287 - Personalizing PSA Follow-Up after Prostate Radiation Therapy: Developing a Patient Specific AI Powered Predictive Model for Clinically Meaningful Recurrence

Purpose/Objective(s): Phoenix Criteria remains the standard definition of biochemical recurrence in prostate cancer (PC) following definitive RT. However, with modern PSMA PET imaging and novel biomarkers, this criteria may delay detection of clinically significant recurrence. We developed a machine learning model using patient factors and PSA kinetics to predict recurrence after definitive RT.

Materials/Methods: Patients status-post definitive RT for PC with a rising PSA >1 ng/dl absolute value were included. Patients were excluded if they had <3 post-RT PSAs, incomplete data, or insufficient follow-up. Clinical, treatment, cancer outcomes, and longitudinal PSAs were retrieved. Primary endpoint was clinical meaningful recurrence, defined as radiographic or biopsy-confirmed progression or any progression requiring treatment. An XGBoost model was trained on a per-PSA approach predicting the primary endpoint within a 5-year window. An 80/20 split was used for training and 5-fold cross-validation for hyperparameter tuning. The probability threshold (0.79) for binary classification was prespecified on cross-validation folds to target =90% specificity. Evaluation of model performance was also completed on a per-patient basis with a patient classified as positive if any PSA-level prediction exceeded the prespecified threshold, compared against Phoenix Criteria as reference standard. Performance was assessed using overall and time-binned AUC with 95% bootstrapped confidence intervals (1,000 iterations), calibration plots, and lead time analysis. To allow for proper estimations, patients with >24 months gap between PSAs were excluded from testing set (n=51).

Results: Of 10,480 PC patients with available PSA data, 1,684 met selection criteria (1,347 training, 286 testing), with 11,530 evaluable PSAs (mean 7 PSAs/patient). The XGBoost model achieved an overall AUC of 0.879 on the per-PSA test set. At the prespecified threshold, sensitivity was 90.7% and specificity 91.1%, compared to 77.1% and 97.6% respectively, for Phoenix criteria. The model achieved an overall AUC of 0.975 on the per-patient test set. Performance generally improved as time from treatment increased, with a time-binned AUC of 0.85 for <1 year to 0.92 for >5 years from RT completion. The XGBoost model resulted in a lower PSA at prediction of recurrence than the recalculated Phoenix criteria (median 1.7 vs 3.1 ng/dL) and greater lead time (median 0.55 vs 0.21 years and mean 1.00 vs 0.41 years, respectively) for predicting clinically meaningful recurrence. SHAP analysis identified PSA velocity, risk category, and most recent PSA value as dominant predictive features.

Conclusion: The model demonstrated promising performance for personalized dynamic recurrence risk assessment after RT in PC patients, providing earlier detection and longer lead team. Due to the retrospective design, further validation is needed.