Predicting the Risk of Psychological Disorders Using Digital Behavioral Markers: An Artificial Intelligence and Behavioral Science Framework

Abstract

This study aimed to identify the components, requirements, and mechanisms of a culturally responsive, ethically governed, and person-centered framework for predicting the risk of psychological disorders using digital behavioral markers, artificial intelligence, and behavioral science principles. This qualitative study was conducted using conventional content analysis. The participants were 22 clinical psychologists, psychiatrists, artificial intelligence and data science specialists, and digital health professionals working in Tehran, Iran. They were selected through purposive maximum-variation sampling. Data were collected exclusively through in-depth semi-structured interviews. Theoretical saturation was reached after the nineteenth interview, followed by three additional interviews to confirm the stability and completeness of the categories. The interviews were transcribed verbatim and analyzed through initial coding, constant comparison, aggregation of conceptually similar codes, and development of subcategories and main categories. NVivo software was used to organize, manage, retrieve, and compare the qualitative data. Credibility and dependability were enhanced through member checking, peer debriefing, maximum-variation sampling, documentation of the analytical process, and examination of alternative interpretations. The analysis generated five main categories: “multimodal architecture of digital behavioral markers,” “dynamic and person-specific prediction,” “understandable artificial intelligence with human oversight,” “ethical governance, data security, and algorithmic justice,” and “clinical translation and cultural localization.” Participants emphasized that no single digital marker could reliably indicate psychological risk. Changes in mobility, sleep, circadian regularity, communication, smartphone interaction, language, voice, and physiological activity needed to be interpreted relative to the individual’s baseline and social context. The proposed system was therefore conceptualized as a probabilistic early-warning mechanism rather than an autonomous diagnostic instrument. Reliable psychological risk prediction requires multimodal longitudinal data, individualized temporal models, explicit uncertainty estimation, explainability, fairness auditing, dynamic consent, data minimization, and continuous professional oversight. The proposed framework may support early screening, prioritization of referrals, monitoring of symptom trajectories, and timely preventive interventions. Nevertheless, algorithmic outputs should be treated as decision-support information and should not replace clinical interviews, professional judgment, contextual assessment, or the therapeutic relationship.