IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 44, NO. 3, MARCH 1997

159

A Real-Time Microprocessor QRS Detector System with a 1-ms Timing Accuracy for the Measurement of Ambulatory HRV Antti Ruha,* Member, IEEE, Sami Sallinen, Member, IEEE, and Seppo Nissil¨a Abstract— The design, test methods, and results of an ambulatory QRS detector are presented. The device is intended for the accurate measurement of heart rate variability (HRV) and reliable QRS detection in both ambulatory and clinical use. The aim of the design work was to achieve high QRS detection performance in terms of timing accuracy and reliability, without compromising the size and power consumption of the device. The complete monitor system consists of a host computer and the detector unit. The detector device is constructed of a commonly available digital signal processing (DSP) microprocessor and other components. The QRS detection algorithm uses optimized prefiltering in conjunction with a matched filter and dual edge threshold detection. The purpose of the prefiltering is to attenuate various noise components in order to achieve improved detection reliability. The matched filter further improves signal-to-noise ratio (SNR) and symmetries the QRS complex for the threshold detection, which is essential in order to achieve the desired performance. The decision for detection is made in real-time and no search-back method is employed. The host computer is used to configure the detector unit, which includes the setting of the matched filter impulse response, and in the retrieval and postprocessing of the measurement results. The QRS detection timing accuracy and detection reliability of the detector system was tested with an artificially generated electrocardiogram (ECG) signal corrupted with various noise types and a timing standard deviation of less than 1 ms was achieved with most noise types and levels similar to those encountered in real measurements. A QRS detection error rate (ER) of 0.1 and 2.2% was achieved with records 103 and 105 from the MIT-BIH Arrhythmia database, respectively. Index Terms— Biomedical monitoring, electrocardiography, matched filters, signal detection, timing jitter.

I. INTRODUCTION

T

HE goal of this project was to develop a complete heart rate variability (HRV) measurement system able to measure, store, and post-process the HRV data obtained from a subject. The system consists of a host computer and a QRS detector device. The host computer is used to setup the measurement parameters and to retrieve the measurement results from the QRS detector. The QRS detector detects the waves of the electrocardiogram (ECG) signal and calculates Manuscript received June 19, 1995; revised September 13, 1996. Asterisk indicates corresponding author. *A. Ruha is with the University of Oulu, Department of Electrical Engineering, Oulu FIN-90570 Finland (e-mail: [email protected]). S. Sallinen was with the University of Oulu, Department of Electrical Engineering, Oulu FIN-90570 Finland. He is now with Varian-Dosetek Oy, Tietajantie, 14 Espoo FIN-02130 Finland. S. Nissil¨a is with the University of Oulu, Department of Electrical Engineering, Oulu FIN-90570 Finland. Publisher Item Identifier S 0018-9294(97)01468-7.

the intervals between two successive waves to form beatto-beat (commonly abbreviated as RR) interval data set. The RR interval data is post-processed in the host computer to calculate statistical figures and visualize them for analysis. The post-processing algorithm is also capable of correcting to some extent the errors occurred in QRS detection by filtering out the false detections caused by noise in the measurement. The false detections are filtered out by using rules that successive RR intervals do not differ more than a certain percentage or they are regarded as noise and discarded [1]. An important part of this work was to develop a QRS detection algorithm and microprocessor-based detector unit of a reasonable size and power consumption, which achieves good timing accuracy in QRS detection even in noisy measurements. In the measurements of ambulatory HRV immunity to noise is of greater importance than in HRV measurements at rest. The HRV measurements at rest are usually affected only by low levels of noise, consisting mainly of amplitude modulation of the QRS complex and mains coupling, whereas the measurements of ambulatory HRV are affected by greater levels of noise, for example, due to motion artifacts. Also in the HRV measurement during physical exercise good timing accuracy is desired as the HRV level reduces with increasing exercise level and can be few milliseconds with high intensity levels [2]. In order to achieve small timing error with respect to the biological HRV good tolerance to noise and good timing accuracy are required and these requirements are considered as the main objectives in this work. The QRS detection algorithm in this work is based on optimized filtering and on a simple decision algorithm as the optimized filtering is considered to be the key to achieve good timing accuracy. At this stage, the QRS algorithm has been developed purely for the accurate detection of the QRS complex and the decision algorithm is kept quite simple to reduce computational load. The decision algorithm is performed in real time, no search-back is used, and it is based on an amplitude comparison to an adaptive threshold. More complex decision rules can be used later if required, and they will provide their maximum advantage in conjunction with optimized filtering. II. SYSTEM DESCRIPTION A. QRS Detector Design Considerations 1) Optimized Filtering: The QRS complex contains signal components in a relatively wide frequency band from about 2

0018–9294/97$10.00  1997 IEEE

160

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 44, NO. 3, MARCH 1997

to 100 Hz with peak at 10–15 Hz. In ambulatory measurements the ECG signal is corrupted with noise due to motion artifacts, mains coupling etc. which must be taken into account in designing the frequency response of the QRS filter. The purpose of the signal filtering is to attenuate noise and enhance those features of the signal used for detection as this leads to an increased probability for correct decisions. The matched filter is a known solution to maximize signal amplitude-to-noise ratio so it is the best solution to be used in QRS detection based on amplitude comparison. The matched filter frequency response for a signal corrupted with normal distributed white noise (Gaussian) is [3] (1) is a constant, is the signal spectra, and corresponds to the filter delay. The amplitude response is identical to the amplitude spectra of the signal, but the phase response is opposite in sign. This means that the impulse response of the matched filter is identical to the signal waveform reversed in time domain. The matched filter for colored noise is a better solution for QRS filter due to noise types encountered in QRS detection. The matched filter frequency response for a signal corrupted with colored noise is [3]

where

(2) is the noise power spectrum. The frequency where response can be calculated and the impulse response obtained using inverse Fourier transform, if the signal and the noise spectra are known. the practical method is to use a whitening prefilter and a matched filter for the prefiltered signal. In such case, the impulse response for the matched filter can be defined from the prefiltered signal waveform. For the ultimate solution the QRS detection the properties of the matched filter should adapt continuously to temporal variations in the QRS complex and noise properties. This, however, leads to increased complexity in realization. 2) Decision Rule: Principles from communications theory can be applied in defining the threshold level. In communications theory the optimum threshold level is calculated for the on-off signal corrupted by Gaussian distributed noise. If equal probabilities of false negatives and false positives are desired, for the on-off signal is [4] the optimum threshold level (3) and are the probabilities of two signal values, where is the signal amplitude, and the noise variance (mean noise power). In binary data transmission where the probability of ones and zeros are equal, the optimum threshold level is 50% of the symbol “one” amplitude. Similarly, in QRS detection the QRS complex (or preprocessed QRS complex) can be thought of as symbol “one” and the silent segment between the two QRS complexes as

Fig. 1. The HRV interval measurement system.

the symbol “zero” and the corresponding probabilities are the duration of each “symbol” divided by the duration of the complete cycle. For example, at a heart rate of 60 beats per minute (bpm) the durations are roughly 80 and 920 ms and with a signal-to-noise ratio (SNR) of five (QRS amplitude 1, noise voltage rms value 0.2, thus mean noise power 0.04) the threshold should be around 60% according to (3). If the number of false positives is to be minimized, the threshold should correspondingly be set higher and vice versa for minimized false negatives. In HRV measurements the threshold should be at a lower level (30–40%) in order to decrease the amount of false negatives (missed beats). This is at the expense of an increase in false positives, however, which can be corrected more easily in the postprocessing than can false negatives. No fixed threshold level can be used, but it must adapt to varying signal levels in order to remain at the same relative level and maintain the desired detection properties. B. General Description of the HRV Measurement System The complete HRV measurement system consists of a personal computer (PC) as a host computer and the batteryoperated microprocessor-based QRS detector, which can be connected to the PC with an optoisolated RS-232 link, Fig. 1. During the measurement the detector unit functions independently of the host computer, but an optically isolated RS-232 communications link was included to provide communications to a host computer for the measurement setup and results retrieval while the ECG electrodes are connected to the patient to guarantee patient safety. The PC provides a graphical interface to configure the detector device and download the measurement data from it using an interactive user-interface program. The configuration task includes the individual sampling of the patient’s QRS complex to set up the matched filter impulse response. The detector device detects the QRS complex, with the aid of an algorithm described later, and calculates RR intervals from two successive detections and stores the RR interval data in its on-board memory. The PC unit is then used to download the RR data stored in the detector unit and to post-process the RR interval data into statistical figures of the HRV. C. The QRS Detector 1) The Algorithm: a) Gain control and prefiltering: The selected QRS detection algorithm employs signal processing in both the analog

RUHA et al.: REAL-TIME MICROPROCESSOR QRS DETECTOR SYSTEM

161

Hz) was chosen so that a typical noise spectra will be whitened, allowing the full use to be made of a matched filter in the following stage. The bandpass filter is realized using a complex-resonator/comb-filter pair which uses the current and two delayed input samples and two delayed output samples to calculate the output

Fig. 2. The QRS detection method.

and digital domains. The block diagram for the signal processing path is shown in Fig. 2. The device was fitted with an amplifier stage with automatic gain control (AGC) and a fourth-order analog prefilter which attenuates the components of the measured signal outside the 0.5–35-Hz frequency band. The frequency response is shown in Fig. 3(a). This enhances the SNR enough so that the ECG signal can be amplified to approximately 75% of the maximum input range of the analog/digital (A/D) converter without baseline variation or motion artifacts driving the amplifier or A/D converter out of its dynamic range. The band-limited and amplified signal is then fed to a 10-bit A/D converter which samples the signal at 500 samples/s. The AGC algorithm is shown below. Variable definitions: HIGH: positive envelope LOW: negative envelope ACG TD: AGC decay time constant (2 s.) AGC MAX, AGC MIN: AGC limit constants (0.75 and 0.25 of the ADC range) For (“each input sample at the ADC output”) If ( HIGH) then HIGH = . If ( LOW) then LOW = . If (“AGC TD seconds have passed since previous update of HIGH or LOW”) then HIGH = 0.9*HIGH, LOW = 0.9*LOW. If (HIGH ACG MAX) or (LOW AGC MAX) then (“decrease the preamplifier gain by a step”). If (HIGH AGC MIN) and (LOW AGC MIN) then (“increase the preamplifier gain by a step). b) Filtering: The initial sampling speed of 500 Hz and interpolation to 2 kHz in subsequent stages was selected because it was believed to provide sufficiently accurate timing information. As this frequency is an integral multiple of the 50-Hz power-line frequency being used in Europe, it allows for a simple implementation of a 50-Hz notch filter which removes any residual power-line interference from the sampled signal. The notch filter is realized using a comb filter, which introduces a transmission zero at 50 Hz by summing the current and a 10-ms delayed sample together. The amplitude response in shown in Fig. 3(b). The second digital filter stage after the comb filter is a bandpass filter which attenuates the low-frequency noise caused by, for example, motion artifacts. The passband (15–40

(4) This stage also introduces a transmission zero at a frequency near 60 Hz to attenuate the mains-noise of 60 Hz used in some countries. The amplitude response is shown in Fig. 3(c). Complex-resonator/comb-filter was chosen for easy implementation. More sophisticated filters based on infinite impulse response (IIR) or finite impulse response (FIR) structures could also be used and these would allow more freedom in choosing the transfer function for the whitening filter. All the multiplications and divisions required in the comb and complex resonator filters are by a power of two so they can implemented with bit shifting. This allows for the filters to be implemented very efficiently in assembly language. The combined amplitude response is shown in Fig. 3(d). The final filtering stage is a matched filter providing an optimal SNR and, more importantly, a symmetrical output pulse waveform. The matched filter output for the filter impulse 128 is calculated as response length

(5)

are input samples and is an output where sample of the matched filter. The matched filter impulse response is optimized for each patient by selecting the filter coefficients in the beginning of a measurement by selecting a good representative of the bandpass-filtered QRS complex from an interactive display on a personal computer. For use as an impulse response for the matched filter, the sampled QRS complex is then further processed by removing its dc component, windowing it, and normalizing it to obtain a gain of one for the matched filter stage. A typical impulse response is shown in Fig. 4. While the matched filter is computationally the most expensive operation used by the developed device, it is nevertheless crucial for the performance of the system. Typical filtered waveforms for almost noiseless and noisy ECG signals are shown in Figs. 5 and 6. The symmetrical output pulse waveform of the matched filter is clearly seen in Fig. 5(c), which is utilized in the threshold detection. The noisy ECG signals in Fig. 6(a) was recorded while the subject was running on the spot and represents a typical ECG signal in ambulatory measurement, in which the noise due to motion artifacts is dominant. The signal quality improvement due to bandpass filtering in Fig. 6(b) is probably adequate for QRS detection, but the matched filter is required to obtain still improved timing accuracy.

162

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 44, NO. 3, MARCH 1997

(a)

(b)

(c)

(d)

Fig. 3. (a) The amplitude responses of the analog prefilter, (b) digital comb filter, (c) digital bandpass filter, and (d) the combination of these.

In order to provide a better timing-resolution than provided by the 500 Hz sampling frequency, the signal at the matched filter output is linearly interpolated to four times the original sampling rate to 2000 Hz. The linear interpolation is performed as (6) (7) (8) (9) , is the original (2 ms), and where is the interpolated sampling period (500 s). Ideally, the original sampling rate could be higher and interpolation would be unnecessary, but in that case the matched filter impulse response length would be 512, i.e., also four times larger than originally, resulting in a too power-hungry operation. As the signal is already band limited very well under the Nyquist frequency of 250 Hz in this system, this interpolation is feasible. c) Detection: QRS detection from the filtered signal is subsequently performed by an adaptive threshold detector

Fig. 4. The matched filter impulse response for a prefiltered QRS complex.

which uses a threshold of about 40% of the maximum value which has occurred in the filter stage’s output over the last 1.5 s. For 200 ms after each QRS detection the threshold is raised to 90% of the maximum value previously present to prevent false detection due to T-wave. The detection and the threshold adaption algorithm can be described using a

RUHA et al.: REAL-TIME MICROPROCESSOR QRS DETECTOR SYSTEM

163

(a)

(b)

(c)

Fig. 5. (a) Almost noiseless ECG signal, which is (b) bandpass filtered, and (c) matched filtered.

pseudo-code as follows. Variable definitions: ENV: matched filter output envelope THR: threshold constant (0.4–0.6) THRES: threshold coefficient DET: binary value detection signal ETR: envelope rise rate constant (2 s) ETD: envelope decay rate constant (5–15 s) EHC: envelope hold time constant (2 s) T LASTP: detection threshold keep time (200 ms) For (“each output sample from the matched filter”) If ( THRES * ENV) then DET else DET . If ( ENV) then ENV = ENV + ETR* . If ( ENV) and (“EHC seconds has passed since the previous update”) then ENV = 0.9*ENV. If (DET ) and (DET ) then THRES = 0.9. If (“more than T LASTP has passed since the previous detection”) and (THRES THR) then THRES = 0.9*THRES. As the matched filter has the additional property of providing a symmetrical output for the desired inputs, dual edge threshold detection can be applied successfully, Fig. 7. In dual edge detection the vertical center-line of the signal exceeding as the threshold level value is calculated by being the fiducial point for the QRS complex. This is known to reduce the timing error by reducing the sensitivity to lowfrequency additive and multiplicative noise, i.e., the baseline and amplitude modulation caused by breathing [5]. With such

noise the fiducial point is not affected as and are equal. According to our experience, amplitude variation can be up to 50% of the QRS complex amplitude when the subject under monitoring is breathing heavily. When compared to systems which only use rising or falling edge detection of the QRS complex, this arrangement provides a better accuracy for QRS detection. 2) Hardware: The preamplifier and prefilter are realized with analog hardware and the signal processing for QRS detection is realized in digital domain. The preamplifier utilizes a commercially available instrumentation amplifier (AMP-04, Analog Devices) which has good performance in common mode rejection. In addition, the gain and upper frequency limit can be set with a minimum amount of external components (one resistor and capacitor). The preamplifier is combined with an analog fourth-order bandpass filter (0.5–35 Hz) to attenuate noise outside the QRS complex spectrum already in the front-end of the signal processing path. After amplification and filtering but before analog-to-digital conversion the ECG signal is fed through an AGC stage to keep the signal amplitude at an optimum level of the A/D converter (ADC) input range. The AGC is realized with a conventional digitalto-analog (D/A) converter (DAC) by using it as a multiplier, where the analog signal is connected to the reference input and the gain-controlling digital word to the data inputs of the DAC. The digital signal processing (DSP) blocks of the system were realized using a Motorola 68HC16 DSP microcontroller, which provides a general purpose 8/16-bit processor core, a DSP engine with an internal ADC and some advanced system integration and power conservation functions. The processor includes a software-controlled clock synthesizer so that its speed can be adjusted in order to minimize power drain. The software was written using the HC16 assembly language used

164

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 44, NO. 3, MARCH 1997

(a)

(b)

(c)

Fig. 6. (a) Noisy ECG signal, which is (b) bandpass filtered, and (c) matched filtered. TABLE I NOISE TYPES USED IN TESTS

Fig. 7. The dual edge detection principle with amplitude modulated pulse.

in this microcontroller. The power consumption of the device was found to be a little over 25 mA at a clock speed of 3.41 MHz. At this speed the device is still capable of running the software with a matched filter tap count of 128. This equates to more than 35 h of operating time on standard 1000-mAh Ni–Cd batteries. III. TEST RESULTS In this study, verification of the QRS detection performance achieved with the developed system was carried out by two separate methods. The timing accuracy was tested with an artificial QRS complex purposely corrupted by various noise types, and the QRS detection reliability was tested with ECG records taken from the MIT-BIH database [6]. Timing accuracy must be tested with an artificial or a single sampled QRS complex repeated periodically to achieve a precise time reference for the test. This cannot be achieved with natural ECG recordings as the time reference is unknown. Also, the annotations in the MIT-BIH records cannot be used as time references as the time resolution in those records is low due to the low sampling frequency of 360 Hz used. The artificial ECG signal was generated according to the recommendation in the

Association for the Advancement of Medical Instrumention (AAMI) standard [7] shown in Fig. 8 and fed to the system being tested with the aid of a (D/A) converter. Several types of noise were added to the signal to evaluate timing accuracy under noisy conditions. Noise types used are shown in Table I. All measurements were carried out using four different SNR’s and four different QRS repetition frequencies, 50, 100, 150, and 200 bpm, and the total number of pulses given were 100, 100, 150, and 150, respectively. The corresponding ECG signal periods were 1199, 601, 399, and 301 ms to prevent synchronization of the sine noise signals with the QRS complex. Testing with different QRS rates is important when adaptive thresholds are used since the adaptation algorithm may affect detection performance at different pulse rates. The standard deviation (sd) and maximum range of the timing error was calculated from each result set. Detections with timing error more than 30% of the pulse period were excluded in the analysis of some test results, but the number of discarded detections was also low in these cases (always smaller than 2.3%). The number of discarded results are shown in the individual test results. The 0.5-Hz additive sine noise simulates the baseline variation due to breathing. The detector shows good tolerance to

RUHA et al.: REAL-TIME MICROPROCESSOR QRS DETECTOR SYSTEM

165

TABLE IV TEST RESULTS WITH MIT-BIH RECORDS 103

=

Fig. 8. The artificial ECG signal (a 0.5–5 mV [typ. 1 mV], d ms, dt 180 ms, at 0.4–1.2 mV and qt 350 ms) [7].

=

=

=

= 70–120

TABLE II THE QRS DETECTION TIMING ACCURACY WITH 5-HZ ADDITIVE NOISE

TABLE III THE QRS DETECTION TIMING ACCURACY WITH ADDITIVE GAUSSIAN NOISE

this type of noise, the detection timing error is less than 0.9 ms (sd) with pulse rates of 50–200 bpm and noise amplitudes up to 80% (160% peak-to-peak) of QRS complex amplitude. The maximum timing error of all detections was 2 ms. The 5-Hz sine wave represents a noise signal with a frequency in the upper frequency range of noise due to motion artifact. The timing error of less than 1 ms (sd) with maximum error of 2 ms was achieved with sine noise amplitude of 40% (80% peak-to-peak) relative to QRS complex amplitude. Greater noise amplitude of 80% increased timing error (sd 4 ms) at highest pulse rate of 200 bpm markedly, Table II. The detector exhibits good tolerance to mains interference as the comb filter effectively removes the 50-Hz mains-noise, the timing error is less than 0.5 ms at noise levels up to 80% (sd) at all pulse rates tested. The maximum timing error of all 1 ms. detections was 2 Gaussian noise simulates noise due to electromyogram (EMG) signals and motion artifacts. With a Gaussian noise level of 25% (rms) of QRS complex amplitude the timing error is less than 1 ms (sd), maximum timing error for all detections at this noise level is 3 3 ms. With higher levels of noise the timing error increased, the results are shown in Table III. The detector is insensitive to amplitude modulation of the QRS complex which is important even in HRV measurement at rest due to amplitude modulation caused by respiration. Amplitude modulation up to 30% allows the QRS complex to be detected with timing error less than 0.7 ms (sd). The

AND

105

maximum timing error was 2 1 ms for 99.5% of detections with pulse frequencies of 50–200 bpm. Higher modulation increased timing error noticeably. In order to assess the reliability of the QRS detections, the system was also tested with clean and noisy ECG signals obtained from the MIT-BIH arrhythmia database [6]. The same measurements were carried out on a commercially available heart-rate detector [8] targeted for sports training to obtain a reference on the validity of the results. The measurement setup consisted of the MIT-BIH database on CD-ROM, a SUN workstation with a CD-ROM drive, and a 16-b D/A data conversion card, so that the setup could be used as a signal source with selected signals from the database. The detections given by the system were compared to the annotations on the records with software in the host computer. Two records from the MIT-BIH database were used in testing. Record number 103 contains a signal with little noise, and the system was able to detect practically all of the 2082 recorded QRS complexes. The record 105 contains some very noisy sections due to motion artifacts and the system detected the QRS complexes with small number of errors, Table IV. The detection error rate (ER) is according to (10) 0.1 and 2.2%, correspondingly Detection error rate Number of false positives Number of false negatives Total number of QRS complexes (10) IV. DISCUSSION A. QRS Detection Timing Accuracy In [9] a sampled ECG signal was used to test the QRS detection performance of nine published algorithms. The noise types were quite similar to the ones used in this work including mains-noise, EMG, baseline variation due to respiration, and motion artifacts. However, no timing accuracy performance was reported in that work. The QRS detection timing accuracy achievable with this system and typical noise levels to be encountered in real measurements are estimated in Table V. The figures are obtained from the results of the tests with the artificial ECG signal corrupted with noise of various types. These tests indicate that a timing accuracy of 1 ms can be achieved in real measurements corrupted with typical levels of noise. The test with 5-Hz additive sine noise representing the higher end of the spectrum of typical noise artefact resulted in a timing error of 0.9 ms (sd) at noise amplitude of 40%. Although the noise level due to motion artifacts in real measurements can be larger in amplitude its spectral content

166

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 44, NO. 3, MARCH 1997

TABLE V THE DETECTION TIMING ACCURACY WITH TYPICAL NOISE LEVELS AND HEART RATES OF 50–200 BPM

is located mainly at lower frequencies than 5 Hz [10] and is thus attenuated by the filtering used in the detector. Thus, an accuracy of 1 ms and good tolerance to noise due to motion artifacts can be expected in real measurements even with higher levels of noise than indicated by the 5-Hz sine noise amplitude. In [11] a timing error of 1.65 ms (sd) was achieved with an SNR of 10 dB (noise rms is 30% of QRS amplitude) in software tests. The noise band was 0–50 Hz, but the noise distribution is not reported. The error sd of 1.8 ms was achieved in [12]. This is probably too optimistic a result as the portion of detections with a timing error >10 ms was 38% of all detections and they were excluded from the sd calculation of detection timing error. The timing error of the developed system was 1 ms (sd) with Gaussian noise at rms value of 25%. The developed system was less affected by 50-Hz noise (sd 0.5 ms) than the algorithm in [12] (sd 2.5 ms). Although these tests are probably not directly comparable due to different test setups the overall performance of the developed system is expected to be favorable in comparison with the performance of [11] and [12]. B. QRS Detection Reliability The detector was able to detect practically all of the QRS complexes of a recorded real signal with little noise (MIT-BIH record 103), and the test with a noisy signal (MIT-BIH record 105) resulted in a detection ER of 2.2% with respect to the total number of QRS complexes in this record. The detection errors with record 105 occur mostly during the noisy sections (total of 4 min) of the record. Record 105 was also used to test the QRS detectors in other published papers so performance comparisons can be made. The test results with other published results on the MIT-BIH recordings are summarized in the Table VI. The performance of the developed system with an ER of 2.2% and number of failed detections (FD) of 56 was better than with linear bandpass filtering method [15] as expected (ER 3.5%, FD 89). The performance also exceeds that obtained by fixed bandpass filtering and threshold detection 2.9%, FD 75), even when a search-back algorithm (ER was used [14]. The performance level achieved is between the performance of linear adaptive filtering (ER 2.4%, FD

TABLE VI TEST RESULTS WITH MIT-BIH RECORDS 105 (TOTAL OF 2572 QRS COMPLEXES)

62) and neural network-based adaptive matched filtering (ER 0.5%, FD 14) [13]. Adaptive filtering is effective in cases where the signal and the corrupting noise are stationary (e.g., mains-noise) or their characteristics alter so slowly that the system is able to adapt to these variations. This is not the case, however, with noisy ECG signals where the noise characteristics, mainly motion artifacts, can vary considerably and adaptive filter probably fails to be the optimal filtering method [16]. The good performance of the developed system can also be partially explained by the fact that the developed method is “adaptive” to each patient’s QRS complex. It is interesting to note that the commercially available heart rate monitor based on a highly optimized linear bandpass filtering and threshold detection achieves a quite high level of performance (ER 2.8%, FD 71) which is comparable with results achieved with a similar method equipped with additional search-back. Only a software implementation of a method based on a neural network was able to offer better performance [13]. However, a QRS detector for ambulatory use imposes requirements for low power consumption and physical size and a microprocessor implementation of a system based on neural networks, depending on network topology and layer and neuron number would probably lead to an unacceptable power consumption due to the high arithmetic capability required. A digital ASIC solution might be feasible although it would probably be expensive in terms of the silicon area usage. Analog very large scale integration (VLSI) implementation offers on advantages in area and power consumption and promising results of analog implementation of neural networks are reported. The analog solutions, however, have disadvantages as they are sensitive to circuit-level nonidealities, offset, noise in electrical circuits, etc., and they require high expertise in design. V. SUMMARY The design and performance of an ambulatory QRS detector are presented in this work. The QRS detector is intended for the accurate measurement of HRV and reliable QRS detection in both ambulatory and clinical use. The QRS detection algorithm uses prefiltering in conjunction with a matched filter and dual edge threshold detection. The purpose of the prefiltering is to attenuate various noise components in order to achieve improved detection reliability. The matched filter further improves SNR and symmetries the QRS complex for

RUHA et al.: REAL-TIME MICROPROCESSOR QRS DETECTOR SYSTEM

the threshold detection which is essential to reduce sensitivity to low-frequency additive and multiplicative noise. The decision for detection is made in real-time and no search-back method is used. The QRS detection timing accuracy and detection reliability of the detector system was tested with an artificially generated ECG signal and with records from the MIT-BIH arrhythmia database. The use of an artificial ECG signal in timing tests allowed the timing reference for the QRS complex to be known and also allowed the accurate control of the noise characteristics. The artificial signal was corrupted with noise which simulate the noise generated by EMG, respiration, motion, and power-line interface. The tests indicate that a timing error less than 1 ms (sd) can be achieved in real measurements even in noisy conditions. In QRS detection reliability tests records from the MIT-BIH database were used as test signals. The QRS detection reliability is better than that of more complicated methods based on linear filtering and search-back or adaptive filtering and detection ER of 0.1 and 2.2% was achieved with records 103 and 105 from the MIT-BIH Arrhythmia database, respectively.

167

[12] K. G. Lindecrantz and H. Lilja, “New software QRS detector algorithm suitable for real time applications with low signal-to-noise ratios,” J. Biomed. Eng., vol. 10, pp. 280–284, May 1988. [13] Q. Xue, Y. H. Hu, and W. J. Tompkins, “Neural-network-based adaptive matched filtering for QRS detection,” IEEE Trans. Biomed. Eng., vol. BME-39, pp. 317–329, Apr. 1992. [14] P. Hamilton and W. Tompkins, “Quantitative investigation of QRS detection rules using the MIT/BIH arrhythmia database,” IEEE Trans. Biomed. Eng., vol. BME-33, no. 12, pp. 1157–1165, 1986. [15] J. Pan and W. Tompkins, “A real time QRS detection algorithm,” IEEE Trans. Biomed. Eng., vol. BME-32, no. 3, pp. 230–236, 1985. [16] N. V. Thakor and Y.-S. Zhu, “Applications of adaptive filtering to ECG analysis: Noise cancellation an arrhythmia detection,” IEEE Trans. Biomed. Eng., vol. BME-38, pp. 785–794, Aug. 1991.

Antti Ruha (S’88–M’88) received the Diploma Engineer and Licentiate of Technology degrees in electronics from the University of Oulu, Finland, in 1988 and 1993, respectively. He is currently developing ambulatory biomedical instrumentation, especially heart rate and heart rate variability measurement instruments. His current research also include applications of low-power VLSI circuit technology for biomedical applications.

REFERENCES [1] M. Malik, T. Cripps, T. Farrell, and A. J. Camm, “Prognostic value of heart rate variability after myocardial infarction. A comparison of different data-processing methods,” Med. Biol. Eng. Comput., vol. 27, pp. 603–611, 1989. [2] Y. Yamamoto, R. L. Hughson, and J. C. Peterson, “Autonomic control of heart rate during exercise studied by heart rate variability spectral analysis,” J. Appl. Physiol., vol. 71, no. 3, pp. 1136–1142, 1991. [3] P. Z. Peebles, Probability, Random Variables, and Random Signal Principles. New York: McGraw-Hill, 1980, ch. 9, pp. 257–267. [4] M. Schwarz, Information Transmission, Modulation, and Noise. New York: McGraw-Hill, 1970, ch. 8, pp. 577–584. [5] G. J. H. Uijen and J. P. C. de Weerd, “Accuracy of QRS detection in relation to the analysis of high-frequency components in the electrocardiogram,” Med. Biol. Eng. Comput., vol. 17, pp. 492–502, 1979. [6] MIT-BIH Database Distribution, Massachusetts Institute of Technology, 77 Massachusetts Avenue, Room 20A-113, Cambridge, MA 02139. [7] “Standard for Cardiac Monitors, Heart Rate Meters and Alarms (draft),” Association for the Advancement of Medical Instrumention (AAMI), Aug. 1981. [8] Polar Sport Tester manufactured by Polar Electro Oy, Professorintie 5, FIN-90440 Kempele, Finland. [9] G. M. Friesen et al., “A comparison of the noise sensitivity of nine QRS detection algorithms,” IEEE Trans. Biomed. Eng., vol. 37, no. 1, pp. 85–98, 1990. [10] N. V. Thakor et al., “Estimation of QRS complex power spectra for design of a QRS filter,” IEEE Trans. Biomed. Eng., vol. BME-31, pp. 702–705, Nov. 1984. [11] V. Barbaro, P. Bartolini, and M. Fierli, “New algorithm for the detection of the ECG fiducial point in the averaging technique,” Med. Biol. Eng. Comput., vol. 21, pp. 129–135, 1991.

Sami Sallinen (S’92–M’93) received the Diploma Engineer degree from the University of Oulu, Finland, in 1994. His professional interests include computing and especially embedded computing and electronics in medicine. He is currently involved in developing software for radiation therapy treatment planning at Varian-Dosetek Oy, Espoo, Finland.

Seppo Nissil¨a received the Diploma Engineer, Licentiate of Technology, and Doctor of Technology degrees in electronics engineering from the University of Oulu, Finland, in 1987, 1990, and 1995, respectively. He is an Acting Professor of Optoelectronics in the Department of Electrical Engineering, University of Oulu. His current research interests include electrooptical measurements, especially fiber-optics; and biomedical measurements and devices, especially noninvasive, ambulatory heart rate, and blood pressure instruments.