Automatic Sleep Staging with Multi-Scored Datasets via Deep Learning

Sleep is a key point in our physical and mental well-being. To date, Polysomnography (PSG) is the gold standard in the analysis of sleep disorders. Sleep scoring procedure has always been affected by intra and inter-operator variability. Introducing AI into clinical practice could save physicians time and increase the agreement between different sleep centres. This Thesis aims to investigate alternative methods in the use of multi-scored datasets in Automatic Sleep Staging. The model used is an existing Deep Learning (DL) architecture: DeepSleepNet-Lite. Two experiments have been performed on two open-source multi-scored datasets. Experiment-1 introduces a new method called Empirical Label Smoothing. It is a variant of Uniform Label Smoothing which uses an empirical distribution derived from consensus among a cohort of physicians. Several analyses are performed to assess its validity, such as performance and model calibration. Monte Carlo Dropout and Query Procedure are employed to estimate and analyse the model uncertainty. Hypnodensity Graphs Analysis is exploited to quantify the similarity between predicted probabilities and the clinicians’ consensus. Experiment-2 investigates the influence of physicians’ agreement on the training of a DL architecture. The same model is trained first on all data, then only on epochs with a high degree of agreement. Performance, calibration, and uncertainty of the model are analysed. Experiment-1 shows no significant performance improvement in training a model with Empirical Label Smoothing compared to Uniform Label Smoothing. However, a calibration enhancement of the model is achieved. In addition, the Hypnodensity Graphs Analysis shows a higher degree of similarity. Experiment-2 proves that training a model only on certain epochs can improve performance and calibration. In the future, both experiments should be performed on the same database to compare the results.