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

Balasundaram Kadirvelu, Teresa Bellido Bel, Aglaia Freccero, Martina Di Simplicio, Dasha Nicholls, A. Aldo FaisalJournal of Medical Internet Research 2025 · DOI

TL;DR

This study demonstrates that combining active (self-reported) and passive (smartphone sensor) data with advanced machine learning can effectively predict mental health risks in adolescents. This approach shows promise for scalable, early detection and intervention in community settings, addressing a critical gap in adolescent mental healthcare.

What Problem Does This Paper Address?

Adolescence is a critical period for the onset of mental health disorders, yet many young people don't seek help. Digital phenotyping, using smartphone data, offers a scalable way to detect risks early. However, its application in general, nonclinical adolescent populations has been limited, leaving a gap in prevention efforts. This paper aims to bridge this gap by exploring the feasibility and utility of using machine learning to predict mental health risks (internalizing/externalizing difficulties, eating disorders, insomnia, suicidal ideation) in school-going adolescents by integrating both active and passive smartphone data.

Methods

The researchers developed a deep learning framework that integrates active and passive smartphone data. A key innovation is the use of contrastive pretraining with triplet margin loss. This technique helps the model learn stable, user-specific behavioral patterns by teaching it to distinguish between data from the same user on different days (positive examples) and data from different users (negative examples). This "fingerprinting" process helps ignore daily noise and capture genuine behavioral shifts. After pretraining, the model is fine-tuned using supervised learning to classify individuals into high-risk or low-risk categories for specific mental health outcomes. They also compared this approach to standard machine learning models like CatBoost and a multilayer perceptron (MLP) without pretraining.

Datasets

The study used data from 103 adolescents recruited from 3 secondary schools in northwest London. Participants used the "Mindcraft" app for 14 days, providing daily self-reports (active data) and allowing continuous passive sensor data collection (passive data). An independent external validation cohort of 45 adolescents from 5 additional schools was also used to assess generalizability.

Key Insights and Ideas

  • Multimodal data is key: Combining active and passive data significantly outperformed single-modality models, highlighting the complementary nature of subjective self-reports and objective sensor data.
  • Contrastive learning for stability: The contrastive pretraining approach was crucial for stabilizing user-specific behavioral patterns, making the model more robust to the inherent daily variability in smartphone data. This is a significant methodological advancement for digital phenotyping.
  • Passive data's low burden: Despite declining engagement with active data over time, passive data collection remained more consistent, suggesting its potential for scalable, long-term monitoring with lower participant burden.
  • "Digital Fingerprinting": The contrastive learning method effectively creates a unique "digital fingerprint" for each user, allowing for more personalized and accurate risk prediction.

Key Results

  • The combined active and passive data model achieved mean balanced accuracies ranging from 0.67 (insomnia) to 0.77 (suicidal ideation).
  • This combined approach significantly outperformed unimodal models (active or passive data alone) for most outcomes. For example, for suicidal ideation, the combined model achieved 0.77 balanced accuracy, compared to 0.71 for active data and 0.62 for passive data.
  • The contrastive pretraining approach led to higher balanced accuracy (0.67) compared to models without pretraining (0.65) and a CatBoost benchmark (0.64).
  • External validation on an independent cohort showed similar performance patterns, with balanced accuracies ranging from 0.63 to 0.72, indicating good generalizability.
  • SHAP analysis identified clinically relevant features, such as negative thinking (active) and location entropy (passive), as important predictors.

Limitations and Future Work

  • Sample size and generalizability: The study was conducted in 3 London schools, which may limit generalizability to other socioeconomic and regional contexts. The gender imbalance (due to a girls-only school) might also bias findings. Future work should include more diverse samples.
  • Heterogeneous sensor activation: Not all participants enabled all sensors, leading to incomplete data. While sentinel values were used, this missingness could still introduce bias. Future work should explore fairness-aware modeling.
  • Dampened sensitivity: Cumulative median aggregation, while robust to outliers, might reduce sensitivity to sudden, clinically significant behavioral changes. Hybrid approaches combining cumulative data with volatility-sensitive features are suggested.
  • Limited contextual factors: The study did not include broader factors like genetics, SES, or family history, which can influence mental health. Future research could integrate these.
  • Engagement decline: Active data engagement decreased significantly over the 14-day period. Future app iterations should focus on context-aware adaptive sampling, gamification, and personalized feedback loops to improve long-term engagement.
  • Interpretability for users: While SHAP values provide technical interpretability, translating these into actionable insights for adolescents, parents, and educators is crucial for real-world adoption.

Summary updated

Abstract

This study demonstrates the feasibility and utility of smartphone-based digital phenotyping for predicting mental health risks in nonclinical, school-going adolescents. By integrating active and passive data with advanced machine modeling techniques, this approach shows promise for early detection and scalable intervention strategies in community settings.
BibTeX
@article{kadirvelu2025digital,
  author = {Balasundaram Kadirvelu and Teresa Bellido Bel and Aglaia Freccero and Martina Di Simplicio and Dasha Nicholls and A. Aldo Faisal},
  title = {{Digital Phenotyping for Adolescent Mental Health: Feasibility Study Using Machine Learning to Predict Mental Health Risk From Active and Passive Smartphone Data}},
  journal = {Journal of Medical Internet Research},
  year = {2025},
  doi = {10.2196/72501},
}
Figure