Characterizing Benzodiazepine Use in a Large National Study via Wearables and Deep Learning
Franklin Ruan, Stephen Adjei, Adaobi Amanna, George D. Price, Michael V. Heinz, Nicholas C. Jacobson · 2024
Behavior-based pharmacological side effects, including those of benzodiazepines, are often assessed through self-reported surveys, introducing response bias. This study aimed to address this limitation by leveraging deep learning and wearables in a large-scale, naturalistic setting to detect and characterize benzodiazepine psychomotor effects objectively. Using data from the 2005-2006 National Health and Nutrition Survey, 7162 participants wore physical activity monitors for seven days, with 137 reporting benzodiazepine use (55% of prescriptions were verified). We constructed and cross-validated several machine-learning models to predict benzodiazepine use from movement. The top-performing ConvLSTM model achieved a test AUC of 0.71 [95% CI, 0.71-0.71] with a Cohen’s d of 0.78 [95% CI, 0.78-0.78]. Hence, we find a moderate-to-large correspondence between the predicted and known benzodiazepine use activity. The analysis accounted for potential demographic confounders, suggesting the movement differences are specific to benzodiazepine use. Participants on benzodiazepines showed reduced physical activity upon waking and in the late afternoon, deeper sleep, and increased sedentary behavior. The study shows promise in using passively collected movement data to detect and characterize benzodiazepine psychomotor side effects and supports the use of deep-learning passively collected data to create objective metrics for psychotropic medication effects.