Explainable deep learning in healthcare: A methodological survey from an attribution view

Area of Science:

• Computational Biology
• Artificial Intelligence in Medicine
• Biomedical Informatics

Background:

• Electronic Health Record (EHR) data and deep learning (DL) advancements fuel interest in AI-driven clinical decision support systems.
• The 'black-box' nature of DL hinders its adoption in real-world healthcare settings.
• Interpretable DL is crucial for end-users to validate AI model predictions and recommendations.

Purpose of the Study:

• To provide a comprehensive review of interpretability methods for DL models in healthcare.
• To serve as a methodological reference for researchers and clinicians.
• To guide the selection and application of interpretability techniques in medical AI.

Main Methods:

• Introduction to various DL interpretability methods.
• Discussion of the advantages, disadvantages, and suitable scenarios for each method.
• Analysis of the adaptation and application of general interpretability methods to healthcare problems.

Main Results:

• Detailed overview of DL interpretability techniques.
• Comparative analysis of methods for healthcare applications.
• Insights into how interpretability aids physicians in understanding AI technologies.

Conclusions:

• Enhancing DL interpretability is key to its successful integration into clinical practice.
• This survey equips AI and clinical professionals with knowledge to choose optimal interpretability methods.
• Facilitating trust and understanding in AI-powered healthcare solutions.