Machine Learning-Driven Stroke Prediction Using Independent Dataset

Abstract

The incidence of stroke cases has witnessed a rapid global rise, affecting not only the elderly but also individuals across all age groups. Accurate prediction of stroke occurrence demands the utilization of extensive data pre-processing techniques. Moreover, the automation of early stroke forecasting is crucial to prevent its onset at the initial stage. In this study, stroke prediction models are evaluated to estimate the likelihood of stroke based on various symptoms such as age, gender, pre-existing medical conditions, and social variables. The machine learning techniques employed include Linear Support Vector Classifier, Extreme Gradient Boosting Classifier, Multilayer Perceptron, Adaptive Boosting Classifier, Bootstrap Aggregating Classifier, and Light Gradient-Boosting Machine. The purpose of this paper is to optimize the hyperparameters of machine learning approaches in developing stroke prediction models. The goal was achieved through a comprehensive comparison of three different sampling techniques for handling imbalanced datasets and evaluating their performance by using various metrics. The most effective model is identified, which is the Adaptive Boosting Classifier utilizing the Tomek Links, with a cross-dataset accuracy of 99% which demonstrated a reliable performance and generalization as evidenced by high cross-validation scores and accuracy on an independent dataset. The next stage of this endeavor entails looking into multiple ways to forecast the development of new dangerous diseases such as breast cancer and skin disorders. In the long run, the aim of subsequent work is to build a powerful toolset that is obtainable to all medical practitioners, allowing for the pre-emptive diagnosis of all potentially hazardous illnesses.