JMIR Publications

ABSTRACT

Background:

Stroke is a common cardiovascular disease in neurology. Stroke is mainly caused by cerebral artery thrombosis, which is harmful to human body. At present, the diagnosis of stroke is mainly dependent on imaging detection and neuropsychological score. These methods are not highly sensitive to early brain injury and subjective. Compared with traditional statistics, machine learning in artificial intelligence system can predict diseases more accurately and be applied to the diagnosis and treatment of neurology. At present, there are few researches on machine learning and stroke diagnosis.

Objective:

The study aimed to explore the predictive value of artificial intelligence system for common neurological stroke diseases, and to provide reference for the application of artificial intelligence system in the field of medical neurology.

Methods:

Univariate and multivariate Cox and Logistic regression model algorithms were used to predict the recurrence of stroke, and the related factors were analyzed. The receiver operating characteristic (ROC) curve was used to detect the accuracy and sensitivity of Cox and logistic models. According to the Support Vector Machines (SVM) algorithm in machine learning, the relevant prediction model of stroke recurrence was established. The mean value method, median method, linear regression method, and normalized EM were used to fill and preprocess the data. The influencing factors were selected by conservative means method and the risk factors of stroke recurrence were predicted by SVM model. The prediction results of stroke risk factors by Cox model, Logistic regression model, and SVM model were compared.

Results:

Cox model was used to analyze the risk factors of stroke recurrence. Finally, it was concluded that family history of stroke, systolic blood pressure, total cholesterol, disease station construction, and history of hypertension were the main risk factors of stroke recurrence. The Area Under Curve (AUC) value of the ROC Curve of Cox model was 0.913, the prognostic index was 2.314, the sensitivity of Cox model was 0.754, and the specificity was 0.805. In the multivariate logistic model, four factors including history of hypertension, history of heart disease, family history of stroke, and dietary habits can be used as the individual risk factors for stroke recurrence. The AUC value of the ROC curve of the logistic model was 0.889> and 0.7. Among the four different data filling algorithms, after data filling, the ROC curve of the median method for the prediction of stroke recurrence had the highest AUC, which was 0.874. The conservative means method was used to select the influencing factors. The top 10 risk factors predicted by the SVM model for stroke patients were age, maximum inflation level, systolic blood pressure, total drug therapy, hypertension, general health, abnormal electrocardiogram, ESRS score, history of diabetes, and creatinine. The prediction results of three different models for stroke risk factors were compared, which was 0.873SVM>0.861Logistic>0.853Cox.

Conclusions:

In the risk analysis model of recurrent stroke, the SVM model has advantages over the Logistic and Cox models.