CardioNet: A Robust CRNN Network in Cardiac Abnormality Diagnosis from Phonocardiogram

The global rise in cardiovascular disease highlights the urgent need for early and accurate detection strategies to mitigate its impact. Phonocardiogram (PCG) signals, being non-invasive, cost-effective, and convenient, hold significant potential for diagnosing cardiac conditions. However, traditional methods often face challenges such as high computational complexity, sensitivity to noise, and reliance on extensive datasets, which can limit their practical application. To address these limitations, this study proposes a lightweight deep learning model named as "CardioNet" which combines convolutional neural network (CNN) and bidirectional long short-term memory (Bi-LSTM) layers to effectively capture both spectral and temporal features from PCG signals. The model processes raw PCG signals by extracting Mel-frequency cepstral coefficients (MFCCs) as input features. CardioNet achieves high accuracy of 99.50% in quiet conditions with only 0.18 million learnable parameters, 390282 FLOPS and a small memory requirement of 2.76 MB which demonstrates its’ computational efficiency. Furthermore, the model exhibits robust performance, with accuracy ranging from 96.8% to 98.4% across unseen datasets of varying noise levels (at SNR of 0 dB to 10 dB). This adaptability to noisy environments, combined with its lightweight architecture, makes CardioNet ideal for integration into wearable devices in future, to enable real-time heart sound monitoring and effective cardiovascular disease detection in both clinical and remote settings.