Application of Recurrent Neural Network for the Prediction of Target Non-Apneic Arousal Regions in Physiological Signals

Abstract

This work presents a new method for detection of target non-apneic arousals by applying a recurrent neural network architecture on the various specified polysomno-graphic (PSG) signals. The proposed two stage architecture uses sequences of instantaneous frequencies and spectral entropies of the chosen PSG signals as feature vectors. At the first stage, these feature vectors are used to train several long-short term memory (LSTM) models. The LSTM networks can learn long-term relationships between time steps of time-frequency based sequences obtained out of physiological signals. As a second stage, some quadratic discriminant (QD) layers are modelled and appended to the trained LSTMs in groups. Subsequently, the outputs of all the QD layers are averaged for making final prediction. The models are trained using features obtained from one minute windows of the signals. However, the decision making on test signals involves inputs of one minute windows with half minute overlapping. When evaluated with 2018 PhysioNet/CinC Challenge dataset, the experimental outcomes demonstrate overall AUROC and AUPRC scores of 0.85±0.10 and 0.50±0.15 respectively for the training data. The generated test results indicate the AUROC and AUPRC scores of 0.624 and 0.10 respectively on a random subset of the test data.