Artificial Intelligence to Analyze Auscultation Signals to Detect PDA Heart Defect

This paper presents the application of Convolutional Neural Networks (CNN) for classifying auscultation signals, with a focus on diagnosing Patent Ductus Arteriosus (PDA) in patients. A dataset of 161 labeled spectrograms is utilized in a supervised machine learning (ML) framework to perform binary classification of spectrogram signals as either (i) normal or (ii) indicative of PDA. The auscultation signals undergo pre-processing through: (i) a Butterworth filter to enhance signal-to-noise ratio, (ii) ShortTime Fourier Transform (STFT) for temporal and frequency analysis, (iii) a gammatone filter to simulate cochlear auditory behavior, and (iv) spectrogram labeling. Five CNN architectures are evaluated by varying: (i) the number of hidden layers, (ii) convolutional layers, (iii) dense layers, and (iv) epochs. The addition of dense layers improves the model's ability to capture global patterns, resulting in a validation accuracy of approximately 9 6%.