Exploring clinical determinants and machine learning-based prediction of disease characteristics in Graves ophthalmopathy.

CLINICAL RELEVANCE: Improved risk stratification in Graves ophthalmopathy (GO) may support earlier recognition of severe disease patterns and more individualised clinical monitoring. BACKGROUND: GO exhibits heterogeneous clinical behaviour, and predicting disease severity and course remains challenging. This study aimed to evaluate clinical determinants of disease phenotype and severity in patients with established GO using conventional statistical analyses and machine learning (ML) approaches. METHODS: Medical records of 153 patients with GO were retrospectively reviewed. Demographic characteristics, clinical findings, thyroid function, autoantibody levels, smoking status and MRI-defined tissue phenotype were analysed. Patients were classified according to disease severity, activity, onset pattern, and tissue predominance. In addition to classical statistical tests, logistic regression, random forest, support vector classifier, and k-nearest neighbours algorithms were trained and evaluated using accuracy, F1-score, and AUC metrics. RESULTS: Male sex and active smoking were significantly associated with higher disease severity and activity. Muscle-predominant GO was associated with older age, male sex, smoking, higher severity, and higher activity. No significant differences were observed among clinical subgroups regarding thyroid hormone levels or autoantibody titres. Among ML models, tissue predominance showed the strongest predictive performance, followed by ophthalmopathy severity, whereas disease activity and ophthalmopathy onset showed lower discrimination. Explainability analyses identified sex, age, and smoking as the most relevant contributors across models. CONCLUSION: Demographic and clinical factors contributed more consistently to classification performance than thyroid related biochemical parameters. ML models provided variable but clinically interpretable discrimination across different dimensions of GO and may support future risk stratification approaches.