Early prediction of PSA nadir in metastatic castration-sensitive prostate cancer with machine learning approaches.

Abstract

222 Background: Early achievement of PSA nadir is a well-established prognostic marker in mHSPC. Understanding PSA kinetics is crucial for optimizing treatment and disease management. This study aimed to develop and externally validate a machine learning model predicting the probability of achieving a PSA nadir ≤0.2 ng/mL at sixth month, based on baseline and 3th-month PSA measurements. Methods: This retrospective study included 211 patients with mHSPC (Training set:168 [80%], Test set:43 [20%]).The primary endpoint was achieving a PSA nadir ≤0.2 ng/mL at 6.-month after treatment initiation. Clinical and laboratory variables—including treatment type, Gleason score, metastatic volume and pattern, comorbidity status, age, baseline and 3.-month PSA, ALP, LDH, and hemoglobin—were used as predictors. Data preprocessing involved iterative imputation for missing values, variance thresholding, and feature scaling. Two predictive models, logistic regression and gradient boosting (GB), were developed using a nested cross-validation framework to prevent data leakage and overfitting. Hyperparameters were optimized via grid search, and model performance was assessed on an independent test set (n=43).Calibration, Brier score, bootstrap and decision curve analyses were conducted to evaluate clinical applicability. Model interpretability was examined through SHAP analysis, and external validation was performed on two independent cohorts. Results: In nested cross-validation, the GB model outperformed logistic regression in all metrics except recall and was selected as the final model. In the test set (n = 43), the GB model achieved an accuracy of 86.1%, ROC-AUC of 0.88, sensitivity of 82.6%, and specificity of 90.0%. SHAP analysis revealed that 3.-month PSA, ALP, LDH, and hemoglobin were the most influential features. The model demonstrated good calibration (Brier score = 0.12) and provided positive net clinical benefit between 20% and 78% risk thresholds. In the two center external validation cohort (n = 66), the model maintained strong predictive performance (ROC-AUC = 0.85, accuracy = 86.4%, sensitivity=75.0%, specificity=94.7%). Conclusions: Our GB based model accurately predicted 6.-month PSA nadir achievement in mHSPC and demonstrated robust generalizability. This model may assist clinicians in early identification of patients unlikely to achieve a PSA nadir, supporting timely treatment intensification and personalized disease management.