Abstaining ECG Classifiers Through Explainable Prototypical Spaces

Deep learning models have been deployed for var-ious healthcare tasks owing to their success in ingesting dif-ferent modalities of raw multimedia data. Despite the recent advancements, their widespread adoption in the critical domain of healthcare is restricted due to lack of trust in black-box models. Furthermore, the shortcomings of the existing popu-lar evaluation metrics neither incites confidence in the model predictions nor quantifies model performance capacity towards responsibility for incorrect decisions. In this paper, we propose an abstention method for Convolutional Neural Network (CNN) models by quantification of the embedding space of the networks as prototypical space for each class to explain model capacity, improve performance through responsibility. The prototypical space for each class is the most confident space where the samples are correctly classified with minimal error consequently improving confidence and reliability in the model predictions. We demonstrate improvement in performance up to 18.17% and 12.10% through abstention of model outcomes by identification of outcomes which can be confidently trusted based on the prototypical spaces for each class and network architecture invariance of our proposed approach with two different 2D CNN architectures on the public MIT-BIH ECG dataset. The improvement in performance through our explainable abstaining method mitigates risk in crucial healthcare tasks and improves reliability in model predictions. We also discuss future improvements and applications of our proposed prototypical spaces in the domain of healthcare.