1.

INTRODUCTION

Coronary artery disease (CAD) is the leading cause of death worldwide. In Europe alone in 2021 its prevalence was 100 million and it was responsible for 3.6 million deaths [1]. The resulting cost of health care, and lost production was around €282 billion per year [2]. Existing diagnostic methods apart from ECG require skilled operators and expensive equipment.

Recent advances in coronary computed tomography angiography (cCTA) make it possible to detect and localise occlusive lesions with high accuracy and precision, whilst minimising the use of contrast media, albeit at the cost of greater exposure to ionising radiation. Of equal or greater concern is the fact that approximately 2/3 of symptomatic patients are found, on investigation, to be occlusion-free [3], representing a great waste of resources and thus a loss of productivity for healthcare systems and the patients themselves, to say nothing of their distress on facing a false positive diagnosis and the risks associated with contrast injection and exposure to unnecessary ionising radiation. Therefore, there is a need for a low cost screening device, initially to be used in a hospital environment such as a cardiac clinic, but ideally in the future in a primary care setting.

Previous work on acoustic localisation of coronary arterial stenosis, summarised in [4, 5] has shown that, in principle, the detection at the skin of the shear and transverse waves generated at the vessel wall by disturbed flow may be used in the diagnosis of arterial disease. With this in mind, we have developed a novel device, in the form of a stick-on patch, containing miniature accelerometers and microphones, designed to detect the sounds produced by disturbed (or possibly turbulent) blood flow downstream of occlusive lesions in stenosed coronary arteries. In this paper we present the results of measurements on a chest phantom, using discrete sensors and the aforementioned patch.

2.

MATERIAL AND METHODS

Following preliminary experiments to characterise the viscoelastic properties of a tissue mimicking gel [6] and to model the wave propagation within them [7-9], we constructed a simple model of the chest. It comprises a cuboidal acrylic container filled with a soft-tissue-mimicking viscous gel (Aquasonic, Parker Labs, Fairfield, NJ, USA. Length 400 mm, width 150 mm, height 100 mm), with an embedded latex tube (i.d. 6.5 mm, wall thickness 0.25 mm) representing a coronary artery (depth of centreline below gel surface, 13.3 mm). The upper surface of the gel is covered with a thermoplastic polyurethane sheet 0.1 mm thick, (Platilon, Epurex Films, Bomlitz, Germany), to mimic the skin. The system was water-filled, pressurized from a header tank (figure 1, left panel). Pressure in the tube upstream and downstream of the stenosis was measured with 2 catheter-tip manometers (6f gauge, Gaeltec, Dunvegan, Scotland) and flow, with an ultrasonic transit time device (Transonic TS403, Ithaca NY, USA) fitted with a cannulating probe. Surface movement was monitored at various positions ranging from 30 mm upstream of the stenosis to 60 mm downstream, by up to 6 miniature 3-axis MEMS accelerometers (ADXL 337, Analog Devices, Norwood, MA, USA.) and miniature microphones (SiSonic, Knowles Precision Devices, Norwich, UK). Signals were recorded and digitised at a rate of 1 kHz using a 16 channel acquisition system (Powerlab 16/35; AD instruments, Oxford, UK), processed and displayed in real time by associated software (Labchart v8). The latex tube was fitted with non-axisymmetric stenoses having area reductions of 60%, 75% and 90% and a 90% axisymmetric stenosis, with the unstenosed tube as a control. Recordings were made with steady flow rates of zero, 150, 250, 350 and 450 ml/min (encompassing the physiological and pathological range for the major coronary arteries).

Figure 1.

left panel, perfusion system. Right panel, sensor patch mounted on the skin of the model chest and connected through ribbon cables to the Zigbee wireless transmission module. One line of sensors were positioned over the tube (seen as a yellow outline below the surface) and the remaining 4 are offset to the side by 10 mm. The patch positioned was adjusted to give 12 measurement sites, 6 over the tube and 6 offset to the side. (More details in the text.)

In subsequent experiments, the microphones and discrete accelerometers were replaced by a custom-built sensor patch, consisting of 8 pairs of accelerometers (ADXL335 Analog Devices, Norwood, MA, USA) and microphones (SiSonic, Knowles Precision Devices, Norwich, UK) fitted to a flexible printed circuit board organised as shown in fig. 1, right panel, taped onto the skin surface of the model chest. The patch is also fitted with a built-in reference microphone to compensate for ambient noise.

The patch is connected to a transmitter module (92 x 82 x 34mm, 3V Li Ion battery, PCIA interface and amplifier, Zigbee wireless transceiver and antenna), which amplifies the analog patch signals, having subtracted the microphone reference signal from those of the 8 other microphones. The resulting signals are transmitted wirelessly to a receiver unit (92 x 82 x 34mm, with the same wireless hardware and protocol). The receiver module is connected by wire to the data acquisition system and recordings were made with the patch in one of three positions arranged to give 12 measurement sites, six directly above the tube, aligned with its long axis at distances of 30 and 15 mm upstream from the centre of the stenosis, one directly over the stenosis and three downstream at distances of 15, 30 and 60 mm from the centre of the stenosis. The remaining six positions, corresponding to the first six, were offset to one side of the tube by 10 mm. The recorded time domain data were transformed to the frequency domain, using the built-in fast Fourier transform of LabChart, with 1000 data points in each transform, Hann windowing, and 50% overlap. The spectra were normalized by subtracting data collected under no-flow conditions from the finite flow data. To quantify the results, the area under the curves (AUCs) were calculated for each of these normalised spectra over frequency intervals of width 20Hz, in the range of 0-500Hz. AUC values were expressed as means and standard deviations for each frequency band across 3 repeated measurements.

3.

RESULTS

Typical results for the discrete sensors are shown for an axisymmetric stenosis with 75% cross sectional area reduction, with the measurement point between 20 and 30 mm downstream from the centre of the stenosis, figure 2. For the accelerometers, there was a clear local maximum at around 250 Hz for flow rates above 100 ml/min, with a monotonic relationship between peak power and flow rate. The microphone spectra showed a similar fall at low frequencies to those of the accelerometer with a local maximum at approximately half the frequency, followed by a smaller peak at the same frequency as that seen in the accelerometer spectra. However for this second peak the monotonic relationship between peak height and flow rate was not seen and, given that it occurs at an integral multiple of the mains frequency, it is possibly an artefact.

Figure 2.

Frequency spectra under steady flow conditions with sensors positioned between 20 and 30mm downstream from the centre of a 50% axisymmetric stenosis. Left panel, accelerometer; right panel microphone. (Note that these were sequential rather than simultaneous recordings.)

These preliminary proof of concept experiments were followed by measurements with the sensor patch (fig. 1 right) replacing the discrete sensors, using the same perfusion system and stenoses. The results for the patch experiment were complex, with differences seen between the two sensor types and stenosis severity in their response to changing flow rate, position with respect to the stenosis and stenosis severity/geometry. Some representative examples are shown in figs. 3 and 4. For the microphone data (fig. 3) at frequencies below 100 Hz, the negative AUC values were associated with the upstream positions (i.e. AUC, lower than the zero flow values). At higher frequencies, the AUC values from some of the more distal positions, also became negative, whereas most of the energy was found directly over the stenosis and also at 15 mm and 30 mm downstream from the stenosis (i.e positions 4, 5, 8, 9 and 10). It is also worth noting that the AUC values at position 10 (30 mm downstream) are consistently higher than those at position 8 (15 mm).

Figure 3.

Relationship between AUC and frequency for microphone data, at a flow rate of 150 ml/min (typical in-vivo value for a major coronary artery) through a 90% non-axisymmetric stenosis. Bar colour indicates measurement position as follows; Even numbers starting from 2 at positions -30, -15, 0, 15, 30, 60 mm from stenosis centre, all directly above the tube long axis. Odd numbers at the same axial positions, parallel to the tube axis but offset 10 mm to the side.

Figure 4.

Relationship between AUC and frequency for accelerometers, at a flow rate of 150 ml/min (typical in-vivo value for a major coronary artery) through a 90% non-axisymmetric stenosis. Bar colour indicates measurement position as follows; Even numbers starting from 2 at positions -30, -15, 0, 15, 30, 60 mm from stenosis centre, all directly above the tube long axis. Odd numbers at the same axial positions, parallel to the tube axis but offset 10 mm to the side.

For the accelerometer data (fig. 4) there was less variation with frequency with the consistently high AUC values seen at the 15 and 30 mm upstream positions (1,2 and 3) as well as at positions 7 and 10, i.e. 15 mm downstream (off-axis) and 30 mm (on axis).

Consistent differences in AUC were seen between the sensors mounted over the tube and those offset to the side. These were confirmed by a paired t-test and the results, (table 1), show significant differences for both sensor types, with the on axis microphones producing relatively higher signals than their off-axis counterparts, whereas the reverse was seen for the accelerometers.

Table 1.

Paired t-test comparing on and off-axis signals for each sensor type.

To assess the individual contributions of frequency, flow rate, sensor position and stenosis type as well as to reduce the data and simplify the results, we set up a multivariable regression model with the independent variables, listed in table 2 and AUC for each frequency band as the dependent variable.

Table 2.

Summary of multivariable regression analysis. For both sensor types, the dependent variable was the area under the FFT spectrum split into bands of width 20 Hz. The independent variables are listed in the shaded cells. (B and β are the raw and standardised regression coefficients, respectively.)

The results show that both sensor types detected significant effects of each independent variable on AUC. From this it follows that, in principle, it is possible to determine not only the presence or absence of a stenosis, but also its position and severity. The values of the individual standardised regression coefficients (β in table 2) reveal that, for the microphones, overall AUC falls with frequency, while the values for the other factors increase, whereas for the accelerometers the inverse is seen. It is hard to explain this difference in physical terms and it may be due the fact that the positive increase in AUC with frequency in the zero flow microphone spectra, outweigh that seen at finite flow rates leading to a fall in the normalised AUC values with increasing frequency.

4.

DISCUSSION

From the preliminary discrete sensor experiments we had expected that their position would be strongly associated with the acoustic signal strength and surface displacements. This was broadly the case for most of the measurements and in the statistical model in which sensor position is a significant variable. The paired t-tests revealed a highly significant difference between the signals from the microphones placed directly on the tube and those in the off-axis position. Given that the sensors placed on top the tube axis were closer to the disturbed flow and the resulting signal reaching the ‘skin’. However, for the accelerometers the off-axis signals were stronger. Although we can find no obvious physical explanation for this difference. Nevertheless, the accelerometers are a beneficial addition to the device because they provide complementary data to that generated by the microphones and can thus help to locate the stenosis. Currently due to limited bandwidth in the wireless protocol, we have only recorded z-axis accelerometer signals (i.e movement normal to the skin surface). However, due to bending of the surface as it is displaced, it is likely that the z-axis signal of an accelerometer attached to this surface will record some lateral movement in addition to movement perpendicular to the plane, whereas a microphone will respond primarily to air pressure changes caused by net skin movement in the z-direction. Detailed measurements and modelling of the strain field under and around the patch, beyond the scope of this study, would be required to understand this behaviour. The work described here has several other major limitations. Firstly, the in-vitro measurements are currently confined to steady flow only. Secondly, the chest model is primitive, lacking ribs, lungs, a beating heart and the consequent sounds it produces. Although the amplitude of the heart sounds is many times greater than the bruit caused by the disturbed flow in the coronary arteries, most of this is generated as the ventricles contract during systole, whereas flow in the coronary arteries is maximal during diastole [10] and is therefore not subject to this background noise. Nevertheless, in spite of these limitations, the measurements have proved the principle that the sensor patch can detect weak signals and that the complementary information provided by the two types of sensor arranged in a grid can detect the presence of a stenosis, as well as its location, at least in a straight tube with a single stenosis.

Multivariable regression is cumbersome and time consuming, and the limited results presented here are hard to interpret. However, the data lends itself to analysis using machine learning (ML) techniques. Preliminary work has shown that a recurrent neural network trained on in silico data could predict the location of the source of time varying surface signals with good accuracy. Motivated by that work, a follow up study [11] has demonstrated that a selection of ML tools could effectively localize the effect from its cause. To mitigate the risks associated with the ethical dangers around ML and artificial intelligence (AI) we will use tools from the rapidly developing area of Explainable AI (XAI) and so use a clear and transparent methodology with as far as is possible, explainable outcomes.

We have recently completed the design and construction of an improved sensor patch (fig. 5) which is lighter and very much more flexible than the original shown in fig. 1.

Figure 5.

2^nd generation patch prototype and disposition on the chest for in-vivo measurement

Although there is still much in-vitro work to be done, essentially a repeat of the study outlined here with the new patch and the inclusion of pulsatile flow to provide a large quantity of ML training data, we are now preparing an in-vivo study in which we will assess the diagnostic sensitivity and specificity of the system by comparing the patch predictions in patients undergoing elective diagnostic angiography – the reference method. We note that although in a device such as this on which clinical decisions will have to depend, it essential that specificity is at least as good as the gold standard of coronary angiography; the sensitivity being somewhat less critical. Given that approximately 60% of patients presenting at a typical heart clinic with chronic chest pain who then undergoing elective angiography [3], this will provide the opportunity to test the patch and train the ML model on patients with coronary artery stenosis as well as those free of occlusive disease.

5.

CONCLUSION

The results to date suggest that our acoustic patch can detect displacements of the skin of a model chest associated with disturbed or turbulent flow. We have also shown that, using a multi-variable regression model it is possible to detect differences in flow rate and stenosis severities. Furthermore, differences in the spectra associated with the proximity of the sensors to the stenosis suggest that, with adequate training, a machine learning approach could infer the presence or absence of a stenosis as well as its location. Accelerometer data showed some similarities to that of the microphones, although in some cases, the AUC values for the accelerometers showed the opposite trend to that expected. For the future, additional in-vitro experiments with the new patch are in progress and a first-in human trial on patients undergoing elective coronary angiography is imminent. If the specificity and sensitivity are comparable with angiography, the patch will provide a powerful, low cost screening tool with the potential to reduce the cost and resource demands currently associated with the diagnosis of coronary artery disease.