Robust Automatic Heart Disease Detection From Audio Signals Using Entropy-Spectrograms and a 2D Convolutional Neural Network

The most common diseases worldwide are those that are heart failure related, every year 17.9 million people die from this problem, many of the deaths might have been prevented if they had been diagnosed at early stages. In order to detect some heart diseases, medical doctors auscultate the patient by hearing the heart beats with a stethoscope. The proper diagnosis depends on the skills of the professional, but frequently the doctor has to listen in areas of high ambient noise. We propose an automatic diagnosing method which consists in producing images with the spectrograms of entropy (i.e. entropygrams) from the audio signal of the heart beats of the patients, and using a 2D convolutional neural network for predicting if the entropygram comes from a healthy individual or from a patient with a heart disease. To reproduce a noisy environment we mixed the original audios with white noise achieving a Signal to Noise Ratio up to -10 dB. Entropygrams showed indeed robustness to noise and classification achieved an accuracy of 99.43 % with the noisiest audios.