Acoustic Assessment of the Physical Integrity of Bjo¨rk-Shiley Convexo-Concave Heart Valves
Background Several lines of evidence indicate a two-stage failure mode for the Bjo¨rk-Shiley convexo-concave (C/C) heart valve, in which one of the two outlet strut legs separates from the flange before the other, potentially providing an opportunity to identify and prophylactically replace failure-prone valves. Radiographic single leg separation (SLS) detection, although successful, is subjective and skill intensive, implying a need for both an objective preliminary screen and subsequent corroboration of the radiographic findings.
Methods and Results We developed a time-windowed, power density analysis of C/C valve closing sounds to detect the vibrational resonance that characterizes the presence of an intact outlet strut in clinically functioning, 29-mm-flange size C/C valves. Recordings from more than 800 patients enrolled in radiographic SLS detection studies were analyzed, and the assessment algorithm was evaluated through a blinded test of 32 study valves for which the true status became known consequent to an autopsy or surgical explantation. Valves were objectively scored on a 0-to-1 scale, with 1 being assuredly intact and scores of <0.50 indicating a probable SLS. All except five valves (incorrectly designated probable SLS) were classified correctly, for a sensitivity of 1.00 (95% confidence interval, 0.79 to 1.00) and a specificity of 0.69 (0.41 to 0.89).
Conclusions This level of accuracy is sufficient to serve as an effective preliminary screen, potentially allowing a threefold concentration of SLS prevalence among the C/C valves of patients undergoing radiographic assessment. The value of acoustic classification in avoiding unnecessary operations prompted by false-positive radiographs is less certain.
When mechanical failure of the Bjo¨rk-Shiley C/C heart valve occurs, it usually appears to be a two-stage process in which one leg of the two-legged outlet strut separates from the valve flange, leaving a variable interval during which the valve continues to function normally until the second leg becomes sufficiently weakened to permit escape of the disc occluder. Valves with intentionally made SLSs implanted in sheep have continued to function for as long as 58 months, even though one valve failed after 3 days.1 The radiographically detectable SLS prevalence in humans suggests a median persistence in the order of 2 years.1 2 Although the interval between the onset of SLS and complete valve failure is likely to be variable, it appears that the period is often sufficient to permit identification of SLS valves, which are presumably more prone to failure than others with similar epidemiologically defined fracture risks.1 3 As a consequence, individual C/C valve fracture potential can now be defined by epidemiologically derived risk groups, taking into account valve size, implant position, manufacturing period (weld date), welder group, patient age at implant, and patient sex, and by the presence or absence of a radiographically detectable SLS.1 2
The cost and skill-intensive nature of radiographic SLS detection imply that considerable efficiency could be achieved by having an objective preradiographic screening assessment, and the difficulty and subjectivity involved in the radiographic interpretations argue for having corroborative data with which to support the radiographic diagnoses. With these needs in mind, the radiographic studies were planned to include acoustic recordings at classic ascultatory positions on the chest as part of each patient visit, even though acoustic assessment of the physical integrity of C/C valves was only in an early developmental stage when the radiographic studies began.
Acoustic assessment of C/C valves was focused from the onset on 29-mm-flange–size valves, which include nominal size valves of 29, 31, and 33 mm because these sizes have identical metal and carbon components, differing only in the widths of their cloth sewing rings. These larger valves generally are associated with higher strut fracture risks, particularly when implanted as mitral replacements in younger individuals.1 3 Finite element analysis predicted that the fundamental resonance of an intact outlet strut for 29-mm C/C valves would be in the range of 7 to 8 kHz.4 The presence of this resonance was verified with strain-gauge measurements in vitro and by analysis of acoustic data recorded from sheep implanted with 29-mm mitral C/C valves.5 6 The aim of our study was to enhance the detectability of this marker of outlet strut intactness to the extent that its absence in a C/C valve recording with an adequate signal-to-noise ratio would indicate the presence of an SLS.
The acoustic data were obtained using a computer-controlled (486 PC architecture) data-acquisition system of our own design.7 A contact microphone (model 106B50, PCB Piezotronic) was placed on the chest wall over the fifth intercostal space near the left sternal border for recording mitral valves and over the second interspace adjacent to the right sternal border for recording aortic valves. The recording technique was an integral part of the documents approved by the institutional review board or ethics committee of each site and was clearly described in the informed consent signed by each patient. Contemporaneous acoustic interpretation was excluded by the radiographic protocols.
The analog signal from the microphone was amplified with automatic gain control that permitted simultaneous recording of the gain values so that the original signal amplitudes could be recovered.7 The data were passed through a 22-kHz antialiasing low-pass filter before digitization with a 16-bit analog-to-digital converter (model FAST 16-1-1, Analogic Corp) at a sampling rate of 100 kHz. A discrete 15-ms record encompassing the C/C valve closing sound for each cardiac cycle was captured and stored; the recording session was continued until at least 300 closing events were acquired.
Simultaneous continuous recordings of the acoustic signal and the ECG were made on a digital audio tape recorder (model DTC-690, Sony Corp) to serve as a backup for the time-windowed data. When a patient had both aortic and mitral valves, the digital audio tape recordings served as the primary acquisition, using the ECG r wave signal amplitude as a time reference to create separate 15-ms windows for the closing events for each valve.
Early observations indicated that the outlet strut resonance could be detected best by selecting a 5-ms window beginning when ≈95% of the acoustic energy had dissipated, equating to subsidence of the transient sound associated with the disc (the “click”).5 During the development of assessment algorithms based on this processing technique, it was observed that the intact outlet strut resonance was often better represented in time windows earlier or later than the one specified by ≈95% energy dissipation.7 8 Consequently, a more generalized approach evolved using multiple overlapping time windows to sample the entire acoustic event.
The resulting TFR technique allowed us to detect the presence or absence of the intact outlet strut tonal component through a peak density function that emphasized the consistency of spectral peaks with respect to both time and frequency. Closing events passing signal-to-noise ratio and zero-crossing threshold acceptance criteria were time aligned to a reference point corresponding to the maximum smoothed data amplitude.9 Analysis windows of 5.5 ms were defined relative to this point, with successive windows having ≈93% overlap. Spectral estimates were computed for all windows from a given closing event and for all satisfactory events in a recording session with the use of short-time Fourier transformation and were ensemble averaged. This power spectrum was then combined with a measure of peak density within the target band to form a single TFR value that was representative of the entire recording session. Figs 1⇓ and 2 illustrate these two principal contributors to the TFR value for an intact valve. The power spectrum in Fig 1⇓ shows dominant resonant structure in the 7-to-8–kHz band. The peak density display in Fig 2⇓ shows consistent maxima in the 7.75-to-7.80–kHz cells in time windows of −1.2 to 0.4 over 300 beats. SLS valves typically had neither of these features.
Initially, the features and the algorithm used to develop the TFR values were heuristic because of the rarity of confirmed-status valve recordings early in the radiographic studies. As known-status valves became available, recordings from 13 SLS and 23 intact valves were used for training to refine individual feature definitions and the algorithm used for combining the features into a single TFR value. After the design was fixed, the training data were processed, yielding the TFR values shown in Fig 3⇓. Threshold values of <0.25 indicated an SLS valve, and values of ≥0.5 indicated intactness. In keeping with the objective of developing a prescreen for radiographic SLS detection, the threshold of <0.5 was selected to indicate a probable SLS.
TFR Acoustic Classification Testing
TFR values were provided to the Shiley Heart Valve Research Center for ≈800 radiographic study patients as of October 31, 1995. In turn, the center revealed the true status of 32 valves not common to the training set. The TFR values for the initial recording of each of these test-set valves were used to evaluate the accuracy of TFR values as a prescreening tool. In addition, recordings for the one known false-negative and four false-positive radiographic examinations were matched with the corresponding radiographic ratings to examine the possibility that contemporary knowledge of the acoustic classification might have had a mitigating influence on the actions taken.
Table 1⇓ shows the true physical status of each test-set valve and the corresponding TFR values in increasing order. Table 2⇓ shows the application of the <0.5 threshold defined from the training set, identifying 16 of 16 SLS test-set valves correctly and 11 of 16 intact test-set valves as being intact. A TFR value of <0.5 as a putative prescreen for radiographic SLS detection would have had a sensitivity of 1.00 (95% CI, 0.79 to 1.00) and a specificity of 0.69 (95% CI, 0.41 to 0.89).10 Fig 4⇓ shows the test-set results as a receiver operating characteristic curve, encompassing all potential decision thresholds permitted by the data. The area under the curve is 0.92, which equates to the probability that a randomly chosen SLS valve will have a lower TFR value than a randomly chosen valve with an intact outlet strut.11
Data bearing on the potential impact of contemporary knowledge of the TFR values for the five known radiographic misinterpretations are shown in Table 3⇓. The TFR values might have been seen as being cautionary in one of the four false-positives but appeared to endorse the others. The TFR value for the only verified false-negative radiographic interpretation correctly identified the valve as having an SLS. The radiographic ratings corresponding to the five acoustic false-positives in Table 1⇑ were “apparently normal” in only two instances; the radiographs for the other three valves were “suspicious” or “minimally suspicious” for the radiographic appearance of an SLS.
Clinical acoustic detection of findings portending sudden failure of a prosthetic heart valve began with the work of Hylen et al12 in 1968. They were able to identify potentially dangerous poppet ball variances by the absence of phonocardiographic peak frequencies of >1.3 kHz in patients with early, Model 1000 Starr-Edwards aortic valves and to confirm their findings by prophylactic explantations.13 In 1987, Stein et al14 were the first to look for an acoustic signature that could mark a C/C valve as having an SLS. They analyzed the opening sound of one intact valve and one valve with a “through-and-through” break in one strut leg with the use of a pulse duplicator and observed different dominant frequencies and a poor correlation between the frequency spectra of the two valves, causing them to suggest that analysis of C/C valve opening sounds might be clinically useful.
Several years later, Walker and Scotten15 studied C/C valve closing sounds under similar circumstances and reached the same conclusion with regard to closing sound analyses. They appear to have been the first to recognize that complete separation at the break site, which had previously been the model of an SLS, was not really representative of the clinical setting because the two SLS surfaces in clinical valves typically stay in apposition and, in fact, grate against each other.1 2 4 This distinction proved to be important because the change in fundamental vibration frequency from that characterizing an intact outlet strut to that of an open SLS is variably reduced when the SLS surfaces are in apposition. Although Stein and others16 17 revisited the opening sound by analyzing recordings in humans, most researchers were attracted to the closing sound because of its higher energy level and greater signal-to-noise ratio.5 6 7 8 9 15 18
Although an open SLS has a fundamental resonance of 2.2 kHz, Chia4 identified 4.3 kHz as the dominant resonance for a contacting SLS and made pulse duplicator strain-gauge measurements that showed variations from 3.0 to 6.5 kHz. We searched >200 clinical recordings for consistent energy peaks in this range without success. Energies were identified occasionally in the 3-to-6–kHz range, but they bore no relation to a valve being intact or SLS, and their source remains unknown. Our signal processing algorithms, like those of others, were designed to emphasize repetitive consistent events and to denigrate random energy. It may be that beat-to-beat varying contact of the SLS surfaces causes corresponding variations in apposed-surface SLS dominant frequencies that would then be undetectable by our processing.
As first reported by Plemons et al,5 what emerged was the consistent presence of resonance in the 7-to-8–kHz range marking an intact outlet strut. In parallel with the results of Hylen et al,12 13 the presence of an abnormality was now marked by the absence of a normal finding. Initially, we exploited the durability of the intact outlet strut resonance, which persisted, in most cases, after 95% of the disc energy had dissipated. In classifying training valves, however, we found that better accuracy was achieved using multiple overlapping time windows over the entire acoustic event. A multistep, TFR technique evolved that yielded a single TFR value, with 0.00 (undetectable intact strut resonance) indicating an SLS and 1.00 ensuring intactness.
Eberhardt et al18 also recognized the importance of consistent energy in the 7-to-8–kHz range. Using data from this and other selected influential subbands, they achieved perfect blinded classification of nine clinical recordings, consisting of seven intact and two SLS valves. TFR values were equally accurate when applied to eight times as many SLS valves in the blinded 32-valve test set but achieved a specificity of only 0.69 with intact valves (95% CI, 0.41 to 0.89). Reanalysis of the five false-positives showed that these valves had either weak outlet strut resonance or moderate strut resonance with obscuring nearby disc energy.
The receiver operating characteristic curve encompassing the spectrum of decision thresholds for TFR acoustic values as descriptors of outlet strut integrity has an under-the-curve area of 0.92 (Fig 4⇑). Receiver operating characteristic curves must pass through both 0-0 and 1-1 intercepts, and the area under the curve always has a value between 0 and 1. This area is a measure of the comparative discriminatory power of tests and models in a variety of fields, including clinical medicine.11 19 A perfect discriminator would have an under-the-curve area of 1.00, whereas a worthless test with no discrimination is defined by an area of 0.50. Clinically useful tests have under-the-curve areas ranging from 0.70 to 0.95, implying a high likelihood of clinical usefulness for TFR acoustic classification.19
Although greater specificity would be desirable, the accuracy of TFR acoustic screening is sufficient to prescreen higher-risk valve patients for radiographic study. One aortic and 25 mitral SLS valves were radiographically identified in a study of 964 C/C valves in 842 patients with a >0.1% annual fracture risk.1 Four false-positives and one verified false-negative were incurred in the process. TFR prescreening would have potentially effected a threefold SLS prevalence concentration in the C/C valves being assessed with radiography. Radiographic false-positives, which have proved likely to provoke an otherwise unneeded operation, would be expected to have been diminished by restricting the radiographic examinations to a higher-prevalence population.20
It had been assumed that different confounders should affect acoustic and radiographic assessments of the physical status of a C/C valve. A difficult-to-resolve radiograph interpretation therefore might coincide with a definitive TFR value, and vice versa. If the TFR values had been known contemporaneously, they would have refuted only one of the four false-positive radiographic interpretations. Similarly, the radiographic ratings for three of the five acoustic false-positives were at least “minimally suspicious” for the presence of an SLS. It may be that some valves or, more likely, some patients have characteristics that impede recognition of intactness by both methods. In keeping with the higher sensitivity of the TFR classification, it would have declared the known missed SLS to be an SLS valve. More data must be accumulated to permit a meaningful assessment of the corroborative interaction of the two tests.
Selected Abbreviations and Acronyms
|SLS||=||single leg separation|
This work was funded by an unrestricted grant from the Shiley Heart Valve Research Center, Irvine, Calif.
- Received June 5, 1996.
- Revision received September 26, 1996.
- Accepted October 7, 1996.
- Copyright © 1997 by American Heart Association
O'Neill WW, Chandler JG, Gordon RE, Bakalyar DM, Abolfathi AH, Castellani MD, Hirsch JL, Wieting DW, Bassett JS, Beatty KC, Soltis MA, Timmis GC, Grines CL. Radiographic detection of strut separations in Bjo¨rk-Shiley convexo-concave mitral valves. N Engl J Med. 1995;333:414-419.
Chia R. Finite element analysis of vibrations of the Bjo¨rk-Shiley convexo-concave heart valve. Proc Annu IEEE Symp Computer-Based Med Sys. 1994;7:48-52.
Plemons TD, Hovenga M, Reeder H, Dow JJ, Inderbitzen B, Chia R, Wieting D. Classification of the Bjo¨rk-Shiley convexo-concave valve status by detection of the intact outlet strut resonant frequency. Proc. Annu Int Conf IEEE Eng Med Biol Sci.. 1993;15:389-390.
Plemons TD, Schreck SG, Inderbitzen RS. Stability of the BSCC heart valve outlet strut resonant frequency under changing physiological conditions and observation times. Proc Annu IEEE Symp Computer-Based Med Sys. 1994;7:119-124.
Plemons TD, Dow JJ, Wieting DW, Chandler JG. Multiple feature acoustic classification of BSCC artificial heart valves. Proc Annu IEEE Symp Computer-Based Med Sys. 1994;7:125-130.
Scarbrough K, Inderbitzen RS. Application of time-frequency analysis to the classification of Bjo¨rk-Shiley convexo-concave heart valve condition. Proc Annu IEEE Symp Computer-Based Med Sys. 1994;7:131-138.
Kendall MG, Stuart A. The Advanced Theory of Statistics: Volume 2. 2nd Ed. New York, NY: Hafner Publishing Co; 1967:98-120.
Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Diagn Radiol. 1982;143:29-36.
Hylen JC, Kloster FE, Herr RH, Hull PQ, Ames AW, Starr A, Griswold HE. Phonocardiographic diagnosis of aortic ball variance. Circulation. 1968;38:90-102.
Hylen JC, Kloster FE, Herr RH, Starr A, Griswold HE. Sound spectographic diagnosis of aortic ball variance. Circulation. 1969;39:849-858.
Stein PD, Sabbah HN, Albert DE, Snyder JE. Spectral signature of the opening sound of the Bjo¨rk-Shiley convexoconcave valve as a potential indicator of strut fracture. Am J Noninvas Cardiol. 1987;1:369-372.
Durand LG, Grenier MC, Inderbitzen R, Wieting DW, Stein PD. Low frequency analysis of opening sound for detection of single leg separation of Bjo¨rk-Shiley convexo-concave heart valves. J Heart Valve Dis. 1995;4(suppl I):S32-S37.
Swets JA. Measuring the accuracy of diagnostic systems. Science. 1988;240:1285-1293.
Sox HC, Blatt MA, Higgins MC, Marton KI. Medical Decision Making. Stoneham, Mass: Butterworth-Heineman; 1988:67-101.