Deep neural network based missing data prediction of electrocardiogram signal using multiagent reinforcement learning

Abstract

Listen

Translate article

ObjectiveClinical morphology of electrocardiogram (ECG) signal is compulsory to analyze the cardiac activity. During long term measurement, missing of data is a common factor, caused by sensor loosening, resulting in devastating feature extraction procedure; hence, prediction of those missing data is indispensable. MethodsIn this work, bidirectional long short-term memory recurrent neural network (LSTM-RNN) based prediction of missing segment of ECG signal is accomplished, governed by reinforcement learning (RL) using multiagent. The LSTM-RNN has internal gate architecture, which can predict the upcoming sequence using the past features of ECG. Each agent of the multiagent system (MAS), which represents various fictitious domain of ECG, has independent learning module to predict respective domain using a ‘critic’ network, which was conducted by a multilayer perceptron neural network (MLPNN). The MAS finally predicted the missing segment by ‘cooperative learning’ algorithm using a ‘coordinator agent’. ResultsThe proposed algorithm was tested on MIMIC-II ECG database, available in Physionet. Experimental result shows that, using the RL, the trained LSTM-RNN was able to predict the missing segments more precisely with correlation coefficient higher than 0.9 and low root mean squared error. ConclusionThe proposed method is applicable to any single channel ECG signal, and the quality of predicted signal ensures a wide application in medical applications as well as telecardiology system. SignificanceA comparative study with previously published works showed an improved performance, related to ECG missing data prediction, implemented on MIMIC-II records.