Abstract
OBJECTIVE: The diagnosis of prostate cancer (PCa) relies on prostate-specific antigen (PSA) in blood, but the specificity of PSA remains inadequate. This study aims to develop and validate a highly efficient and accurate diagnostic model based on urinary exosomes combined with machine learning (ML) techniques, to identify PCa patients at an early stage and provide support for clinical decision-making. METHODS: This study included 287 patients from Shanghai Changhai Hospital, Shanghai Shibei Hospital, and Taizhou People's Hospital, consisting of 89 PCa patients and 198 benign prostatic hyperplasia (BPH) patients. Urinary exosomes were collected from these patients. LASSO regression was used to screen key variables, and nine ML algorithms (including XGBoost, random forest, and logistic regression) were employed to construct the diagnostic model. Model performance was evaluated using AUC, learning curves, calibration curves, and decision curve analysis (DCA), and the contributions of key predictors were visualized using the SHAP method. RESULTS: Among the 11 clinical features included, 3 key features were selected: u-PSA, u-PSMA, and u-AMACR. Among the nine algorithms, the GBDT model performed best, achieving an AUC of 1.000 in the training cohort and 0.987 in the validation cohort. SHAP analysis showed that u-PSA, u-PSMA, and u-AMACR were the most important predictors for PCa. The learning curves indicated that the model fit well and remained stable, while DCA demonstrated significant clinical net benefit, and the calibration curves indicated good diagnostic performance. CONCLUSION: The urinary exosome-based PCa diagnostic model demonstrates high diagnostic efficacy and can assist physicians in better identifying PCa, reducing unnecessary biopsies.