Exploratory machine-learning evaluation of clinical and dosimetric features in late prostate radiotherapy toxicity

Prediction of late radiotherapy toxicity remains challenging, as routinely available clinical and dosimetric variables may incompletely capture individual radiosensitivity. This study evaluated interpretable machine-learning models for predicting baseline- corrected late gastrointestinal (GI) and genitourinary (GU) toxicity following prostate radiotherapy and examined model-derived feature contributions.

Data from 229 men with prostate cancer enrolled in the REQUITE cohort were retrospectively analysed. Late toxicity was defined as an increase of ≥1 CTCAE grade from baseline occurring ≥12 months following radiotherapy. Prespecified clinical and dosimetric predictors were evaluated separately for GI and GU toxicity, with correlated dose-volume parameters represented using organ-specific equivalent uniform dose (EUD). L1-regularised logistic regression, Elastic Net and XGBoost were evaluated using repeated nested stratified cross-validation. Performance was assessed using out-of-fold predictions, bootstrap confidence intervals and outcome-label permutation testing. SHAP analysis assessed feature contributions and stability across repeats.

GI and GU toxicity worsening occurred in 107 (46.7%) and 125 (54.6%) patients, respectively. GI ROC-AUCs were 0.573, 0.571 and 0.587 for L1, Elastic Net and XGBoost, respectively, with permutation testing indicating above-chance discrimination (p = 0.036, 0.047 and 0.022). Rectal EUD was the most consistently influential GI predictor. GU discrimination was below chance, with no evidence of discrimination beyond the null distribution. GU feature importance was more heterogeneous and model dependent.

Routine clinical and dosimetric variables provided limited predictive information for late GI toxicity but did not reliably discriminate GU toxicity. The findings support investigation of additional dosimetric and biological markers for toxicity prediction.