Comprehensive Summary
Koutsouleris et al. explored whether psychosis transitions in young individuals with clinical high-risk (CHR) syndromes or recent-onset depression (ROD) could be better predicted by combining clinical, neurocognitive, neuroimaging, and genetic data with expert assessments using machine learning. The study followed 334 patients with CHR or ROD and 334 matched healthy controls across five European countries. Over a follow-up period of 9 to 36 months, 26 patients transitioned to psychosis. Models were tested using a range of data sources and validated on three external cohorts to ensure generalizability. The study compared clinicians’ predictions of psychosis risk with various machine learning models, including those focusing on clinical-neurocognitive data, polygenic risk scores (PRS) for schizophrenia, and structural MRI. Clinicians achieved a balanced accuracy of 73.2%, excelling in specificity (84.9%) but showing limited sensitivity (61.5%). Machine learning models showed higher sensitivity (76%-88%) but lower specificity (53.5%-66.8%). A combined “cybernetic” risk calculator integrating clinician and algorithm predictions improved balanced accuracy to 85.5%. The models decreased the false-negative rate from 38.5% to 15.4%.Additionally, the researchers designed a sequential workflow that reduced diagnostic burden while maintaining a balanced accuracy of 85.9%. External validation confirmed the generalizability of the models across different populations. The findings also identified significant predictors of psychosis, such as motor disturbances, childhood trauma, and reduced facial emotion recognition, as well as structural brain changes identified via MRI. The study addressed potential confounders, including image quality, treatment, and follow-up duration, and demonstrated that predictions were consistent regardless of these factors. The models were further optimized for clinical use by streamlining the number of variables required for predictions, reducing complexity while preserving accuracy.
Outcomes and Implications
The study demonstrates how combining machine learning with expert input could transform psychosis prediction and improve outcomes for young people at risk. By integrating multiple data sources and refining workflows, the approach enhances the precision of risk assessment, particularly in complex cases involving CHR and ROD. Early and accurate detection allows for timely interventions, potentially reducing the long-term impact of psychosis. The sequential workflow introduced in the study is particularly valuable for clinical practice, as it reduces diagnostic burden while maintaining high accuracy. This makes the approach scalable and accessible for widespread implementation. The successful external validation suggests that these methods could be adopted in various healthcare settings, paving the way for more personalized and effective mental health care. Future research should focus on large-scale deployment and real-world application to ensure this approach benefits broader populations.