Advancing deep learning for automated stroke detection: a review

Abstract

Stroke remains a leading cause of death and disability worldwide, necessitating improved diagnostic tools for early detection and classification. Machine learning (ML) techniques have shown promise in addressing this critical healthcare challenge by enabling efficient analysis of stroke-related data. However, the lack of standardized datasets, limited real-time clinical applicability, and the complexity of model interpretability hinder broader adoption. This review critically examines 34 research articles published between 2014 and 2025, focusing on traditional ML, deep learning, transfer learning, and hybrid approaches for stroke detection and classification. Key findings highlight that Traditional ML models such as Support Vector Machines (SVM) and Random Forests (RF) have been widely used but show limitations in high-dimensional medical imaging tasks. Conversely, advanced deep learning models, such as EEG-DenseNet and ResNet50, excel in stroke segmentation and classification tasks, while hybrid methods demonstrate potential for improving accuracy through ensemble strategies. The review also underscores the challenges of dataset scarcity, ethical concerns, and integration barriers in clinical settings. Recommendations for future research include developing more representative datasets, advancing explainable AI methods, and exploring real-time implementation frameworks to bridge the gap between research and clinical practice.