Scalable Biomarkers for Psychiatry via Digital Phenotyping, Sensorimotor Modeling, and Neuroimaging

Abstract

Psychiatric illnesses present a formidable challenge due to their heterogeneity, subjective diagnosis, and limited objective biomarkers. As the field moves toward precision psychiatry, there is a pressing need to identify scalable, mechanistically grounded, and behaviorally relevant biomarkers that can be deployed across diverse populations. This dissertation addresses that need by integrating methods from digital phenotyping, sensorimotor modeling, and neuroimaging to uncover new avenues for scalable psychiatric biomarker development. In Chapter 1, we address clinical heterogeneity in schizophrenia through digital phenotyping using smartphone-derived ecological momentary assessments (EMA) and passive sensor data. By employing a rigorous data imputation and clustering pipeline across an international cohort, we identify three data-driven subtypes that show distinct symptomatology and functional profiles. These clusters correspond to different symptom domains (e.g., affective vs. non-affective presentations) and reveal meaningful clinical correlations that may inform personalized interventions. Chapter 2 focuses on scalable sleep estimation using smartphone sensors, tackling the limitations of traditional actigraphy and self-report methods. We introduce a Bayesian hidden Markov model that estimates sleep states using screen-state and accelerometer data alone, enabling continuous sleep tracking without the need for wearables. We validate the model using simulated datasets and self-reported EMA surveys from multiple cohorts, highlighting its robustness and potential utility in psychiatric contexts, particularly where sleep variability is a key feature. In Chapter 3, we explore the cognitive architectures underlying motor learning through the lens of explicit and implicit adaptation. Using carefully designed behavioral experiments, we disentangle the feedback-dependent dynamics of reward-based and error-based learning. Our findings show that reward feedback engages explicit strategies and indirectly drives implicit learning through a serial pathway, whereas error feedback drives both systems in parallel. These architectures provide a novel mechanistic framework for understanding motor learning deficits in psychiatric populations. A meta-analysis of schizophrenia studies reveals impairments in both implicit recalibration and explicit strategy use, suggesting dysfunction in cerebellar and prefrontal systems and highlighting motor adaptation as a candidate biomarker. Finally, Chapter 4 investigates the impact of chronic adolescent THC exposure on large-scale brain network connectivity in nonhuman primates. Using resting-state functional MRI and dual regression analysis, we uncover non-linear, dose-dependent effects on the default mode, salience, and central executive networks. These alterations, especially within the central executive network, persist even after THC discontinuation and implicate adolescent exposure in long-lasting network reconfiguration. These findings underscore the utility of translational neuroimaging models in probing the neurodevelopmental origins of psychiatric vulnerability. Together, this dissertation offers a multifaceted yet integrative approach to psychiatric biomarker discovery. By leveraging scalable technologies and biologically grounded frameworks, we advance the potential for objective, continuous, and individualized assessment tools in psychiatry. The combined insights from digital phenotyping, sleep modeling, motor learning, and functional connectivity provide a robust foundation for future biomarker-based interventions and personalized mental health care.