Automated Detection of Sleep Arousals From Polysomnography Data Using a Dense Convolutional Neural Network

Matthew Howe-Patterson1, Bahareh Pourbabaee2, Frederic Benard3
1OMsignal Inc., Algorithm Developer, 2OMsignal Inc., Data Scientist, 3VP of Engineering


In this work, a dense convolutional neural network (DenseNet) was constructed to detect arousal using relevant measurements including three EMGs from the chin, chest and abdomen, respiratory airflow, and three EEGs comprising two central and one posterior derivations, out of 13 Polysomnography (PSG) channels provided in 2018 Physionet challenge database. To extend the receptive fields of our proposed DenseNet, dilated convolutions are deployed, where the dilation rates are increased exponentially with the depth of the network. Our model structure is initialized with one batch normalization and a convolutional layer that is followed by six DenseNet blocks as well as one long short-term memory (LSTM) layer. Finally, a softmax function is used to compute the arousal probability per input sample. Two DenseNet models are respectively trained using cross-entropy loss function (DenseNet1) and the sum of focal and dice losses (DenseNet2) via the adaptive learning optimization method (Adadelta). The dataset includes 994 subjects with annotated PSG data that were partitioned into separate training (594), validation (200) and testing (200) records for each fold of cross-validation. The database is severely unbalanced with very few examples provided for positive class (arousal regions). Hence, to avoid achieving unsatisfactory results, a downsampling was also applied on majority class data to partially correct for the data imbalance seen during training. Our model performance was evaluated using two metrics: the area under the precision-recall curve (AUPRC) and the area under the receiver operating characteristic curve (AUROC). To measure our model generalization, 3-fold cross-validation process was also employed using 6-minute data windows during training. Finally, the average AUPRC and AUROC values were 0.416 and 0.891, respectively on our testing dataset using DenseNet1 model. Having models DenseNet1 and DenseNet2 output probabilities averaged, the AUPRC and AUROC on testing data were improved to 0.441 and 0.897, respectively on a single fold.