Automated Diagnosis of Blepharoptosis Using Frontalis Electromyography-Based Machine Learning

Background: Blepharoptosis is characterized by abnormal drooping of the upper eyelid. Conventional diagnostic approaches often rely on subjective assessments, which may limit diagnostic accuracy. Therefore, this study developed a machine learning-based diagnostic model for blepharoptosis using frontalis electromyography (EMG). Methods: The study included 24 patients with blepharoptosis (marginal reflex distance 1 ≤2 mm) and 19 healthy controls (marginal reflex distance 1 ≥ 4 mm). Frontalis EMG signals were recorded during visual tasks at various gaze angles. Time-, frequency-, and time–frequency domain features were extracted and analyzed using Radial Basis Function Support Vector Machine (SVM), Linear SVM, and Logistic Regression algorithms. Models were validated using stratified group fivefold cross-validation. Feature selection and Shapley Additive Explanations (SHAP) analysis were performed to identify informative features for classification. Results: Using EMG features extracted from the 10–20-s primary gaze interval, the Linear SVM model achieved the highest area under the curve of 0.9187. SHAP analysis identified Wavelet Entropy as the most influential factor. Lower Wavelet Entropy values tended to shift predictions toward the patient group, whereas higher values shifted predictions toward the healthy control group. Root mean square and mean absolute value were the next most influential features, whereas Wavelet Energy and Variance contributed less to model predictions. Conclusion: Frontalis EMG-based machine learning models demonstrated potential for distinguishing patients with blepharoptosis from healthy controls. Linear SVM model exhibited the highest discriminative performance, while SHAP analysis identified Wavelet Entropy, root mean square, and mean absolute value as the most influential features contributing to model prediction.