Predict disabling severe attacks in neuromyelitis optica spectrum disorders: warning symptoms and innovative machine learning models.

BACKGROUND: Severe disabling attacks neuromyelitis optica spectrum disorders (NMOSD) severely affect patients’ quality of daily life and life safety. METHODS: This retrospective study enrolled consecutive Chinese patients suffering from NMOSD who visited the China-Japan Friendship Hospital (Beijing, China) between October, 2010 and February, 2023. Correlation analysis was used to perform feature selection. The prediction model was constructed using the support vector machine (SVM) and extreme gradient boosting (XGBoost) algorithm. RESULTS: A total of 356 patients (mean [SD] age, 34.45[15.22] years) and 1291 NMOSD attacks were eligible for this study. ON and age were positively and linearly correlated, and TM and age were negatively and linearly correlated. Throbbing headache and neuralgia showed a significant linear relationship with disabling episodes, and circumventricular organ- area postrema syndrome (CVO-APS) showed a significant linear relationship with disabling episodes only in a few cases. We select top three high correlation variable (HCV) for predicting ON and TM models. Then, we constructed prediction models based on the XGBoost using age as a feature, and HCV and warning symptoms (WS) as features respectively. The ML models showed reasonable predictive performance for ON (Age + WS: AUC, 0.809; Age + HCV: AUC, 0.787) and TM (Age + WS : AUC, 0.817; Age + HCV: AUC, 0.854) We also constructed the first prediction model about severe NMOSD attacks using XGBoost and SVM. Among them, the fundamental model incorporated the 16 features with AUC of Xgboost: 0.830 and SVM: 0.775. On the basis of the fundamental model, we also incorporated the results of nadir and remission the expanded disability status scale (EDSS) or visual outcome scale (VOS) from the last time attack as features to construct 12 optimized models. As a whole, the optimized models showed higher predictive performance for severe attack. The AUC for the model adding the nadir EDSS scores as features were XGBoost: 0.862, SVM: 0.741, adding the nadir VOS scores as features were XGBoost: 0.870, SVM: 0.723, adding the remission EDSS scores as features were XGBoost: 0884, SVM: 0.806, adding the remission VOS scores as features were XGBoost: 0.9998, SVM: 0.700, adding both nadir and remission EDSS scores as features were XGBoost: 0.899, SVM: 0.794, adding both nadir and remission VOS scores as features were XGBoost: 0.905 and SVM: 0.705. CONCLUSION: This study innovatively identified associations between disabling attacks of NMOSD and age, warning symptoms and developed an easy-to-use, less costly and less invasive machine learning model for predicting NMOSD disabling attack symptoms and severe attacks.