Tissue outcome prediction in hyperacute ischemic stroke: Comparison of machine learning models

Area of Science:

• Neurology
• Radiology
• Artificial Intelligence

Background:

• Acute ischemic stroke (AIS) management relies on predicting final infarct volume to guide treatment decisions.
• Machine learning (ML) offers potential for improved tissue fate prediction compared to existing clinical methods.
• Accurate prediction of infarct evolution is crucial for optimizing patient care and treatment strategies.

Purpose of the Study:

• To benchmark three ML models against a clinical diffusion-perfusion thresholding method for voxel-based infarct prediction in AIS patients.
• To evaluate the predictive performance of Gradient Boosting, Random Forests, and U-Net models using baseline MRI data.
• To determine the most effective ML approach for predicting final infarct volume in AIS prior to recanalization therapy.

Main Methods:

• Retrospective analysis of MRI data from 394 AIS patients before recanalization treatment.
• Utilized baseline MRI (MRI0) including diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC), and Tmax maps as model inputs.
• Trained Gradient Boosting, Random Forests, and U-Net models to predict final infarct volume, validated against manually segmented 24-hr follow-up DWI lesions (MRI24h).

Main Results:

• Gradient Boosting demonstrated superior predictive performance with a median Dice Score of 0.53 [0.29-0.68].
• This performance was significantly better than U-Net (0.48 [0.18-0.68]), Random Forests (0.51 [0.27-0.66]), and the clinical thresholding method (0.45 [0.25-0.62]) (P < 0.001).
• ML models, particularly Gradient Boosting, showed enhanced accuracy in predicting final infarct extent.

Conclusions:

• Gradient Boosting significantly outperforms other ML models and the current clinical method for infarct prediction in AIS.
• This ML model shows promise for enhancing clinical decision-making and patient management in acute ischemic stroke.
• The findings support the integration of advanced ML techniques into routine stroke care for improved outcomes.