Functional Outcome Prediction in Acute Ischemic Stroke Using a Fused Imaging and Clinical Deep Learning Model

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate prediction of long-term clinical outcomes in acute ischemic stroke patients is vital for patient management and clinical decision-making.
• Current prediction methods often rely on subjective assessments and time-consuming image analysis.
• This study addresses the need for objective and efficient outcome prediction in stroke care.

Purpose of the Study:

• To develop and validate a Deep Learning model for predicting the 90-day modified Rankin Scale (mRS) score in acute ischemic stroke patients.
• To fuse diffusion-weighted imaging data with acute clinical information for enhanced prediction accuracy.
• To reduce subjectivity and user burden associated with traditional outcome assessment methods.

Main Methods:

• A cohort of 640 acute ischemic stroke patients with available MRI and 90-day mRS data was used, randomly split into training, validation, and internal testing sets.
• An external validation cohort from Lausanne University Hospital (n=280) was included to assess model generalization.
• Performance was evaluated using accuracy for ordinal mRS, accuracy within ±1 mRS category, mean absolute prediction error, and prediction of unfavorable outcomes (mRS >2).

Main Results:

• Fused clinical-imaging Deep Learning models significantly outperformed clinical-only and imaging-only models in both internal and external cohorts.
• The top fused model achieved an area under the curve (AUC) of 0.92 for unfavorable outcome prediction in the internal test cohort.
• In the external cohort, the best fused model demonstrated an AUC of 0.90 and outperformed others with a mean absolute error of 0.90.

Conclusions:

• A Deep Learning model integrating diffusion-weighted imaging and clinical variables provides a robust method for predicting 90-day stroke outcomes.
• This fused approach enhances prediction accuracy while minimizing subjectivity and the need for extensive postprocessing.
• The model demonstrates strong generalizability, offering a valuable tool for clinical prognostication in acute ischemic stroke.