Prediction of in-hospital mortality in patients with cardioembolic stroke and atrial fibrillation using machine learning methods

Cardioembolic stroke (CES) is a common and severe variant of ischemic stroke. Cardioembolic stroke (CES) is a common and severe form of ischemic stroke, the main risk factor for which is atrial fibrillation. Among deceased patients with acute cerebrovascular accident, 40 % die in the first 48 hours of inpatient treatment. This demonstrates the importance of the in-hospital mortality (IHO) indicator and the search for its predictors with the construction of prognostic models using machine learning methods. The aim of the study was to develop prognostic models of IHO in patients with CES and AF based on modern ML methods. Material and methods. The study included 259 patients with CES and AF who were admitted to the regional vascular center No. 1 of Novosibirsk from November 2022 to December 2023. Logistic regression, random forest, stochastic gradient boosting and categorical boosting were used to develop prognostic models. Results and discussion. During the study, statistically significant differences were found between those who died and those who were discharged in terms of age, the presence of nosocomial and communityacquired pneumonia, chronic kidney disease, the number of comorbidities, as well as the values of scales and indices upon admission (the sum of points on the NIHSS (The National Institutes of Health Stroke Scale), Rankin scale, RRS (rehabilitation routing scale), and the Rivermead index. The prognostic significance of these parameters was assessed using univariate logistic regression. Prognostic models of IHO were developed based on machine learning methods. The models based on NIHHS and RRS had the greatest predictive potential, which was shown in a comparative analysis of the ROC curves of the developed models. The contribution of the predictors of these models to the implementation of the end point was estimated using the Shapley method. Conclusions. In our study, the highest prognostic value for CES and AF was demonstrated by the NIHSS, Rankin scale, and RRS scores, the Rivermead index at admission, the presence of nosocomial and community-acquired pneumonia, chronic kidney disease, and age. The resulting models, when widely implemented, will enable personalized management of the risk of IHO in patients with CES and AF.