Predicting early clinical recovery in first-episode psychosis: development and external validation of a clinically interpretable multivariable model.

Abstract

Identifying patients with first-episode psychosis (FEP) who are unlikely to achieve early clinical recovery (ECR) is critical for personalised intervention and resource allocation. ECR - defined as the concurrent achievement of symptomatic and functional remission - represents a clinically meaningful outcome that captures both illness control and functional reintegration. To develop and externally validate prediction models for ECR using clinical, cognitive and genetic data. We analysed two large, independent Spanish cohorts: the primeros episodios psicóticos cohort (<i>N</i> = 335), for model development and internal validation, and the Programa Asistencial a las Fases Iniciales de Psicosis cohort (<i>N</i> = 668), for external validation. Forty-seven baseline clinical and cognitive variables and 87 polygenic risk scores (PRSs) were examined. Predictors were selected using penalised logistic regression. Logistic regression and three machine learning algorithms were compared for discrimination, calibration and clinical utility. The best-performing model was a logistic regression using six routinely collected clinical and cognitive predictors (duration of untreated psychosis, days of treated psychosis, baseline functioning, insight, executive function and cognitive reserve), with an optimism-corrected area under the receiver operating characteristic curve of 0.73 in development and 0.63 in external validation. PRS models showed limited external generalisability and did not improve prediction. Machine learning algorithms offered no advantage over regression models. A simple, interpretable logistic regression model based on routine clinical and cognitive variables can predict early recovery in FEP with acceptable generalisability. These findings support the use of transparent, clinically grounded models in early psychosis care and highlight the current limitations of genetic predictors for individualised treatment.