A machine learning–based case-control classification model for first-episode schizophrenia using routine blood biochemical markers

To construct and compare the predictive performance of different machine learning algorithms for distinguishing first-episode schizophrenia patients from healthy controls using routine peripheral blood biochemical markers, with the aim of developing an internally validated case-control classification model with interpretable presentation. This single-center retrospective case-control study enrolled 316 first-episode schizophrenia patients diagnosed according to ICD-10 criteria at Shaoxing Seventh People’s Hospital between June and December 2025, and selected 474 healthy controls who underwent routine health check-ups during the same period. All participants were Chinese individuals from the same regional medical setting. Demographic characteristics and fasting peripheral blood biochemical parameters were extracted from the hospital information system. The complete dataset had no missing values for the included variables and was randomly split into a training set ( n = 553) and a validation set ( n = 237) in a 7:3 ratio. Feature screening and hyperparameter tuning were performed only within the training set. Univariate logistic regression with Benjamini-Hochberg false-discovery-rate correction, L1-regularized logistic regression, and recursive feature elimination (RFE) were used for feature selection. Eight machine learning algorithms were trained, and validation-set performance was evaluated using AUC, sensitivity, specificity, accuracy, precision, and F1 score. Following feature selection, high-density lipoprotein (HDL), total bilirubin (T-Bil), blood glucose (GLU), albumin-to-globulin ratio (A/G), total cholesterol (TC), and blood urea nitrogen (BUN) were identified as predictors for distinguishing schizophrenia patients from healthy controls. In the validation set, the Gradient Boosting Machine (GBM) algorithm demonstrated the highest AUC (0.885). However, considering the training-validation performance gap, potential overfitting in more complex models, and the need for model transparency, the multivariate logistic regression model was selected as the final case-control classification model. A nomogram and online probability calculator were constructed to visualize the logistic regression equation. This study integrates machine learning algorithms with routine peripheral blood biochemical markers to establish an internally validated case-control classification model for differentiating first-episode schizophrenia patients from healthy controls. The final logistic regression model showed good discrimination and calibration in internal validation, but external validation and clinically realistic comparator groups are required before clinical diagnostic use can be inferred.