Interpretable machine learning for early neurological deterioration prediction in atrial fibrillation-related stroke

• Published in: Scientific reports 2021-10, Vol.11 (1), p.20610-20610, Article 20610
• Main Authors: Kim, Seong-Hwan, Jeon, Eun-Tae, Yu, Sungwook, Oh, Kyungmi, Kim, Chi Kyung, Song, Tae-Jin, Kim, Yong-Jae, Heo, Sung Hyuk, Park, Kwang-Yeol, Kim, Jeong-Min, Park, Jong-Ho, Choi, Jay Chol, Park, Man-Seok, Kim, Joon-Tae, Choi, Kang-Ho, Hwang, Yang Ha, Kim, Bum Joon, Chung, Jong-Won, Bang, Oh Young, Kim, Gyeongmoon, Seo, Woo-Keun, Jung, Jin-Man
• Format: Article
• Language: English
• Subjects: 631/378; 692/617; Algorithms; Atrial Fibrillation - epidemiology; Atrial Fibrillation - physiopathology; Cardiac arrhythmia; Fibrillation; Humanities and Social Sciences; Humans; Learning algorithms; Logistic Models; Machine Learning; Models, Theoretical; multidisciplinary; Myocardial Infarction - complications; Myocardial Infarction - epidemiology; Myocardial Infarction - physiopathology; Neurologic Examination - methods; Prediction models; Prognosis; Prospective Studies; Registries; Republic of Korea - epidemiology; Risk Assessment - methods; ROC Curve; Science; Science (multidisciplinary); Stroke; Support Vector Machine
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-021-99920-7

We aimed to develop a novel prediction model for early neurological deterioration (END) based on an interpretable machine learning (ML) algorithm for atrial fibrillation (AF)-related stroke and to evaluate the prediction accuracy and feature importance of ML models. Data from multicenter prospective stroke registries in South Korea were collected. After stepwise data preprocessing, we utilized logistic regression, support vector machine, extreme gradient boosting, light gradient boosting machine (LightGBM), and multilayer perceptron models. We used the Shapley additive explanation (SHAP) method to evaluate feature importance. Of the 3,213 stroke patients, the 2,363 who had arrived at the hospital within 24 h of symptom onset and had available information regarding END were included. Of these, 318 (13.5%) had END. The LightGBM model showed the highest area under the receiver operating characteristic curve (0.772; 95% confidence interval, 0.715–0.829). The feature importance analysis revealed that fasting glucose level and the National Institute of Health Stroke Scale score were the most influential factors. Among ML algorithms, the LightGBM model was particularly useful for predicting END, as it revealed new and diverse predictors. Additionally, the effects of the features on the predictive power of the model were individualized using the SHAP method.