Abstract

Introduction: Stroke occurs due to interrupted brain blood flow, leading to cell death. AI and computer vision aid diagnosis, prediction, and patient management. Technologies like CT, MRI, and PET enhance stroke assessment. However, ML-based stroke diagnosis is underexplored in developing countries, including Nigeria, with limited model comparisons. Aim: To systematically review the existing machine learning models for stroke diagnosis, and identify their strengths and weaknesses. Materials and Methods: A systematic review of 880 Google Scholar articles on Machine Learning and Stroke was conducted. After applying PRISMA criteria, 44 studies were selected. Results: The search returned 880 articles. After screening and removal of duplicates, the number of articles was reduced to 489. Out of these 391 papers were excluded based on title, keywords and abstract, 391 relevant studies met the inclusion criteria, out of these 98 articles were eligible. 54 articles were excluded after further screening and 44 papers that met the criteria for inclusion were reviewed. We found that the most commonly used ML models were random forest (10 studies), support vector machine (6 studies), neural networks (24 studies), and logistic regression (4 studies). The accuracy of machine learning algorithms ranged from 0.58 to 0.97. Conclusion: We discovered that there are increasing research efforts on Machine Learning Models and stroke prediction but with a very few studies done in developing countries. The performance of the existing machine learning models is good but can be improved upon. Major improvements and validations are required for stroke models' adoption into clinical practice. Our future plan is to develop a homemade machine learning model for stroke diagnosis.

Keywords: Machine Learning, Stroke, and Diagnosis