Use of machine learning techniques for phenotyping ischemic stroke instead of the rule-based methods: A nationwide population-based study (Preprint)

Abstract

In South Korea, many studies have evaluated stroke using claims data; most of these studies have defined ischemic stroke by using an operational definition following the rule-based method, assigning the I63 diagnosis code for admission, and incorporating the imaging or drug claims data. Rule-based methods tend to overestimate the number of patients with ischemic stroke.

We aimed to identify an appropriate algorithm for phenotyping stroke by applying machine learning (ML) techniques to analyze the claims data.

We obtained the data from the Korean National Health Insurance Service database, which is linked to the National Health Insurance Ilsan Hospital database (n=30,897). The ML techniques were constructed based on the results of a previous chart review (gold standard), and the demographic variables such as qualifications, examination/screening, medical utilization, and total medical costs were used as features. The performance of prediction models (extreme gradient boosting [XGBoost] or long short-term memory [LSTM]) was evaluated using the area under the receiver operating characteristic curve (AUROC), the area under precision-recall curve (AUPRC), and calibration curve.

In total, 30,897 patients were enrolled in this study, 3,145 of whom (10.18%) had ischemic stroke. XGBoost, a tree-based ML technique, had the AUROC was 93.63% and AUPRC was 64.05%. LSTM, such as neural network based deep learning methods, showed results similar to those of the rule-based method. The F1 score was 70.01%, while the AUROC was 97.10% and AUPRC was 85.70%, which was the highest.

We proposed recurrent neural network based deep learning techniques to improve stroke phenotyping. We anticipate the ability to produce rapid and accurate results by utilizing the National Health Information Database.