Enhanced Wearable ECG Monitoring with LSTM Autoencoder-Based Anomaly Detection

Abstract

Cardiovascular diseases (CVDs) remain as the leading cause of death worldwide as they cause close to 18 million deaths in a year. Early and precise diagnosis of heart defects like arrhythmia, ischemia, and heart failure is still an important consideration in minimizing the mortality rate and enhancing the quality of life. Smartwatches, chest straps and patches are wearable body sensors that turned out to be revolutionary in terms of continuous and non-invasive cardiac health monitoring. Such devices produce large amounts of Electrocardiogram (ECG) in real-time, which serves as a useful source of anomaly detection. Nonetheless, the proper analysis of these data is complicated by such issues as temporal complexity, variability due to the human activity, inconsistency of the sensor locations, and imbalance in the number of classes where the abnormal cardiac events are underrepresented significantly. The conventional anomaly detection methods (rule-based thresholds, statistical models, etc.) find it difficult to cope with those issues, which results in high false alarm rates and low clinical usability. In order to overcome these shortcomings, this paper proposes an anomaly detection model using a Long Short-Term Memory (LSTM) Autoencoder on ECG signal. The framework takes advantage of the temporal learning of LSTM networks in an encoder-decoder framework to learn normal cardiac patterns and recognize the anomalies based on the reconstruction error. PhysioNet ECG dataset was used as an evaluation dataset and preprocessing activities were applied to include data cleansing, data normalization, time-series segmentation, as well as imbalanced classes. To measure the performance of the offered approach, the following classical machine learning algorithms were implemented and tested on the basis of the following measurements: precision, recall, F1-score, ROC-AUC, and the confusion matrices: Local Outlier Factor (LOF), Elliptic Envelope, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), K-Means Clustering and Logistic Regression.

As it can be seen with the use of experimental findings, the proposed LSTM AutoEncoder has a major advantage, as it works much better than traditional methods and provides 99.45% accuracy, with even better precision and recall. In contrast to supervised methods, the model can be trained solely on normal ECG data so it can be more easily scaled and can be applied in a real-world environment where there are limited labeled anomaly data. This research identifies the promise of unsupervised models based on the deep learning approach in improving the accuracy, reliability, and real-time capacity of the wearable cardiac monitoring system. Moreover, the results open the door to the incorporation of the custom deviation identification structures into telemedical systems so that interventions could be implemented in time and the burden on healthcare services could be lessened.