[home] [Personal Program] [Help]

---

Session: Poster Session 1 (Even numbers)
Session starts: Thursday 26 January, 16:00
Presentation starts: 16:00

---

Jaap van der Aar (Technische Universiteit Eindhoven)

Abstract:
Sleep stage scoring, an essential component of sleep disorder diagnostics, classifies 30-second recordings as either Wake, N1, N2, N3, or rapid-eye-movement (REM) sleep. Conventionally, polysomnograhy (PSG) recordings are manually scored by clinicians, a labour-intensive process which can suffer from inter-rater variability. In the recent years, a plethora of automated neural networks for sleep scoring have been developed for single-electrode electroencephalography (EEG), including SleepNet, SeqSleepNet, and DeepSleepNet. Although showing high performance, training of these models is often limited to specific electrode locations and large heterogeneous datasets of young, healthy subjects. In contrast, the clinical practice deals with a range of sleep disorders, heterogeneous patient characteristics, and a limited amount of data. Also, recordings from the specific electrode location are not always available, either because of a different setup or bad signal quality. Therefore, we study to which extent these well-trained models can be utilized when data characteristics change. First, TinySleepNet (a computational lighter version of DeepSleepNet) is trained on a dataset of 94 healthy subject, using the F3-M2 EEG electrode derivation and 10-fold cross validation. Next, we repeat the experiment while changing either the electrode location (C3-M2; Cz-Fz; F3-F4), the population (age-matched mild-to-moderate obstructive sleep apnea; age-matched psychophysiological insomnia), or any combination of above-mentioned electrode locations and populations. For each changing characteristic, we study whether the pre-trained model can sufficiently predict sleep stages in the new data, and if not, whether partial retraining with a limited amount of new data is possible, or full retraining of the model is required. Results show to which extent a widely used sleep stage model, trained on a healthy and heterogenous population, can be used in the clinical setting. We show under which conditions the model can be used directly, when a small sample of additional data is required, and when full retraining is necessary. Furthermore, the implications of this study can pave the road for more accurate automated sleep staging in new wearable EEG devices, where electrode locations differ, signal quality can suffer, and data availability is limited.