Abstract
Background: Attention-deficit/hyperactivity disorder (ADHD) is characterized by fluctuating symptom severity that is challenging to capture with traditional cross-sectional assessments. Smartphone-based digital phenotyping offers a promising approach for continuous, ecologically valid monitoring. This study aimed to investigate whether passive smartphone sensor data can predict within-day and between-day fluctuations in ADHD symptoms in adults.Methods: Thirty-six adults with clinically diagnosed ADHD (mean age 28.4 years, 58% female) participated in a 4-week observational study. Smartphone sensors (accelerometer, gyroscope, GPS, screen state) were continuously recorded via a custom app. Ecological momentary assessments (EMAs) of inattention and hyperactivity/impulsivity were administered 6 times daily. Machine learning models (random forest, gradient boosting) were trained to predict symptom levels from sensor-derived features (e.g., activity counts, location entropy, screen time).Results: Random forest models achieved moderate accuracy in predicting inattention (R² = 0.31, RMSE = 0.89) and hyperactivity/impulsivity (R² = 0.27, RMSE = 0.94) from sensor features. Key predictors included screen time variability, location entropy, and accelerometer mean amplitude. There was significant inter-individual variability in model performance (R² range 0.08–0.55).Conclusions: Smartphone sensors can capture meaningful variance in ADHD symptom fluctuations, supporting the feasibility of digital phenotyping for ADHD. Personalized models may enhance predictive accuracy, but further work is needed to improve generalizability and clinical utility.