BioSync integrates cardiac, neural, behavioral, and speech data into the BioSync Index (BSI), a continuous digital biomarker defined under the BEST framework. The model employs multi-head self-attention on modality tokens and includes a linear branch for feature concatenation.
The architecture is motivated by latent-variable measurement theory and the idea that joint observations may contain unique information. Evaluation was conducted on synthetic cohorts involving multiple modalities, including HRV, EEG, actigraphy, and speech, as well as a metabolic-autonomic cohort structured around a public wearable schema.
Results showed high AUC scores in the cognitive cohort, with BioSync slightly outperforming simple concatenation. In the metabolic cohort, BioSync achieved accuracy and F1 scores marginally higher than concatenation. The BSI correlated strongly with latent severity in both cohorts.
Additional tests demonstrated the model's robustness under modality-dropout training, outperforming concatenation at various corruption rates. The architecture's properties were compared with other fusion designs, though validation on real cohorts remains necessary.
Source: https://arxiv.org/abs/2609.04504