Sztuczna inteligencja w przewidywaniu padaczki po udarze mózgu: przegląd badań z modelami przewidywania opartymi na obrazach oraz zastosowanie w afrykańskich ośrodkach zdrowia

BACKGROUND: Poststroke seizures (PSS) and poststroke epilepsy (PSE) are important neurological complications associated with poorer functional recovery, increased mortality, and reduced quality of life after stroke. Neuroimaging plays a central role in identifying patients at increased risk, and recent advances in AI have enabled the development of imaging-based prediction models that may improve individualized risk stratification. However, the current evidence remains fragmented, and the extent to which these approaches are applicable to African health care settings is unclear. OBJECTIVE: This scoping review aimed to chart the existing global evidence on the development of imaging-based prediction models that can help in predicting PSS and PSE, especially AI-based ones. The purpose was to look into the imaging techniques, computational algorithms, and validation methodologies used and the representation of African populations. METHODS: This scoping review was performed using the Joanna Briggs Institute methodological framework, and the results were presented based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) framework. Articles searched using databases such as PubMed, Embase, Scopus, Web of Science, IEEE Xplore, and Google Scholar involved the application of computed tomography (CT), magnetic resonance imaging (MRI), angiography imaging techniques, radiomic markers, or multimodal neuroimaging techniques in statistical and AI-based prediction models for seizure-related outcomes after stroke. RESULTS: Twenty-three studies published between 2005 and 2026 met the eligibility criteria. The evidence demonstrated a clear methodological evolution from conventional imaging-based prediction models to radiomics, machine learning, deep learning, explainable AI, and multimodal approaches. MRI and CT were the predominant imaging modalities, with cortical involvement, lesion location, infarct volume, hemorrhagic transformation, and quantitative radiomic features emerging as the most frequently reported imaging predictors. Most AI models relied on internal validation, while relatively few underwent external validation. The evidence base was concentrated in East Asia, Europe, and North America, with no eligible imaging-based AI models having been developed or externally validated using African datasets. CONCLUSIONS: Imaging-based AI for predicting PSS and PSE has progressed rapidly, but important methodological and geographical gaps remain. Limited external validation, heterogeneous reporting, and the absence of African development and validation cohorts restrict confidence in the generalizability of existing models. Future research should prioritize the development of explainable imaging-based AI models using representative African neuroimaging datasets and their external validation in multicenter studies.