Anomaly Detection in ECG Signals using Deep Long Short Term Memory Networks

Abstract

Electrocardiography (ECG) signals are largely accessed to monitor the health condition of the human heart, and the resultant time series signals are analysed manually by the
medical professionals to detect if there are any kind of anomalies such as arrhythmia.Manual Diagnosis of ECG Signals is often Prone to Errors.Past work in Automating the Analysis requires extensive Pre-Processing which is time consuming and cumbersome. It takes a significant time of Heart Patients in their Precarious Condition. There is a requirement of Computational Analysis which is fast and efficient. Some of the analysis develop the marked features and design a classifier for discriminating between the healthy ECG signals and those which contains Arrhythmia. This method requires knowledge and the relevant data
of the various different types of Arrhythmia for training the model. However, there can be many different and different new types of Arrhythmia can occur which previously were not a part of the original training set. Thus, it may be more wise to adopt an anomaly detection approach towards analysing them. In this Project, we are utilizing A deep recurrent neural network (RNN) architecture with the Long Short Term Memory Network (LSTM) units for developing a predictive model from the healthy ECG signals. The probability distribution of the prediction errors from the models, using Maximun Likelihood Estimate (MLE) are used
for indicating anomalous or non-anomalous behaviour. The main advantage of using LSTM networks is that the ECG signals be directly applied into the network without any extensive pre-processing as used by other Detection techniques. No Prior information of abnormal signals makes it worthwhile, as it needs to be trained only on normal data. MIT-BIH Arrhythmia Physionet Database has been used to obtain ECG time series data for both non-anomalous periods and anomalous periods. Both the Stateful and Stateless Modes of LSTM are proposed and implemented. The results are compared with the already developed models such as NN Classsifier, RNN and GRU. The Results from the Stateful LSTM Model show Precision of 99.31% and a TPR/FPR ratio of 144.38. Results are promising and indicate that the Deep-Stacked Long Short Term Memory Networks (LSTM) models are feasible for detecting anomalies in ECG signals within a short span of time.