Optimizing Stroke Detection Using Evidential Networks and Uncertainty-Based Refinement

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Conventional stroke impairment assessments lack precision, often missing subtle deficits.
• Robotic systems offer objective kinematic data for motor function evaluation.
• Early detection of neurological impairments is vital for effective treatment and disability management.

Purpose of the Study:

• To develop and validate a deep learning model for detecting subtle motor impairments post-stroke and in transient ischemic attack (TIA) patients.
• To leverage uncertainty estimation to refine the model's sensitivity in identifying minimally impaired individuals.
• To assess the clinical utility of AI-driven kinematic analysis for stroke and TIA detection.

Main Methods:

• Analysis of kinematic data from 337 stroke patients and 368 healthy controls using the Kinarm Exoskeleton system.
• Application of an evidential deep learning network to distinguish impaired from healthy participants and quantify prediction uncertainty.
• Iterative retraining and test set refinement based on model uncertainty to enhance detection sensitivity.

Main Results:

• The deep learning model successfully distinguished stroke patients from controls.
• Sensitivity for detecting subtle impairments in minimally impaired stroke patients increased from 0.55 to 0.75 after uncertainty-based refinement.
• Detection accuracy for transient ischemic attack (TIA) impairments improved from 0.86 to 0.92.

Conclusions:

• Deep learning models integrating uncertainty estimation show significant potential for detecting subtle neurological impairments.
• This approach enhances the early identification of individuals with stroke-related deficits, including those with TIA.
• The findings suggest a pathway toward more personalized and effective post-stroke rehabilitation strategies.