Evaluating Convolutional and Recurrent Neural Network Architectures for Respiratory-Effort Related Arousal Detection During Sleep

This work evaluates the performance of convolutional and recurrent neural networks on the task of detecting Respiratory Effort-Related Arousals (RERAs). Feature time-series were extracted from EEG, EOG, CHIN, CHEST, ABDOMINAL, AIRFLOW, SaO2, and ECG and normalized on a per-subject basis. Next, multi-timescale windows from these time-series were associated with the presence or absence of RERA during the window forming the data for model training. More than 1 million RERA-windows and 17 million no-arousal windows were used for model training, and more than 200K RERA-windows and 4 million no-arousal windows were used for testing and validation. Google Cloud ML Engine was used to select model hyperparameters using the validation data. The model with the best hyperparameter combination evaluated on the test set achieved an AUC-ROC score of 0.916 and AUC-PR score 0.573.