Transferring Structured Knowledge in Unsupervised Domain Adaptation of a Sleep Staging Network

Yoo, Chaehwa; Lee, Hyang Woon; Kang, Je-Won

doi:10.1109/jbhi.2021.3103614

Cited by 27 publications

(32 citation statements)

References 41 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In this study, we both used the full Sleep-EDF-SC database (referred to as Sleep-EDF-SC-78) and a subset of PSG records from 20 subjects (referred to as Sleep-EDF-SC-20), which corresponds to an earlier version of the Sleep-EDF-SC database that has been extensively studied in the literature. As recommend in [4], [22], [32], only the in-bed parts (from lights off time to light on time) of the PSG recordings were used in our study. Following [4], we converted sleep annotations to meet the AASM standard.…”

Section: Sleep-edf-scmentioning

confidence: 99%

“…The source models in our experiments were trained following the non-meta-learning setup (baseline 1) in the original paper. The source models are subsequently adapted to each subject from the target domain by DSA [22] most transfer tasks using EEG. On UCD and MASS SS2, the accuracy gains by AdaDSA over MetaSleepLearner are 9.83% and 6.56%, respectively.…”

Section: Compared With Related Workmentioning

confidence: 99%

“…To reduce the labeling cost on target domains, a few unsupervised domain adaptation approaches have been proposed [20]- [22]. In principle, these approaches enhance the source models' performance on target domains by learning domaininvariant features through a set of adversarial losses.…”

Section: ) Compared To Previous Domain Adaptation Workmentioning

confidence: 99%

“…In comparison, the proposed methods are source-free and have minimal parameters to optimize. As one of the state-of-the-art, Yoo et al [22] used a similar experimental setting with our study (transferring pre-trained deepSleepNet+ on MASS to Sleep-EDF-SC-20 and Sleep-EDF-ST dataset), providing a benchmark for us to directly compare with. As shown in table V, both DSA and AdaDSA have attained higher performance than that reported in [22].…”

Section: ) Compared To Previous Domain Adaptation Workmentioning

confidence: 99%

“…The efficacy of this framework on state-of-the-art networks still needs to be further validated. Unsupervised Domain adaptation approaches using adversarial learning were proposed to learn domain-invariant representations of sleep samples from different databases [20]- [22]. In these works, specific network architecture was designed and trained simultaneously using samples from source and target datasets.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Unsupervised Domain Adaptation by Statistics Alignment for Deep Sleep Staging Networks

Fan

Zhu

Jiang

et al. 2022

IEEE Trans. Neural Syst. Rehabil. Eng.

View full text Add to dashboard Cite

Deep sleep staging networks have reached top performance on large-scale datasets. However, these models perform poorer when training and testing on small sleep cohorts due to data inefficiency. Transferring welltrained models from large-scale datasets (source domain) to small sleep cohorts (target domain) is a promising solution but still remains challenging due to the domain-shift issue. In this work, an unsupervised domain adaptation approach, domain statistics alignment (DSA), is developed to bridge the gap between the data distribution of source and target domains. DSA adapts the source models on the target domain by modulating the domain-specific statistics of deep features stored in the Batch Normalization (BN) layers. Furthermore, we have extended DSA by introducing cross-domain statistics in each BN layer to perform DSA adaptively (AdaDSA). The proposed methods merely need the well-trained source model without access to the source data, which may be proprietary and inaccessible. DSA and AdaDSA are universally applicable to various deep sleep staging networks that have BN layers. We have validated the proposed methods by extensive experiments on two stateof-the-art deep sleep staging networks, DeepSleepNet+ and U-time. The performance was evaluated by conducting various transfer tasks on six sleep databases, including two large-scale databases, MASS and SHHS, as the

show abstract