Application of an Interpretable Machine Learning for Estimating Severity of Graves' Orbitopathy Based on Initial...

Area of Science:

• Ophthalmology
• Endocrinology
• Medical Informatics

Background:

• Graves' orbitopathy (GO) presents with varying severity and phenotypes.
• Accurate prediction of GO severity and type is crucial for timely intervention.
• Initial ophthalmic findings are key indicators for disease progression.

Purpose of the Study:

• To construct and validate risk prediction scores for moderate-to-severe and muscle-predominant Graves' orbitopathy.
• To identify initial clinical and ophthalmic factors predictive of GO severity.
• To establish a risk stratification tool for Korean patients with GO.

Main Methods:

• Machine learning-based automatic clinical score generation algorithm applied.
• Development of the Score for Moderate-to-Severe type of GO risk Prediction (SMSGOP).
• Development of the Score for Muscle-predominant type of GO risk Prediction (SMGOP).
• Utilized data from 400 GO patients with at least 6 months follow-up.

Main Results:

• 44.8% of patients had moderate-to-severe GO, with 12.5% being muscle-predominant.
• SMSGOP incorporates age, diplopia, thyroid stimulating immunoglobulin, NOSPECS, CAS, and muscle area ratio.
• SMGOP includes age, diplopia, eye deviation, FT4, and GD-GO interval.
• Predictive thresholds identified as ≥46 for SMSGOP and ≥49 for SMGOP.

Conclusions:

• This study introduces the first risk prediction scores for GO severity and type based on initial findings in a Korean population.
• The developed scores (SMSGOP and SMGOP) offer valuable tools for early risk assessment.
• Initial ophthalmic and clinical data are significant predictors for GO outcomes.