Heart failure diagnosis and ejection fraction classification via feature fusion model using non-contact vital sign signals

Comprehensive Summary

In this study, Feng et. al. developed a fusion model that combines manual features (MFs) and deep features (DFs) for the improved diagnosis of heart failure (HF) and left ventricular ejection fraction (LVEF) classification in a manner that utilizes non-contact ballistocardiography (BCG)/respiratory signals. The need for this model emerged because although ballistocardiography has gained recognition for its user-friendliness and cost-effectiveness as a tool that monitors HF, its MFs are unable to adequately characterize LVEF dynamics for its users. Based on clinical diagnoses, 83 participants were recruited and assigned to a binary (healthy and HF) and ternary (healthy, LVEF ≥ 40% HF, and LVEF < 40% HF) classification group. Piezoelectric sensors were used to record non-contact vital signs, and BCG and respiratory signals were isolated to use for the creation of a model that integrated MFs and DFs. A multi-scale ResNet-BiLSTM network model was subsequently created for DF extraction from the BCG/respiratory signals. Results displayed that the feature fusion model outperformed conventional methods and achieved accuracies of 98.20% and 98.76% for the binary and ternary classifications, respectively. Overall, the hybrid model that combines MFs with DFs may facilitate and improve the at-home diagnosis of HF and LVEF classification.

Outcomes and Implications

This study is especially significant to communities in which cardiac imaging is scarce or inaccessible as the piezoelectric sensor-based solution is portable and convenient. Since the hybrid model can be adapted for the preliminary screening of HF in out-of-hospital settings, this technology can catalyze early detection and treatment for patients.