Homer Babbidge Library

Abstract: Depression is a prevalent mental health disorder, and identifying effective treatment early remains a key challenge. Traditional assessments rely on clinical questionnaires, which are costly, infrequent, and prone to biases. With the rise of smartphones, new opportunities have emerged to monitor mental health using active self-reports and passive sensor data. To address the shortcomings of traditional assessments, we explore the use of smartphone-collected data—specifically, self-reported daily mood and anxiety ratings, medication survey and passive location-based sensory data—to predict depression treatment outcomes. In the first part, we evaluated 5-point Likert-scale mood and anxiety self-ratings as alternatives to clinical questionnaires. These self-reports significantly correlate with clinical scores (QIDS), achieving an F1 score of 0.65 using machine learning models like SVM, XGBoost, and Random Forest, compared to 0.71 using clinical questionnaires. In the second part, we investigate the use of passively collected location data for predicting treatment outcomes. To address the heterogeneity between Android and iOS data collection, we apply a domain adaptation technique (Correlation Alignment - CORAL) to align data from both platforms, significantly improving prediction performance. Our results show that combining location features with baseline clinical scores yields an F1 score of up to 0.67—comparable to periodic self-reported questionnaires—highlighting location data as a promising predictor of depression treatment outcomes.   In the third part, we examine whether smartphone data collected early in treatment, specifically during the initial 2–4 weeks, can reliably predict outcomes at the 12th week (the end of the treatment course). Early prediction of treatment outcomes is critical for enabling timely clinical decisions. This study integrates weekly medication surveys, daily mood and anxiety self-ratings, and location-based sensory data to predict depression treatment outcomes 8 to 10 weeks before treatment completion with help of deep learning models (GRU-D and BRITS). Our findings indicate that smartphone data collected in the early phases of treatment can achieve prediction accuracies comparable to weekly clinical questionnaires.

Conferences & Speakers