International Journal of Scientific Research in Engineering and Management (IJSREM) Volume: 08 Issue: 05 | May - 2024 SJIF Rating: 8.448 ISSN: 2582-3930 © 2024, IJSREM | www.ijsrem.com DOI: 10.55041/IJSREM33939 | Page 1 Advanced Deep Learning for ECG Anomaly Detection in Imbalanced Data Archana Ratnaparkhi Department of Electronics and Telecommunication Vishwakarma institute of information Technology Pune,Maharashtra,India Affiliated to Savitribai Phule Pune University archana.ratnaparakhi@viit.ac.in Pallavi Deshpande Department of Electronics and Telecommunication Vishwakarma Institute of Information Technology Pune,Maharashtra,India pallavi.deshpande@viit.ac.in Ketaki Kshirsagar Department of Electronics and Telecommunication Vishwakarma Institute of Information Technology Pune,Maharashtra,India ketki.kshirsagar@viit.ac.in Minal Deshmukh Department of Electronics and Telecommunication Vishwakarma Institute Of Information Technology Pune,Maharashtra,India minal.deshmukh@viit.ac.in Shradha Habbu Department of Electronics and Telecommunication Vishwakarma Institute of Information Technology Pune,Maharashtra,India shradha.habbu@viit.ac.in Gauri Ghule Department of Electronics and Telecommunication Vishwakarma Institute of Information Technology Pune,Maharashtra,India gauri.ghule@viit.ac.in Abstract—Segmentation of ECG to obtain significant and relevant features has been an inevitable step to reduce the dimensionality of dataset in automated heart disease diagno- sis systems.Accurate and speedy classification of heart beats is required to reduce high mortality rate which is prevalent due to cardiovascular diseases(CVD). Nonstationarity and high variability exhibited by ECG signal leads to increase in com- plexity of analysis in time and frequency domain.Challenges in processing are further enhanced due to imbalanced and vague datasets.Deep learning based methods have been used in litera- ture to combat the problem of imbalanced datasets.This paper employs an effective recurrent neural network with long short term memory layers(LSTM) to classify the heart beats into two classes.It has been observed that LSTM network can effectively extract the sequential timing information in the input ECG samples.To remove the imbalance in the datasets,oversampling and focal loss based weight balancing techniques have been used which eventually enhance the accuracy of classification. MIT- BIH database has been used for experimental evaluation.The proposed approach ,LSTM network with oversampling tech- nique,provides an accuracy of 99.54% which is far better as compared to the traditional approaches which yield accuracy around 95%.Moreover this method is insensitive to quality of ECG signals due to the involvement of fuzzification procedure in the initial steps.Deployment of the proposed method for biosignal telemetry or pharmaceutical research to assist the physicians in their work is a possible future advancement in this domain. Index Terms—ECG, RNN, LSTM,Classification, imbalance. I. INTRODUCTION The electrocardiogram (ECG) stands out as a paramount gauge of heart health. According to the World Health Or- ganization, approximately 18% of global deaths stem from cardiovascular ailments yearly, with many succumbing due to delayed treatment. Swift and automated diagnosis is impera- tive for timely intervention. Diverse ECG classification techniques encompass time do- main analysis, which scrutinizes intervals and amplitudes for feature generation, and frequency domain methods, which leverage significant frequencies aiding in heartbeat detection. Attempts at segmentation via Markov models have proven in- adequate, necessitating semi-Markov models. Time-frequency analysis emerges as highly effective in extracting precise fre- quency data. Classifiers like Support Vector Machines (SVM) and Multilayer Perceptrons (MLP) have been utilized, often enhanced by various search algorithms. Many conventional methods rely on rigidly predefined features, leading to elevated false positive rates and conse- quent misdiagnoses. We adopt deep learning for automated classification, overcoming these challenges. Long Short-Term Memory (LSTM), an advanced technique for time series pro- cessing, retains pertinent information while discarding noise. Recent applications of LSTM in ECG analysis have yielded remarkable accuracies, notably a 99.86% accuracy in temporal feature extraction and 91% accuracy in arrhythmia detection. Models like the LSTM-based autoencoder and error profile modeling showcase the versatility and efficacy of LSTM networks in ECG analysis. Evaluation metrics like F1 scores