Interpretable Deep Learning Models for Medical Diagnosis: A Case Study on Cardiac Arrhythmia Classification

Abstract

An in-depth study on the creation of interpretable deep learning models for precise classification of cardiac arrhythmias is presented here. In this study, cutting-edge deep learning techniques are used to the well-known MIT-BIH Arrhythmia Database. These techniques include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and their hybrid architecture. The electrocardiogram (ECG) recordings in this dataset have been annotated, making it possible for models to learn complicated patterns and categorise a wide variety of cardiac arrhythmias. To aid in the interpretation of ECG signals, attention mechanisms are integrated. The suggested hybrid CNN-RNN model performs exceptionally well, with an accuracy of 94.56%. In addition to enhancing the model's interpretability, visualising the attention weights reveals useful insights into the decision-making procedure. The use of interpretable deep learning models and the illumination of the mechanisms driving accurate predictions are two major contributions of this thesis to the field of cardiac arrhythmia classification. This study demonstrates the promise of such models to improve healthcare by assisting physicians in making accurate diagnoses and crafting effective treatment plans.