PREDICTING POST-STROKE COGNITIVE IMPAIRMENT WITH MACHINE LEARNING: A REVIEW OF NEUROIMAGING, CLINICAL, AND BLOOD BIOMARKER APPROACHES

Abstract

Cognitive decline following stroke — ranging from subtle attentional deficits to frank vascular dementia — affects between 30 and 40 percent of survivors and is a major but frequently underrecognised determinant of long-term functional outcome. Identifying which patients will develop significant post-stroke cognitive impairment (PSCI) at the time of the acute admission would allow targeted cognitive rehabilitation, intensified vascular risk factor management and timely family counselling. Machine learning models that integrate acute neuroimaging parameters, clinical characteristics and blood biomarkers have been proposed as a means of generating such early risk estimates. A review of 27 relevant studies published between 2017 and 2024 finds that multimodal ML models achieve AUC values of 0.82–0.93 for six-month PSCI prediction, substantially exceeding the 0.68–0.76 range associated with NIHSS-based scoring. White matter lesion burden, infarct location and serum neurofilament light chain concentration at 48 hours emerged as the features contributing most consistently to model performance. These findings have direct relevance for stroke services in Uzbekistan, where the burden of stroke-attributable cognitive impairment is high and dedicated cognitive rehabilitation infrastructure is limited.