KMAN Publication Inc. (KMANPUB) Journal of Adolescent and Youth Psychological Studies (JAYPS) 2981-2526 7 Serial Number 50 2026 10 10 Detecting Adolescent Emotional Distress from Daily Smartphone Behavioral Patterns Using Long Short-Term Memory Networks Detecting Adolescent Emotional Distress from Daily Smartphone Behavioral Patterns Using Long Short-Term Memory Networks 1 19 10.61838/kman.jayps.6354 EN Journal Article 2026 07 21 <table> <tbody> <tr> <td> <p><strong>Objective:</strong> This study aimed to determine whether daily smartphone-derived behavioral patterns could accurately detect high emotional distress among Colombian adolescents using Long Short-Term Memory (LSTM) networks and whether sequential deep learning outperformed alternative predictive models.</p> <p><strong>Methods and Materials:</strong> A prospective longitudinal observational study was conducted among 312 adolescents aged 13–18 years recruited from public and private secondary schools in Bogotá, Medellín, and Cali, Colombia. Participants were monitored for eight consecutive weeks using a custom Android smartphone application that passively recorded non-content behavioral metadata related to screen activity, application use, communication, mobility, and temporal routines. Daily emotional distress was assessed through ecological momentary assessment. After preprocessing, 15,884 valid participant-days and 13,912 seven-day behavioral sequences were available for analysis. Data were divided at the participant level into training, validation, and independent test sets. An LSTM model was developed and compared with logistic regression, random forest, gradient boosting, and feed-forward neural network models.</p> <p><strong>Findings:</strong> The LSTM model achieved a ROC-AUC of 0.886, PR-AUC of 0.774, accuracy of 82.3%, sensitivity of 81.1%, specificity of 82.8%, balanced accuracy of 81.9%, and F1 score of 0.737 in the independent test set. It outperformed logistic regression (ROC-AUC = 0.724), random forest (0.801), gradient boosting (0.827), and feed-forward neural network (0.842). Temporal-routine features showed the strongest single-domain predictive performance, followed by mobility features. The seven-day sequence yielded better discrimination than three-day, five-day, and ten-day alternatives. Increased nighttime use, later final smartphone interaction, reduced mobility, greater time spent at the primary location, and reduced routine regularity were among the most influential predictive features.</p> </td> </tr> </tbody> </table> <p><strong>Conclusion:</strong> Daily smartphone behavioral patterns provide a meaningful, privacy-preserving signal for detecting adolescent emotional distress, and LSTM networks offer superior performance by capturing temporal dependencies across multiple behavioral domains.</p> https://journals.kmanpub.com/index.php/jayps/article/download/6354/12117