KEY POINTS
- This retrospective analysis used data from 229 patients with prostate cancer enrolled in the prospective multicenter REQUITE cohort. Investigators evaluated new or worsening late gastrointestinal (GI) and genitourinary (GU) toxicity occurring at least 12 months after radiotherapy, defining toxicity as an increase of ≥1 CTCAE grade relative to baseline, rather than using a fixed grade ≥2 threshold. Late GI deterioration occurred in 107 patients (46.7%), and GU deterioration in 126 (55.0%).
- Three machine learning approaches were compared: L1-regularized logistic regression, Elastic Net regression and XGBoost. Candidate predictors included age, BMI, PTV volume, diabetes, previous abdominal surgery, pelvic nodal irradiation, prior urological procedures, and rectal or bladder equivalent uniform dose (EUD). The use of EUD reduced redundancy between correlated dose-volume metrics while retaining a whole-organ summary of dose exposure.
- Model development used five-fold nested stratified cross-validation repeated five times, with hyperparameter optimization confined to training data. Performance was evaluated using out-of-fold predictions, while uncertainty was assessed through 2,000 patient-level bootstrap resamples and 1,000 outcome-label permutation tests. SHAP analyses were performed on held-out predictions to examine feature contributions without using the same observations for model fitting and explanation.
- Predictive performance for late GI toxicity was limited, with mean ROC-AUC values of 0.573 for L1 regression, 0.571 for Elastic Net, and 0.587 for XGBoost (95% CI 0.520–0.653). The best-performing model achieved sensitivity of 0.658 and specificity of only 0.487, indicating that its ability to identify individual patients developing toxicity was insufficient for reliable clinical risk stratification despite modest above-chance discrimination.
- Late GU toxicity prediction performed even worse, with ROC-AUC values of 0.465, 0.455 and 0.466 for L1 regression, Elastic Net and XGBoost, respectively. Permutation testing found no significant evidence of discrimination beyond chance (p=0.150–0.412). The below-0.5 AUC values should therefore not be interpreted as demonstrating a reproducible inverse association that could simply be reversed to obtain a clinically useful predictor.
- SHAP analysis consistently ranked rectal EUD as the strongest contributor to GI toxicity predictions across all three algorithms and validation repeats, with previous abdominal surgery also contributing. For GU toxicity, bladder EUD, PTV volume, BMI and age showed variable importance depending on the model. These rankings describe how the models generated predictions; they do not establish independent causal effects or validate the predictive usefulness of individual parameters.
- Important limitations include the relatively small 229-patient cohort, absence of independent external validation, reliance on planned rather than accumulated delivered dose, and use of whole-organ EUD rather than spatially resolved or substructure dosimetry. The ≥1-grade worsening endpoint also includes relatively mild symptoms, and no molecular or genomic radiosensitivity markers were available. Thus, the study demonstrates the inadequacy of the evaluated feature set rather than proving that machine learning cannot predict late radiation toxicity.
CLINICAL TAKEAWAY
Routine clinical characteristics and whole-organ dose metrics were insufficient for accurate individualized prediction of late GI and GU toxicity after prostate radiotherapy, despite rigorous internal validation and interpretable modeling. Rectal EUD remained a reproducible contributor to GI risk estimates, but future models will likely require more detailed anatomical dosimetry, accumulated treatment dose and patient-specific biological information before clinical implementation.