Machine learning models for enhanced stroke detection and prediction

Abstract

Stroke detection plays a vital role in medical diagnostics, where timely and accurate identification can improve patient outcomes. This research evaluates the performance of three machine learning models—OzNet-mRMR-NB, Logistics Regression, and an Ensemble CNN—using medical images for stroke prediction. The OzNet-mRMR-NB model integrates VGG19 for feature extraction, mRMR for feature selection, and Naive Bayes for classification, while Logistic Regression processes flattened feature vectors. The Ensemble CNN, leveraging ResNet and VGG19, outperforms the other models with a testing accuracy of 92.43 %, an AUC score of 0.92, precision of 0.93, and an F1-score of 0.92. Additionally, recall for both the Ensemble and OzNet models was 0.93, highlighting the Ensemble model’s capacity to sustain a robust balance between specificity and sensitivity. These results highlight the advantages of combining diverse CNN architectures for improved accuracy and generalizability. This research advances automated stroke detection, with potential clinical applications for timely and informed decision-making. Future work will refine the ensemble approach for broader clinical adoption across diverse patient populations.