LAUSR

Background: The present study was performed to evaluate the brain regional characteristics related with development of post-stroke delirium using the machine learning and statistical analysis. Method: We used clinical and radiological data prospectively collected from 675 acute ischemic stroke patients, who were admitted in stroke unit from August 2017 to July 2018. Delirium occurrence in the patients was screened with Confusion Assessment Method and finally diagnosed using the criteria of the Diagnostic and Statistical Manual of Mental Disorders (fifth edition). Three machine learning models, Support Vector Machine (SVM), Random Forest (RF) and Tree-based Gradient Boosting (XGBoost), were applied for the prediction of post-stroke delirium with the clinical and radiologic data. And logistic regression analysis was performed to evaluate the significance of the brain regional parameters included in the importance features which were obtained from the XGBoost result. Results: Post-stroke delirium occurred in 66 (9.8%) of the total patients. On the comparison of the test accuracy to predict delirium occurrence, RF (94%), XGBoost test (93%), and SVM (89%) showed similar prediction rates. Of the brain regional parameters included in the top 30 feature importance, right side cerebral hemisphere, non-lacunar infarction, severity of periventricular white matter changes, acute temporal lobe lesion, cerebellum, brain stem, and previous lesions developed on right side cerebral hemisphere, and in temporal or frontal lobe. Conclusion: The present study shows that the brain regional characteristics related with the post-stroke delirium are shown to be significant when controlling the other features using statistical analysis with machine learning. Even though we need more studies to validate the relationships between post-stroke delirium and brain regional characteristics, the present brain regional characteristics could provide significant evidences to predict post-stroke delirium for the acute ischemic stroke patients.