Patient selection for proton therapy in lung cancer: Updating and internal-external cross-validation of a normal tissue complication probability model for acute oesophageal toxicity.

PURPOSE: To externally validate and update a published normal tissue complication probability (NTCP) model for the prediction of Grade ≥ 2 acute oesophageal toxicity (AOT) in lung cancer patients receiving (chemo-)radiotherapy (predictors: concurrent chemotherapy, gender, tumour stage, mean oesophagus dose (MOD)).

METHODS AND MATERIALS: Patient cohorts from four centres were available. Initial external validation was performed by assessing model performance. Heterogeneity of performance across cohorts was evaluated using an internal-external cross validation (IECV) method. Based on the IECV procedure, parameters were added or removed from the model to lower heterogeneity in model performance across cohorts with the aim to create a model that performed uniformly well across centres. Ridge regression was used to create the final model and bootstrapping techniques were applied to perform further internal validation.

RESULTS

Out of 937 patients, 444 (47%) had Grade ≥ 2 AOT. External validation indicated model recalibration was necessary for 1 cohort and model revision for another. IECV showed heterogeneity of performance of the original model across cohorts. To lower heterogeneity in model performance, the predictors concurrent chemotherapy, gender, and tumour stage were removed from the model, and overall treatment time (OTT) was added. The final model containing MOD and OTT as predictors showed low levels of heterogeneity in model performance (I2 of 0% for all performance measures) with optimism-corrected AUC of 0.75, calibration intercept of -0.01 and slope of 1.26.

CONCLUSION

The validation and updating of an NTCP-model for AOT resulted in a prediction model performing well in varying patient cohorts. Therefore, this updated model was added to the national indication protocol to select lung cancer patients for proton radiotherapy.