Interpretable machine learning for severity classification of thyroid eye disease using orbital anatomical features

The study presents the development of an interpretable machine learning system for objective severity classification of thyroid eye disease (TED) using MRI measurements of orbital structures. Researchers analyzed 1,054 orbital MRI units from 443 patients and created two datasets - one with all available scans and another with only first-visit scans to reduce temporal bias. Six machine learning models were tested, with Random Forest using class weighting achieving the best performance with an AUC of 0.811. Feature importance analysis consistently showed that ocular protrusion was the most important predictor, followed by rectus muscle thickness and orbital geometric parameters. The results emphasize that controlling for longitudinal redundancy and inter-patient correlations significantly impacts model performance and generalizability.