State-of-the-art sleep arousal detection evaluated on a comprehensive clinical dataset.

Aiming to apply automatic arousal detection to support sleep laboratories, we evaluated an optimized, state-of-the-art approach using data from daily work in our university hospital sleep laboratory. Therefore, a machine learning algorithm was trained and evaluated on 3423 polysomnograms of people with various sleep disorders. The model architecture is a U-net that accepts 50 Hz signals as input. We compared this algorithm with models trained on publicly available datasets, and evaluated these models using our clinical dataset, particularly with regard to the effects of different sleep disorders. In an effort to evaluate clinical relevance, we designed a metric based on the error of the predicted arousal index. Our models achieve an area under the precision recall curve (AUPRC) of up to 0.83 and F1 scores of up to 0.81. The model trained on our data showed no age or gender bias and no significant negative effect regarding sleep disorders on model performance compared to healthy sleep. In contrast, models trained on public datasets showed a small to moderate negative effect (calculated using Cohen's d) of sleep disorders on model performance. Therefore, we conclude that state-of-the-art arousal detection on our clinical data is possible with our model architecture. Thus, our results support the general recommendation to use a clinical dataset for training if the model is to be applied to clinical data.