude, and the width of the data developed by Pan and Tompkins, showed improved signal detection performance (Tan, Chan and Choi, 2000). Based on selected features, various pattern classification methods have been developed for heart disease diagnosis including cardiac arrhythmia classification using autoregressive modelling (Zhang, Jiang, Ge and Xiang 2004) and heart rate variability (HRV) analysis for cardiovascular disease (Ahuja, Raghavan, Lath and Pillai 2004). (Ogawa and Togawa 2000) showed more advanced diagnosis research results, considering patient’s age and gender for diagnosis. However, it is difficult for most classification algorithms to find a general rule base for an automated diagnosis system because of wide variations of signal. Practically, it is nearly impossible to construct a standardised diagnosis rule for heart disease, since ECG signals are different from patient to patient and vary with measured time and patient’s condition, even for the same patient. For perfect classification, the classifier needs to learn every case of ECG signals for all patients, but it is not possible or not cost effective. This problem can be solved if we can group various types of signal characteristics (similarity of signal, gender, age, and etc.) for all patients before diagnosis and can develop a more accurate patient adaptable ECG diagnosis algorithm. For this, the genetic algorithm (GA) finds an optimal set of preprocessing, feature extraction and classifier for each group of signal types. The software level version of the proposed diagnosis system has shown acceptable diagnosis results for clinical records of the MIT-BIH arrhythmia database (Kim and Kim 2005). It was the first generally available set for evaluation and is widely used in research literature (Moody and Mark 2001). Conventionally, the ECG signal is measured on static condition since the appearance of heartbeats varies considerably, not only between patients, but also with movement, respiration and modifications in the electrical characteristics of the body. However, our diagnosis system aims for continuous diagnosis in a dynamic ubiquitous environment, not a static environment. For a dynamic environment, various types of noise including muscle artifact noise and electrode moving artifact noise are coupled and they are continuously changed. To solve this problem, we need complex and fast computation capability for adaptable noise reduction and evolvable hardware was selected because it looks one of the best solutions to meet this requirement. This paper presents a stand-alone ubiquitous evolvable hardware (u-EHW) system to process the adaptable heart disease diagnosis algorithm and detect heart disease. The u-EHW system consists of a reconfigurable evolvable hardware (rEHW) chip, an evolvable embedded processor, and a PDA. The rEHW chip can effectively find the optimal set of preprocessing, and processes the adaptive digital signal processing (DSP). That is, the rEHW chip processes the low-pass, band-pass, and high-pass FIR filter algorithms with various frequencies. The evolvable embedded processor operates GA, feature extraction, and a classifier for each group of signal type. Currently, many DSP and smart health care applications are implemented on digital signal processors and embedded processors by software. Effective DSP algorithms require computing architectures that are less complicated, highly flexible, and more cost-effective for

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

International Journal of Electronics

639

ubiquitous health care applications. Currently, complex and fast computation can be performed by dedicated hardware instead of a digital signal processor since dedicated hardware can operate in parallel. The concept of reconfigurable hardware and evolvable hardware has been studied actively (Higuchi, Iwata and Liu 1996, Tsuda 2000, Stoica et al. 2001). Since evolvable hardware can evolve in real time, it can maintain optimal performance in a variety of unexpected environments. The configurations that optimise the device output responses are combined for better configurations until an optimal architecture can be realised (Faugman 1985). Evolvable hardware continues to reconfigure itself in order to achieve better performance. In contrast to the conventional ECG signal measurement system, the proposed u-EHW system can measure and analyse the ECG signal in a ubiquitous environment. For this, a GA based rEHW chip effectively reconfigures the filter components to remove noise components. Filtered signals are sent to a PDA for automated heart disease diagnosis, and diagnosis results with filtered signals are sent to the medical doctor’s computer where a final decision can be made. Through this, patient customised health care service in ubiquitous environments can be realised. The rest of the paper is organised as follows. Section 2 describes the adaptable heart disease diagnosis algorithms for proposed u-EHW system. Section 3 proposes the u-EHW system for ubiquitous healthcare applications. In Section 4, implementation and results are presented. Finally, conclusions are provided in Section 5.

2. Adaptable heart disease diagnosis algorithm Figure 1 shows the block diagram of a developed EHW based adaptable heart disease diagnosis system. The developed system uses dual mode evolvable hardware (DM-EHW) architecture, which has two operation modes, action mode (AM) and evolution mode (EM). As its name indicates, DM-EHW can select appropriate operation mode between AM

Figure 1. Block diagram of adaptable heart diseases diagnosis system.

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

640

T.S. Kim et al.

and EM. For example, if the noise context estimation results showed that noised context from input data are close to stored knowledge of an evolvable knowledge accumulator (EKA), then DM-EHW operated on AM. For AM operation, a control module determines optimal filter configuration using EKA and sends filter configuration information to rEHW. After that, filtered ECG signals from rEHW are fed to the diagnosis module for heart disease diagnosis. In this case, configured filters of rEHW are fixed until the decision module requires operation change to EM. The decision module selects EM operation if the fitness values of filtered signals are lower than acceptable. In EM operation, a set of chromosomes is reproduced and evaluated. For genetic operation, each chromosome has filter configuration information and genetic operation is executed using a genetic algorithm processor (GAP). If the new filter configuration results are found through genetic evolution, then they are stored in an EKA for possible filter configuration. Physically, the developed system consists of three parts an rEHW chip, a ARM based evolvable embedded processor, and a hand-held display. Also, in terms of functionality, it is mainly divided by six modules; noise context estimation module; control module for optimal configuration of filter block at AM; rEHW module for reconfigurable 3 stage filter based signal processing; genetic algorithm processing module for filter design at EM; decision module for selection of AM and EM using fitness calculation results; and neural network based diagnosis module for heart disease diagnosis.

2.1. Noise context estimation module Accurate estimation of ECG signal noise in a dynamic environment is impossible since signal artifacts from respiration, motion activity, and their coupled signal noise are not predictable. For this reason, noisy signals are grouped into several categories by environmental context estimation. Baseline wander noise and muscle noise are then quantised based on environmental context information. Conventionally, 1  50 Hz IIR band pass filter was widely used to remove baseline noise, but it made changes on important features including signal intervals and vertices. Noised ECG signals are low pass filtered with 2 Hz cut-off frequency, and the gradient of filtered signal is calculated at every 100 samples to quantise the baseline wander noise component estimation. The muscle noise is one of the most difficult noise types to remove since it has nonstationary and non-linear nature. Typically, the Gaussian model is used to model muscle noise and recent research results showed that the impulsive noise model is more accurate to model real-life muscle noise than the conventional Gaussian model (Kim and Kim 2005). Noised signals are high pass filtered with 50 Hz cut-off frequency and then impulse type components are considered as muscle noise components. The number of vertices above the heuristic determined threshold line is counted and this quantised information is used as muscle noise context information.

2.2. Neural networks based control module and filter block In order to identify the adaptable filter composition, the outputs of the noise context estimation module are fed to the inputs of the control module. The control module consists of two parts, an evolutionary knowledge accumulator (EKA) and a neural network for the filter design. The EKA stores the optimal filter design results for six-clustered dynamic

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

International Journal of Electronics

641

measurement environments. For network construction, two types of outputs from the context estimation module are used as network inputs and six outputs are used to cluster the dynamic measurement environment into six groups. The 6 best filter combinations for each cluster are a 0.8  45 Hz band pass filter, a 0.8 Hz high pass filter, a 50 Hz low pass filter, a combination of a 0.8 Hz high pass filter and a 50 Hz low pass filter, a combination of a wavelet interpolation filter (WIF) and a 0.8  45 Hz band pass filter, and a combination of WIF, a 0.8 Hz high pass filter and a 50 Hz low pass filter. For clustering of environmental noise, a three layered feed-forward error back-propagation neural network (BPNN) is applied and network weights are updated after each training vector (by a ‘‘training by vector’’ method).

2.3. Genetic algorithm based decision module and evolution module For a continuous performance update, every filtered output from the 3-stage filter block is evaluated by a fitness function at a decision module. The fitness function (F) is made up of summation of detection rates for five ECG signal features as shown in Equation (1), F¼

n X

fi

ð1Þ

i¼1

where n and f represents the number of features and detection rate of each feature, respectively. In this work, P-R interval, QRS duration, Q-T interval, R-R interval and peak of S wave are considered for feature values for ECG signal characterization. To calculate detection rate, the template matching method is applied (Kim and Kim 2005). If the fitness value is over 0.8, ECG feature extraction results are considered as acceptable values or the system assumes that designed filter cannot effectively remove noise, so GAs search a new set of filter block for adaptable noise reduction. Based on GAs search results, EKA is updated and the neural network for environmental context based clustering task is retrained.

2.4. Diagnosis module The five feature extraction values from decision modules are fed to an input of diagnosis module to classify disease type. In this work, four types of heart diseases, sinus bradycardia (SB), sinus tachycardia (ST), right bundle branch block (RBBB) and left bundle branch block (LBBB) were considered as possible disease types. Also, mixed types of disease including STþLBBB, STþRBBB are also considered. For pattern classifier, three types of neural network classifiers including BPNN, competitive neural network (CNN) and probabilistic neural network (PNN) are tested. To evaluate diagnosis performance, classification accuracy, uncertainty, sensitivity, specificity, positive predictive accuracy and negative predictive accuracy are calculated. After performance evaluation BPNN showed best diagnosis results among three classifiers and finally BPNN is selected for the diagnosis module.

3. Ubiquitous evolvable hardware system The overall u-EHW system consists of a rEHW chip, an evolvable embedded processor, and a hand-held terminal, as shown in Figure 2. The input data of the rEHW and

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

642

T.S. Kim et al.

Figure 2. Block diagram of ubiquitous evolvable hardware (u-EHW) system.

evolvable embedded processor is the measured ECG signal data, which may contain muscle noise, baseline noise, in-coupling noise, and 60 Hz power line noise.

3.1. Reconfigurable evolvable hardware chip The rEHW chip processes the DSP algorithms, i.e., low-pass, band-pass, and high-pass filter algorithms with various frequencies. This chip consists of a reconfigurable processing unit, a configuration manager, and coefficient memory, as shown in Figure 3. The reconfigurable processing unit has a 3-stage reconfigurable filter block in which the optimal FIR filter function can be searched and selected by GA chromosome data obtained using a GA running on an embedded processor. Each reconfigurable filter block includes 20 reconfigurable processing modules (RPMs). The GA has been used to evolve digital circuits. A chromosome represents a component of order. The 30-bit GA chromosome data has 3-stage 30-bit chromosome data, which defines the type of optimal filter, cutoff frequency, and filter order, as shown in Figure 4. The first 2-bit data decides the optimum filter type, such as a low-pass, bandpass, or high-pass filter. If a high-pass filter or low-pass filter is selected by the GA, cutoff frequency #1 or #2 has the data, respectively. If the band-pass filter is selected, cutoff freqeuncy #1 and #2 have the data. The last 3-bit data defines the filter order, which decides the number of filter taps, from 6- to 20-taps. The configuration manager consists of the order distributor, stage selector, and memory address decoder. The order distributor analyses the filter order information and the stage selector selects the filter stage to be used. Coefficient memory stores the coefficients, which are used in each filter. The rEHW chip is responsible for providing the best solution for realisation and autonomous adaptation of FIR filters, and is used to

643

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

International Journal of Electronics

Figure 3. Block diagram of (a) reconfigurable evolvable hardware and (b) reconfigurable filter block.

process the optimal DSP algorithms for noise removal operation prior to feature extraction and classification steps. Generally, low-pass, band-pass, and high-pass filters are used separately in many preprocessing algorithms (Kim and Kim 2005).

3.2. Evolvable embedded processor Before the software is ported to an embedded processor for the algorithm verification step, developed algorithms are tested and evaluated on a commercialised Linux OS based PC. The noise context estimation module, control module, decision module, diagnosis module, and software level rEHW modules are separately developed and tested. After that, each module is integrated to one system and tested through AM and EM operations. After finishing PC level verification of integrated adaptable heart disease diagnosis system modules, programming codes of system modules ported to the embedded processor.

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

644

T.S. Kim et al.

To morph software level algorithms on a hardware platform, data format conversion with optimal truncation is required. At software level, the measured ECG signals (raw data) were operated as double precision floating point numbers and they are truncated to 16-bit fixed point numbers to use FPGA based reconfigurable filter blocks at hardware level implementation. To find optimal bit truncation, truncated signals are compared with original signals to keep the signal feature characteristics. Signal feature values are agreed at the required data precision level. After completion of program porting to the embedded processor, the software level rEHW module is replaced by the FPGA based rEHW chip. After FPGA level algorithm verification, the FPGA based rEHW chip is implemented on the ASIC chip. After that, TCP/IP wireless mobile communication module is added to a developed system for application in a ubiquitous healthcare environment. The evolvable embedded processor implements a noise context estimation module, decision module, and control module to process the feature extraction and classifier algorithms by software. Also, the embedded processor processes the GA, which is a search procedure inspired by populating genetics, and has excellent search capabilities for finding a good solution to a problem without a priori information about the nature of the problem (Sakanashi, Iwara and Higuchi 2004). For a continuous performance update, every filtered output from the rEHW chip is evaluated by a fitness function at the decision module. To determine the adaptable filter composition, the outputs of the noise context estimation module are fed as inputs of the control module. The fitness function is defined as the detection percentage of five ECG signal features: P-R interval; QRS duration; Q-T interval; R-R interval; and S peak point, using So and Chan’s and Kim’s feature extraction method (Kim and Kim 2005). In order to confirm the fitness function, the filtered signals are fed into the noise context estimation module again. If the fitness value is greater than 0.8, the ECG feature extraction results are considered to be acceptable values. The noise context estimation module can estimate the noise context using a neural network for

Figure 4. GA chromosome data definition.

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

International Journal of Electronics

645

filter design. For this, pre-determined optimal filter design results for six-clustered dynamic measurement environments are stored. For network construction, two types of outputs from the noise context estimation module are used as network inputs and six outputs are used to cluster the dynamic measurement environment into six groups. For clustering of environmental noise, a three layered feed-forward error back-propagation neural network (BPNN) is applied and network weights are updated after each training vector (by the ‘‘training by vector’’ method).

4. Implementation and result The rEHW architecture was modelled in Verilog HDL and functionally verified using a ModelSim simulator. The output data from the Verilog coded architecture was validated against a bit-accurate MATLAB model. The rEHW chip was implemented using standard 0.18 mm CMOS technology. Figure 5 shows the rEHW chip layout and Table 1 shows the rEHW chip summary.

Figure 5. rEHW chip layout.

646

T.S. Kim et al.

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

Table 1. rEHW chip summary. Chip summary Technology

Standard 0.1 mm CMOS 1.8 V core, 3.3V I/O

Die size Package Gate count Clock speed

3.75  3.75 mm2 64-pin LQFP 582,182 80 MHz

The rEHW chip consists of 582,182 gate excluding memories and the operating clock frequency is about 80 MHz. There are five control/status registers in the rEHW chip, a control register, interrupt status register, interrupt clear register, and two chromosome registers. Besides these control/status registers, 4 Kbytes of input and output buffer memory are allocated for data transfer between the rEHW chip and embedded processor. The ECG data consist of 1024 samples and they are processed by the rEHW chip. The chromosome used by the rEHW chip is selected by the GA algorithm in the embedded processor and set in the chromosome register. The rEHW chip processes 1024 items of data every 15 msec. After processing, the data the rEHW chip generates interrupt the embedded processor in order to update processed data A u-EHW test-bed system for adaptable heart disease diagnosis was implemented as a stand-alone embedded system that includes an rEHW chip and an evolvable embedded processor. Besides a stand-alone embedded system, a PDA and DB server were also developed for a demonstration. The noise context estimation module, control module, decision module, and GA have been implemented by a bit-accurate C-model and implemented on an embedded processor. Hardware of the implemented test-bed consists of an ATMEL AT91RM9200 ARM processor, of which the performance is 200 MIPS, a rEHW chip, external memories, and a LAN interface, as shown in Figure 6. To avoid coupling noise when running at a high speed of 200 MHz, the whole system is divided into two sections of a 6-layer PCB; a CPU board and a communication/rEHW chip board. The test-bed has three kinds of memories: NAND flash memory; SDRAM; and EEPROM, which are used for program memory, data memory, and configuration storage, respectively. The interface between the rEHW and the main processor was implemented as an asynchronous memory interface that enables the rEHW to be used with a general embedded processor even the data transfer bandwidth is somehow limited. Since the rEHW lacks non-volatile memory such as flash memory inside, the configuration data are stored in the serial EEPROM and initialised by the main processor before processing. To provide a communication feature, it has two channels of standard serial ports, a USB port, and 10/100 Mbps Ethernet channel. Although serial channels and Ethernet are enough for the development purpose, to be used as a ubiquitous health care device, it should have a wireless communication capability. To provide a wireless connectivity, it uses a USB-based wireless LAN and a broadband modem over which an Internet connection can be used. Figure 6(b) shows the implemented u-EHW test-bed including the embedded processor board and rEHW chip board. The system control software was implemented using the embedded Linux environment. As an operating system, embedded Linux version 2.6.12 is ported. On top of the embedded Linux, control application software such as the noise context estimation module, control

647

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

International Journal of Electronics

Figure 6. u-EHW system test-bed, (a) block diagram of u-EHW test bed, and (b) implemented u-EHW system test-bed.

module, and decision module are implemented. Since all of the application software is developed in standard ANSI-C, they are easily ported onto other platforms as well. A standard Linux-hosted Pentium PC is used as a development environment. GCC 3.4 is used for cross compiling and targeting to the embedded processor system. Software of the implemented test-bed are grouped into three categories: system software that includes rEHW control software and heart disease diagnosis software; user interface software that is implemented on a hand-held device, PDA; and database management software in a DB server, as shown in Figure 7. To demonstrate the u-EHW system, a set of sample ECG data were gathered from various patients and stored in the embedded processor board. The raw data and diagnosis results were transmitted to the hand-held terminal (PDA) in real-time over a wireless LAN. The PDA is used to display ECG data and diagnostic results. It displays the raw data wave and the processed wave as well as one out of four diagnostic

T.S. Kim et al.

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

648

Figure 7. u-EHW system platform for automated heart disease diagnosis applications, (a) block diagram, and (b) implemented prototype of u-EHW system.

results; SB, ST, RBBB, and LBBB. Raw data and diagnosis results are also transmitted to the DB server for further diagnosis by doctor in conjunction with the patient’s history. Figure 7(b) shows the stand alone prototype of u-EHW system platform. For validation of implemented rEHW chip, clinical records of the MIT-BIH arrhythmia database are used. It has been widely used in many research literatures for performance evaluation and it was also used for validation of our software level rEHW

649

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

International Journal of Electronics

Figure 8. MIT-BIH ECG signal and filtering results (a) original ECG signal, (b) baseline wander noise, (c) muscle artefact noise, (d) noise added ECG signal, (e) filtered signal by conventional filter design, and (f) filtered signal by proposed method.

Table 2. Execution time of u-EHW systems for different rEHW technologies. Operation mode AM EM

Processing time (sec)

ARMþrEHW (FPGA)

ARMþrEHW (ASIC)

ARM þ rEHW(S/W) 5.03 51.32

4.84 45.84

3.04 28.36

algorithm (Kim and Kim 2005). Figure 8 shows the one example of filtering performance of rEHW. As shown in Figure 8, baseline wander noise and muscle artifact noise are added to the original ECG signal. The added noise signals also came from MIT-BIH noise stress test database. For performance comparison, filtering results of the conventional method are shown in Figure 8(e). For conventional standard filter design, a combination of 0.8 Hz 20-tab FIR high pass filter and 45 Hz 20-tab FIR low pass filter is used. As shown in these figures, both methods remove baseline wander noise, but the conventional filter design method showed bigger S-T segment distortion because of the remaining muscle artifact noise. This result is exactly matched with our software level rEHW result, which was previously verified. To show the superiority of rEHW chip and u-EHW system performance, several versions of u-EHW systems were implemented using the different rEHW technologies (e.g., software, FPGA, and ASIC chip) and the execution time of u-EHW systems are compared as shown in Table 2. In Table 2, ARMþrEHW(S/W) represents at pure software version of u-EHW system which is executed on a 180 MHz ARM board. The rests also used an ARM board, but they implemented rEHW algorithm using FPGA and ASIC ways. As shown in this table, the rEHW ASIC chip based u-EHW system showed superior execution time compared with the FPGA based system for both AM and EM operation modes. However, significance of hardware implementation can be found at the

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

650

T.S. Kim et al.

EM operation mode, since the EM mode requires complex and iterative computation for GA based filter reconfiguration works. Therefore, the rEHW ASIC chip can provide highperformance that cannot be achieved using regular DSP or FPGA solutions while remaining flexible enough to be used in different designs.

5. Conclusions In this paper, we present a ubiquitous evolvable hardware (u-EHW) system, which implements a stand-alone adaptable heart disease diagnosis system. The u-EHW system consists of an rEHW chip, an evolvable embedded processor, and a hand-held terminal (PDA). A flexible and reconfigurable signal processing rEHW ASIC architecture has been developed, simulated and implemented in the chip. The rEHW chip is used for adaptive digital signal processing and the embedded processor implements feature extraction and a classifier for each group of signal type. The proposed stand alone u-EHW system performs well, especially in changing noisy environments, since it can adapt itself to the external environment. The proposed u-EHW system is a promising solution for various applications such as DSPs, communications, and ubiquitous healthcare systems.

Acknowledgements This work was supported partly by the Ministry of Commerce, Industry and Energy, and partly by the MIC (Ministry of Information and Communication), Korea, under the ITRC support program supervised by the IITA (IITA-2006-C1090-0603-0019).

References Ahuja, N., Raghavan, V., Lath, V., and Pillai, S. (2004), ‘‘Heart Rate Variability and its Clinical Application for Biofeedback,’’ Proc. Computer-Based Medical Systems, 24–25 June 2004, 263–266. Faugman, J. (1985), ‘‘Uncertainty Relation for Resolution in Space, Spatial Frequency, and Orientation Optimization by Two-dimensional Cortical Filters,’’ Journal Opt. Soc. Amer., 2(7), 675–676. Higuchi, T., Iwata, M., and Liu, W. (1996), Evolvable Systems: From Biology to Hardware, Springer: Tsukuba. Kim, T.S., and Kim, H.-D. (2005), ‘‘Context-Aware Computing Based Adaptable Heart Diseases Diagnosis Algorithm,’’ KES 2005, LNAI3682, 3682, 284–290. Kohlerm, B., Hennig, C., and Orglmeister, R. (2002), ‘‘The Principles of Software QRS Detection,’’ IEEE Engineering in Medicine and Biology, 21(1), 42–57. Marshall, A., Stansfield, T., Kostarnov, I., et al. (1999), ‘‘A Reconfigurable Arithmetic Array for Multimedia Applications,’’ Proceedings of the ACM/SIGDA Seventh International Symposium on FPGAs, 21–23 February 1999, 135–143. Moody, G., and Mark, R. (2001), ‘‘The Impact of the MIT-BIH Arrhythmia Database,’’ IEEE Engineering in Medicine and Biology, 20(3), 45–50. Ogawa, M., and Togawa, T., (2000), Attempts at Monitoring Health Status in the Home,’’ IEEEEMBS Special Topic Conference Microtech. In Medicine & Biology, Oct., 552–556. Ramli, A., and Ahmad, P., (2003), Correlation Analysis for Abnormal ECG Signal Features Extraction IEEE 4th National Conference on Telecommunication Technology, 232–237. Sakanashi, H., Iwara, M., and Higuchi, T. (2004), ‘‘Evolvable Hardware for Lossless Compression of Very High Resolution Bi-level Images,’’ IEE Proc.-Comput. Digit. Tech., 151, 277–286.

Downloaded By: [2007-2008 Inha University] At: 15:04 1 August 2008

International Journal of Electronics

651

Stoica, A., Zebulum, R., Keymeulen, D., Tawel, R., Daud, T., and Thakoor, A. (2001), ‘‘Reconfigurable VLSI Architectures for Evolvable Hardware: From Experimental Field Programming Transistor Arrays to Evolution-Oriented Chip,’’ IEEE Trans. on VLSI Systems, 9(1), 227–232. Tan, K., Chan, K., and Choi, K. (2000), ‘‘Detection of the QRS Complex, P wave and T wave in Electrocardiogram,’’ IEEE Advances in Medical Signal and Information Processing, 4–6 September 2000, 41–47. Tsuda, N. (2000), ‘‘Fault-Tolerant Processor Arrays Using Additional Bypass Linking Allocated by Graph-Node Coloring,’’ IEEE Trans. Computers, 49, 431–442. Zhang, Z., Jiang, H., Ge, D., and Xiang, X. (2004), ‘‘Pattern Recognition of Cardiac Arrhythmias Using Scalar Autoregressive Modelling,’’ World Congress on Intelligent Control and Automation, 6, 15–19.