KMAN Publication Inc. (KMANPUB) Journal of Adolescent and Youth Psychological Studies (JAYPS) 2981-2526 7 Serial Number 49 2026 09 10 Forecasting Daily Stress in Adolescents from Mood Variability, Sleep Quality, Social Interaction, and Smartphone Use Using Long Short-Term Memory Networks Forecasting Daily Stress in Adolescents from Mood Variability, Sleep Quality, Social Interaction, and Smartphone Use Using Long Short-Term Memory Networks 1 19 10.61838/kman.jayps.5672 EN Journal Article 2025 12 10 <table> <tbody> <tr> <td> <p><strong>Objective:</strong> This study aimed to forecast next-day stress among Canadian adolescents using seven-day temporal sequences of mood variability, sleep quality, social interaction, and smartphone use through a long short-term memory network.</p> <p><strong>Methods and Materials:</strong> A prospective intensive longitudinal study was conducted with 312 adolescents aged 13–18 years recruited from Ontario, British Columbia, Alberta, and Quebec, Canada. Participants completed baseline assessments followed by 30 consecutive days of ecological momentary assessment and passive smartphone monitoring. Daily perceived stress was assessed each evening, mood was recorded four times daily, sleep quality was measured through morning sleep diaries, and social interaction was evaluated using daily self-reports and communication metadata. Passive smartphone indicators included total screen time, device unlocks, session duration, application-category use, and nighttime smartphone activity. Seven-day sliding windows were used to predict the following day’s stress. Data were divided at the participant level into training, validation, and independent test sets. The LSTM model was compared with persistence, multiple linear regression, random forest, and feedforward neural-network models.</p> <p><strong>Findings:</strong> The LSTM model demonstrated the best predictive performance in the independent test set, with a mean absolute error of 1.42, root mean square error of 1.89, coefficient of determination of .62, and correlation of .79 between observed and predicted stress. It reduced mean absolute error by 27.6% compared with the persistence model, 20.2% compared with linear regression, 11.8% compared with random forest, and 7.8% compared with the feedforward neural network. Ablation analyses showed that removing mood, sleep, social-interaction, and smartphone-use variables increased prediction error, with mood variability producing the greatest decline in performance. Negative-mood variability, subjective sleep quality, nighttime smartphone use, loneliness, and previous-day stress were the most influential predictors.</p> <p><strong>Conclusion:</strong> Sequential patterns of mood, sleep, social experiences, and smartphone behavior can meaningfully forecast adolescents’ next-day stress, and LSTM networks offer greater predictive accuracy than conventional and nonsequential models.</p> </td> </tr> </tbody> </table> https://journals.kmanpub.com/index.php/jayps/article/download/5672/11786