KEY POINTS
- The model was developed from the prospective FFCD-ANABASE cohort of 1,015 patients treated at 60 French centres. Median age was 65 years, 75% were women, 43% had T1–T2N0 disease, and 57% had T3–T4 or node-positive disease.
- All patients received radiotherapy, with a median primary-tumour dose of 60 Gy. Concurrent chemotherapy was given to 77%, most commonly a mitomycin-based regimen, which accounted for 88% of chemotherapy-treated patients.
- Data were separated before feature selection into 606 training/validation patients, 200 calibration patients, and an untouched 209-patient test set. The final XGBoost accelerated-failure-time model used 14 clinical, biological, and treatment variables and retained censored patients through survival-specific methods.
- In the independent test subset, the concordance index was 0.735 (95% CI 0.65–0.82) and the three-year time-dependent AUC was 0.755 (0.66–0.85). The calibrated three-year Brier score was 0.135, suggesting reasonable agreement between predicted and observed risk.
- At the prespecified 37% predicted-risk threshold, sensitivity was 64%, specificity 74%, positive predictive value 39%, negative predictive value 89%, and overall accuracy 72%. The high negative predictive value suggests greater potential for identifying lower-risk than truly high-risk patients.
- Three-year local relapse was 11% in the AI-defined low-risk group (n=137) versus 39% in the high-risk group (n=72; log-rank p=2.78×10⁻⁵). Decision-curve analysis showed greater modelled net benefit than treating everyone, treating no one, or using T3–T4/node-positive status alone around the selected threshold.
- SHAP analysis ranked WHO performance status, tumour size, and age as the most influential inputs. Poorer performance status and larger tumours increased predicted risk; older age appeared protective, potentially because the endpoint was local relapse rather than a composite incorporating competing mortality.
- Testing used a holdout sample from the same national cohort rather than a genuinely independent population. The positive predictive value remained modest, the model has not demonstrated that risk-guided treatment changes improve outcomes, and external validation and regulatory assessment are required.
CLINICAL TAKEAWAY
The model may eventually help identify patients with a sufficiently low relapse risk for de-escalation studies or less intensive surveillance. It is not ready to determine dose reduction or treatment intensification in routine care: the clinically decisive next step is external validation followed by prospective testing of model-guided decisions.