Abstract DP098: SocialBit: Validation of a Lightweight Machine Learning Algorithm for Measuring Social Interactions in Real-Time in Stroke Survivors with Diverse Neurological Abilities

Introduction: Social isolation has adverse biological effects on the brain that negatively impact recovery after brain injuries such as stroke. Social interaction, via its complex cognitive demands and effects on brain plasticity, may support stroke recovery. However, no clinically tested tool for real-time measurement of social interaction exists. We developed SocialBit, a lightweight, privacy-preserving machine learning algorithm that detects social interactions using ambient audio features on a commercial smartwatch without storing raw audio or using natural language processing. Methods: In a prospective longitudinal validation study, we examined the validity of SocialBit in a dataset comprising 88,918 minutes (~1,482 hours) of observation from 153 hospitalized stroke survivors with diverse neurological abilities, including aphasia and cognitive impairment. Participants wore the device for up to 8 days during routine care while human observers recorded social interactions via livestream (ground truth). Results: SocialBit achieved high overall classification accuracy (sensitivity: 0.87, specificity: 0.88, Area under the Curve, AUC: 0.94) ( Figure 1 ) and maintained accuracy in patients with aphasia (AUC 0.93). SocialBit’s performance was robust to environmental variation including background TV, side conversations, care setting, and smartwatch models. The algorithm’s performance varied by social interaction characteristics with higher sensitivity for conversations in English, with more speakers, and of more than minimal depth. Finally, SocialBit detected that patients with greater stroke severity engaged in less social interaction, mirroring patterns observed in human-coded ground truth. Each 1-point increase in NIH Stroke Scale score corresponded to a 0.9% decrease in SocialBit-estimated interaction time (r = –0.19, p = .029) and 1.1% human-coded interaction time (r = –0.22, p = .008) ( Figure 2 ). Conclusion: These findings support SocialBit as the first universally accessible and privacy preserving social sensor for stroke survivors with diverse neurological abilities. As a digital biomarker, it may enable new interventions to preserve or improve cognition, social engagement, and quality of life.