Social
Detection of Loneliness and Social Connectedness from Passive Sensor Data Using Machine Learning Afzalbek Fayzullaev* Afzalbek Fayzullaev Fayzullaev Fayzullaev Fayzullaev Fayzullaev Fayzullaev Fayzullaev Fayzullaev School of Medicine, Boston University
Passive sensor data from wearable devices collect information on everyday life and behavioral patterns, making them a promising tool for detecting loneliness and social connectedness. This study evaluated the feasibility of using Fitbit data to identify loneliness and low social support.
Data were from 17,127 respondents to the Social Determinants of Health (SDOH) Survey from the All of Us Research Program who were aged ≥ 18 years old with valid Fitbit data (defined as at least 10 hours of wear time and 100-45,000 daily steps per day, collected during the 7 days preceding SDOH survey administration). The sample was randomly split into training (80%) and test (20%) sets. Loneliness was measured using the UCLA 8-Item Loneliness Scale (range: 8-32); participants with a score ≥ 21 were classified as lonely. Lack of social support was calculated using the RAND Social Support Survey Instrument (range: 8-40); participants with a score ≥ 26 were classified as lacking social support. Extreme gradient boosting (XGBoost) models predicted loneliness and lack of social support using mean step count, age, gender, race and ethnicity, income, education, and marital status. Model performance was evaluated in the test set using sensitivity, specificity, and SHAP values for variable importance.
The average age was 53.9 years (SD=15.0); 70.3% were female and 78.5 % were non-Hispanic White. The XGBoost model for loneliness achieved sensitivity of 0.58 and specificity of 0.71, whereas the models for predicting social support achieved a sensitivity of 0.59 and specificity of 0.70. According to the SHAP values, mean step count was the most influential predictor after marriage and income.
In one of the largest Fitbit-based studies to date, models using mean daily step counts showed modest performance in predicting loneliness and social connectedness. Future work will evaluate alternative algorithms, address class imbalance, and extend the observation window.
