KIT's FF-TRUST Beats Corrupted Sleep Labels Across Clinical Domains by 7 Points
A new benchmark and framework called FF-TRUST tackles sleep staging when labels are noisy and training data comes from mismatched clinical sources.
- New NL-DGSS benchmark tests noisy-label learning under multi-source domain shift for sleep staging across 5 datasets.
- Existing methods like DISC, ELR, and FedDSHAR degrade sharply when label noise and domain shift coexist.
- FF-TRUST adds time and frequency EMA-based early learning regularization plus a forward-free confidence diversity loss.
- At 0.6 asymmetric noise, FF-TRUST beats FedDSHAR by 4.42% ACC and SleepDG by 6.94% ACC.
- Model keeps SleepDG's 6.5M-parameter backbone; gains come entirely from the new regularization objectives.
- Code and benchmark promised at github.com/KNWang970918/FF-TRUST; accepted to MLESP 2026 at IEEE BIBM.
Automated sleep staging maps EEG and EOG recordings to five classes: Wake, N1, N2, N3, and REM. Published models often assume consistent recording hardware, similar patient populations, and reliable human annotations. Clinical datasets violate all three assumptions, with particularly high scorer disagreement around transitional stages such as N1.
Researchers from Karlsruhe Institute of Technology, ETH Zurich, Hunan University, and INSAIT address these conditions in a new paper. The work introduces a benchmark for noisy-label, multi-source domain generalization in sleep staging and proposes FF-TRUST, a training method designed for physiological time series. The paper was accepted to MLESP 2026 at IEEE BIBM.
When label noise meets domain shift
Sleep-stage annotations are costly and subjective, and agreement between human scorers varies by class. Combining recordings from several institutions adds differences in sensors, acquisition protocols, demographics, and clinical conditions. Label errors therefore occur alongside shifts in the underlying signal distribution.
The authors call this setting Noisy Labels in Multi-Source Domain-Generalized Sleep Staging, or NL-DGSS. A model trains on several source datasets and is evaluated on a held-out dataset that represents an unseen clinical domain.
The benchmark adapts five noisy-label methods to the SleepDG framework: NPN, JAL, DISC, ELR, and FedDSHAR. These methods were largely developed for independent and identically distributed image data. Their performance falls under the benchmark’s combination of corrupted labels, multimodal biosignals, class imbalance, and cross-dataset distribution shift.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
