BioSync: Transformer-Based Cross-Modal Fusion for a Multimodal Physiological Digital Biomarker

BioSync proposes a Transformer-based multimodal physiological digital biomarker, integrating electrocardiogram, neurological, behavioral, and speech data collected by wearable devices into the BioSync Index (BSI). The model uses a multi-head self-attention mechanism to process modality labels and introduces linear branches for feature concatenation. In two literature-guided synthetic cohorts, BioSync achieved an AUC of 0.928 in the mental decline cohort, and accuracy and F1 scores of 0.764 and 0.766, respectively, in the metabolic autonomic nervous system cohort, both outperforming feature concatenation methods; the correlation coefficient between BSI and potential severity was 0.91 and 0.68, respectively, in both cohorts. A ablation experiment showed that the performance improvement stemmed from the wide-depth combined architecture, and BioSync performed better than the concatenation method under five modality-absent training conditions. Although its AUC in the cognitive cohort was higher than five published digital biomarker reference values, differences in dataset and task limitations restrict direct comparative conclusions; comparisons across six predefined criteria revealed modal fusion…