Enhancing explainability in clinical deep-learning models: Latent-space variable decoding is superior to...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning models are increasingly used in cardiology for clinical decision support.
• The
• black box
• nature of these models hinders physician trust and validation.
• Explainability techniques like Grad-CAM lack rigorous assessment of reliability and reproducibility.

Purpose of the Study:

• To rigorously assess the explainability of Grad-CAM in cardiology.
• To compare Grad-CAM with alternative saliency methods from intrinsically explainable deep learning models.

Main Methods:

• A cohort of 1930 hypertrophic cardiomyopathy (HCM) patients with electrocardiographic data was analyzed.
• Novel deep learning models were developed to predict left ventricular apical aneurysm and massive LV hypertrophy.
• Saliency analysis was performed using Grad-CAM and latent-space variable decoding (LSVD).

Main Results:

• Deep learning models showed comparable predictive performance for both techniques (e.g., C-statistic 0.95 vs 0.93 for LV apical aneurysm).
• Grad-CAM yielded variable attention maps with limited insight into decision-making.
• LSVD enabled direct visualization of differentiating electrocardiographic characteristics and assessed model overfitting risk.

Conclusions:

• LSVD provides more robust explainability for deep learning models in cardiology than Grad-CAM.
• LSVD facilitates better understanding of model predictions, enhancing clinical utility.