Processing of arterial pressure waves with a digital computer

Processing of arterial pressure waves with a digital computer

COMPUTERS AND BIOMEDICAL RESEARCH Processing with C. FRANK STARMER,~ 6,90-96 (1973) of Arterial a Digital PHILIP A. MCHALE, Pressure Computer*...

478KB Sizes 0 Downloads 47 Views

COMPUTERS

AND

BIOMEDICAL

RESEARCH

Processing with C. FRANK STARMER,~

6,90-96 (1973)

of Arterial a Digital

PHILIP A. MCHALE,

Pressure Computer*

Waves

AND JOSEPH C. GREENFIELD,JR.$

Departments of Medicine (Division of Cardiology) and Physiology-Pharmacology, Duke University Medical Center and Veterans Administration Hospital and Department of Computer Science, Duke University, Durham, North Carolina (27710) Received August 28, 1972 A series of algorithms to identify the time of onset and end of ventricular ejection using a central aortic blood pressure and the electrocardiogram are presented. The reliability of these algorithms was evaluated by comparing the index points found by this method to those obtained by an algorithm which uses aortic blood flow. For the 659 cardiac cycles examined in four dogs with widely varying stroke volumes and arterial pressures, the mean difference between the two methods was both positive and negative with a maximum value of 9 f 1 msec (mean & SE). In addition, 153 cardiac cycles were manually identified and compared to the pressure-oriented algorithm identifications. The mean difference in this case was 1 & 1 msec for both the onset and end of ventricular ejection. An analog catheter simulation technique was used to define the minimum frequency response characteristics required of the blood pressure catheter system. A system with a damping factor of 0.4 and a natural resonant frequency of 11 Hz, which yielded a frequency response curve flat to 8 Hz and peaking at 10 Hz, was the minimum system consistent with reliable identification of the onset and end of ventricular ejection. INTRODUCTION

The increased availability of digital computer systems has provided cardiovascular investigators with a powerful tool which is capable of analyzing large volumes of data rapidly and accurately. One of the primary applications consists of identifying the salient features in a particular time-varying waveform. Numerous pattern recognition algorithms for processing electrocardiographic data have been documented (Z-3). However, there is a paucity of algorithms which can be used with hemodynamic waveforms (4, 5). In numerous cardiovascular studies it is useful to identify the time of onset and end of ventricular ejection to allow the data to be divided into its systolic and diastolic components. * This research supported in part by Contract PH-43-67-1440 from the National Heart and Lung Institute and by NIH Grant HL-09711 from the United States Public Health Service. t Recipient of Research Career Development Award l-K4-HL-70,102 from the United States Public Health Service. $ Recipient of Research Career Development Award l-K3-HL-28,112 from the United States Public Health Service. Copyright 0 1973 by Academic Press, Inc. 90 All rights of reproduction in any form reserved.

PROCESSING

ARTEiRIAL

PRESSURE

WAVES

91

Benson (5) recently published an algorithm which uses phasic ascending aortic blood flow to identify these ejection indices, but in most situations, measurement of aortic blood flow is not feasible. However, the central aortie blood pressure waveform is commonly measured. It can be used to define the beginning and end of ejection if the initial rapid upstroke and dicrotic notch can be reliably iden~fied for each cardiac cycle. The purpose of this paper is to present a series of algorithms for recognition of the onset and end of ventricular ejection from the central aortic blood pressure waveform and to define the minimum fidelity characteristics required of the catheter system for reliable identification of these two events.

Analog recordings of central aortic blood pressure, aortic blood flow, and the electrocardiogram were obtained from four awake mongrel dogs which had been chronically instrumented with an electromagnetic flow meter probe and a high fidelity pressure catheter. Wide ranges of aortic blood pressure and flow rates had been obtained by pacing the heart at rates varying from 80 to 1SO~ats~min during the course of several pharmacologic interventions. Five second blocks of la-bit data samples were digitized at a rate of 2OO/sec and entered into an IBM 1130 digital computer for processing. Two tasks must be performed to reliably identify the initial rapid upstroke and the dicrotic notch from the digitized blood pressure data: (1) a segment which has a high probability of containing a single pressure wave must be located and (2) the detailed features within that segment must be identified. Detection of the R wave of the electrocardiogram provides one reliable technique for locating this data segment since a ventricular contraction can be expected to occur 50 to 200 msec later. An alternative method of segment location is to simply search the aortic pressure data for an event which appears to represent the onset ofejection. Wowever, the relatively low amplitude pressure waveforms associated with premature ventricular contractions or ventricular alternans can be mistaken for a noise event and not properly identified. For this reason, electrocardiogram R-wave detection was employed in this study to locate the pressure data segment to be subjected to detailed analysis. Location of the R wave is accomplished by finding a point in the electrocardiographic data where the absolute value of the first derivative exceeds a preset threshold. A least-squares estimate of the first derivative is obtained from 4

m C-G*-2

-Y&l

+Yi+l

+

2Yi+2Y10dt,

(1’)

where the y,s are the indicated EKG data points, dt is the sampling interval, and di is the value of the first derivative at the ith data point. The first 2 set of EKG

92

STARMER,

MCHALE,

AND

GREENFIELD

data are scanned to locate the maximum absolute value of the first derivative. The derivative threshold is set at 50 % of this maximum value (1) and this same threshold is used throughout the analysis of a complete 5-set data block. A new threshold is computed for each new data block. Onset of Ejection Algorithm

The onset of ejection is defined as that point in the pressure wave which is followed by four consecutive first derivatives greater than 10 mm Hg/sec and which occur between 30 and 200 msec after the R wave of the EKG. The derivatives are computed using Eq. 1, where the yis now represent the pressure data points. If no such ejection onset is found, either no mechanical event was associated with the electrical event identified as an R wave, or the identified R wave was artifact. In either case, the EKG data is scanned for the next point to be considered as an R wave, as indicated by the first derivative exceeding the preset threshold, and the above onset of ejection algorithm is repeated. This repeat EKG scan is started 80 msec after the last identified R wave. Peak Systolic Pressure Algorithm

The peak systolic pressure point is defined as the maximum filtered pressure obtained during ejection. This point is identified by searching for the maximum filtered pressure which occurs between the onset of ejection and a point 200 msec after the onset of ejection. Filtered pressure,f,, is computed from a 5-point moving average filter where ,fi z (Pi-2 + Pi-1 + Pi + Pi+, i- Pi+,)/5.

(2)

The filtered values are computed only during the 200 msec search bracket. Dicrotic Notch Algorithm

The dicrotic notch is defined as the point of local minimum pressure occurring in the region of the maximum second derivative. The maximum second derivative is sought between 30 msec after the systolic pressure point and 300 msec after the onset of ejection. A least-squares estimate of the second pressure derivative is obtained from Si = (2Pi-2 -- Pi. 1 - 2Pi - Pi+ 1 + 2Pi+,)/ 14Ar,

(3)

where the Pis are the pressure data points, At is the sampling interval and si is the value of the second derivative at the ith data point. To find the dicrotic notch point, the minimum pressure which occurs within 12 data points (for a 5-point formula) of the maximum second derivative is determined. This point is used to mark the end of ejection. To identify the next pressure wave, an EKG R-wave search is initiated at the point following the last identified dicrotic notch and the series of algorithms are repeated.

PROCESSING

ARTERIAL

PRESSURE

WAVES

93

The reliability of the pressure waveform recognition algorithm was evaluated by comparing the ejection interval index points obtained by the method described above with those obtained from aortic blood flow data. The flow-defined index points were determined using an algorithm previously described by Benson (5). The data from another chronically instrumented animal were used to estimate the fidelity of the catheter system which is necessary for this algorithm to function. An analog computer representation of a catheter system, as described by Fry (6), was implemented on a Simulator model 240 computer. The aortic blood pressure measured with the Statham SF-I catheter-tip manometer was fed into this analog catheter representation and the appropriate coefficients were adjusted to simulate various catheter frequency response characteristics. The modified aortic blood pressure data were then entered into the digital computer, along with the simultaneously measured aortic blood flow and electrocardiogram for processing by the pattern recognition algorithms. The minimum fidelity characteristics required of the catheter system were determined by noting the analog coefficient settings which modified the aortic pressure signal to such an extent that the pressure waveform algorithm no longer reliably identified the onset and end of ventricular ejection. RESULTS

A typical recording of the aortic blood pressure and flow waveforms are shown in Fig. 1. It is assumed that the aortic flow, as obtained from an electromagnetic flowmeter probe, accurately reflects the ventricular ejection phases. Table 1 shows the results of comparing the ejection indices as determined by the aortic pressure algorithm to those obtained from the aortic flow algorithm of Benson (5). Columns 5 and 6 are the mean difference in milliseconds between the ejection indices using the two computer methods. These differences were both positive and negative and the largest difference found was 9 i 1 msec (mean i SE). In addition, the algorithms presented here identify the index points more rapidly than the flow-oriented algorithm. The individual digitized data points for the 153 cardiac cycles obtained in dog 4 were also displayed and the index points were manually identified and compared to those found by the two algorithms. Using aortic pressure, the manual and computer methods differed by only 1 & 1 msec at both the onset and end of ejection. With the aortic blood flow, the manual and computer methods differed by 5 & 1 msec and 7 f 3 msec for the two ejection index points. Thus, it can be seen that the aortic blood pressure-EKG algorithm described above reliably identifies the onset and end of ejection under a wide variety of hemodynamic conditions. By modifying the recorded pressure signal with analog catheter simulation prior to entry into the digital computer, it was possible to define the minimum frequency response characteristics required for these algorithms to work reliably. A simulated system with a damping factor of 0.4 and a natural resonant frequency of 11 Hz,

94

STARMER,

MCHALE,

AND

GREENFIELD

1 sec.

FIG. 1. Example of waveforms used in identification EE = end of ejection.

algorithms. BE = beginning of ejection;

which yields a frequency response curve flat to 8 Hz, peaking at 10 Hz then decreasing at a rate of 12 dB per octave, was the minimum system consistent with accurate identification of the ventricular ejection indices. DISCUSSION

The algorithms presented in this report provide a reliable method for identifying the time of onset and end of ventricular ejection using the central aortic blood TABLE COMPA~~NOFFLOW-ORIENTED

1

PRESSURE-ORIENTED AL~RI~MS

AND

Stroke volume range

Mean aortic pressure range

Begin ejection difference

End ejection difference

(cc)

(mm&9

(msec) 9+ 1”

(msec) 211

Dog no.

No. of beats

1

149

lo-80

98-231

2 3 4

198 164 153

lo-52

56-155 56-179 62-226

a Mean f SE,

BEAT~DENTIFICAT~~N

lo-40

11-34

-4 ri; 1

--2 + 1 2t2

--J+

1

I*2 1 rr2

PROCESSING ARTERIAL PRESSURE WAVES

95

pressure waveform. In the series of 659 cardiac cycles presented in Table 1, the largest error was 9 * 1 msec which, at a data sampling rate of 200/set, is less than 2 data points. The pressure waveforms associated with premature ventricular contractions or ventricular alternans can be troublesome if the identification algorithm is based solely on aortic pressure and not on both the EKG and pressure. The relatively small pressures associated with these events can appear similar to diastolic noise and may not be identified properly. Alternatively, a pressure algorithm which is sensitive to small-amplitude pressure waves will sometimes mistake catheter whip artifact during diastole as a true pressure wave. The simultaneous use of EKG R-wave detection to locate segments in the aortic pressure data which are likely to contain a true pressure wave provides a simple and reliable way to identify systolic pressure events of widely varying magnitude. Digital filtering of the pressure data was used during the search for the peak systolic pressure to minimize the effects of noise which may sometimes be associated with the pressure waveform. The derivative formulas used for locating the onset and end of ejection also have inherent filtering. These techniques increase the reliability of the algorithm and minimize the problems associated with systolic noise. Use of selective filtering, as contrasted with filtering the complete pressure wave, minimizes the computer time required for pattern recognition. The frequency response characteristics of the catheter system used to obtain the aortic blood pressure data play an important role in the success or failure of any pattern recognition algorithm. Overdamping the catheter system (damping factor >l) in the range of normally available resonant frequencies (5-15 Hz) tends to smooth the onset of ejection and dicrotic notch and may cause the algorithms to miss these points. Underdamping factor (damping factor ~1) is desirable in this situation since this tends to emphasize the onset of ejection and the dicrotic notch. However, too much underdamping (damping factor ~0.2) introduces severe oscillations following the dicrotic notch which make reliable identification of this index point difficult. The algorithm presented here requires a catheter system with a damping factor of 0.4 with an undamped resonance frequence of 11 Hz. This represents the minimum response characteristics consistent with reliable ventricular ejection index point identification. ACKNOWLEDGMENTS The authors wish to thank the Medical Illustration Service of the Durham Veterans Administration HospitaI for processing the illustrations and Mrs. Rosa B. Ethridge for typing the manuscript. REFERENCES 1. GEDDES,J. S., AND WARNER, H. R. A PVC detection program. Comput. Biomed. Res. 4,493-508 (1971). 2. BONNER, R. D., AND SCHWETMAN, H. D. Computer diagnosis of electrocardiograms. II. A computer program for EKG measurements. Comput. Biomed. Res. 1,366-386 (1968). 4

96

STARMER, MCHALE, AND GREENFIELD

3. WARTAK, J., ~~ILLIKEN, J. A., AND KARCHMAN, J. Computer program for pattern recognition of electrocardiograms. Comput. Biomed. Res. 4, 344-374 (1970). 4. ARMENT, B. E., HIGGENS, L. S., AND PECKHAM, C. G. Automatic data reduction and processing for the cardiopulmonary research laboratory. Comput. Biomed. Res. 3, 108-123 (1970). 5. BENSON, D. W. An algorithm for defining the cardiac cycle using ascending aortic blood flow. Comput. Biomed, Res. 4,216-233 (1971). 6. FRY, D. L. Physiologic recording by modern instruments with particular reference to pressure recording. Physiol. Rev. 40,753-788 (1960).