Digital Phenotyping for Adolescent Mental Health: Feasibility Study Using Machine Learning to Predict Mental Health Risk From Active and Passive Smartphone Data.

Integration of active and passive smartphone data outperformed single-modality models across all four mental health outcomes.

• Combined active and passive data achieved mean balanced accuracies of 0.71 (SD 0.03) for SDQ-high risk, 0.67 (SD 0.04) for insomnia, 0.77 (SD 0.03) for suicidal ideation, and 0.70 (SD 0.03) for eating disorder.
• Single-modality models (active-only or passive-only) performed below these levels across all outcomes.
• Performance was assessed using leave-one-subject-out cross-validation (LOSO-CV) with balanced accuracy as the primary metric.
• The study included n=103 participants with mean age 16.1 years (SD 1.0) from 3 UK secondary schools.

The contrastive pretraining approach with triplet margin loss improved representation stability and predictive robustness compared to models without pretraining.

• The deep learning model incorporated contrastive pretraining with triplet margin loss to stabilize user-specific behavioral patterns.
• Comparative analyses were conducted using CatBoost and multilayer perceptron (MLP) models without pretraining.
• The contrastive learning approach outperformed both CatBoost and MLP baseline comparators.
• The pretraining step was designed to capture individual behavioral baselines before supervised fine-tuning for binary classification.

The smartphone-based digital phenotyping approach demonstrated generalizability in an independent external validation cohort.

• An independent external validation cohort of n=45 participants was used.
• Balanced accuracies of 0.63-0.72 were achieved across outcomes in the external validation cohort.
• These results suggest the approach generalizes to new settings beyond the original training population.
• The external validation cohort was separate from the primary study sample of n=103.

SHAP analysis identified clinically relevant active and passive features as most important for mental health prediction.

• Negative thinking (an active/self-reported feature) was highlighted as a key predictor by SHAP analysis.
• Location entropy (a passive/sensor-based feature) was identified as an important predictor, underscoring the complementary value of combining subjective and objective data.
• Shapley Additive Explanations (SHAP) values were used to assess feature importance across the model.
• Correlation analyses confirmed meaningful associations between key digital features and clinical scale scores.

Adolescents from nonclinical school populations used the Mindcraft smartphone app for 14 days providing both active self-reports and continuous passive sensor data.

• Participants (n=103; mean age 16.1 years, SD 1.0) were recruited from 3 UK secondary schools.
• Active data included daily self-reports such as mood, sleep, and loneliness.
• Passive data included continuous sensor streams such as location, step count, and app usage.
• The study duration was 14 days.
• Four binary classification outcomes were targeted: SDQ-high risk, insomnia, suicidal ideation, and eating disorder.

Over 75% of lifetime mental disorder cases emerge before age 25, yet most young people with significant symptoms do not seek support.

• Adolescents are described as particularly vulnerable to mental disorders.
• The statistic that over 75% of lifetime cases emerge before age 25 is cited as motivation for the study.
• Most young people with significant symptoms do not seek support, creating a gap in community-based prevention.
• The study targets nonclinical, school-going adolescents to address this gap.