Machine learning helps predict early onset psychosis with serum protein biomarkers, neuropsychometry, and clinicodemographic data.

Early-onset psychosis presents diagnostic challenges due to overlapping clinical presentations and complex comorbidities, typically requiring specialized tertiary care with extensive neuroimaging, neuropsychometric testing, and multidisciplinary evaluation. This case-control study investigated whether machine learning could integrate multiple diagnostic modalities to create an objective diagnostic framework for early-onset psychosis. We recruited 45 patients with early-onset psychosis and 34 healthy controls from a tertiary referral centre. Participants underwent comprehensive assessment including serum protein biomarker analysis (brain-derived neurotrophic factor, proBDNF, p75 neurotrophin receptor, S100B), neuropsychometric testing (Iowa Gambling Task, Simple Response Time, Zabor Verbal Task), and demographic evaluation. Four machine learning algorithms (logistic regression, support vector machine, random forest, XGBoost) were trained on five feature combinations using nested cross-validation with hyperparameter optimization. XGBoost demonstrated superior performance, achieving optimal classification with the complete multimodal dataset (accuracy: 0.91 ± 0.08, precision: 0.92 ± 0.08, area under curve: 0.97 ± 0.04). Feature importance analysis revealed cognitive measures, particularly Zabor Verbal Task errors and response time parameters, as most discriminative, with brain-derived neurotrophic factor pathway components showing highest biomarker importance. Machine learning effectively integrated neuropsychometric and protein biomarker data for high-accuracy early-onset psychosis classification, with multimodal approaches outperforming single-domain assessments.