The study examined the use of complete blood count (CBC) indicators and machine learning algorithms to predict the risk of mental disorders. They enrolled 1,379 university volunteers in September 2024, collecting data on age, sex, and 22 CBC variables; the dependent variable was the binary result according to the SCL-90 scale, where there were 1,023 negative and 356 positive cases. They used SMOTEtomek hybrid sampling to address data imbalance and random forest to select features, which identified 14 optimal variables. They compared four models: XGBoost, AdaBoost, Random Forest, and GBDT, with XGBoost performing best with an AUC of 0.860 on the training set and 0.827 on the test set. SHAP analysis confirmed that basophil percentage (BASO %), basophil count (BASO#) and mean corpuscular hemoglobin (MCH) were the main contributors. A logistic regression nomogram confirmed these findings. Fifteen volunteers had missing data for four indicators.