Prediction of Psychosis in Children and Adolescents: a Machine Learning Approach

Aims: Because community and clinical studies indicated an impact of development on the early detection of psychosis, we studied the development of psychosis over two years in 8- to 17-year-olds. Methods: Within a Swiss-German naturalistic longitudinal study, we examined clinical predictors of conversion to psychosis within 2 years in 69 clinical high-risk patients (35% male), 147 inpatients not suspected to develop psychosis (47% male) and 110 community subjects (46% male) using a multistep machine-learning approach. Participants were examined with the Schizophrenia Proneness Instrument-Child & Youth version (SPI-CY), the Structured Interview for Psychosis-Risk Syndromes (SIPS), the Structured Clinical Interview for DSM-IV and a neurocognitive battery. Additionally, an external validation was performed with a 13- to 17-year-old CHR sample of an independent prospective naturalistic Swiss study (N=80, 14 converters) that used the same diagnostic instruments. Results: Thirteen participants converted to psychosis in the model developing sample, 12 of them had met CHR criteria at baseline. Fifty-one significant predictors were identified, predominately SPI-CY and SIPS items. Using these predictors only, BAC increased to 88.5% (sensitivity: 84.6%, specificity: 92.3%). External validation in terms of an out of sample cross validation revealed a sufficient generalizability of BAC = 63.1% determined by a high sensitivity (92.9%) and a low specificity (33.3%). Conclusions: The high number of predictors, compared to similar analyses in adult samples, and low specificity in the older validation sample indicate that prediction of psychosis in children and adolescents might be more complex than in adults, likely for the ongoing development