230 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-32, NO.

3, MARCH 1985

A Real-Time QRS Detection Algorithm

JIAPU PAN AND WILLIS J. TOMPKINS, SENIOR MEMBER, IEEE

Abstract-We have developed a real-time algorithm for detection of a derivative, and a moving window integrator. The nonlinear
the QRS complexes of ECG signals. It reliably recognizes QRS com- transformation that we use is signal amplitude squaring.
plexes based upon digital analyses of slope, amplitude, and width. A Adaptive thresholds and T-wave discrimination techniques pro-
special digital bandpass filter reduces false detections caused by the var-
ious types of interference present in ECG signals. This filtering permits vide part of the decision rule algorithm.
use of low thresholds, thereby increasing detection sensitivity. The algo- The slope of the R wave is a popular signal feature used to
rithm automatically adjusts thresholds and parameters periodically to locate the QRS complex in many QRS detectors [5]. An
adapt to such ECG changes as QRS morphology and heart rate. For analog circuit or a real-time derivative algorithm that provides
the standard 24 h MIT/BIH arrhythmia database, this algorithm cor- slope information is straightforward to implement. However,
rectly detects 99.3 percent of the QRS complexes.
by its very nature, a derivative amplifies the undesirable higher
frequency noise components. Also, many abnormal QRS com-
INTRODUCTION plexes with large amplitudes and long durations are missed in
THERE are many uses for a reliable QRS recognition algo- a purely derivative approach because of their relatively low R-
rithm. Computer interpretation of the 12-lead ECG is a wave slopes. Thus, R-wave slope alone is insufficient for proper
popular technique. Coronary care units now use arrhythmia QRS detection. To achieve reliable performance, we must ex-
monitors extensively. Widely used Holter tape recording re- tract other parameters from the signal such as amplitude, width,
quires a Holter scanning device that includes a QRS detector and QRS energy [6], [7].
to analyze the tapes much faster than real time. Currently It is very important to evaluate a QRS detector algorithm
under development are arrhythmia monitors for ambulatory using a standard arrhythmia database. There are now two such
patients which analyze the ECG in real time [1] -[31. When databases available: MIT/BIH [8] and AHA (American Heart
an arrhythmia appears, such a monitor can be programmed to Association) [9] . The performance of an algorithm on a data-
immediately store an interval of the abnormal ECG for sub- base is not the ultimate answer as to its utility in a clinical en-
sequent transmission to a central station where a physician can vironment, but it provides a standardized means of comparing
interpret it. Such a device requires a very accurate QRS recog- the basic performance of one algorithm to another.
nition capability. False detection results in unnecessary trans- ALGORITHM OVERVIEW
mission of data to the central station or requires an excessively We implemented the QRS detection algorithm in assembly
large memory to store any ECG segments that are unnecessarily language. It operates on either a Z80 (Zilog) or an NSC800
captured. Thus, an accurate QRS detector is an important part (National Semiconductor) microprocessor. All the processing
of many ECG instruments. is done with integer arithmetic so that the algorithm can op-
QRS detection is difficult, not only because of the physio- erate in real time without requiring excessive computing power.
logical variability of the QRS complexes, but also because of The database provides two simultaneous ECG channels. We
the various types of noise that can be present in the ECG signal. attempted two-channel analysis, but abandoned this approach.
Noise sources include muscle noise, artifacts due to electrode Due to the way that the electrode positions are orthogonally
motion, power-line interference, baseline wander, and T waves placed in Holter recording, a high-quality signal on one channel
with high-frequency characteristics similar to QRS complexes. normally implies a low-amplitude ECG with a poor signal-to-
In our approach, digital filters reduce the influence of these
noise ratio on the second channel. The only way that two-
noise sources, and thereby improve the signal-to-noise ratio. Of channel algorithms will yield improved performance for most
the many QRS detectors proposed in the literature, few give
serious enough attention to noise reduction.
patients is by adopting a new way of electrode placement that
will provide usable signals in both channels.
Software QRS detectors typically include one or more of Fig. 1 shows signals at various steps in digital signal processing.
three different types of processing steps: linear digital filtering, First, in order to attenuate noise, the signal passes through a
nonlinear transformation, and decision rule algorithms [4]. We digital bandpass filter composed of cascaded high-pass and low-
use all three types. Linear processes include a bandpass filter, pass filters. Fig. l(b) shows the output of this filter. The next
Manuscript received July 27, 1984; revised October 24, 1984. This
process after filtering is differentiation [see Fig. 1 (c)] , followed
work was supported in part by the National Institutes of Health under by squaring [see Fig. 1 (d)], and then moving window integra-
Grant HL00765. tion [see Fig. 1 (e)]. Information about the slope of the QRS
J. Pan is with the Department of Biophysics, Shanghai Second Medical is obtained in the derivative stage. The squaring process in-
College, Shanghai, People's Republic of China.
W. J. Tompkins is with the Department of Electrical and Computer tensifies the slope of the frequency response curve of the deriv-
Engineering, University of Wisconsin, Madison, WI 53706. ative and helps restrict false positives caused by T waves with

00 18-9294/85/0300-0230$0 1.00 © 1985 IEEE

PAN AND TOMPKINS: REAL-TIME QRS DETECTION ALGORITHM 231

(a)
(a)
(b)
(b)
(c)
(c)
(d)

(d)
(e)

(f) (e)

(g) Fig. 2. QRS detection algorithm processing steps for a noise-contam-

inated ECG from the MIT/BIH database. (a) Original signal. (b) Out-
put of bandpass filter. (c) Results of moving-window integration.
(d) Original ECG signal delayed by the total processing tine. (e)
Fig. 1. QRS detection algorithm processing steps for a normal ECG from Output pulse stream.
the MIT/BIH database. (a) Original signal. (b) Output of bandpass
filter. (c) Output of differentiator. (d) Output of squaring process.
(e) Results of moving-window integration. (f) Original ECG signal
delayed by the total processing time. (g) Output pulse stream. ECG with this digital bandpass filter improves the signal-to-
noise ratio and permits the use of lower thresholds than would
be possible on the unfiltered ECG. This increases the overall
higher than usual spectral energies. The moving window inte- detection sensitivity. The detection thresholds float over the
grator produces a signal that includes information about both noise that is sensed by the algorithm. This approach reduces
the slope and the width of the QRS complex. Fig. 1(f) is the the number of false positives caused by types of noise that
same as the original ECG in Fig. 1(a) except delayed by the mimic the characteristics of the QRS complex.
total processing time of the detection algorithm. Fig. l(g) The algorithm uses a dual-threshold technique to find missed
shows the final output stream of pulses marking the locations beats, and thereby reduce false negatives. There are two sepa-
of the QRS complexes after application of the adaptive rate threshold levels in each of the two sets of thresholds. One
thresholds. level is half of the other. The thresholds continuously adapt
Fig. 2 illustrates a set of signals similar to those in Fig. 1 for to the characteristics of the signal since they are based upon
a noise-contaminated ECG in the database. The algorithm is the most-recent signal and noise peaks that are detected in the
able to correctly detect QRS complexes in the presence of the ongoing processed signals. If the program does not find a QRS
severe noise typical of the ambulatory ECG environment. We complex in the time interval corresponding to 166 percent of
based all judgments of correctness upon the annotations in the the current average RR interval, the maximal peak detected in
database. Each annotation on the location and morphology that time interval that lies between these two thresholds is con-
of a beat was determined by arbitration between two cardiol- sidered to be a possible QRS complex, and the lower of the
ogists who had to be in agreement on all beats in order for an two thresholds is applied. In this way, we avoid requiring a
ECG data segment to be placed in the database. long memory buffer for storing the past history of the ECG,
The algorithm is divided into three processes: learning phase and thus require minimal computing time to accomplish the
1, learning phase 2, and detection. Learning phase 1 requires search-back procedure to look for a missing QRS complex.
about 2 s to initialize detection thresholds based upon signal and Unfortunately, the dual-threshold technique is only useful
noise peaks detected during the learning process. Learning if the heart rate is regular. In abnormalities such as bigeminy
phase 2 requires two heartbeats to initialize RR -interval average or trigeminy, we cannot find the missed beat by searching back
and RR-interval limit values. The subsequent detection phase through the same time interval as for regular heart rates. For
does the recognition process and produces a pulse for each QRS the case of irregular heart rates, both thresholds are reduced
complex. The thresholds and other parameters of the algorithm by half in order to increase the sensitivity of detection and to
are adjusted periodically to adapt to changing characteristics avoid missing valid beats.
of the signal. Once a valid QRS complex is recognized, there is a 200 ms
We use two sets of thresholds to detect QRS complexes. One refractory period before the next one can be detected since
set thresholds the filtered ECG, and the other thresholds the QRS complexes cannot occur more closely than this physio-
signal produced by moving window integration. By using thresh- logically. This refractory period eliminates the possiblity of a
olds on both signals, we improve the reliability of detection false detection such as multiple triggering on the same QRS
compared to using one waveform alone. Preprocessing the complex during this time interval. When a QRS detection
232 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-32, NO. 3, MARCH 1985

occurs following the end of the refractory period but within Frequency (Hz)
360 ms of the previous complex, we must determine if it is a
valid QRS complex or a T wave. In this case, we judge the
waveform with the largest slope to be the QRS complex.
To be reliable, a QRS detection algorithm must adapt each
of its parameters with time so as to be able to operate properly .L;
for ECG's of different patients as well as for ECG morphology (r)
changes in a single patient. In our algorithm, each threshold
automatically adapts periodically based upon peak values of L

signal and noise. When a QRS must be found using second

(lower) thresholds, threshold readjustment occurs twice as fast
as usual. In the dual-threshold technique, the RR-interval aver- E4EL
age must be updated for each heartbeat.
Two separate measurements of the average RR interval are
maintained. One RR-interval average is the mean of all of the Fig. 3. Amplitude response of the digital bandpass
(3
most recent eight RR intervals. A second RR -interval average dB) is 5-11 Hz.
is the mean of the most recent eight beats that fell within the
range of 92-116 percent of the current RR-interval average. Low-Pass Filter
Without this first average, this approach would be suitable only The transfer function of the second-order low-pass filter is
for a slowly changing and regular heart rate. When the heart rate
62
suddenly changes, the first RR-interval average substitutes for = (1
(I .
-

the second one. The algorithm adapts rapidly to a changing H(Z) z 1)2
(1)
signal. It can even adapt rapidly when switching from one The amplitude response is
patient's ECG to another without requiring special learning
phases. sin2 (3oT)
METHODS
HwT =sin2 (wTI2) (2)
An analog filter bandlimits the ECG signal at 50 Hz. An ana- where T is the sampling period. The difference equation of
log-to-digital converter (ADC) samples the ECG at a rate of 200 the filter is
samples/s. The resulting digital signal passes successively y(nT) = 2y(nT - T) - y(nT - 2 T) + x(nT)
through a sequence of processing steps that includes three
linear digital filters implemented in software. First is an integer- - 2x(nT- 6T)+x(nT- 12T) (3)
coefficient bandpass filter composed of cascaded low-pass and where the cutoff frequency is about 11 Hz and the gain is 36.
high-pass filters. Its function is noise rejection. Next is a filter The filter processing delay is six samples.
that approximates a derivative. After an amplitude squaring
process, the signal passes through a moving-window integrator. High-Pass Filter
Adaptive thresholds then discriminate the locations of the QRS The design of the high-pass filter is based on subtracting the
complexes. output of a first-order low-pass filter from an all-pass filter (i.e.,
Bandpass Filter the samples in the original signal). The transfer function for
such a high-pass filter is
The bandpass filter reduces the influence of muscle noise,
60 Hz interference, baseline wander, and T-wave interference. (1+z1)) (4)
The desirable passband to maximize the QRS energy is approx- H(z)
=

imately 5-15 Hz [10], (111. Our filter is a fast, real-time re- The
cursive filter in which poles are located to cancel zeros on the amplitude response is
unit circle of the z plane [12] . This approach results in a filter
design with integer coefficients. Since only integer arithmetic |H(wT)I = [256 +cossin2(wT/2)
(16cT)] 1/2 (5)
is necessary, a real-time filter can be implemented with a simple
microprocessor and still have available computing power left The difference equation is
to do the QRS recognition task. y(nT) = 32x(nT - 16 T) - [y(nT - T)
This class of filters having poles and zeros only on the unit
circle permits limited passband design flexibility. For our + x(nT) - x(nT - 32 T)] . (6)
chosen sample rate, we could not design a bandpass filter di- The low cutoff frequency of this filter is about 5 Hz, the gain
rectly for the desired passband of 5-15 Hz using this specialized is 32, and the delay is 16 samples.
design technique. Therefore, we cascaded the low-pass and
high-pass filters described below to achieve a 3 dB passband Derivative
from about 5-12 Hz, reasonably close to the design goal. Fig. 3 After filtering, the signal is differentiated to provide the QRS-
shows the overall frequency response. complex slope information. We use a five-point derivative with
PAN AND TOMPKINS: REA L-TIME QRS DETECTION ALGORITHM 233

the transfer function Frequency (Hz)

LI
aU-
H(z) (l/8 T) (-z-2 - 2z- + 2z' + Z2) (7)
The amplitude response is
|H(wT)j = (1/4T) [sin (2coT) + 2 sin (cwT)]. (8 ) 1:J
The difference equation is [7] ~L
,,

y(nT) = (1/8 T) [-x(nT - 2 T) - 2x(nT - T)

L
+ 2x(nT + T) +x(nT+ 2T)]. (9) 1.
I
D
-0
Fig. 4 shows that the frequency response of this derivative is --I

nearly linear between dc and 30 Hz (i.e., it approximates an E

ideal derivative over this range). Its delay is two samples.

V
Squaring Function Fig. 4. Amplitude response of the digital derivative filter.
After differentiation, the signal is squared point by point.
The equation of this operation is
y(nT)= [x(nT)]2. (10)
This makes all data points positive and does nonlinear amplifi-
cation of the output of the derivative emphasizing the higher
frequencies (i.e., predominantly the ECG frequencies).
Moving-Window Integration
The purpose of moving-window integration is to obtain wave- (W-QS)
form feature information in addition to the slope of the R wave. Fig. 5. The relationship of a QRS complex to the moving integration
It is calculated from waveform. (a) ECG signal. (b) Output of moving-window integrator.
QS: QRS width. W: width of the integrator window.
y(nT) = (1/N) [x(nT- (N - 1) T) +x(nT- (N - 2) T)
+ * * + x(nT)]
-
(11)
where N is the number of samples in the width of the integra- The higher of the two thresholds in each of the two sets is
tion window. used for the first analysis of the signal. The lower threshold
Fig. 5 shows the relationship between the moving-window is used if no QRS is detected in a certain time interval so that
a search-back technique is necessary to look back in time for
integration waveform and the QRS complex. The number of
samples N in the moving window is important. Generally, the the QRS complex. The set of thresholds initially applied to
width of the window should be approximately the same as the the integration waveform is computed from
widest possible QRS complex. If the window is too wide, the SPKI = 0.125 PEAKI + 0.875 SPKI
integration waveform will merge the -QRS and T complexes
together. If it is too narrow, some QRS complexes will produce (if PEAKI is the signal peak) (12)
several peaks in the integration waveform. These can cause dif- NPKI = 0.125 PEAKI + 0.875 NPKI
ficulty in subsequent QRS detection processes. The width of
the window is determined empirically. For our sample rate (if PEAKI is the noise peak) (13)
of 200 samples/s, the window is 30 samples wide (150 ms). THRESHOLD Il = NPKI + 0.25 (SPKI - NPKI) (14)

Fiducial Mark THRESHOLD I2 = 0.5 THRESHOLD Il (15)

The QRS complex corresponds to the rising edge of the inte- where all the variables refer to the integration waveform:
gration waveform. The time duration of the rising edge is equal
PEAKI is the overall peak,
to the width of the QRS complex. A fiducial mark for the
SPKI is the running estimate of the signal peak,
temporal location of the QRS complex can be determined from NPKI is the running estimate of the noise peak,
this rising edge according to the desired waveform feature to
THRESHOLD Il is the first threshold applied, and
be marked such as the maximal slope or the peak oftheR wave.
THRESHOLD 12 is the second threshold applied.
Adjusting the Thresholds A peak is a local maximum determined by observing when the
The thresholds are automatically adjusted to float over the signal changes direction within a predefined time interval. The
noise. Low thresholds are possible because of the improve- signal peak SPKI is a peak that the algorithm has already estab-
ment of the signal-to-noise ratio by the bandpass filter. lished to be a QRS complex. The noise peak NPKI is any peak
234 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-32, NO. 3, MARCH 1985

that is not related to the QRS (e.g., the Twave). The thresh- RR AVERAGE2 = 0.125 (RR'_7 + RR'?6 + * * + RR') -

olds are based upon running estimates of SPKI and NPKI. That (25)
is, new values of these variables are computed in part from their
prior values. When a new peak is detected, it must first be where RR' is the most recent RR interval that fell between
classified as a noise peak or a signal peak. To be a signal peak, the acceptable low and high RR -interval limits. The RR -inter-
the peak must exceed THRESHOLD I1 as the signal is first val limits are
analyzed or THRESHOLD 12 if searchback is required to find RR LOW LIMIT = 92% RR AVERAGE2 (26)
the QRS. When the QRS complex is found using the second
threshold-, RR HIGH LIMIT = 116% RR AVERAGE2. (27)
SPKI = 0.25 PEAKI + 0.75 SPKI. (16) RR MISSED LIMIT = 166% RR AVERAGE2 (28)
The set of thresholds applied to the filtered ECG is determined If a QRS complex is not found during the interval specified
from by the RR MISSED LIMIT, the maximal peak reserved between
the -two established thresholds is considered to be a QRS
SPKF = 0.125 PEAKF + 0.875 SPKF candidate.
(if PEAKF is the signal peak) (17) If each of the eight most-recent sequential RR intervals that
are calculated from RR AVERAGE1 is between the RR LOW
NPKF = 0.125 PEAKF + 0.875 NPKF LIMIT and the RR HIGH LIMIT, we interpret the heart rate
(if PEAKF is the noise peak) (18) to be regular for these eight heart beats and
THRESHOLD F1 = NPKF + 0.25 (SPKF - NPKF) (19) RR AVERAGE2>-RR AVERAGE1. (29)
THRESHOLD F2 = 0.5 THRESHOLD Fl (20) This is the case for normal sinus rhythm.
where all the variables refer to the filtered ECG: T-Wave Identification
PEAKF is the overall peak, When an RI interval is less than 360 ms (it must be greater
SPKF is the running estimate of the signal peak, than the 200 ms latency), a judgment is made to determine
NPKF is the running estimate of the noise peak, whether the current QRS complex has been correctly identified
THRESHOLD Fl is the first threshold applied, and or whether it is really a T wave. If the maximal slope that
THRESHOLD F2 is the second threshold applied. occurs during this waveform is less than half that of the QRS
waveform that preceded it, it is identified to be a Twave; other-
When the QRS complex is found using the second threshold,
wise, it is called a QRS complex.
SPKF = 0.25 PEAKF + 0.75 SPKF. (21)
EVALUATION
For irregular heart rates, the first threshold of each set is re- We used the MIT/BIH arrhythmia database to evaluate the
duced by half so as to increase the detection sensitivity and to QRS detection algorithm [8]. The database consists of 48
avoid missing beats: half-hour recordings for a total of 24 h of ECG data. The data-
THRESHOLDII v- 0.5 THRESHOLD II (22) base is on four-channel FM magnetic tape. Channels I and 2
are the two-channel ECG signals. Channel 3 is an annotation
THRESHOLD Fl 0.5 THRESHOLD Fl. (23)
channel recorded in a standard binary format, and channel 4
To be identified as a QRS complex, a peak must be recognized is a binary-recorded timing track.
as such a complex in both the integration and bandpass-filtered Fig. 6 shows the experimental setup for evaluation of the
waveforms. QRS algorithm. It includes two four-channel FM tape re-
Adjusting the Average RR Interval and Rate Limits corders, two Z80-based laboratory microcomputer systems, a
dc amplifier, and a nonfade oscilloscope.
Two RR-interval averages are maintained. One is the average Using the QRS detection algorithm, one of the microcom-
of the eight most-recent beats. The other is the average of the puter systems samples and analyzes the database ECG played
eight most-recent beats having RR intervals that fall within cer- back from one of the FM recorders. If it detects a QRS com-
tain limits. The reason for maintaining these two separate aver- plex, it generates a pulse. Simultaneously, the second micro-
ages is to be able to adapt to quickly changing or irregular heart computer monitors the encoded beat annotation channel of
rates. The first average is the mean of the eight most-recent the recorder and generates a pulse coincident with the fiducial
sequential RR intervals regardless of their values. mark annotated in the database.
RR AVERAGE1 = 0.125 (RRIn-I7 +RRIn-6 + -+RRn) A second four-channel FM tape recorder records the original
ECG waveform being analyzed on one channel and the pulses
(24) from each of the microcomputer systems on two other chan-
where RR n is the most-recent RR interval. nels. All of these signals appear on a nonfade display oscillo-
The second average is based on selected beats. scope or a chart recorder for further visual evaluation.
PAN AND TOMPKINS: REAL-TIME QRS DETECTION ALGORITHM 23 5

The maximal amplitude of the ECG from the FM recorder is

less than +1 V, so a dc amplifier amplifies the signal by a gain
of 2.5 to ensure that the signal uses the full range of the A/D
converter (the input voltage range is ±2.5 V).
The 24 h MIT/BIH database contains more than 116 000
beats. Table I summarizes the performance of our algorithm
Fig. 6. Experimental setup for evaluating the QRS detection algorithm for this database. It produced 507 false positive beats (0.437
using the MIT/BIH database. 1: ECG signal from MIT/BIH database percent) and 277 false negative beats (0.239 percent) for a total
channel 1. 2: Binary annotation from MIT/BIH database channel 3.
3: Sampled and reconstructed ECG signal. 4: Pulse generated by detec- detection failure of 0.675 percent.
tion algorithm. 5: Pulse obtained from annotation track fiducial mark. The problem tapes are characterized, in general, by stretches
of noise, baseline shifts, and artifacts. Tape 108 has unusually
tall, peaked P waves quite uncharacteristic in morphology of
TABLE I
RESULTS OF EVALUATING THE REAL-TIME QRS DETECTION ALGORITHM typical P waves. Particularly at the beginning and end of this
USING THE MIT/BIH DATABASE tape, these P waves are classified as QRS complexes because of
their high slopes. This leads to a high false positive count on
Total Failed Failed this tape. Tape 222 also has some non-QRS waves with highly
Tape (No. FP FN Detection Detection
(No.) Beats) (Beats) (Beats) (Beats) (%) unusual morphologies that lead to false positives. A false posi-
tive detection for one of these bizarre waveshapes can some-
100 2273 0 0 0 0 times lead to a false negative on the following QRS complex
101 1865 5 3 8 0.43 because the algorithm includes a latency time from the time a
102 2187 0 0 0
103 2084 0 0 0
0
0
QRS is detected to the next permitted detection.
104 2230 1 0 1 0.04
SUMMARY
105 2572 67 22 89 3.46
106 2027 5 2 7 0.05 We have developed an on-line real-time QRS detection algo-
107 2137 0 2 2 0.09 rithm and implemented it in Z80 assembly language. This
108 1763 199 22 221 12.54
109 2532 0 1 1 0.04 algorithm reliably detects QRS complexes using slope, ampli-
111 2124 1 0 1 0.05 tude, and width information. A bandpass filter preprocesses
112 2539 0 1 1 0.04
113 1795 0 0 0 0
the signal to reduce interference, permitting the use of low-
114 1879 3 17 20 1.06 amplitude thresholds in order to get high detection sensitivity.
115 1953 0 0 0 0 In the algorithm, we use a dual-thresholds technique and search-
116 2412 3 22 25 1.04
117 1535 1 1 2 0.13 back for missed beats.
118 2275 1 0 1 0.04 The algorithm periodically adapts each threshold and RR in-
119 1987 1 0 1 0.05 terval limit automatically. This adaptive approach provides
121 1863 4 7 11 0.59
122 2476 1 1 2 0.08 for accurate use on ECG signals having many diverse signal char-
123 1518 0 0 0 0 acteristics, QRS morphologies, and heart rate changes. In evalu-
124 1619 0 0 0 0
ations using the MIT/BIH arrhythmia database, the algorithm
200 2601 6 3 9 0.35
201 1963 0 10 10 0.51 failed to properly detect only 0.675 percent of the beats.
202 2136 0 4 4 0.19 REFERENCES
203 2982 53 30 83 2.78
205 2656 0 2 2 0.08 [1] N. V. Thakor, J. G. Webster, and W. J. Tompkins, "Design, im-
207 1862 4 4 8 0.43 plementation, and evaluation of a microcomputer-based portable
208 2956 4 14 18 0.60 arrhythmia monitor," Med. Biol. Eng. Comput., vol. 22, pp. 151-
209 3004 3 0 3 0.10 159, 1984.
210 2647 2 8 10 0.38 [21 R. G. Mark, G. B. Moody, W. H. Olson, S. K. Peterson, P. S.
212 2748 0 0 0 0 Schulter, and J. B. Walters, Jr., "Real-time ambulatory arrhyth-
213 3251 1 2 3 0.09 mia analysis with a microcomputer," Comput. Cardiol., pp. 57-
214 2262 2 4 6 0.26 62, 1979.
215 3363 0 1. 1 0.03 [3] L. Patomaki, J. Forsti, V-M. Nokso-Koivisto, Y. Jokinen, and E.
217 2208 4 6 10 0.45 Lansimies, "On line recording and analysis of the ECG in ambula-
219 2154 0 0 0 0 tory patients," Comput. Cardiol., pp. 173-175, 1981.
220 2048 0 0 0 0 [4] 0. Pahlm and L. Sornmo, "Software QRS detection in ambulatory
221 2427 2 0 0 0.08 monitoring-A review," Med. Biol. Eng. Comput., vol. 22, pp.
222 2484 101 81 182 7.33 289-297, 1984.
223 2605 1 0 1 0.04 [5] M. L. Ahlstrom and W. J. Tompkins, "Automated high-speed
228 2053 25 5 30 1.46 analysis of Holter tapes with microcomputers," IEEE Trans.
230 2256 1 0 1 0.04 Biomed. Eng., vol. BME-30, pp. 651-657, Oct. 1983.
231 1886 0 0 0 0 [61 M. Nygards and L. Sornmo, "A QRS delineation algorithm with
232 1780 6 1 7 0.39 low sensitivity to noise and morphology changes," Comput.
233 3079 0 1 1 0.03 Cardiol., pp. 347-350, 1981.
234 2753 O- 0 0 0 [71 A. Ligtenberg and M. Kunt, "A robust-digital QRS-detection algo-
rithm for arrhythmia monitoring," Comput. Biomed. Res., vol.
48 patients 116 137 507 277 784 0.675 16, pp. 273-286, 1983.
236 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. BME-32, NO. 3, MARCH 1985

[8] "MIT/BIH arrhythmia database-Tape directory and format spec- and the second year was at the University of Wisconsin, Madison, where
ification," Document BMEC TR00, Mass. Inst. Technol., Cam- he completed the work reported here.
bridge, 1980. Database is available from Bioengineering Divi-
sion KB-26, Beth-Israel Hospital, 330 Brookline Avenue, Boston,
MA 02215.
[9] "AHA database." Database is available from Emergency Care Re-
search Institute, 5200 Butler Pike, Plymouth Meeting, PA 19462.
[10] H. G. Goovaerts, H. H. Ros, T. J. vanden Akker, and H. Schneider,
"A digital QRS detector based on the principle of contour lim- f Willis J. Tompkins (S'61-M'66-SM'77) received
iting,"IEEE Trans. Biomed. Eng., vol. BME-23, p. 154, 1976. _ the B.S. and M.S. degrees in electrical engi-
[11] N. V. Thakor, J. G. Webster, and W. J. Tompkins, "Optimal QRS
detector," Med. BioL Eng. Comput., vol. 21, pp. 343-350, 1983.
>
| lmW
neermg from the University of Maine, Orono,
in 1963 and 1965, respectively, and the Ph.D.
[12] P. A. Lynn, "Online digital filter for biological signals: Some fast degree in biomedical electronic engineering from
designs for a small computer," Med. Biol. Eng. Comput., vol. 15, *; £0i the University of Pennsylvania, Philadelphia, in
pp. 534-540, 1977. 1973.
From 1965 to 1968 he was an Electrical Engi-
neer at Sanders Associates, Inc., Nashua, NH,
where he worked on research and development
of data storage systems. He was employed from
1973 to 1974 at the Hospital of the University of Pennsylvania as a Bio-
medical Engineer. Since 1974 he has been on the Faculty of the Uni-
Jiapu Pan received the M.D. degree in 1962 with versity of Wisconsin, Madison, where his teaching specialty is on the
a specialty in biophysics in medicine from the topic of computers in medicine, an area in which he has developed two
Shanghai Second Medical College, Shanghai, new courses. His research interests include on-line biomedical computing
People's Republic of China. applications. He is co-editor, with J. G. Webster, of the textbookDesign
Since that time, he has been on the Faculty of Microcomputer-Based Medical Instrumentation (Englewood Cliffs,
of the Department of Biophysics of the Shanghai NJ: Prentice-Hall, 1981) and co-editor, with J. G. Webster, A. M. Cook,
Second Medical College. His background in- and G. C. Vanderheiden, of the textbook Electronic Devices for Re-
cludes research in medical applications of lasers, habilitation (New York: Wiley, 1984).
flow cytometry for single cell analysis and Dr. Tompkins is a member of the IEEE Engineering in Medicine and
sorting, artificial arm control using electromyo- Biology Society (currently Region IV Administrative Committee Rep-
graphic signals, body surface mapping in electro- resentative and Vice President for Technical Activities), the IEEE Com-
cardiography, and microcomputer-based medical instrumentation. Re- puter Society, the Association for the Advancement of Medical Instru-
cently he spent two years in the United States as a visiting scholar. The mentation, and the Biomedical Engineering Society. He is a Registered
first of these was at Case Western Reserve University, Cleveland, OH, Professional Engineer in the State of Wisconsin.