Using LSTM Networks for Multiparameter Physiological Signal Reconstruction to Reduce Training Time

Abstract

Signal corruption or dropout can present issues in continuous patient monitoring in the intensive care unit. As a result, the ability to accurately reconstruct absent or corrupt signals can greatly enhance critical patient care. The 2010 PhysioNet/Computing in Cardiology Challenge required participants to develop algorithms to reconstruct a missing 30-second portion of a signal. The Challenge dataset consisted of 300 multiparameter records, each containing six to eight 10-minute, continuous physiological signals. Among the highest-scoring algorithms were neural networks and adaptive/Kalman filtering. Although both algorithms scored well in the competition, these methods used significant amounts of training data from each record, 8 minutes and 5.5 minutes, respectively. Different techniques, such as Recurrent Neural Networks, have been proposed in the literature since the Challenge for multiparameter signal reconstruction, with varying success. We analyzed the performance of these existing algorithms and developed a new Long Short-Term Memory (LSTM) network that produces reconstructions with a relatively short training time (8 seconds). The LSTM network performed comparably well to these algorithms in terms of reconstruction accuracy for three out of the four signal types evaluated (arterial blood pressure, plethysmograph, and respiratory signals), and also had the advantage of drastically reducing the training time needed to achieve accurate signal reconstructions to a mere 8 seconds.