Background: Recently, it has been reported that stroke registry data analyzed by machine learning (ML) can predict patients’ outcomes with high accuracy. However, there have been few reports that examined large-scale acute stroke registry data using ML, and there is no system that can adapt the results to individual cases.
Aims: The purpose of this study is developing outcome prediction formulas using ML analysis of the Japan Stroke Data Bank (JSDB), which is a nationwide acute stroke registry with individual data and have accumulated approximately 260,000 cases from more than 100 facilities nationwide over the past 23 years.
Methods: The derivation cohort used data of acute stroke patients registered in JSDB from 2016 to 2019. ML algorithm was used to develop a formula for predicting stroke outcomes at discharge. The results were compared with manual-made formulas using a logistic regression (LR) model and validated using JSDB data in 2020. Favourable functional outcome was defined as mRS 0-2.
Results: ML algorithm was able to predict favourable functional outcome with an area under curve (AUC) 0.907 [95%CI 0.903-0.913], sensitivity 81.5%, specificity 84.0%, and positive predictive value (PPV) 83.9%. Each statistical value was significantly higher than that calculated using LR model (AUC 0.868 [0.862-0.875], sensitivity 78.1%, specificity 79.9%, and PPV 79.9%). In validation, each value was comparable with that of derivation. Premorbid mRS, Japan coma scale (JCS), NIHSS total score, NIHSS subs-core of lower limbs, stroke classification, and age were the top six important factors for ML algorithm.
Conclusion: ML algorithm was able to predict favourable functional outcome in acute stroke patients at discharge with higher accuracy than LR model. Premorbid mRS, stroke classification, JCS, NIHSS total score, NIHSS sub-score of lower limbs at admission, and age were important factors for ML method.