Abstract TP14: SocialBit: A Novel Smartwatch Sensor to Detect Social Interaction Frequency in Stroke Survivors

Abstract

Introduction: Stroke survivors experience high levels of social isolation which hinder their rehabilitation and well-being. A gap in the field that impedes intervention development is a lack of real-time social interaction measurement tailored for this population. In response, we developed SocialBit, a smartwatch-based machine learning algorithm to discreetly and passively detect social interactions through privacy-preserving acoustic analysis. In this observational study, we estimated the accuracy of SocialBit to detect social interactions compared to human observers in stroke survivors. Methods: SocialBit was built as a neural network machine learning algorithm using features from YAMNet and a Transformer classifier. It was then trained and tested in stroke survivors for up to 8 days in hospital. Independently, human observers tallied the presence or absence of social interactions every minute to establish the ground truth. SocialBit performance of detecting social interactions was compared against the ground truth. Results: We analyzed 1137.7 hours of ambient audio data in 90 patients within 14 days of stroke onset. 16% of patients had aphasia, 25% reported symptoms of clinical depression, and 28% reported moderate to severe loneliness. SocialBit had an accuracy of 0.83 (SD = 0.2), sensitivity of 0.86 (SD = 0.2), specificity of 0.81 (SD = 0.02), and an area under the curve (AUC) of 0.90 (SD = 0.2) (Figure 1). Conclusion: SocialBit accurately detected the frequency of social interactions in stroke survivors in a hospital setting. These results will drive future planned studies to examine SocialBit's performance in patients’ home communities. SocialBit shows the potential to be the first real-time social sensor to quantify social interactions and social isolation in stroke survivors.