Predicting cognitive decline in cognitively impaired patients with ischemic stroke with high risk of cerebral hemorrhage: a machine learning approach

Eun Namgung¹Young Sun Kim²Sun U Kwon³Dong-Wha Kang^3*

• ¹Asan Institute for Life Sciences, Seoul, Seoul, Republic of Korea
• ²Nunaps Inc., Seoul, Republic of Korea
• ³Department of Neurology, Asan Medical Center, College of Medicine, University of Ulsan, SONGPA-GU, Republic of Korea

Background and objective: Cognitive decline progresses rapidly in stroke patients, increasing risks of stroke recurrence. Predicting deterioration within a year in patients with poststroke cognitive impairment (PSCI) could guide targeted interventions for dementia prevention and better prognosis. In this PreventIon of CArdiovascular events in iSchemic Stroke patients with high risk of cerebral hemOrrhage for reducing cognitive decline substudy, machine learning on clinical and imaging data was used to predict cognitive decline over 9 months in PSCI patients.Methods: This retrospective study included 109 patients with acute ischemic stroke and high-risk cerebral hemorrhage with PSCI (baseline Korean-Mini Mental Status Examination [K-MMSE] < 24), along with baseline clinical imaging and K-MMSE assessments at baseline and after 9 months. Four machine learning algorithms were trained, Categorical Boosting (CatBoost), Adaptive Boosting (AdaBoost), eXtreme Gradient Boosting (XGBoost), and logistic regression, to predict cognitive decliners, defined as a decline of ≥3 K-MMSE points over 9 months, and ranked variable importance using the SHapley Additive exPlanations methodology.Results: CatBoost outperformed the other models in classifying cognitive decliners within 9 months. In the test set, CatBoost achieved a mean area under the curve (AUC) of 0.897, with an accuracy of 0.873; other models performed as follows: logistic regression (AUC 0.775), AdaBoost (AUC 0.767), and XGBoost (AUC 0.722). Higher baseline K-MMSE scores (total, language, orientation to place, and recall), longer interval between stroke and baseline K-MMSE, initial National Institutes of Health Stroke Scale scores, and lesion volume ratio were identified as key predictors of cognitive decline in CatBoost. Cognitive decliners showed longer interval between stroke onset and pharmacotherapy initiation than nondecliners.Conclusion: CatBoost effectively recognized patients with ischemic stroke at high risk of cognitive decline over 9 months. Recognizing these high-risk individuals and their risk and protective factors allows for timely and targeted interventions to improve prognosis in PSCI patients.