A Machine Learning Framework for Valve Disease Detection and Cardiac Parameter Estimation Using Seismocardiogram Signals

Effective early screening is essential for preventing the progression of valvular heart diseases, such as stenosis and regurgitation. Diagnostic tools such as echocardiography, which are costly and require trained experts, often limit early detection in low-resource settings. This study explores the potential of seismocardiogram (SCG) signals as a cost-effective, user-friendly alternative for cardiac health monitoring. We propose a machine learning framework utilizing SCG signals to detect valve diseases and estimate key cardiac parameters: ejection fraction (EF), left ventricular end-diastolic dimension (LVEDD), and left ventricular posterior wall thickness (LVPW). Our novel feature extraction approach integrates Berthil cepstrum with spectral and wavelet-based methods to process artifact-free 2-second SCG segments. These features are used to train XGBoost classifiers and regressors. Evaluated on a cohort of 16 stenosis and 13 regurgitation patients from a publicly available SCG database, our model demonstrates remarkable performance improvements over existing methods. Using the Root Mean Squared Logarithmic Error (RMSLE) objective function, XGBoost regressor models achieve significant improvement compared to those using Mean Squared Error (MSE). Feature importance analysis reveals that the top 20 features retain the most relevant information, making the model computationally efficient with minimal performance loss. This study presents a promising and reliable tool for the early detection of stenosis and regurgitation, along with accurate estimation of cardiac parameters.