Priv-Aug-Shap-ECGResNet: Privacy Preserving Shapley-Value Attributed Augmented Resnet for Practical Single-Lead Electrocardiogram Classification

Abstract

We aim to build an effective automated single-lead Electrocardiogram (ECG) classification system to enable remote and timely screening of critical cardio-vascular diseases like Heart attack. However, the expenses associated with cardiologist-intervened ECG annotation limits the number of training instances. While conventional deep learning models require large set of training examples for accurate classification, we propose Priv-Aug-Shap-ECGResNet which demonstrates that deep learning algorithm (for e.g., residual network or ResNet) with ablation of unimportant features from the given training dataset can ensure consistently better classification performance over relevant state-of-the-art algorithms. Additively perturbed training augmentation with Shapley attribution finds out the right feature subset with the assistance of the axioms of transferable utility, namely "efficiency" and "null player" on which Shapley value game is defined. Priv-Aug-Shap-ECGResNet is enabled with novel data privacy preservation feature through differential privacy technique to provide measured obfuscation to render ZeroR classification equivalent knowledge gain to the adversary.