Using Auxiliary Loss To Improve Sleep Arousal Detection With Neural Network

Our pipeline consists of a hand-crafted preprocessor and a neural network classifier. We applied transformations on the physiologic signals to gain features in both time- and frequency domains. The proposed algorithm was trained on 994 annotated records of polysomnographic signals. Most of the features were generated from the EEG signal such as power spectral density, and entropy. We extracted features from the EOG, EMG, airflow, and ECG signals too. All the features were normalized.These 68 features were resampled in 21 non-continuous moments around the current timestamp, and fed into a 3-layer neural network in order to assign a probability of arousal at each second. Arousal samples were enriched during training to battle data imbalance. Additional (auxiliary) losses can guide the network to learn high-level concepts, even though they will not be evaluated. We used sleep stages as additional training targets, which were easier to learn than arousals despite being multi-class. This approach slightly increased arousal AUPRC.Our submitted results for the entire test set were evaluated: AUPRC=0.42.Our 10-fold cross validation results for the AUPRC are the following: [0.47110, 0.41672, 0.44305, 0.42842, 0.44644, 0.47969, 0.45082, 0.49320, 0.45913, 0.41278] averaging 0.450.