Hybrid Approach for Electrocardiogram Images Classification: Residual Network and Random Forest

This paper presents a hybrid approach for classifying electrocardiogram (ECG) papers in the George B. Moody PhysioNet Challenge 2024 by integrating Deep Residual Networks (ResNet) and Random Forest (RF) algorithms.Two distinct pipelines were developed to enhance classification accuracy and efficiency.The first pipeline (RRF) utilizes ResNet50 to directly extract features from ECG paper, which are subsequently classified using an RF model.The second pipeline (YRRF) employs the YOLO object detection model to segment ECG papers into 12 individual leads before feature extraction with ResNet50, followed by classification with RF.The accuracy, F1 score, and area under the receiver operating characteristic curve (AUROC) were 79.8%, 78.26%, and 0.89, and 73.74%, 71.11% and 0.81 for the RRF and YRRF models respectively.These were trained on two classes (normal and abnormal).This demonstrated that the leadspecific feature extraction approach was not helpful in this study because it caused a reduction in the performance of the hybrid model.Our team, Leicester Fox had an F_measure score of 0.068 for an official phase using the RRF classification model when trained on 12 classes.This study underscores the effectiveness of hybrid modelling techniques in medical image analysis and sets the stage for future research in automated cardiovascular diagnostics.