Prediction of recovery grade in Bell's palsy using machine learning models

While most cases of Bell's palsy resolve successfully, some patients do not fully recover despite medical treatment. Various tools have been developed to assess the status of facial nerve palsy (FNP), enhance prognostic accuracy, and guide early rehabilitation or more intensive therapies. However, existing evaluation tools often struggle with accurately predicting moderate cases of FNP. Recently, machine learning algorithms have shown promise in predicting clinical outcomes across various diseases by analyzing historical medical data. This study aimed to establish machine learning models for predicting acceptable recovery and House-Brackmann (H-B) grades at each visit using Support Vector Machine (SVM) and Long Short-Term Memory (LSTM). By leveraging data from serial H-B grades for each clinic visit and other clinically significant features, we demonstrated that changes in FNP grades can be predicted with high accuracy, achieving 0.903 accuracy for forecasting acceptable recovery and a Mean Absolute Error (MAE) of approximately 0.46 per H-B grade for each visit. Given the substantial social and emotional impact of FNP, early intervention based on prognosis is crucial. Our machine learning model can aid clinicians in predicting outcomes and providing appropriate consultation and treatment for FNP patients.