A retrospective study developed a diagnostic model for first-episode schizophrenia using machine learning algorithms based on routine blood tests and demographic data. 180 first-episode schizophrenia patients (January–August 2024) and 214 healthy controls participated. Data on age, gender and blood parameters were divided into a training set (70%, n=275) and a validation set (30%, n=119). Significant predictors were identified by univariate logistic regression (p<0.1), Boruta and LASSO: Arg, TP, ALP, HDL, UA and LDL. Random Forest performed best with an AUC of 1.00 on the training set and 0.877 on the validation set. Finally, a multivariate logistic regression model was chosen due to the risk of overlearning and a nomogram was created. They validated the model on an external set and a differential diagnosis set with good performance. This tool can serve as a diagnostic aid for schizophrenia.