Evaluation of machine learning methods to stroke mortality risk prediction

Background and aims: Stroke is the second leading cause of mortality in the world and the leadingcause of acquired disability in adults. Also, two-thirds of the stroke-induced burden occurs indeveloping countries. Identifying patients’ mortality risk at admission can contribute to valuableclinical care modifications by identifying high-risk patients with poor outcomes who require moreintensive resources. Nowadays, Artificial Intelligence (AI) and machine learning as a main fieldof it, is gaining much attention. Using machine learning, may improve diagnostic accuracy, speedup decision-making, and be very useful in prediction. The aim of this study was to apply variousmachine learning (ML) methods to predict in-hospital mortality risk among stroke patients.Method: This cross-sectional study used clinical and demographic data of ۱۲۸۱ hospitalizedstroke patients from February ۲۰۱۸ to March ۲۰۱۹ to develop a mortality risk assessment model.Data were preprocessed and cleaned, important features were selected, and features and caseswith more than ۲۰% missing values were removed. Sampling with replacement was done in orderto have equivalent groups. Data imputation was also done to replace the remaining missing values.Data were analyzed through ۵-fold cross-validation using models including Random Forest,Support Vector Machine, Decision Tree, Neural Network, Naïve Baysian, and XGBoost repeated۵ times. Their results were evaluated and compared and the best one was selected. Models weredeployed in R Studio software using packages including “randomForest”, “caret”, “e۱۰۷۱”, “neuralnet”,“naivebayes”, and “xgboost”, and evaluation metrics were recorded.Results: Twenty-eight out of thirty features for ۱۲۰۸ samples were extracted. After feature selection,nine features with the highest importance value were selected: NIHSS, CT result, treatmentreceived, Length of Stay (LOS), MRS, Age, Weight, Systolic Blood Pressure and Blood Glucose.The reported accuracy for the Random Forest, SVM, Decision Tree, Neural Network, NaïveBayes and XGBoost were ۰.۹۷۹, ۰.۹۲۵, ۰.۹۰۳۳, ۰.۵۰۰, ۰.۸۵۰ and ۰.۹۶۳, respectively. RandomForest model reported the best accuracy was ۰.۷۹ (۹۵% CI of ۰.۹۶۱۱, ۰.۹۹۱۳); and receiver operatingcharacteristic (ROC) was drawn.Conclusion: Among all five machine learning algorithms used in this study, random forest algorithmhad the better performance in predicting in-hospital mortality stroke patients, thus furtherresearch should be conducted on random forest algorithm. The prediction models could be usedfor early risk assessment of patients with stroke. Identification of patients at high risk for mortalityimmediately after admission has the potential of enabling early discharge planning.