KEY POINTS
- This Dutch multicenter study externally validated and updated a published normal tissue complication probability (NTCP) model for grade ≥2 acute esophageal toxicity in lung cancer patients receiving curative-intent radiotherapy. Data from 1,155 patients across four institutions were available, including 937 patients with complete information for validation of the original model and 1,071 patients for development of the final updated model, of whom 498 (47%) developed grade ≥2 acute esophageal toxicity.
- The original model included mean esophageal dose (MOD), concurrent chemotherapy, sex, and clinical T stage. Initial external validation revealed substantial differences in discrimination and calibration across institutional cohorts, with I² values of 72–94% during internal-external cross-validation. Differences in radiotherapy schedules, chemotherapy regimens and toxicity assessment contributed to concerns about transporting the original prediction model between centers.
- The model was simplified by removing concurrent chemotherapy, sex and T stage, while adding overall treatment time (OTT). The final ridge-regression model therefore uses only mean esophageal dose and overall treatment time, with the fitted linear predictor S = −3.634 + 1.496 × ln(MOD) − 0.0297 × OTT, where MOD is expressed in Gy and OTT in days; predicted complication probability is calculated as 1 / (1 + e^(−S)).
- Following updating, model discrimination reached an AUC of 0.75 (95% CI 0.72–0.77), with calibration-in-the-large of −0.01 (95% CI −0.14 to 0.12) and calibration slope of 1.26 (95% CI 1.06–1.46). The slope indicates some overestimation at lower predicted risks and underestimation at higher risks, despite satisfactory average calibration.
- The updated model demonstrated much more consistent performance across participating cohorts, with I² reduced to 0% for discrimination and calibration measures in the internal-external validation procedure. Bootstrap assessment found minimal apparent optimism, but the estimated probability of meeting all predefined good-performance criteria was only 53.1%, underscoring that improved model transportability does not eliminate individual prediction uncertainty.
- The updated model was incorporated into the Dutch national indication protocol for proton therapy in lung cancer. Under this model-based selection approach, proton treatment may be indicated when the predicted absolute reduction in grade ≥2 acute esophageal toxicity is at least 10 percentage points compared with an individual patient's photon plan. This threshold guides modality selection using predicted risks rather than directly observed comparative toxicity outcomes.
- Important limitations include the use of different toxicity grading systems (RTOG versus CTCAE), differences in fractionation and dose normalization across centers, and the absence of prospective patient-reported swallowing outcomes. The model was evaluated using Dutch radiotherapy cohorts, while clinical toxicity reduction after model-selected proton treatment was not directly tested; independent validation in additional populations and prospective proton-treated cohorts remains necessary.
CLINICAL TAKEAWAY
The updated Dutch NTCP model provides a relatively simple, clinically applicable approach to estimating acute esophageal toxicity from mean esophageal dose and overall treatment time, supporting individualized photon-versus-proton planning comparisons. Its inclusion in national proton selection criteria is a meaningful implementation step, but predicted risk reductions must not be interpreted as demonstrated clinical benefit, and continued prospective validation remains essential.