Development and validation of a machine learning model of radiation-induced hypothyroidism with clinical and dose-volume features.

BACKGROUND AND PURPOSE

Radiation-induced hypothyroidism (RIHT) is a common but underestimated late effect in head and neck cancers. However, no consensus exists regarding risk prediction or dose constraints in RIHT. We aimed to develop a machine learning model for the accurate risk prediction of RIHT based on clinical and dose-volume features and to evaluate its performance internally and externally.

MATERIALS AND METHODS

We retrospectively searched two institutions for patients aged >20 years treated with definitive radiotherapy for nasopharyngeal or oropharyngeal cancer, and extracted their clinical information and dose-volume features. One was designated the developmental cohort, the other as the external validation cohort. We compared the performances of machine learning models with those of published normal tissue complication probability (NTCP) models.

RESULTS

The developmental and external validation cohorts consisted of 378 and 49 patients, respectively. The estimated cumulative incidence rates of grade ≥1 hypothyroidism were 53.5% and 61.3% in the developmental and external validation cohorts, respectively. Machine learning models outperformed traditional NTCP models by having lower Brier scores at every time point and a lower integrated Brier score, while demonstrating a comparable calibration index and mean area under the curve. Even simplified machine learning models using only thyroid features performed better than did traditional NTCP algorithms. The machine learning models showed consistent performance between folds. The performance in a previously unseen external validation cohort was comparable to that of the cross-validation.

CONCLUSIONS

Our model outperformed traditional NTCP models, with additional capabilities of predicting the RIHT risk at individual time points. A simplified model using only thyroid dose-volume features still outperforms traditional NTCP models and can be incorporated into future treatment planning systems for biological optimization.