Multivariate multimodal modeling improves non-invasive classification of echocardiographic dysfunction beyond single electromechanical markers

Background Individual multimodal electro- and mechanocardiographic markers derived from synchronized electrocardiography (ECG), phonocardiography (PCG), impedance cardiography (ICG), and seismocardiography (SCG) are associated with cardiac dysfunction. However, it remains unclear whether multivariate integration of complementary multimodal features provides consistent incremental discrimination beyond the strongest individual marker across different echocardiographic abnormalities. Objective To determine whether multivariate integration of synchronized multimodal ECG/PCG/ICG/SCG features improves discrimination beyond the strongest individual multimodal electromechanical marker across predefined echocardiographic abnormalities. Methods We analysed 1069 participants from a prospective single-centre diagnostic cohort who underwent synchronized ECG/PCG/ICG/SCG acquisition as well as standard ECG, and comprehensive transthoracic echocardiography. A multimodal feature space was constructed including electrical intervals, systolic time fractions, morphology descriptors, and cross-modal electromechanical coupling indices. Predefined targets were tricuspid annular plane systolic excursion (TAPSE) below 17 mm, impaired right ventricular (RV) strain, left ventricular (LV) systolic dysfunction, advanced diastolic dysfunction, severe left atrial (LA) enlargement, and moderate-to-severe tricuspid regurgitation. For each target, cross-validated logistic regression models were trained using (1) an integrated multimodal feature set and (2) individual multimodal candidate features. To reduce comorbidity-related overlap, the controls were restricted to a predefined healthy reference subgroup excluding major comorbidities such as ischaemic heart disease, hypertension, and diabetes. Results Across all predefined echocardiographic targets, multivariate multimodal models outperformed the strongest individual multimodal electromechanical markers. AUC-ROC improved from 0.77–0.83 for univariate models to 0.85–0.88 for multivariate models, corresponding to absolute gains of +0.03 to +0.09. The largest improvements were observed for left atrial enlargement (0.767 to 0.859) and diastolic dysfunction (0.788 to 0.865), while consistent gains were also seen for right ventricular strain (0.810 to 0.861), TAPSE ≤17 mm (0.826 to 0.878), tricuspid regurgitation (0.816 to 0.854), and left ventricular systolic dysfunction (0.815 to 0.849). Conclusion Multivariate integration of synchronized multimodal ECG/PCG/ICG/SCG features provides consistent incremental discrimination beyond the strongest individual multimodal electromechanical markers across all evaluated echocardiographic targets. The findings support the presence of non-redundant pathophysiological information across electromechanical domains and justify prospective multicentre evaluation of integrated multimodal phenotyping as a non-invasive screening approach.