Validated prediction of xerostomia in a real-world population: a step toward model-guided radiotherapy.

Abstract

Background and purpose: The aim of this study is to validate an Normal Tissue Complication Probability (NTCP) model for xerostomia in a large quality-registry cohort, enabling its future use in individualized NTCP-based treatment planning.

Material and methods: A model predicting grade ≥ 2 xerostomia (6 months post-radiotherapy) was selected for validation, including the mean dose to both the parotid and the submandibular glands, in addition to the baseline score for xerostomia, as predictors. Our local validation cohort consisted of 674 patients (204 events), treated between 2012 and 2024, with a median follow-up of 10.3 months (range 5-24). A closed testing procedure was performed to investigate the need for model updating, and the performance of the models was assessed with calibration curves, discrimination, the Brier score, and the Hosmer-Lemeshow test.

Results: The calibration curve demonstrated that the model predicted the dose-response relationship well. The validation cohort showed a slightly stronger dose response, with a slope of 1.16. The calibration intercept of -0.12 revealed an overestimation of xerostomia. However, the closed testing procedure indicated that a recalibration of the model was needed, and the HL-test showed a significant deviation. The recalibrated model showed perfect calibration but still limited discrimination (Area Under the Curve (AUC) 0.62).

Conclusion: The validated model performed well in our real-life dataset despite the differences between the training and validation cohorts, particularly considering the lack of baseline score in our cohort. This highlights the potential for improved performance with baseline inclusion but still suggests that an individualized NTCP-based treatment-planning protocol can be developed using the recalibrated published model.