Anomaly Detection with Deep Long Short Term Memory Networks

In this study, a robust anomaly detection technique for ECG signals is developed using deep gated recurrent neural networks (GRNN) with Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM) unit. By training deep GRU and LSTM networks on normal ECG data acquired from healthy subjects, a robust prediction model that learns to predict future time steps of ECG time series is developed. The prediction errors are modeled using Multivariate Gaussian Distribution and the estimations of optimum parameters were performed via Maximum Likelihood Estimation (MLE) method. By using probability distributions of prediction errors and optimum threshold values, the classification of normal and abnormal time series is performed. The results of the study show that deep LSTM networks with stacked recurrent hidden layers can learn higher-level temporal features in ECG time series without prior knowledge of the data and can robustly model normal time series behaviors. The performance results of the proposed deep learning and Gaussian-based statistical anomaly detection technique over the European ST-T database show that the technique provides the reliable diagnosis of cardiovascular diseases by performing the robust detection of anomalies in ECG time series.