Development of a machine learning-based mortality prediction model for patients with mental disorders and COVID-19

Introduction Patients with mental disorders are at increased risk of adverse outcomes from COVID-19, but prognostic evidence specific to this population remains limited. This study aimed to develop and validate machine-learning models for predicting 31-day mortality among hospitalized patients with mental disorders and laboratory-confirmed COVID-19. Methods Data were retrospectively collected from 439 hospitalized patients across 10 hospitals in Henan Province, China. Patients were randomly divided into a training cohort (n = 308) and an independent test cohort (n = 131). Oversampling was applied during model development to address class imbalance. Candidate predictors were selected using LASSO, Boruta, and random forest methods, and eight machine-learning algorithms were trained. SHAP analysis was used for model interpretation, and Kaplan–Meier analysis compared survival between model-defined risk groups. Results Patients were generally older, 63.1% were female, and comorbidities were common. Several models showed good discrimination in the training cohort, although some showed overfitting. In the test cohort, the neural network model with LASSO-selected features performed best, with an AUC of 0.911 (95% CI: 0.832–0.990). SHAP analysis identified concomitant hormone therapy, alkaline phosphatase, and lymphocyte count as the leading predictors. The high-risk group had significantly higher cumulative mortality than the low-risk group (log-rank P < 0.0001). Discussion A machine-learning model based on routine clinical and laboratory variables may support short-term mortality risk stratification in this regional multicenter cohort.