External Validation of 8 Normal Tissue Complication Probability Models for Radiation-Associated Primary Hypothyroidism Using Long-Term Thyroid Function Outcomes

Abstract

Primary hypothyroidism is a common late complication following neck irradiation. The reported incidences were as high as 20-50% among head and neck cancer survivors. Several normal tissue complication probability models for radiation-associated primary hypothyroidism have been proposed, but none has been externally validated using long-term thyroid function outcome. We undertook a systematic review of the literature and identified 8 distinct normal tissue complication probability (NTCP) models for radiation-associated primary hypothyroidism (6 logistic regression models and 2 nomograms). We retrospectively contoured and reviewed the thyroid gland dosimetry of all consecutive head and neck cancer patients who underwent definitive radiotherapy between January 2013 and December 2015 in two tertiary oncology centers. Patients with abnormal pre-radiotherapy thyroid function were excluded. Thyroid function was monitored at least yearly after radiotherapy. Primary hypothyroidism was defined as serum thyroid-stimulating hormone levels above the upper limit of normal. Euthyroid controls were defined as patients who had normal thyroid function at least 5 years beyond the completion of radiotherapy. Diagnostic properties (sensitivity, specificity and accuracy) of the 8 NTCP models were determined. Discrimination and calibration of the models were assessed using c-statistics, discrimination slopes, calibration curves and Brier score. With a median follow-up of 7.0 years, we identified 216 cases of post-radiation primary hypothyroidism and 120 euthyroid controls within the study period. All four logistic regression models (Boomsma 2012, Rønjom 2013, Shen 2021 and Huang 2021) which were based on mean thyroid dose and thyroid volume demonstrated satisfactory diagnostic performances. The accuracy, c-statistic, discrimination slope and Brier score of these four models were > 0.70 (0.711 - 0.732), > 0.75 (0.763 - 0.765), > 0.20 (0.203 - 0.295) and < 0.20 (0.187 - 0.198), respectively. Calibrations of prediction were excellent at ranges of high predicted probabilities. At ranges of low predicted probabilities, these models underestimated the risk of primary hypothyroidism by 10-20%. Other models or nomograms which utilized alternative radiation dose-volume parameters (e.g., V30, V35 or V50), or without factoring in the variation in thyroid volume, had significantly worse performance in terms of both discrimination and calibration. This large external validation study showed that NTCP models that included thyroid volume and mean thyroid dose had satisfactory performances in predicting long-term thyroid function outcomes following neck irradiation. Refinement is required to optimize calibration in patients deemed at low risk of post-radiation primary hypothyroidism.