Abstract
OBJECTIVE: To develop a machine learning-based model incorporating prostate-specific antigen (PSA) levels and prostate ultrasound video clips for diagnosing prostate cancer. METHODS: The study enrolled 928 participants, of whom 429 had prostate cancer and 499 other non-prostate cancers. Univariate and multivariate analyses of serological indices were conducted to detect significant variables. From this cohort, 742 patients were randomly chosen for model validation, while the other 186 were employed to evaluate the accuracy and reliability of the model. Seven features were extracted from ultrasound video clips and combined with PSA and other clinical indicators. Predictive models were established using six machine learning algorithms and receiver operating characteristic (ROC) curves were used to determine the optimal model. SHapley Additive exPlanations (SHAP) was utilized to visualize feature importance in the best-performing model. RESULTS: All six of the evaluated machine learning models performed favorably, with area under the ROC curve (AUC) values in the test set ranging from 0.800 to 0.881. Of these models, the XGBoost model achieved the most promising performance, significantly surpassing that of the other models (P < 0.05). SHAP visualization revealed that PSA, prostatic volume(PV), age, wavelet.LHL.firstorder. Median, wavelet.HLH.glszm.ZoneEntropy, and original.shape.MinorAxisLength were the most influential features in the XGBoost model. CONCLUSION: The developed machine learning models demonstrated significant potential for prostate cancer diagnosis. Among them, the XGBoost model outperformed the others, highlighting its superior predictive capability.