Deep Learning and Dual-Domain Fusion of Temporal and Spectrotemporal SCG for Valvular Heart Disease Classification

Heart valve disease (HVD) is a growing cause of cardiovascular morbidity, yet accessible screening tools remain limited. Tri-axial seismocardiogram (SCG) signals capture mechanical signatures of both stenosis and regurgitation, but their non-stationary nature and subtle inter-lesion morphological differences challenge robust classification from a single representation. This study proposes an electrocardiogram (ECG)-free dual-domain deep learning framework that combines a 1D convolutional-recurrent encoder of raw SCG waveforms with a 2D EfficientNet encoder of continuous wavelet transform scalograms, integrating temporal and spectrotemporal representations through feature-level fusion. The framework was evaluated on primary valve lesion classification in 94 patients, with an additional exploratory analysis of multi-valvular, aortic-stenosis-dominant disease. Under patient-level stratified group cross-validation, the fused model achieved the highest mean accuracy (65.3%; 95% CI, 59.8-70.7%), exceeding both unimodal encoders and the majority-class (35.2%) and covariate-only (39.1%) baselines. The same mean performance ordering was preserved when the image backbone was replaced, although the confidence intervals overlap at this cohort size. Under the segment-level protocol used in prior studies, the framework achieved 99.63% accuracy, with faster training convergence and more compact, separable feature clusters than either unimodal encoder. These findings support the value of combining temporal and spectrotemporal SCG representations and demonstrate its potential for ECG-free HVD classification under patient-independent evaluation.