ECG Arrhythmia Classification a Lightweight, Explainable, and Web-Deployable Deep Learning Framework Using Knowledge-Distilled GAN Features and Flask

Abstract

A Cardiac Arrhythmias remain a leading cause of sudden cardiac events worldwide, and timely automated screening from Electrocardiogram (ECG) signals can materially improve patient outcomes. Recent literature has explored Generative Adversarial Network (GAN) based architectures that achieve very high offline classification accuracy, but these designs are typically large, computationally heavy, and unsuitable for direct deployment in low-latency clinical or web-based settings. This work proposes an extension to that line of research: a two-stage pipeline in which a cycle-consistent generative model would be used only during the offline training phase to balance rare arrhythmia classes and to enrich the feature space, while a separate, knowledge-distilled lightweight Convolutional Neural Network (CNN) would be trained to reproduce the larger network's decision behavior using soft-label distillation. The distilled classifier is designed to be small enough to run interactively, wrapped inside a Python Flask web application that accepts a single ECG sample, returns the predicted arrhythmia class, and renders a bar chart of class-wise prediction confidence. This document presents the proposed architecture, the planned evaluation on the MIT-BIH Arrhythmia Data-base and the PTB-XL 12-lead ECG dataset, to benchmarking methodology predictions