Clinical Report: Interpretable Machine Learning Approaches for Classifying Severity in Thyroid Eye Disease
Overview
This study presents an interpretable machine learning framework for assessing the severity of thyroid eye disease (TED) using orbital anatomical characteristics. The Random Forest model demonstrated the highest performance metrics.
Background
Thyroid eye disease (TED) is a prevalent orbital condition linked to autoimmune thyroid disease, characterized by significant clinical manifestations and potential vision loss. Current severity assessments are subjective and variable.
Data Highlights
Model
AUC
Recall
F1-score
Specificity
Random Forest (class weighting)
0.811
0.669
0.648
0.815
Key Findings
Random Forest with class weighting achieved the highest AUC of 0.811.
Random Forest with SMOTE achieved the highest recall (0.669), F1-score (0.648), and specificity (0.815).
Feature importance analysis ranked ocular protrusion as the top predictor of TED severity.
Controlling for longitudinal scan correlations is crucial for accurate model evaluation.
Dataset B, which included only first-visit scans, reduced visit-related confounding.
Clinical Implications
The findings indicate that machine learning models can enhance the objectivity of TED severity assessments by integrating anatomical MRI features.
Conclusion
The study highlights the use of interpretable machine learning frameworks in assessing thyroid eye disease severity.
An artificial intelligence–based optical coherence tomography pathway met noninferiority criteria for false-positive diabetic macular edema referrals and was associated with fewer referral decisions in a randomized clinical trial.
