Classifying Acute Ischemic Stroke Onset Time using Deep Imaging Features.

Area of Science:

• Neurology
• Medical Imaging
• Machine Learning

Background:

• Stroke outcome prediction models guide treatment decisions.
• Accurate time-since-stroke (TSS) classification is crucial for treatment eligibility (e.g., <4.5 hours).
• Limited research exists on classifying unknown TSS using imaging features.

Purpose of the Study:

• To develop and compare machine learning models for classifying time-since-stroke (TSS) <4.5 hours using magnetic resonance (MR) imaging features.
• To propose a deep learning model for extracting hidden representations from MR perfusion-weighted images to enhance classification.
• To evaluate the performance of advanced machine learning methods against current clinical approaches for TSS classification.

Main Methods:

• Construction and comparison of various machine learning classifiers utilizing MR imaging features.
• Development of a deep learning model to extract features from MR perfusion-weighted images.
• Implementation of cross-validation techniques to assess model performance.

Main Results:

• The best machine learning classifier achieved an Area Under the Curve (AUC) of 0.68 for TSS classification.
• This performance significantly surpasses the AUC of 0.58 obtained by current clinical methods.
• Incorporating deep learning-extracted features demonstrated classification improvement.

Conclusions:

• Advanced machine learning methods, particularly deep learning on MR imaging, show significant potential for improving time-since-stroke classification.
• These findings suggest a pathway for more accurate and timely treatment eligibility determination in stroke patients.
• The developed models offer a promising tool to aid clinicians in stroke management.