Usefulness of Heat Map Explanations for Deep-Learning-Based Electrocardiogram Analysis.

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Cardiology Informatics

Background:

• Deep neural networks (DNNs) offer potential for analyzing complex medical data and aiding diagnoses.
• The "black box" nature of DNNs hinders understanding and clinical trust.
• Explainable AI (XAI) methods aim to demystify DNN predictions.

Purpose of the Study:

• To develop a DNN using transfer learning for sex prediction from electrocardiograms (ECGs).
• To evaluate the clinical utility of Grad-CAM heat maps for explaining DNN predictions in ECG analysis.
• To assess whether XAI-generated ECG features are interpretable and useful for medical professionals.

Main Methods:

• Transfer learning was employed to build a DNN model.
• Electrocardiogram (ECG) data was used for sex prediction.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was utilized to generate visual heat maps for model interpretability.
• Medical doctors provided feedback on the usefulness of Grad-CAM heat maps.

Main Results:

• A DNN model was successfully developed to predict sex from ECGs.
• Grad-CAM generated heat maps highlighting ECG regions used by the DNN.
• Medical doctors found the Grad-CAM heat maps not clinically useful or informative for understanding sex prediction.
• The identified ECG features in heat maps were not recognized as discriminative for sex by clinicians.

Conclusions:

• Current Grad-CAM-based XAI methods do not provide meaningful clinical insights for DNN-based sex prediction from ECGs.
• The developed XAI approach is not currently suitable for clinical application in this context.
• There is a need for novel, medically tailored explanation techniques for DNNs in clinical diagnostics.