Multicenter external validation of normal tissue complication probability models for radiation-induced primary hypothyroidism in head and neck cancer survivors with long-term endocrine outcomes.

Abstract

Numerous normal tissue complication probability (NTCP) models have been developed to predict radiation-induced primary hypothyroidism, but their external validity and clinical applicability remain unclear. This multicenter study aimed to externally validate and compare published NTCP models using long-term endocrine outcomes and to assess their predictive performance and clinical utility. A comprehensive systematic literature search was carried out to identify published NTCP models and nomograms for hypothyroidism prediction. External validation was performed using a multicenter cohort of 1,054 HNC patients treated with intensity-modulated radiotherapy at four oncology centers. Primary hypothyroidism was defined as an elevated serum thyroid-stimulating hormone (TSH) level exceeding the upper limit of normal. Clinical data and radiation dose-volume parameters of the thyroid gland were extracted. Cumulative incidence of hypothyroidism was estimated using competing risk analysis with death as a competing event, applying inverse probability of censoring weighting to account for informative loss to follow-up. Model predictive performance was compared using multiple metrics, including accuracy, discrimination, calibration curves, and Brier score. Decision curve analyses were conducted to evaluate clinical utility. Ten candidate NTCP models or nomograms were identified for external validation. Model performance varied, with accuracy ranging from 0.551 to 0.645 and AUC from 0.577 to 0.748 at 8 years of follow-up. Calibration slopes were <1 and calibration intercepts were positive across all models. Models incorporating thyroid mean dose (Dmean) and thyroid function reserve metrics, such as thyroid volume or baseline TSH level, demonstrated superior calibration and discrimination. For models with satisfactory performance, decision curve analysis indicated net benefits over treat-all or treat-none strategies across threshold probabilities between 0.30 and 0.90. This study represents the largest external validation of prediction models for radiation-induced hypothyroidism to date. Clinicians should be aware of potential miscalibration of long-term hypothyroidism risk by current models, particularly at the extreme ends of predicted probabilities. Models that incorporated thyroid Dmean and surrogates of thyroid reserve demonstrated satisfactory performances and showed promise for guiding personalized radiotherapy planning.