Abstract
BACKGROUND: Prostate cancer is one of the most prevalent malignant tumors of the male genitourinary system. The occurrence of metastasis significantly influences treatment strategies and prognosis. However, current risk assessments for metastatic disease primarily rely on single imaging or pathological indicators, which are often limited by suboptimal accuracy and considerable individual variability. OBJECTIVE: This study aimed to develop a high-performance predictive model for prostate cancer metastasis by integrating semiquantitative parameters from [(18)F]PSMA-1007 PET/CTwith key clinicopathological features. METHODS: We retrospectively analyzed data from prostate cancer patients, includingPSMA PET/CT-derived features (SUVmax, SUVmean, PSMA-TVp, TL-PSMAp) and clinical-pathological variables (age, tPSA, Gleason score). Five machine learningalgorithms-Logistic Regression, Support Vector Machine, Random Forest, Naive Bayes, and XGBoost-were evaluated for metastasis prediction performance. Model performance was assessed using accuracy, sensitivity, precision, and area under the ROC curve (AUC). Shapley additive explanations (SHAP) were applied to interpret the most effective model. RESULTS: Among the five algorithms, the XGBoost model achieved an accuracy of 90.32%, sensitivity of 90.0%, specificity of 94.74%, and an AUC of 0.8977. SHAP analysis identified PSMA-TVp, TL-PSMAp as the most important predictors, followed by SUVmax, tPSA, and Gleason score. These findings highlight the key role of PSMA-derived tumor burden in metastasis prediction. Force plots further revealed the individual-level contributions of features, supporting the model's clinical interpretability. CONCLUSION: The XGBoost-based multimodal model integrating PET/CT semiquantitative parameters with clinicopathological data demonstrated excellent accuracy and interpretability in predicting prostate cancer metastasis. This approach has strong potential for clinical application and may provide a valuable tool for personalized treatment decision-making.