Interpretable Deep Models for Cardiac Resynchronisation Therapy Response Prediction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Deep learning (DL) models achieve high accuracy in medical image classification but often lack interpretability, hindering clinical trust and adoption.
• Existing clinical knowledge is valuable for generating explanations but difficult to integrate into standard DL models.
• Interpretability is crucial for understanding model decisions and facilitating clinical translation of AI in healthcare.

Purpose of the Study:

• To develop a novel deep learning framework for interpretable medical image classification.
• To enhance the explainability of DL models by integrating existing clinical knowledge.
• To improve the prediction of treatment response using interpretable AI models.

Main Methods:

• A deep learning framework based on a variational autoencoder (VAE) was proposed for image-based classification.
• The VAE framework allows prediction from the latent space and visualization of decision boundary effects.
• The VAE disentangles the latent space using clinical knowledge, enabling prediction of outputs and explanations.

Main Results:

• The framework demonstrated high sensitivity (88.43%) and specificity (84.39%) in predicting cardiac resynchronization therapy (CRT) response from cardiac MRI.
• The model successfully integrated clinical knowledge to generate explanations for its predictions.
• The approach showed potential for discovering new biomarkers disentangled from existing knowledge.

Conclusions:

• The proposed VAE-based framework enhances the interpretability of deep learning models in medical image classification.
• This interpretable approach can improve clinical trust and facilitate the translation of AI tools in medicine.
• The framework shows promise for predicting treatment response and uncovering novel biomarkers in cardiology.