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

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Thyroid eye disease (TED) severity assessment is subjective and varies between observers.
• Machine learning (ML) models often use non-interpretable radiomic features, underutilizing objective MRI anatomical measurements.
• Longitudinal scans can inflate ML model performance due to patient-specific correlations and data redundancy.

Purpose of the Study:

• Develop an interpretable ML framework for objective TED severity stratification (mild, moderate-to-severe, sight-threatening).
• Quantitatively integrate orbital anatomical parameters with clinical assessment criteria.
• Evaluate how data handling strategies impact model generalizability.

Main Methods:

• Retrospective analysis of 1,054 orbital MRI scans from 443 TED patients.
• Two datasets: Dataset A (all scans) and Dataset B (first-visit scans only) to reduce bias.
• Trained and evaluated six ML models (LR, SVM, KNN, RF, XGBoost, LightGBM) using cross-validation, comparing AUC, F1-score, and recall.

Main Results:

• Random Forest with class weighting achieved the highest AUC (0.811) under class-imbalance strategies.
• Random Forest with SMOTE yielded the highest recall (0.669), F1-score (0.648), and specificity (0.815).
• Ocular protrusion, rectus muscle thicknesses, and orbital geometry were key predictors.

Conclusions:

• Controlling for longitudinal data redundancy and intra-patient correlations is crucial for model evaluation and generalizability.
• Random Forest with class weighting showed the best performance on temporally deduplicated first-visit scans.
• The framework integrates measurable anatomical parameters for predictions, emphasizing standardized quantification and workflow reproducibility in medical AI.