March 26, 2018 11:3 WSPC/WS-JMMB heartsound ...

3 downloads 0 Views 1MB Size Report
10, pp. 233–238, 2009. 26. T. Toni, D. Welch, N. Strelkowa, A. Ipsen, and M. P. Stumpf, “Approximate bayesian computation scheme for parameter inference and ...
March 26, 2018

11:3

WSPC/WS-JMMB

heartsound

Heart diseases detection from noisy recordings of smartphone devices

Wen-Hsien Ho Healthcare Administration and Medical Informatics, Kaohsiung Medical University, Medical Research, Kaohsiung Medical University Hospital, Kaohsiung, Taiwan, ROC Yenming J. Chen∗ Logistics Management, National Kaohsiung University of Sci&Tech, Taiwan, ROC Yuzhen Zhang Cardiology, The First Affiliated Hospital of Soochow University, Suzhou, China Yanyun Tao School of Railway Transportation, Soochow University, Suzhou, China Hsin-Wen Kuo Logistics Management, National Kaohsiung University of Sci&Tech, Taiwan, ROC

This paper aims to develop an algorithm to detect heart diseases through ordinary smartphones without additional equipment for cost accessibility. Among various vital signs emitted by organs, sounds can be easily observed and carry ample information. However, these sounds are small and noisy. Detecting anomalies involves great challenges in signal processing. This study presents a novel method that overcomes noises to estimate cardiovascular health. We use time-scale techniques in time series analysis to extract disease traits and suppress excessive ambient noises. Using datasets from PhysioNet, our model achieved a nearly 100% accuracy in heart disease diagnosis. Our approach also performs well under excessive noises for diseases producing heart murmurs. With heavy noise contaminated signals, training accuracy still closed to 100%, and the testing accuracy still remained around 84%. Keywords: smartphone, noise, phonocardiography (PCG), time-frequency, wavelets, ensemble learning

1. Introduction A smartphone is an ideal device to ubiquitously monitor personal health conditions through easily collected audio and video signals 1 . Frequent and noninvasive car∗ corresponding

author: [email protected]. MOST 104-2410-H-327-018-MY3, MOST 106-

2221-E-037-001 1

March 26, 2018

2

11:3

WSPC/WS-JMMB

heartsound

Wen-Hsien Ho, Yenming J. Chen*, Yuzhen Zhang, Yanyun Tao, and Hsin-Wen Kuo

diovascular monitoring is important in the surveillance of cardiovascular diseases, such as coronary heart disease and hypertension. However, complicated acoustic transmission and noises prevent the accurate detection of valve defects. Moreover, recording valve sounds or phonocardiography (PCG) through consumer devices and by untrained personnel presents additional challenges, such as low devise sensitivity and mismatched organ positioning 2,3,4 . A viable framework to detect chronic heart failure remotely in noninvasive hemodynamic monitoring technologies must overcome those challenges 5 . Without effective noise reduction and internal states estimation techniques, using sounds for diagnosis is difficult 6,7,8 . Therefore, this paper aims to develop an effective algorithm to detect anomaly cardiophysiology situation through ordinary smartphones. The novelty of this study reflects on integrating wavelet denoising algorithms and adaBoost classification algorithm to tackle the above-mentioned challenges. In order to achieve the goal of low costing, we exploited sound signals that only a smartphone can acquire for the purpose of cost accessibility. To assess our noise rejection capability, we tested our algorithm on noise-added samples and the results showed that our method can compensate the simplicity of smartphones and the ambient noises. 2. Related Literature The findings of this paper are related to three streams of literature. To diagnose diseases through heart sound successfully, a set of effective sound features must be exploited. In most diagnoses, potential internal parameters exist between external features and disease states. Therefore, effective classification and estimation algorithms are required to reversely reveal disease states through limited observations. Effective noise suppression must be considered to resist excessive noise in the recorded heart sounds. This study fills in the gaps of existing studies. To extract useful features of medical signals for effective processing, many estimation methods have been developed, from traditional ones, such as short-time Fourier transform 9 , Wigner-Ville distribution 10 , wavelet transform 11,12,13,14 , Hilbert transform 15,16,17 , empirical mode decomposition 18 , artificial neural networks 19,20 , ensemble Kalman filtering 21,22,23 , hidden Markov models 24 , and stochastic subspace-based identification 25 , to contemporary methods for nonstationary signal analysis such as, approximate Bayesian computation 26 , particle Markov chain Monte Carlo (MCMC) 27,28 , synthetic likelihood 29 , wavelet timeseries 30,31 , and stochastic resonance 32,33 . Ref. 34 summarized several common features for further heart disease analysis. Ref. 31 employed synchrosqueezed wavelet transform to extract time-varying features, such as instantaneous frequency (IF) for non-stationary quasi-periodic signals. However, most abnormal heart sounds occur during valve closure. Some studies segmented the duration of valve closure for further analysis. Ref. 24 extended a hidden

March 26, 2018

11:3

WSPC/WS-JMMB

heartsound

Heart diseases detection from noisy recordings of smartphone devices

3

semi-Markov model with logistic regression to segment noisy PCG signals accurately into S1 and S2 sounds. Ref. 35 segmented S1 , S2 from murmurs through the normalized average Shannon energy of the PCG signal. Existing studies either require clear recording of signals or do not attain a satisfactory accuracy. Our method needs only a small number of features and resists to ambient noises. Therefore, our method is suitable to be implemented in a smartphone device. When a sufficient set of features is acquired, classifying good samples from bad ones is possible. Ref. 36 proposed a classification approach using ensemble empirical mode decomposition and correlation dimension. Cardiovascular valve closure produces a significant amount of chirp sounds. Ref. 37 developed a classification method for chirp signals. To estimate the joint probability densities, the hierarchical Bayesian learning and MCMC methods were used in the algorithm. The MCMC methods represent a broad range of integration techniques and can approximate a multidimensional integral by generating samples through an ergodic Markov chain, which ensures that the generated samples converge to a desired distribution according to ergodic theorem. Various methods are available to perform heart sound classification. The studies by 38,39,40,41,19,42,43,44 applied wavelet analysis, ergodic hidden Markov model, neural network, combined wavelet and self-organizing map, combined neural network and wavelets, combined S-transform and neural network, adaptive feature selection, and wavelet packet decomposition tree, respectively, for heart sound classification. Two types of noises exist in medical signals. One is artifact such as crosstalk sounds from the lungs, and the other is random noise such as electric statics and friction between the recording device and skin. To suppress the noises, the processing algorithm may cause a certain degree of distortion and decrease quality for further analysis. Various methods have been used to separate target signals from artifact sounds, such as the separation of heart and respiration sounds by weight scales prediction in time-frequency analysis 45 and singular spectrum analysis 46 . To suppress random noise, time-frequency analysis is able to remove insignificant components in the multiscale domain. Ref. 47 improved the accuracy of wavelet denoising by constructing a hidden Markov model. Ref. 48 reduced the noise of phonocardiograms through the use of optimal wavelet. Ref. 49 improved the thresholding algorithm from the neighboring coefficients of multi-wavelet. Ref. 50 employed a line enhancer with variable step size to cancel out artifacts, such as EMG and lung sounds. Ref. 51 proposed a wavelet denoising technique with a simple thresholding rule, which increases SNR smoothly. Ref. 52 enhanced the denoising performance through a translation-invariant wavelet for tackling Gibbs phenomena around discontinuities. Additionally, stochastic resonances have proven to be effective in reconstructing signals through extremely large background noise. Ref. 33 successfully detected weak signals by using multiscale stochastic resonance.

March 26, 2018

4

11:3

WSPC/WS-JMMB

heartsound

Wen-Hsien Ho, Yenming J. Chen*, Yuzhen Zhang, Yanyun Tao, and Hsin-Wen Kuo

3. Methods To collect PCG signals, requiring external devices other than the phone itself is inconvenient 53 . OS-related real-time capture problems, such as time-lag and buffering, have also been overcome by current technology 54 . Although 55 have developed an electronic stethoscope by coalescing an external microphones with a regular stethoscope, the use of external devices is still considered inconvenient. Fortunately, 56 confirmed that directly using smartphones’ built-in microphones is feasible in collecting valve sounds. For the electronic performance, the built-in microphones of most smartphones are sensitive enough; even the headset microphones can acquire heart sounds with a high signal-to-noise ratio 1 . Therefore, the experiment used a smartphone purchased three Fig. 1. Experiment setup of an ordiyears ago, without sophisticate recording de- nary smartphone with a consumer-level vices. We recorded the heart sounds using both headset. The part contoured with a red circle is headset microphone. a headset microphone under $50 and the builtin microphone (Fig. 1). PCG signals were recorded at several specific spots on the surface of thorax. Without acoustic impedance matching on the recording device, users have to move along the surface of thorax to obtain goodquality heart sounds. Furthermore, the recordings may still contain a high amount of ambient noise, and they are subject to a good noise rejection algorithm. For example, 49,51,52 employ standard thresholding algorithms to eliminate noises under certain threshold values in the time-scale representations. However, the recordings in our simple experiment appear to be of good quality. In Fig. 2, a segment of the recorded PCG signal showed the quality and capability of built-in microphones. The signal quality of an ordinary smartphone is good enough relative to that collected by PhysioNet. This study mainly uses acoustic emission with phonocardiographic symptoms due to its cost-effectiveness in detecting heart anomalies. Although abnormal signs of cardiovascular diseases can also be observed in other media, such as blood pressure, electrocardiogram, we will integrate other media in the future studies. The heart normally produces two prominent sounds, S1 and S2 in each cardiac cycle. S1 occurs at the onset of ventricular systole and the closure of atrioventricular valves, with a duration of around 100200 ms, while S2 marks the end of ventricular systole and the closure of aortic and pulmonic valves 14 . S1 and S2 comprise (M1 , T1 ) and (A2 , P2 ) sub-components, respectively, corresponding to the mitral, tricuspid,

March 26, 2018

11:3

WSPC/WS-JMMB

heartsound

Heart diseases detection from noisy recordings of smartphone devices

Fig. 2. peaks.

5

A sound segment recorded by a low cost smartphone microphone and the detected local

aortic, and pulmonary valve activities 35 . Other sounds, such as S3 and S4 are normally unheard in healthy states. For an ordinary hearing cognition of human beings, heart sounds are heard as “lub dub, lub dub”. In general, “lub” is close to “dub” and far from the next “lub.” A valvular closure problem may create fluid turbulent sounds and distort the pitch of the “lub dub” sound. Therefore, one of our goals in feature selection is to find an invariant indicator that cannot only locate the pathological spot in recorded sounds but also single out the abnormal high-frequency disturbances. The continuous wavelet transform R (CWT) is a generalization of the short-time x(τ )dτ. The function Ψ is called Fourier transform, and C(a, t) = √1a Ψ τ −t a “mother wavelet” and performs dilation in continuous scale a and shift in scale t. However, the continuous representation is redundant in the amount of information for the reconstruction of the original signal. We can reduce the computation by restricting the scale a to a parsimonious dyadic scale, and it becomes discrete wavelet transform (DWT). 4 obtaining the wavelet transforms, we can use the feature of IF ω(t) = By 4 ∂ ln x(t) and instantaneous amplitude (IA) |xa (t)| = x(t) + jH{x(t)}. Evalu= ∂t ating these two quantities directly from their definitions are difficult and therefore wavelet ridge analysis provides a good estimation for the two quantities. The ridge points are taken over the scalogram of a signal x(t). An amplitude ridge point of (a, t) is a time/scale pair (a, t) with the strictly local maximum of