1. Diagnostic Models for Differentiating COVID-19-Related Acute Ischemic Stroke Using Machine Learning Methods.
- Author
-
Gul Ates, Eylem, Coban, Gokcen, and Karakaya, Jale
- Subjects
- *
MACHINE learning , *DIMENSIONAL reduction algorithms , *POST-acute COVID-19 syndrome , *ISCHEMIC stroke , *FEATURE selection - Abstract
Backgrounds: Although COVID-19 is primarily known as a respiratory disease, there is growing evidence of neurological complications, such as ischemic stroke, in infected individuals. This study aims to evaluate the impact of COVID-19 on acute ischemic stroke (AIS) using radiomic features extracted from brain MR images and machine learning methods. Methods: This retrospective study included MRI data from 57 patients diagnosed with AIS who presented to the Department of Radiology at Hacettepe University Hospital between March 2020 and September 2021. Patients were stratified into COVID-19-positive (n = 30) and COVID-19-negative (n = 27) groups based on PCR results. Radiomic features were extracted from brain MR images following image processing steps. Various feature selection algorithms were applied to identify the most relevant features, which were then used to train and evaluate machine learning classification models. Model performance was evaluated using a range of classification metrics, including measures of predictive accuracy and diagnostic reliability, with 95% confidence intervals provided to enhance reliability. Results: This study assessed the performance of dimensionality reduction and classification algorithms in distinguishing COVID-19-negative and COVID-19-positive cases using radiomics data from brain MR scans. Without feature selection, ANN achieved the highest AUC of 0.857 (95% CI: 0.806–0.900), demonstrating strong discriminative power. Using the Boruta method for feature selection, the k-NN classifier attained the best performance, with an AUC of 0.863 (95% CI: 0.816–0.904). LASSO-based feature selection showed comparable results across k-NN, RF, and ANN classifiers, while SVM exhibited excellent specificity and high PPV. The RFE method yielded the highest overall performance, with the k-NN classifier achieving an AUC of 0.882 (95% CI: 0.838–0.924) and an accuracy of 79.1% (95% CI: 73.6–83.8). Among the methods, RFE provided the most consistent results, with k-NN and the ANN identified as the most effective classifiers for COVID-19 detection. Conclusions: The proposed radiomics-based classification model effectively distinguishes AIS associated with COVID-19 from brain MRI. These findings demonstrate the potential of AI-driven diagnostic tools to identify high-risk patients, support optimized treatment strategies, and ultimately improve clinical implications. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF