Comprehensive Assessment and Approaches to Minimize Data Exposure in the Prediction of Hyperacute Stroke

The early identification of hyperacute stroke situations remains essential for fast clinical care and better treatment results. The presence of data leakage continues to be an significant because it produces inflated performance metrics which fail to match actual clinical applications. This research presents a novel framework for identifying and mitigating data leakage in machine learning models for hyperacute stroke detection. This research also investigates current methods before proposing an innovative joint method using Mutual Information (MI), chi-square and ANOVA scores to eliminate data leakage throughout feature selection. Additionally, Synthetic Minority Oversampling Technique (SMOTE) technique is employed for combatting class imbalance and to enhance prediction reliability. Experimental findings establish the proposed framework as robust by providing notable improvements in how consistently classifiers perform. This research fills an important void by evaluating a common interpretability weakness which now provides both reliable deployment methods for real use and clearer assessment outputs.