The Impact of NOACS versus VKAS on Absolute and Relative Cognitive Function Decline over Time: A Machine Learning Approach

Abstract

Background Atrial fibrillation (AF) is the most common arrhythmia in older adults and is associated with an increased risk of cognitive impairment and dementia, even in patients without prior stroke. Nonvitamin K antagonist oral anticoagulants (NOACs) offer a better safety profile than vitamin K antagonists (VKAs), but their cognitive benefit remains uncertain. Aim To assess the impact of NOACs versus VKAs on cognitive decline in elderly AF patients using a machine learning approach. Methods This multicenter prospective cohort study included 983 AF outpatients enrolled between 2008 and 2022 at the Geriatrics Department, University of Catanzaro, and the ProMISE Department, University of Palermo. Stroke and bleeding risks were assessed using CHA2DS2-VASc and HAS-BLED scores. Cognitive function was evaluated using the Mini-Mental State Examination (MMSE). Cognitive decline was defined as a decrease in MMSE score between baseline and follow-up. Patients with prior anticoagulant therapy (OAT), severe dementia, or comorbidities affecting cognition were excluded. Multivariable logistic regression and a random forest classifier were used to assess whether anticoagulant type independently predicted cognitive decline. Class imbalance was addressed using both class-weighted learning and the synthetic minority over-sampling technique (SMOTE), with model performance evaluated through repeated stratified cross-validation and threshold optimization. Results At baseline, cognitive performance was comparable between groups (p=0.11). After a mean follow-up of 7.2 +/- 3.4 years, MMSE scores declined significantly more in VKA-treated patients (-1.7 vs.-0.3 points, p < 0.001). In logistic regression, NOAC use was independently associated with a lower risk of cognitive decline (odds ratio: 0.322; 95% confidence interval: 0.221-0.469; p < 0.0001). The random forest classifier achieved a mean cross-validated AUC of 0.8719 (standard deviation: 0.0273) and a test-set AUC of 0.880. Threshold adjustment and SMOTE improved sensitivity (recall increased: 0.47-0.84), with a precision-recall AUC of 0.763. Permutation importance analysis identified "OAT" as the top predictor. Predicted probabilities of cognitive decline were significantly higher in VKA users (median = 0.70) than in NOAC users (median = 0.09), confirmed by a Kolmogorov-Smirnov test (KS =0.385, p < 0.001). Conclusion NOAC use is associated with a lower predicted probability of cognitive decline, suggesting potential cognitive benefits over VKAs.