Designing AI Models for Predicting Ischemic Stroke Using Administrative Healthcare Database

Abstract

Ischemic stroke (IS) is the most common type of stroke, accounting to about 80% of the stroke, and is a major cause of morbidity and mortality worldwide. The risk factors of IS included older age, male gender, high BMI, smoking habit, hypertension, diabetes, hyperlipidemia, etc. Other factors, such as environmental metal exposure and atrial cardiopathy are also found to be risk factors. In this study, a balanced dataset, consisting of healthcare data of 36,880 IS patients and 36,880 non-IS patients matched with IS patients using indexed date, age, and sex, retrieved from a subset of the National Health Insurance Research Database (NHIRD) of Taiwan, an administrative healthcare database, were adopted for developing the AI models to predict events of ISs. Integrated Genetic Algorithm and Support Vector Machine (IGS) algorithm accompanied with 3 different fitness functions was applied for designing the predictive models. To select the best predictive performance from different AI models, tenfold cross validation were conducted for model training. The predictive performance of the designed models exhibits that predictive accuracy, sensitivity, specificity, and area under ROC curve (AUC) achieve 73.38-73.96%, 73.31-73.91%, 73.03-74.02%, and 0.808-0.813, respectively. The selected features including age, comorbidities, and other comorbidity-related variables are shown to be effective in designing strong predictive models (AUC>0.8) for predicting patients who are more likely to develop IS in the near future. Future works will focus on designing the predictive models using more effective AI technique, such as deep neural network (DNN), and useful variables, such as administrated drugs, to enhance the predictive performance.