Interpretable CNN for ischemic stroke subtype classification with active model adaptation.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Accurate TOAST subtype classification is crucial for ischemic stroke diagnosis and research.
• Manual TOAST classification is challenging due to neurologist experience limitations and time constraints.
• A novel active deep learning architecture is proposed to address these challenges.

Purpose of the Study:

• To develop an effective and efficient method for TOAST subtype classification.
• To simulate the diagnostic process of neurologists using computational methods.
• To improve the accuracy and interpretability of ischemic stroke subtype classification.

Main Methods:

• Feature selection using XGB algorithm to identify relevant indicators.
• Implementation of an active learning framework with a causal Convolutional Neural Network (CNN).
• Adaptive sample uncertainty optimization using a mixed active selection criterion and KL-focal loss.

Main Results:

• A dataset of 2310 patients was used for evaluation.
• Sequential experiments validated the effectiveness of individual contributions.
• The proposed method achieved competitive results, with a notable AUC improvement to 77.4%.

Conclusions:

• A backbone causal CNN was developed, enhancing internal interpretability.
• The model demonstrates potential clinical application value in stroke medicine.
• Future work will focus on incorporating diverse data types and patient cohorts for fully automated classification.