DIGITAL PHENOTYPING USING PASSIVE SMARTPHONE DATA FOR EARLY PREDICTION OF PSYCHOTIC EPISODES IN CLINICALLY HIGH-RISK ADOLESCENTS: A LONGITUDINAL MACHINE LEARNING APPROACH

ANAM NAWAZ MALIK, AYESHA AMIN TARIQ, WAJEEHA NADEEM AND HIRA SHAFIQ

The early identification of adolescents at Clinical High Risk (CHR) for psychosis is critical for preventive intervention. Traditional assessment tools, though valuable, are limited by their reliance on subjective reporting, clinical access, and declining predictive validity. With global increases in digital connectivity, digital phenotyping offers a promising, non-intrusive method to continuously capture behavioral patterns indicative of emerging psychopathology yet remains severely underexplored among youth in non-Western settings. This study aimed to identify and analyze distinct digital behavioral phenotypes using passively collected smartphone data among South Asian adolescents at CHR for psychosis, and to explore their associations with psychotic symptom severity, affective disturbances, and functional outcomes. Using a naturalistic longitudinal design, smartphone usage data including sleep duration, mobility, screen time, and communication metrics were collected from adolescents identified as CHR. Employing unsupervised machine learning (k means clustering), we extracted behavioral subtypes and examined their correlations with validated clinical assessment scales (e.g., SIPS, PANSS, and CGAS). Three distinct digital phenotypes emerged: (1) Isolated Dysregulated (characterized by excessive screen time, erratic sleep, and reduced mobility), (2) Subdued Stable (moderate, consistent behavior patterns), and (3) Externally Engaged Variable (increased mobility and communication with behavioral inconsistency). The Isolated Dysregulated group demonstrated significantly higher levels of subthreshold psychotic symptoms, negative affect, and functional impairment, while the Subdued Stable group showed the most favorable clinical profiles. This study is among the first to demonstrate that passive smartphone data can reveal distinct behavioral risk profiles among CHR adolescents in a South Asian context. These digital phenotypes are not only predictive of symptom severity but also offer culturally attuned insights into real-world functioning. Our findings underscore the transformative potential of digital phenotyping in shifting psychosis prevention from reactive, clinic-based models to proactive, personalized, and scalable interventions, particularly in under-resourced and collectivist societies where early help-seeking is often delayed.