CROSS-REFERENCE TO RELATED APPLICATIONS

The following applications were filed concurrently with this application:

C. M. P. Kierney, et al., "Non-Invasive Blood Flow Measurements Utilizing Cardiac Cycle Synchronization", Ser. No. 730,456, and is assigned-in-part to the same assignee as this application; and

W. T. Hartwell, et al., "Non-Invasive Blood Flow Measurements Utilizing Autoregressive Analysis With Averaged Reflection Coefficients", Ser. No. 730,487, and is assigned to the same assignee as this application.

TECHNICAL FIELD

This invention pertains to the detection of heart rate. In particular, the invention relates to analyzing electrocardiogram signals (EKG) utilizing autoregressive analysis techniques for eliminating noise and for detecting the heart rate using a digital peak detection technique.

BACKGROUND OF THE INVENTION

The electrocardiogram (EKG) is a commonly monitored vital sign. The QRS complex is the dominant feature of the EKG signal and is used in many clinical instruments such as simple cardiotachometers, arrhythmia monitors, and implantable pacemakers. Under normal conditions, in the absence of muscle artifacts and other electrical noise, a large signal-to-noise ratio prevails and techniques exist for the detection of the QRS complex in particular, R-wave detection. However, in some applications, such as ergonomics, a patient may undergo severe physical stress. The EKG signal will be corrupted with a large non-stationary stochastic muscle artifact signal (EMG) together with extraneous transient and continuous noise components due to electromagnetic interferences. The EMG signal results in color noise being introduced into the EKG signal while 60 Hz electromagnetic interference results in white noise being introduced into the EKG signal. Other electromagnetic interference at higher frequencies will also result in white noise.

In addition, there exists medical techniques which require that the precise start of the cardiac cycle be determined from the EKG signals. One such technique is detailed in the copending application of C. M. P. Kierney, et al. The latter application discloses the utilization of autoregressive analysis techniques for analyzing reflected ultrasonic Doppler shifted signals resulting from the flow of blood cells within internal blood vessels. In Kierney, it is necessary to precisely determine the start and end of each cardiac cycle, so that the Doppler shifted signals for a particular cycle can be divided into a predefined number of time segments so that an autoregressive analysis can be performed on each of these time segments. The start and the end of the cardiac cycles must be precisely determined since the results of the autoregressive analysis are averaged over a plurality of cardiac cycles, and the analysis must be performed at the same relative point in time for each of these cycles.

Therefore, there exists a need for a technique for determining a patient's heart rate from an EKG signal that has various noise artifacts in it. In particular, the need exists for a technique that can reliably and accurately extract the heart rate information from an EKG signal that has been corrupted by both color and white noise.

SUMMARY OF THE INVENTION

In an illustrative method and structural embodiment, a departure in the art is achieved by performing an autoregressive analysis of EKG signals to remove color noise, low-pass digital filtering to remove white noise and determining the heart rate and start of cardiac cycles utilizing a peak detection technique.

Advantageously, a cardiac cycle rate detection system comprises an electrocardiogram (EKG) instrument for obtaining an EKG signal from a patient; and after the EKG signal is digitized by an analog-to-digital converter, a first set of program instructions executed by a computer performs an autoregressive analysis filtering of the EKG digitized samples to remove color noise. A second set of program instructions is responsive to the samples resulting from the autoregressive analysis filtering to perform a digital low-pass filtering on the latter samples to remove high-frequency white noise. Finally, a third set of program instructions is responsive to the samples resulting from the low-pass filtering to determine peaks representing the start of cardiac cycles and from the periodicity of those peaks the cardiac cyclic rate.

Advantageously, the autoregressive analysis filtering is performed by calculating linear filter that model the color noise source causing. Results from the autoregressive analysis filtering are residual samples that represent the difference between the signal predicted by the model and each EKG sample.

In addition, the low-pass filtering is implemented by storing an array of filter coefficients defining the low-pass filter and a set of residual samples from the autoregressive filtering analysis, multiplying the array of coefficients times the set of stored residual samples, and summing the multiplication results. The latter sum equals one output sample of the low-pass filter. Next, the set of stored residual samples is updated by another residual sample replacing a present residual sample, and the multiplication and summing process is repeated to produce another output sample of the low-pass filter.

Advantageously, the third set of program instructions locates the peak of maximum amplitude contained within the low-pass filtered samples and then locates peaks of lesser amplitude that are located further from each other and the peak of maximum amplitude by a distance greater than the number of samples equal to the highest expected heart rate. The instructions then measure the distance between adjacent peaks using as a reference the location of the peak of largest amplitude and test for periodicity by comparing successive distance measurements for substantial equality.

A method for determining the heart rate from signals from an electrocardiogram instrument in the absence of significant color or white noise components performs the following steps: digitizing samples from the electrocardiogram instrument, locating the peak of maximum amplitude, locating peaks of lesser amplitude but at a greater distance from each other than the highest heart rate and having an amplitude greater than a predefined percentage of the peak of largest amplitude, measuring the distance between adjacent peaks, testing for periodicity by comparing successive distance measurements for substantial equality, and displaying the periodicity as an indication of the heart rate.

BRIEF DESCRIPTION OF THE DRAWING

FIG. 1 illustrates, in block diagram form, a blood flow analysis system in accordance with this invention;

FIG. 2 illustrates, in flowchart form, the programs executed by computer 106 of FIG. 1 in performing the heart rate detection;

FIGS. 3 and 4 illustrate in greater detail, the program instructions of block 203 of FIG. 2; and

FIG. 5 is a detailed flowchart of block 205 of FIG. 2.

DETAILED DESCRIPTION

A system for analyzing and displaying the heart rate of a patient is illustrated in FIG. 1. EKG 101 is responsive to electrodes 106, 107, and 108 to generate an analog EKG signal containing the QRS complex. This analog EKG signal is first digitized by analog-to-digital converter 102, and then, computer 103 filters and processes the digitized EKG signals and displays the resulting heart rate that is determined from the R-wave of the digitized EKG signal. By executing program instructions, computer 103 first removes the color noise from the data by performing a linear predictive coding (LPC) filtering. This LPC filtering performs the functions of first determining the reflection coefficients which define an inverse filter that models the noise sources and physical structures which have introduced the color noise into the EKG signal. The digitized EKG signals are then filtered by this inverse filter thus removing the color noise. Next, program instructions within computer 103 perform a low pass filtering of the signal resulting from the LPC filtering. Advantageously, this low-pass digital filter can be of the finite impulse response type (FIR). After the LPC and FIR filtering have been performed, computer 103 executes a third set of instructions on the resulting filtered digital signal to determine the position and repetition rate of the R-wave. Once the repetition rate of the R-wave has been determined, it is displayed on video display 104. Advantageously, the resulting information can also be stored on disk storage device 105 for later display.

FIG. 2 illustrates in greater detail the steps necessary for computer 103 to process the digitized EKG signal received from analog-to-digital converter 102. For each set of EKG samples, block 201 determines the number of time channels or frames that the EKG samples are grouped into by dividing the total number of EKG samples by a predefined number which advantageously may be 50. Once the number of channels has been determined, the start and end of each channel within the EKG samples is then determined by block 202. Blocks 203 and 204 perform a LPC filtering on the digitized EKG samples to remove color noise from these digitized signals. This LPC filtering is accomplished by calculating reflection coefficients that model the sources of the color noise and by processing the digitized EKG samples through this LPC filter. The result of the LPC filtering is commonly called the residual which is the difference or error between each sample and the model specified by the reflection coefficients. Block 203 stores the residual/error for each sample. The LPC filtering performed here is done on the basis of time channels or frames; however, it is well known in the art to perform LPC filtering on a sample-by-sample basis.

Once all the digitized EKG samples have been processed by blocks 203 and 204, the resulting residual samples are low-passed filtered by block 205. This low-pass filter has a bandwidth from 0 to 6 Hz and is implemented with a FIR filter. The impulse response of this FIR filter is a

(sin x/X) function.

The low-pass filtering performed by block 205 removes white noise components above 6 Hz and is implemented by doing a digital convolution operation utilizing the previously mentioned function convoluted with the stored residual samples.

Finally, the start of each cardiac cycle and the heart rate is determined by detecting the R-pulse of the QRS complex that is present at the start of each cardiac cycle by block 210; and the results are displayed by block 215. Block 215 also stores the results for later analysis. The R-pulses are detected by implementing a series of operations that first determine the largest pulse present in the data from the FIR filter, next eliminate all samples around this largest pulse by a time equal to the shortest heart beat. Then, other pulses whose amplitudes are greater than 25 percent of the largest pulse are similarly processed. The remaining pulses are removed from the samples. After these remaining pulses have been identified, their periodicity is determined. Greater detail is given later on these operations.

Consider now in greater detail the program illustrated in FIG. 2. The filtering performed by blocks 203 and 204 assumes an filter model given by:

E(z)=X(z) F(z)                                             (1)

where:

E(z) is the z-transform of the driving or LPC residual signal,

X(z) is the z-transform of the input signal, and

F(z) is an all pole filter.

In the sample domain, equation 1 can be written as follows: ##EQU1## which can be rewritten as ##EQU2## where x(n) represents the present time sample, fc represents filter coefficients, and ORDER represents the number of filter elements.

As described in J. D. Markel and A. H. Gray, "Linear Prediction of Speech", Springer-Verlag, Berlin Herdelberg New York, 1980, on page 10, the driving signal e(n) can be interpreted as the prediction error between the actual data sample x(n) and the linear combination of the previous n samples given by ##EQU3## Since the driving function term, can be interpreted as an error, the common analysis procedure is to minimize the sum of the squares of this error as a method for determining the filter coefficients. Once, the filter coefficients have been determined, the difference or error that remains at each time sample between the actual data sample x(n) and the results of the modeling is the output signal from the filter.

Many autoregressive techniques have been developed to minimize this error term; and the one utilized here is the Berg Maximum Entropy method which is described in the paper by L. Marple, "A New Autoregressive Spectrum Analysis Algorithm", IEEE Trans. on Acoustics, Speech, and Signal Processing, Vol. ASSP-28, No. 4, August, 1980, pp. 441-454. The Berg method recursively solves for the filter coefficients, fc(m), by using a forward and backward prediction errors. For any given sample point, the error is calculated by considering the sample points in time preceding the sample point under calculations and sample points prior to the particular point under consideration. The forward error is defined by: ##EQU4## while the backward error is defined by: ##EQU5## where the values of index k range from 1 to (NPTS-ORDER) and fc(0) is defined as 1. NPTS is the number of sample points in the channel being analyzed. Using these two expressions, the problem is to determine a set of filter coefficients that minimize the sum of both the forward and backward errors summed over all sample points subject to a constraint. Thus, the error to be minimized is given by the following equation: ##EQU6## The constraint on the values of fc is that they satisfy the relation ##EQU7## where FC represents the coefficients determined when ORDER-1 terms were used in the prediction equation. This relation is called the Levinson recursion, and the term, rc(ORDER), is frequently referred to as the LPC reflection coefficient.

The filter coefficients fc(k) can be solved for in equation 6 by substituting equations 4 and 5 into equation 6 and taking the partial derivative of this equation with respect to fc(ORDER). The resulting equation is a recursive formula given in terms of the reflection coefficient as follows: ##EQU8##

After the recursive formula in equation 8 has been solved to the desired order of filter, the filtered output of the filter at each sample point is defined by the forward error at that point.

Equation (8) is solved by the fberg routine that is illustrated in FIGS. 3 and 4. This routine calculates equation (8) and verifies that the conditions of equation 7 have been met and implements the Berg method. After the filter has been determined to a sufficient order, the fberg routine defines the values defined by the forward errors as the residual signal, as indicated in block 203.

The fberg routine calculates equation 8 in a recursive manner by first calculating the first reflection filter coefficient for order 1 and then uses this information to calculate a new set of filter coefficients for order 2. Because of this recursive nature, fberg routine is repetitively recalculating equation 8. FIGS. 3 and 4 illustrate the evaluation of equation 8. Blocks 301 through 306, illustrated on FIG. 3, perform the initialization of the various variables used by the subroutine. The initial residual energy, etot, is set equal to the sum of the squares of all the data points of the channel and the backward and forward errors (berr and ferr, respectively) are set equal to corresponding data samples, where the data sample is the digitized Doppler signal. The filter coefficients are initially set equal to 1 by block 306 and the denominator of equation 8 is set equal to the initial residual energy as determined by block 303.

Block 307 determines whether or not the equation has been sufficiently evaluated for the order filter being calculated, and if it has, a return is executed to the program illustrated on FIG. 3 via return block 308. If the filter has not yet been calculated to a sufficient order, then block 309 is executed. The numerator and denominator (num and den, respectively) of equation 8 for this particular order are evaluated by blocks 310 through 313. After the numerator and denominator have been determined, then the reflection coefficient, rc[m], and residual energy for this particular order are evaluated in block 314. Once the reflection coefficient for this order has been determined, then the stepup function is implemented by blocks 315 through 319 to update the previously determined filter coefficients for (ORDER -1) as defined by the Levinson recursion formula and given in equation 7. The highest order filter coefficient is always equal to the reflection coefficient and is set equal to the reflection coefficient by block 320. The past filter coefficients are then updated by blocks 321 through 323 in order for block 317 to evaluate the next set of filter coefficients. The past filter coefficients are designated as FC(n) in equation 7, and as pfc[n] on FIGS. 3 and 4. The forward and backward errors are next updated by blocks 324 through 326 in order to evaluate blocks 311 and 313 in the next iteration. After the forward and backward errors have been updated, control is passed from block 326 to decision block 307 which determines whether or not all the orders have been evaluated. If all the orders have been evaluated, then the contents of the forward error array are stored by block 530 as the residual signals for this particular channel.

Consider now, in greater detail, the low-pass filtering that is performed by block 205 of FIG. 2 as illustrated in greater detail by a flowchart in FIG. 5. The filter implemented by the flowchart of FIG. 5 is of the FIR type having n stages where n may illustratively be 100. The coefficients for each of these stages are stored in elements of the c[ ] array. The input samples are contained in the x[ ] array, and the output samples are stored in the y[ ] array. The output array is later stored in a file for processing by block 210 of FIG. 2. The values of the filtered coefficients stored in c[ ] can be obtained by using the techniques described in the book entitled, Digital Signal Processing, W. D. Stanley, G. R. Dougherty, and R. Dougherty, 1984. The input samples are initially stored in the DATA[ ] array which serves to delay the input signal for latter multiplication by the filter coefficients stored in the c[ ] array. Further, the FIR filter illustrated in FIG. 5 is a linear phase filter.

As illustrated in FIG. 5, blocks 501, 502, and 503 are used to initialize the appropriate variables. As each input sample is read, it is stored in the first element of the DATA[ ] array. The output signal calculations are performed by elements 503, 506, and 507 for all of the elements of the FIR filter. The result of those calculations for each input sample are computed in the variable SUM. The output signal y() is delayed by a number corresponding to half the number of filter stages in relation to the input samples x() by block 508. Once the output sample has been calculated for an input sample, the data array is updated for the next input sample by blocks 509 through 512. Blocks 513 and 514 control the sequencing of the flowchart shown in FIG. 5 so that all of the input samples are processed. After all of the input samples have been processed, the low-pass filtering function is done and an exit is made from the flowchart illustrated in FIG. 5 under control of block 514.

Consider now in greater detail how block 210 performs the rate detection. The program illustrated in Appendix A is a complete implementation of block 210. The routine first identifies and orders all pulses present in the signal from block 205 in descending order down to pulses that are more than 25 percent of the largest pulse amplitude. In addition, any pulses having an amplitude of less than 25 percent of the largest pulse are removed. As the pulses are found, their position within the test data is recorded. After each pulse has been identified, the routine removes all samples on either side of this largest pulse plus or minus the amount of time of the fastest heart beat which advantageously can be 300 beats/second. Then, the routine scans the test data backward in time from the largest pulse until it finds a second pulse. The routine then proceeds to attempt to go further backward in time by an amount equal to the differences between the largest pulse and the second pulse, plus or minus an error amount that is defined as T which advantageously may be 0.04 seconds. If no pulse is found at this location, it is assumed that the second pulse was a noise pulse. The second pulse is eliminated from the test data and the first step is repeated. However, if a third pulse is found at the point where it was predicted to occur based on the time difference between the largest pulse and the second pulse found, this fact indicates that the second pulse is valid and the program steps backward looking for a fourth pulse at the appropriate place using the same technique as previously described. This procedure is then repeated going forward in time. If going first backwards in time fails, then the process is repeated going first forward in time and assuming that the original set of pulses is available. The resulting pulse amplitudes and their locations found by the previous steps define the starting points of the cardiac cycles of the test results and also the heart rate is defined by periodicity of the pulse locations.

It is to be understood that the above-described embodiment is merely illustrative of the principles of this invention; other arrangements may be devised by those skilled in the art without departing from the spirit and scope of the invention. In particular, one skilled in the art could determine that in the absence of substantial amounts of color and white noise that the peak detection technique alone could be used to detect and display the heart rate. ##SPC1##