Abstract Schizophrenia constitutes a chronic and disabling illness. While patients show high rates of response to treatment after a first-episode of schizophrenia, the long-term course of the illness is typically characterized by frequent re- lapses, persistence of symptoms, and enduring cognitive and functional deficits. Despite the prioritization of relapse prevention as a treatment goal, about four out of five patients experience a relapse within the first five years of treatment. Relapses are known to have serious psychosocial, educational, or vocational implications in young adults—a population at high risk of psychosis. However, current psychiatric ability to recognize indicators of relapse in order to prevent escalation of psychotic symptoms is markedly limited. Challenges stem from a lack of availability of comprehensive information about early warning signs, and reliance on fixed time point sampling of cross-sectional data as well as patient or family reported observations, that is subject to recall bias, or on clin- ician sought information, that needs frequent and timely contact. The present proposal seeks to address these gaps in early psychosis treatment, by leveraging patient-generated and patient-volunteered social media data, and developing and validating machine learning approaches for “digital phenotyping” and relapse prediction. Our proposed work is founded on the observation that social media sites have emerged as prominent platforms of emotional and linguistic expression—young adults are among the heaviest users of social media. The work signif- icantly advances the research agenda and extensive pilot investigations of the team, who a) have demonstrated that social media data of individuals can serve as a powerful “lens” toward understanding and inferring mental health state, illness course, and likelihood of relapse, including among young adults with early psychosis; and b) have been involved in examining the role of emergent technologies, like social media, in improving access to and delivery of psychiatric care. Aim 1 will provide theoretically-grounded and clinically meaningful methods for extracting and modeling digital phenotypes and symptoms from social media data of young adult early psychosis patients. Then in Aim 2, we will develop and evaluate machine learning methods that will utilize the extracted social media digital phenotypes to infer patient-specific personalized risk of relapse, and identify its antecedents. Finally, Aim 3 will develop a two-faceted validation framework, to assess the statistical and clinical efficacy and utility of the social media derived inferences of psychosis and relapse in influencing clinical outcomes and in facilitating evidence-based treatment. To accomplish these aims, the project brings together a strong multidisci- plinary team, combining expertise in social media analytics, psychiatry, psychology, natural language analysis, machine learning, information privacy, and research ethics. Our ...