KEY POINTS
- This retrospective multicenter study included 1,407 patients with head and neck cancer treated with definitive radiotherapy across five centers, organized into four cohorts. Model development used 407 patients from two Hong Kong hospitals, with external validation in three independent cohorts of 372, 274, and 354 patients. Radiation-induced hypothyroidism occurred in 694 patients (49.3%), with center-specific event rates ranging from 39.8% to 65.5%.
- Hypothyroidism was defined as the first post-treatment TSH measurement exceeding the local laboratory upper limit (4.2–4.9 µIU/mL), rather than symptomatic hypothyroidism or initiation of thyroid replacement therapy. Among patients developing hypothyroidism, mean time to detection was 32.9 ± 26.2 months, while mean follow-up across the cohort was 40.4 ± 30.2 months. The time-to-event approach was intended to account for the cumulative nature of this late toxicity.
- The final model incorporated 10 predictors: three pretreatment CT radiomics features, four three-dimensional dose-derived dosiomics features, mean thyroid dose, age, and N stage. A Cox proportional hazards framework generated individual cumulative risk estimates at multiple time horizons, rather than a single binary prediction. Thyroid volume was consistently smaller among patients developing hypothyroidism across all four cohorts (all p<0.001).
- Model discrimination varied across institutions. The development cohort achieved a C-index of 0.714 and overall AUC of 0.765, while the three external cohorts achieved C-index/AUC pairs of 0.725/0.800, 0.680/0.754, and 0.627/0.690. The equal-weighted mean external AUC was 0.748, modestly exceeding the strongest evaluated conventional NTCP models, which achieved external mean AUCs of approximately 0.72–0.73.
- Combining predicted three-year and seven-year hypothyroidism risks separated patients into four temporal risk categories. Observed event proportions increased from 21.7% in the low-risk group (83/383) to 46.1% in the intermediate-risk group (147/319), 58.3% in the elevated-risk group (182/312), and 71.8% in the high-risk group (282/393). Hypothyroidism-free survival differed significantly between groups (log-rank p<0.001), although these observed proportions are not fixed-horizon cumulative incidence estimates.
- External calibration revealed an important limitation. In the Xijing Hospital cohort, the model systematically underestimated absolute hypothyroidism risk, and center-specific recalibration reduced the three-year Brier score from 0.465 to 0.196. The cohort differed in age, thyroid volume, and follow-up characteristics; moreover, all its patients had nasopharyngeal carcinoma. These findings emphasize that preserved risk ranking does not guarantee accurate individual risk estimates across institutions.
- The investigators developed an open-source research tool combining planning CT, three-dimensional dose distributions, thyroid segmentation, and clinical information to generate personalized risk curves and thyroid dose summaries. An exploratory simulation suggested a 7.17-percentage-point reduction in model-estimated hypothyroidism risk under one hypothetical thyroid-sparing scenario, but no deliverable treatment plans were generated or validated. The model was not prospectively tested as clinical decision support, and its predominantly Asian, nasopharyngeal cancer-heavy population limits generalizability to other head and neck disease sites.
CLINICAL TAKEAWAY
Integrating thyroid radiomics, spatial dosimetry, and clinical factors may improve individualized prediction of radiation-induced hypothyroidism beyond conventional dose-volume models, potentially supporting future risk-adapted surveillance and treatment planning. However, the incremental predictive gain was modest, absolute-risk calibration differed substantially between centers, and prospective validation is needed before the tool can guide thyroid-sparing decisions or follow-up schedules.