Explainable machine learning for PSMA PET/CT-based clinically significant prostate cancer diagnosis: A dual-center retrospective study
Yinzhao Wang, Yiquan Yang, Yongxiang Tang, Lin Qi, Minfeng Chen, Shuo Hu, Xiaomei Gao, Yu Gan, Shouzhen Chen, Yi CaiBackground
Clinically significant prostate cancer (csPCa) detection is critical to guide treatment while minimizing unnecessary invasive procedures. Noninvasive tools to assist clinical risk stratification require thorough exploration.
Objectives
To develop and validate an explainable machine learning model integrating PSMA PET/CT quantitative features and serum biomarkers for predicting csPCa, and to explore its potential for clinical decision-making and biopsy decision-making.
Design
A dual-center retrospective diagnostic study utilizing patient cohorts consecutively enrolled from prior registered prospective clinical trials.
Methods
We retrospectively analyzed a development cohort (n=340), a temporally independent internal validation cohort (n=93), and an external validation cohort (n=126). Nine machine-learning algorithms were compared and the best model selected. Diagnostic performance was assessed by AUC, sensitivity and specificity; a prespecified PSA gray-zone subgroup (4-20 ng/mL; n=197) was analyzed separately. Clinical decision rate was defined as the proportion of patients classified as low or high risk under predefined high thresholds (PPV or NPV ≥97.5%). SHAP was used for model interpretability.
Results
Among all candidate models, XGBoost demonstrated the best performance. The model achieved AUCs of 0.96 (95% CI: 0.94–0.98) in the training cohort, 0.96 (95% CI: 0.93–1.00) in the testing cohort, 0.90 (95% CI: 0.83–0.96) in the temporally independent internal validation cohort, and 0.93 (95% CI: 0.88–0.97) in the external validation cohort. In the PSA gray-zone subgroup, the model maintained high diagnostic discrimination, with corresponding AUCs of 0.94 (95% CI: 0.90–0.98), 0.96 (95% CI: 0.91–1.00), 0.88 (95% CI: 0.80–0.96), and 0.87 (95% CI: 0.79–0.96), outperforming individual imaging and serological indicators. Under predefined threshold, the XGBoost supported potential clinical decision rate was 51.6% to 62.1% of patients across cohorts while maintaining reliable predictive performance.
Conclusions
An XGBoost model combining PSMA PET/CT metrics and serum indices showed superior diagnostic accuracy and expanded clinical decision performance compared with conventional indicators, including in the PSA gray zone. These findings are encouraging and hypothesis-generating; however, rigorous prospective validation remains mandatory before any clinical translation or therapeutic application.