Our pipeline consists of a hand-crafted preprocessor and a neural network classifier. We applied linear and nonlinear transformations on the physiologic signals in order to gain features in both time- and frequency domain. The proposed algorithm was trained on 994 manually annotated records of polysomnographic records. Most of the features were generated from the EEG signal such as power spectral density, and entropy. We extracted features for the variation in the respiratory pattern from both the respiratory airflow and the chest electromyography signals. The heart rate (BPM) feature was generated from the ECG signal using peak detection. All the features were normalized. These 45 features were resampled in 21 non-continuous (past and future) moments around the current timestamp, and fed into a 3-layer neural network in order to assign a probability of arousal at each second. A 1D layer extracted tendencies, while 2 fully connected layers learned combinations and probabilities. Arousal samples were enriched during training, as they consist of a small part of the data. 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: AUROC=0.822, AUPRC=0.228. Our 10-fold cross validation results show stable performance: [0.243, 0.247, 0.229, 0.225, 0.234, 0.276, 0.246, 0.189, 0.234, 0.242] averaging 0.237 AUPRC.