Objective: Cognitive impairment is a detrimental complication of stroke that compromises the quality of life of the patients and poses a huge burden on society. Due to the lack of effective early prediction tools in clinical practice, many researchers have introduced machine learning (ML) into the prediction of post-stroke cognitive impairment (PSCI). However, the mathematical models for ML are diverse, and their accuracy remains highly contentious. Therefore, this study aimed to examine the efficiency of ML in the prediction of PSCI.Methods: Relevant articles were retrieved from Cochrane, Embase, PubMed, and Web of Science from the inception of each database to 5 December 2022. Study quality was evaluated by PROBAST, and c-index, sensitivity, specificity, and overall accuracy of the prediction models were meta-analyzed.Results: A total of 21 articles involving 7,822 stroke patients (2,876 with PSCI) were included. The main modeling variables comprised age, gender, education level, stroke history, stroke severity, lesion volume, lesion site, stroke subtype, white matter hyperintensity (WMH), and vascular risk factors. The prediction models used were prediction nomograms constructed based on logistic regression. The pooled c-index, sensitivity, and specificity were 0.82 (95% CI 0.77–0.87), 0.77 (95% CI 0.72–0.80), and 0.80 (95% CI 0.71–0.86) in the training set, and 0.82 (95% CI 0.77–0.87), 0.82 (95% CI 0.70–0.90), and 0.80 (95% CI 0.68–0.82) in the validation set, respectively.Conclusion: ML is a potential tool for predicting PSCI and may be used to develop simple clinical scoring scales for subsequent clinical use.Systematic Review Registration: https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=383476.

Background/Objectives: Genetic overlap between any ischaemic stroke (AIS) and coronary artery disease (CAD) can obscure associations less closely tied to coronary risk. We aimed to separate coronary-shared from residual AIS genetic effects. Methods: We applied GWAS-by-subtraction to GIGASTROKE AIS and UK Biobank/CARDIoGRAMplusC4D CAD summary statistics. A Cholesky model separated a CAD-related shared component (F1) from a model-defined residual AIS component (F2). We tested model order, factor covariance, population prevalence, CAD input, LD reference and sample-size parameter then characterised loci using FUMA, SuSiE-RSS, genetic correlation, MAGMA, SMR/HEIDI and targeted colocalisation. Results: AIS and CAD had a genetic correlation of 0.490. F1 and F2 had SNP heritabilities of 0.0251 and 0.00414, and F2 contained 14 FUMA loci. F2 Z scores remained correlated across covariance (minimum r = 0.994), prevalence (minimum r = 0.998) and alternative-CAD analyses (r = 0.984). Reverse ordering produced reallocation of source-trait signal. At rs974819, the T allele was associated positively with CAD and negatively with marginal AIS, producing a model-derived CAD-AIS-discordant F2 association. Source-trait colocalisation showed that NBEAL1, PHACTR1 and PDGFD signals were present in one or both input-trait analyses, whereas the F11 protein signal aligned with AIS and F2. In a potentially overlapping FinnGen cross-implementation analysis, all 13 F2 lead variants showed concordant directions. Within overlapping stroke-subtype analyses, F2 showed stronger alignment with cardioembolic stroke. Conclusions: GWAS-by-subtraction resolved a model-defined residual AIS component that was stable across covariance, prevalence and CAD-input specifications, supporting a component-based view of AIS genetic effects after modelling shared coronary architecture.