Artificial Intelligence-Based Phonocardiogram: Classification Using Cepstral Features

ABSTRACT

When cardiovascular issues arise in a cardiac patient, it is essential to diagnose them as soon as possible for monitoring and treatment would be less difficult than in the old. Paediatric cardiologists have a difficult time keeping track of their patients’ cardiovascular condition. To accomplish this, a phonocardiogram (PCG) device was created in combination with a MATLAB software based on artificial intelligence (AI) for automatic diagnosis of heart state classification as normal or pathological. Due to the safety concerns associated with COVID-19, testing on school-aged children is currently being explored. Using PCG analyses and machine learning methods, the goal of this work is to detect a cardiac condition, whilst operating on a limited amount of computing resources. This makes it possible for anybody, including non-medical professionals, to diagnose cardiac issues. To put it simply, the current system consists of a distinct portable electronic stethoscope, headphones linked to the stethoscope, a sound-processing computer, and specifically developed software for capturing and analysing heart sounds. However, this is more difficult and time-consuming, and the accuracy is lowered as a result. According to statistical studies, even expert cardiologists only achieve an accuracy of approximately 80%. Nevertheless, primary care doctors and medical students usually attain a level of accuracy of between 20 and 40%. Due to the nonstationary nature of heart sounds and PCG's superior ability to model and analyse even in the face of noise, PCG sounds provide valuable information regarding heart diseases. Spectral characteristics PCG is used to characterise heart sounds in order to diagnose cardiac conditions. We categorise normal and abnormal sounds using cepstral coefficients, or PCG waves, for fast and effective identification, prompted by cepstral features’ effectiveness in speech signal classification. On the basis of their statistical properties, we suggest a new feature set for cepstral coefficients. The PhysioNet PCG training dataset is used in the experiments. This section compares KNN with SVM classifiers, indicating that KNN is more accurate. Furthermore, the results indicate that statistical features derived from PCG Mel-frequency cepstral coefficients outperform both frequently used wavelet-based features and conventional cepstral coefficients, including MFCCs. © 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.