Morphological analysis and multimodal deep learning for impedance cardiography feature localization and hemodynamic parameter estimation

Abstract

Aim: This study aimed to reduce the susceptibility of impedance cardiography (ICG) feature-point localization to waveform variability, noise, and inter-individual differences by developing an Electrocardiography (ECG)-ICG-phonocardiography (PCG) multimodal framework for beat-to-beat estimation of left ventricular ejection time (LVET), stroke volume (SV), and cardiac output (CO). Methods: Using the public HeartCycle multimodal dataset, ICG waveform polymorphism and class overlap were characterized using dynamic time warping, K-medoids clustering, and morphological classification. We then developed an end-to-end model with modality-specific one-dimensional convolutional encoders, Transformer-based temporal modeling, and event-query localization. ECG provided beat alignment, and the PCG envelope supplied mechanical-event timing information. Performance was compared with conventional rule-based methods, learning-based baselines, and ablation models under a unified evaluation framework, with paired analyses and leave-one-subject-out (LOSO) validation. Results: Increasing clustering granularity increased morphological overlap and classification difficulty, indicating limitations of local ICG-only rules. On the primary test set, the PCG-integrated model achieved a mean absolute percentage error of 6.51% for SV and 7.35% for CO, lower than those of the conventional rule-based method and the deep-learning model without PCG. Subject-level paired analyses based on the 17-fold LOSO results demonstrated significant reductions in SV and CO estimation errors following PCG integration and further supported the feasibility of subject-independent beat-level estimation. The complete model also yielded lower errors than learning-based baselines and ablation models under current evaluation conditions. These findings reflect the performance of the integrated framework, while its generalizability warrants further validation. Conclusion: ECG-ICG-PCG fusion provides a feasible strategy for more stable beat-to-beat LVET, SV, and CO estimation under the current evaluation conditions. Larger multicenter datasets, diverse patient populations, and prospective longitudinal studies are needed to establish reliable trend monitoring and within-subject tracking of hemodynamic changes.