Toward Generalizable and Clinically Useful Prediction of Occupational Functioning in Schizophrenia Using a Structure-Informed Representation of Cognition

Predicting real-world functioning remains a clinical priority in schizophrenia (SCZ), but existing prediction models face methodological limitations and lack established clinical utility. Cognition is a commonly used predictor and how it is represented may affect predictive performance. The Normative Latent Cognitive Structure (N-LCS) approach provides a structure- informed representation that may address limitations of conventional domain-level scores. Data from the COBRE cohort (163 SCZ, 180 healthy controls [HC]) were used to develop and internally validate ridge regression models for occupational (OF) and social (SF) functioning using N-LCS deviation metrics alongside a priori selected demographic and clinical predictors within a fully nested, optimism-corrected bootstrap framework. The OF model achieved an optimism-corrected AUC of 0.73 and balanced accuracy of 0.71, with near-ideal calibration and net benefit across nearly the full range of threshold probabilities (0-0.99) on decision curve analysis. The SF model showed modest performance (C-index = 0.66, weighted kappa = 0.27). A reduced OF model was externally validated in an independent cohort (NUSDAST; 166 SCZ, 156 HC), maintaining discrimination (AUC = 0.70) and showing improved balanced accuracy (0.56 to 0.69) after recalibration. Compared with the conventional score-based model, the N-LCS OF model achieved comparable predictive performance with fewer predictors and better calibration. Overall, these findings support the potential of the OF model as a generalizable tool for predicting occupational functioning in SCZ.