Non-Contact Heart Sound Measurement Using Independent Component Analysis

A non-contact heart sound measurement method using independent component analysis (ICA) was successfully developed, and the measured heart sound was quantitatively evaluated with the signal-to-noise ratio (SNR). There have been recent developments in the automated diagnosis of cardiovascular diseases based on the measurement of heart sounds. However, measuring heart sounds require physical restraint. Here, we propose a non-invasive and non-constrained method for non-contact measurement of heart sounds using microphones. We successfully demonstrated this method by applying ICA to multi-channel measured sounds with a microphone array. Then, we quantitatively evaluated it by measuring “pseudo heart sounds” at a distance of up to 200 mm from the source using four microphones in an anechoic chamber and a conference room. SNR was improved by increasing the number of microphones. This method could measure pseudo heart sounds even in the presence of artificial noise created by heating, ventilation, and air conditioning systems and human voices. We suggested that measurable distance can be improved by using more microphones. Moreover, we successfully measured actual heart sounds at a distance of 160 mm from the chest wall. This non-contact heart sound measurement method provided valuable information about the heart, such as visual recognition of the first (S1) and the second (S2) heart sounds.

indicating heart health [2]. Abnormal heart sounds indicate 23 various cardiovascular diseases [2]. Although several stud-24 ies have shown the automated diagnosis of cardiovascu- 25 lar diseases by analyzing measured heart sounds [3], [4], 26 [5], these measurement methods usually involve physical 27 restraint such as auscultation by a physician. While a few 28 studies have shown wearable devices for heart sound mon-29 itoring [6], [7], these devices have disadvantages such as 30 The associate editor coordinating the review of this manuscript and approving it for publication was Humaira Nisar . discomfort due to constant wear and the need for intermit-31 tent charging. A microphone device such as a smartphone 32 or a smart speaker placed around the user could monitor 33 their heart sounds daily and provide an effective non-contact 34 method to measure heart sounds. It can enable early and 35 automated diagnosis of cardiovascular diseases. Although 36 previous studies have not explored non-contact heart sound 37 measurement techniques, a few have reported non-contact 38 heart rate measurement methods [8], [9], [10], [11]. The heart 39 rate is also a biological signal similar to the heart sound, 40 but it merely includes the number of heartbeats. Therefore, 41 it does not provide information about abnormalities in the 42 heart's anatomy. A few reports have shown non-contact mea-43 surement of seismocardiography (SCG) using a radar [12], 44 [13]. SCG is the chest wall vibration caused by the heartbeat 45 and can indicate abnormalities in the heart. However, these 46 are physically restrained methods because it requires pinpoint 47 VOLUME 10, 2022 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ targeting of the chest wall by the radar. This paper discusses a non-invasive and non-constrained method for non-contact 49 measurement of heart sounds using microphones. 50 There are two main challenges in developing non-contact 51 heart sound measurement methods. The first is the measure-52 ment issue. As the heart sound is mostly reflected at the 53 boundary between the chest wall and the air due to their 54 acoustic impedance difference, airborne heart sounds can be 55 buried in ambient noise. Once the noise is superimposed, 56 removing these and extracting the heart sound from the 57 observed sound is difficult. Here, we developed a non-contact heart sound measure-111 ment method using ICA and evaluated its performance quan-112 titatively. This is a preliminary report on non-contact heart 113 sound measurement using microphones.

117
The pseudo heart sound, used as the true value for eval-118 uation, was generated as follows. First, the waveform of 119 the airborne heart sound was verified. Even when micro-120 phones were placed in the air, heart sounds could not be 121 captured because they were buried in noise. Therefore, the 122 quasi-periodicity of the heart sound and the ECG were uti-123 lized. The minute potential changes in the ECG can be 124 detected by [25] synchronously averaging it based on its R 125 waves (referred to as R henceforth) to attenuate asynchronous 126 noises. R is the strongest component of the ECG caused by 127 myocardial excitement and is generally used for heart rate 128 detection. We also synchronously averaged heart sounds by 129 simultaneously measuring the R. Heart sounds are quasi-130 periodic signals where the period between one R and the 131 next forms a cycle. Therefore, this process only amplifies 132 the signal synchronizing with the quasi-period while other 133 signals, including environmental noises, were attenuated. 134 The heart sound and ECG were measured in an anechoic 135 chamber. The heart sound is measured using a microphone 136 (MI-1271 and MI-3170, Ono Sokki Co., Ltd., Yokohama, 137 Japan) kept 30 mm away from the chest wall without clothing. 138 A band pass filter (BPF, 10 Hz to 600 Hz, Butterworth, 10th 139 order) then extracts the frequency range of the heart sound. 140 ECG is measured using a 3-point induction method with an 141 ECG monitor (AD8232 SparkFun Single Lead Heart Rate 142 Monitor, SparkFun Electronics, Niwot, Colorado, USA). The 143 measured ECG is preprocessed, including the power supply 144 noise reduction and smoothing, to amplify R. The extracted 145 airborne heart sound, referred to as ''pseudo heart sound,'' 146 retains all the components of the actual airborne heart sound 147 following the ECG. 148 Next, we reproduced the environment to measure the heart 149 sounds by outputting the pseudo heart sound at the same 150 pressure as the actual heart sound using a subwoofer, which 151 was a plane source with a diaphragm (200 mm in diameter).

152
The attenuation characteristics of a plane source and a point 153 source were the same at measurement distances greater than 154 64 mm (≈ 200/π mm). As the heart (specifically, the four 155 valves) is considered closer to a point source than a sub-156 woofer, it is safe to assume that a subwoofer can reproduce 157 the sound pressure of the actual heart sound at a distance of 158 64 mm or more. The amplitude of the pseudo heart sound 159 was adjusted to equal to the sound pressure of actual heart 160 sounds for a measurement distance of 64 mm or more.   This SNR is called quasi-SNR because the modified pseudo 189 heart sound is not strictly a true value. −4.84 dB (for pseudo heart sound extraction) and 1.65 dB 198 (for actual heart sound extraction) is considered a failed 199 heart sound extraction. Conversely, an SNR of −4.84 dB and 200 1.65 dB or higher indicates successful heart sound extraction. 201 This paper mainly focuses on SNR for analysis as it 202 is a reasonable performance measure for signal extraction 203 and includes all other measures. For example, while other 204 measures such as heart rate, heartbeat interval, and domi-205 nant frequency indicate the heart sound extraction efficacy, 206 high accuracy of heart rate detection is not indicative of 207 highly accurate heart sound extraction. This is because a 208 different sound might have a similar cycle rate as the heart 209 sound. The same applies to the others. Conversely, SNR 210 completely corresponds to the accuracy of the waveform and 211 is expected to include all other measures to ensure measure-212 ment accuracy. As other quantitative analyses are included in 213 SNR, it enables quantitative evaluation of heart sound mea-214 surement. Although qualitative evaluation is also important, 215 it requires medical knowledge, which is out of the scope of 216 this paper. Therefore, we focused on quantitative evaluation 217 with SNR.

218
C. SIGNAL PROCESSING FOR HEART SOUND EXTRACTION 219 Fig. 2 shows the block diagram of the proposed heart sound 220 extraction method, including the signal processing workflow. 221 This method uses a microphone array to obtain m-channel 222 sound X [t]. We applied BPF (10 Hz to 600 Hz, Butterworth, 223 10th order) to X [t] to extract the frequency range of the 224 heart sound and obtained Y [t] that could be split into short-225 time segments such as Y i [t] (i = 1, 2, . . .) to reduce the 226 effect of short duration noise. As an extremely short segment 227 reduces its independence, we used the three-second segments 228 empirically. ICA (FastICA [26], sklearn, python) was then 229 applied to Y i [t] to separate them into sources, resulting in the 230 m-channel output, Y i '[t]. SNR was calculated for all channels 231 of Y i '[t], and the one with the highest SNR was selected for 232 the extracted heart sound, yi' [t]. 233 After the extraction process, we evaluated its accuracy. The 234 time period of the sound measured from one experiment is 235 60 seconds, which is split into 20 three-second segments. 236 SNR is calculated for all the segments and expressed as mean 237 VOLUME 10, 2022 normalization for SNR calculation. Specifically, the signal is 239 subtracted by its mean and divided by its standard deviation.     4 shows the experimental attenuation characteristics 291 of the pseudo and actual heart sound. These amplitudes were 292 adjusted so that the sound pressure of the pseudo heart sound 293 output from the subwoofer was equal to that of the actual 294 heart sound. The sound pressures for both heart sounds were 295 plotted against the measurement distances from the source 296 (10 mm to 200 mm). This figure shows that, as expected, the 297 attenuation characteristics were different when the distance 298 was below 70 mm but similar when it was above 70 mm. 299 98628 VOLUME 10, 2022  Based on these results, the measurement distance was set to 300 80 mm or more for the following experiments. in an anechoic chamber. The pseudo heart sound is buried in 306 noise, and heart rate variability is not visible. Fig. 5 (b) shows 307 the waveform of the extracted pseudo heart sound. Although 308 some noise still remains, both S1 and S2 of the pseudo heart 309 sound could be visualized, indicating that the pseudo heart 310 sound is successfully measured. Fig. 6 (a) and (b) show spec-311 trograms of the sound shown in Fig. 5 (a) and (b), respec-312 tively. While Fig. 6 (b) shows visible S1s and S2s, Fig. 6 (a) 313 lacks some information about the pseudo heart sound. This 314 might be due to the presence of plenty of noise at frequencies 315 ranging from 10 Hz to 100 Hz, which is the same as S1s and 316 S2s shown in Fig. 6 (a). Therefore, conventional BPF could 317 not extract the pseudo heart sound. 318 FIGURE 7. Waveforms of the extracted pseudo heart sound at 80 mm, 160 mm, and 320 mm from the source, respectively. FIGURE 8. SNRs of the pseudo heart sound measured in the anechoic chamber and extracted using conventional and the proposed method against the measurement distance. Fig. 7 shows the extracted pseudo heart sound measured at 319 a distance of 80, 160, and 320 mm, respectively. Although 320 the sound at 80 mm has visible S1s and S2s, the one at 321 160 mm has only visible S1s. The one at 320 mm does not 322 have any visible information about the pseudo heart sound. 323 As the distance increases, the pseudo heart sound becomes 324 quieter and gets buried in the ambient noise.    anechoic chamber. This indicates that the effect of small 371 directional noises, which are reduced in an anechoic chamber, 372 does not interfere with non-contact heart sound measure-373 ments. Therefore, measuring the heart sound contactlessly in 374 an ordinary conference room is feasible.

375
The SNR of ''conference room-HVAC'' is under 5 dB 376 even at 80 mm. The four microphones could theoretically 377 remove noises from the three independent sources. Therefore, 378 we expected that ''conference room-HVAC'' has the same 379 SNR as ''m = 3,'' as shown in Fig. 9, if HVAC 1 and 2 were 380 the respective sources. This result, however, shows that the 381 HVACs and voices are not considered a single source as 382 the reverberations from these noises are also regarded as 383 independent sources due to their loudness. In other words, 384 more microphones can improve SNR with loud noises.

386
The actual heart sound was also extracted using the proposed 387 method. Fig. 11 shows the waveforms of the actual heart 388 FIGURE 12. Spectrograms of the extracted actual heart sound measured at 80 mm, 160 mm, and 320 mm from the chest wall, respectively. at 80 mm has the characteristic S1s and S2s of actual heart 392 sounds. For the waveform at 160 mm, only S1 could be visu-393 ally recognized, while that at 320 mm had no information on 394 the actual heart sound. This is consistent with the results for 395 the pseudo heart sound, as shown in Fig. 7, indicating that the 396 reproduced airborne heart sound measurement environment 397 was appropriate, and the proposed method is effective for not 398 only the pseudo heart sound but also the actual heart sound. 399 Fig. 12 shows spectrograms of the extracted actual heart 400 sound, as shown in Fig. 11. The spectrogram at 80 mm had 401 visually recognizable S1s and S2s. However, the S2 is unclear 402 compared to the waveform in Fig. 11. S2 generally varies 403 in intensity and timing with respiration, which might have 404 caused the S2 attenuation at 80 mm seen in Fig. 11. However, 405 we need to increase the number of subjects to verify this 406 possibility. The spectrogram at 160 mm had visually recog-407 nizable S1 with some confusing noises. The spectrogram at 408 320 mm has no information on the heart sound. These results 409 are consistent with the ones shown in Fig. 11. 410 Fig. 13 shows the quasi-SNRs of the actual heart sound. 411 These indicate successful extractions at a distance of 80 mm 412 and 160 mm. For example, the quasi-SNR is 8.46 ± 0.735 dB 413 at 80 mm. These results are consistent with the visual eval-414 uation in Fig. 11 and 12. In addition, the difference between 415 this quasi-SNR at 80 mm and the SNR in Fig. 8 is 5.84 dB. 416 This corresponds to the effect of time variation, including the 417 respiratory variation of the actual heart sound that could not 418 be eliminated by modifying the pseudo heart sound. 420 We successfully developed a non-contact heart sound mea-421 surement method using ICA and quantitatively evaluated the 422 measured heart sound with SNR. We applied ICA to multi-423 channel observed sound with a microphone array to extract 424 minute heart sounds from copious ambient noise. We quanti-425 tatively evaluated this extraction method by reproducing the 426 airborne heart sound measurement environment using pseudo 427 heart sounds.

428
The proposed method could measure the pseudo heart 429 sound at a distance of up to 200 mm from the source by 430 using four microphones, while conventional BPF methods 431 were unable to measure it even at 80 mm. The SNR is 432 15.5 ± 0.792 dB at 80 mm in the conference room, which 433 was increased by 2.34 dB after adding a microphone. This 434 method could measure the pseudo heart sound even in the 435 presence of artificial noise, such as HVAC sounds and voices. 436 We also suggested that more microphones improve the mea-437 surable distance. In addition, this method successfully mea-438 sured actual heart sounds 160 mm from the chest wall in an 439 anechoic chamber. The quasi-SNR is 8.46 ± 0.735 dB at 440 80 mm from the chest wall. This non-contact method pro-441 vided important information about the heart, such as visually 442 recognizable S1 and S2.

443
To our knowledge, this paper is the first to report a 444 non-invasive and non-constrained method for non-contact 445 heart sound measurement using microphones. This method 446 is advantageous for daily continuous health monitoring by 447 eliminating the need for physical constraints, wearing dis-448 comfort, batteries, and dedicated devices compared with 449 the existing methods. Further development of this method, 450 including increasing the number of microphones, will enable 451 daily heart sound monitoring and early automated analysis of 452 cardiovascular diseases.