Toward explainable deep learning in healthcare through transition matrix and user-friendly features.

Area of Science:

• Medical signal and image processing
• Artificial intelligence in healthcare
• Explainable AI (XAI)

Background:

• Deep learning (DL) models in medicine are often "black boxes", limiting transparency and trust.
• Interpreting complex AI decisions is crucial for clinical adoption and ethical considerations.

Purpose of the Study:

• To develop and evaluate a scalable method for enhancing the interpretability of DL models in medical applications.
• To translate DL model decisions into user-friendly and justifiable features for healthcare professionals.

Main Methods:

• A transition matrix approach was used to interpret DL model decisions.
• Interpretable features were defined using clinical guidelines and expert rules.
• The method was validated on electrocardiography (ECG) for arrhythmia detection and magnetic resonance imaging (MRI) for heart disease classification.

Main Results:

• The approach achieved strong agreement with expert annotations, with Cohen's Kappa coefficients of 0.89 for ECG and 0.80 for MRI.
• The results demonstrate the reliability of the method in providing accurate and understandable explanations for DL model decisions.
• The transition matrix method proved scalable and applicable across different medical datasets.

Conclusions:

• The proposed transition matrix approach enhances the interpretability and trustworthiness of DL models in medical AI.
• This method facilitates the alignment of AI outputs with clinical standards, addressing practical and ethical challenges in healthcare.
• The approach shows potential for broad application in medical domains, improving the generalizability of AI in healthcare.