Stroke Detection System Using Machine Learning
TL;DR
This work was the first to attempt a fully integrated clinical and real world application of the Stroke Prediction methodology, and the ideas to be excited about are explainable AI, the integration of IoT wearables, Edge and Cloud computing, and the use of federated learning and ML transformers.
Abstract
The second leading cause of death worldwide is stroke, or cerebrovascular accident. Every year it accounts for approximately 11 of every 100 deaths, resulting in 143 million years of disability-adjusted life. For ischemic stroke, a 4.5 hour thrombolytic window is the best available treatment. Time is critical. The Framingham Stroke Risk Score and the CHA2DS2-VASc score are examples of additive or linear models that do not incorporate machine learning. These models address intersections of multiple stroke risk factors, but at a very low level. Therefore, stroke risk assessment is better adapted to machine learning, and that is what this paper is about. Other areas that this paper focuses on are algorithm selection, dealing with class imbalance, feature selection, explainable ML, and possible applications in the hospital. Evidence from 14 studies was included. The best result in these studies was a Random Forest Classifier from the Stroke Prediction dataset on Kaggle, with predictive accuracy of 96.4% and ROC-AUC of 0.981. This work was the first to attempt a fully integrated clinical and real world application of the methodology. Some of the ideas to be excited about are explainable AI, the integration of IoT wearables, Edge and Cloud computing, and the use of federated learning and ML transformers