The Use of Machine Learning in Stroke Detection: Performance, Medicinal Knowledge, and Future Prospects

Impacting the health of the worldwide population, strokes pose a significant global health challenge, affecting people’s lives on a massive scale. Despite its hazardous nature, strokes are largely preventable; however, due to factors such as awareness and health care access, many suffer an unfortunate fate. Many research endeavors have generated predictive models for stroke; however, less have laid out interpretable methods in the clinical context and optimized models for practicality. This research introduces clear approaches that employ interpretable methods suited to the clinical context. This paper involves engineered models that incorporate ML techniques such as dataset resampling, feature importance and selection, parameterization, and analytical techniques. The approach not only offers comprehensive methods, but also an effective model that sets up the foundations for stroke prediction in the real world. Using the Stroke Prediction Dataset, a Deep Neural Network (DNN) model and a Decision Tree (DT) model were constructed. By applying several techniques, including resampling technique SMOTE, feature selection RFE, and feature importance, both models offered insight into the influences on stroke and into the methods themselves. The multifaceted techniques enabled the DNN to achieve a 71% accuracy and the DT to achieve an astounding 95.4% accuracy, with both having only 3 input features. The high accuracy and low feature count of the DT will build a strong foundation for a cost-effective solution towards stroke prediction in the real world.