From unsupervised to semi-supervised adversarial domain adaptation in electroencephalography-based sleep staging

Abstract

Objective. The recent breakthrough of wearable sleep monitoring devices results in large amounts of sleep data. However, as limited labels are available, interpreting these data requires automated sleep stage classification methods with a small need for labeled training data. Transfer learning and domain adaptation offer possible solutions by enabling models to learn on a source dataset and adapt to a target dataset. Approach. In this paper, we investigate adversarial domain adaptation applied to real use cases with wearable sleep datasets acquired from diseased patient populations. Different practical aspects of the adversarial domain adaptation framework \hl{are examined}, including the added value of (pseudo-)labels from the target dataset and the influence of domain mismatch between the source and target data. The method is also implemented for personalization to specific patients. Main results. The results show that adversarial domain adaptation is effective in the application of sleep staging on wearable data. When compared to a model applied on a target dataset without any adaptation, the domain adaptation method in its simplest form achieves relative gains of 7%-27% in accuracy. The performance on the target domain is further boosted by adding pseudo-labels and real target domain labels when available, and by choosing an appropriate source dataset. Furthermore, unsupervised adversarial domain adaptation can also personalize a model, improving the performance by 1%-2% compared to a non-personal model. Significance. In conclusion, adversarial domain adaptation provides a flexible framework for semi-supervised and unsupervised transfer learning. This is particularly useful in sleep staging and other wearable EEG applications.

Feature matching as improved transfer learning technique for wearable EEG

“…This study uses PSG recordings of 90 patients [23]. It was recorded at the sleep laboratory of the University Hospitals Leuven (UZ Leuven) from January 2021 to September 2022.…”

“…The active learning experiments start from the model obtained with the optimal setting based on the channel and data selection experiments. Using active learning to query informative labels, the model is then finetuned to personalize it to individual patients with the semi-supervised adversarial domain adaptation approach from [23]. The training set of 45 patients is used as the source dataset, and each selected individual recording of the test set as the target dataset for adversarial domain adaptation.…”

“…Disease influences brain activity and, consequently, PSG recordings, creating uncertainty in the data. Moreover, models trained on data acquired with one measurement setup or device may not generalize well to data obtained using a different measurement protocol, resulting in the well-known distribution shift problem [22,23,24,25,26,27]. Hence, measurement factors can also contribute to uncertainty.…”

L-SeqSleepNet: Whole-cycle Long Sequence Modelling for Automatic Sleep Staging

“…78.9 ± 0.6 0.703 ± 0.008 67.9 ± 0.5 67.0 ± 0.3 94.2 ± 0.1 92.0 ± 0.7 25.8 ± 1.4 74.1 ± 0.6 79.7 ± 1.1 68.0 ± 1.5 L-SeqSleepNet 72.3 ± 0.6 0.607 ± 0.008 57.3 ± 0.8 57.8 ± 0.9 92.4 ± 0.2 89.9 ± 0.6 7.5 ± 1.8 65.8 ± 0.9 72.4 ± 1.8 50.8 ± 2.4 SeqSleepNet* [4] 75.0 ± 0.4 0.647 ± 0.006 62.7 ± 0.7 61.3 ± 0.8 93.2 ± 0.1 90.6 ± 0.5 23.1 ± 1.1 71.1 ± 0.5 72.1 ± 0.7 56.6 ± 3.1 SeqSleepNet [4] 70.4 ± 1.5 0.578 ± 0.024 54.1 ± 1.5 55.0 ± 1.7 91.7 ± 0.5 87.4 ± 1.9 0.5 ± 0.5 64.6 ± 1.7 69.3 ± 1.7 48.6 ± 5.0 ADA pers ‡ [47] 72.8 0.618 Secondly, concerning L-SeqSleepNet itself, halving the sequence length to 100 epochs (50 minutes, roughly a half sleep cycle) results in a drop of 0.3% on overall accuracy while doubling it to 400 epochs (200 minutes, roughly two sleep cycles) causes negligible consequence to the performance. This suggests that the sequential information in one sleep cycle is probably all we need for sleep staging.…”

Automated remote sleep monitoring needs machine learning with uncertainty quantification

“…Sleep staging methods are mostly reported on healthy adults, and typically achieve lower performance on diseased subjects [25]. Moreover, sleep staging models trained on data acquired with one measurement setup or device typically do not generalize well to data acquired with a different measurement protocol, because of the well-known distribution shift problem [26,27]. This problem is aggravated for wearable devices, which record data from diverse non-standard locations.…”