Obsessive-compulsive disorders (OCD) are marked by distress, negative emotions, mental processes and behaviors that are reflected in physiological signals such as heart rate, electrodermal activity, and skin temperature. Continuous monitoring of physiological signals associated with OCD symptoms may make measures of OCD more objective and facilitate close monitoring of prodromal symptoms, treatment progress and risk of relapse. Thus, we explored the feasibility of capturing OCD events in the real world using an unobtrusive wrist worn biosensor and machine learning models.
Nine adolescents (ages 10–17 years) with mild to moderate-severe OCD were recruited from child and adolescent mental health services. Participants were asked to wear the biosensor in the lab during conditions of rest and exposure to OCD symptom-triggering stimuli and for up to 8 weeks in their everyday lives and register OCD events. We explored the relationships among physiological data, registered OCD events, age, OCD symptom severity and symptom types. In the machine learning models, we considered detection of OCD events as a binary classification problem. A nested cross-validation strategy with either random 10-folds, leave-one-subject-out, or leave-week(s)-out in both layers was used. We compared the performance of four models: logistic regression, random forest (RF), feedforward neural networks, and mixed-effect random forest (MERF). To explore the ability of the models to detect OCD events in new patients, we assessed the performance of participant-based generalized models. To explore the ability of models to detect OCD events in future, unseen data from the same patients, we compared the performance of temporal generalized models trained on multiple patients with personalized models trained on single patients.
Eight of the nine participants collected biosensor signals totaling 2, 405 h and registered 1, 639 OCD events. Better performance was obtained when generalizing across time compared to across patients. Generalized temporal models trained on multiple patients were found to perform better than personalized models trained on single patients. RF and MERF models outperformed the other models in terms of accuracy in all cross-validation strategies, reaching 70% accuracy in random and participant cross-validation.
Our pilot results suggest that it is possible to detect OCD episodes in the everyday lives of adolescents using physiological signals captured with a wearable biosensor. Large scale studies are needed to train and test models capable of detecting and predicting episodes.