KEY POINTS
- Multicenter external validation study of 10 published prediction models for radiation-induced primary hypothyroidism, using 1,054 head and neck cancer patients treated with intensity modulated radiotherapy at four oncology centers.
- Median follow-up was 9.9 years; 534 of 1,054 patients developed primary hypothyroidism, with cumulative incidence of 35.9% at 5 years and 54.5% at 10 years.
- At 8 years, model accuracy ranged from 0.551 to 0.645, and area under the curve ranged from 0.577 to 0.748.
- All models showed positive calibration intercepts and calibration slopes <1, indicating underestimation at low predicted risks and overestimation at high predicted risks.
- Models incorporating thyroid mean dose and thyroid reserve markers, such as thyroid volume or pretreatment thyroid-stimulating hormone, performed best; the Huang and Wongwattananard models had the highest area under the curve values (0.708 and 0.748) and lowest Brier scores (0.224 and 0.214).
CLINICAL TAKEAWAY
This study supports cautious use of thyroid complication prediction models as planning-support tools, especially when they include thyroid mean dose and a measure of thyroid reserve. It does not support using current models to omit routine thyroid function surveillance, because long-term risk remains common and calibration was imperfect at both low and high predicted probabilities.
SOURCE
International Journal of Radiation Oncology, Biology, Physics