Adapting Sleep Foundation Models for Dense Sleep Spindle Localization
Organizations: Department of Electronics and Informatics (ETRO), Vrije Universiteit Brussel (VUB), Brussels, Belgium · Idiap Research Institute, Martigny, Switzerland · Interuniversity Microelectronics Centre (IMEC), Heverlee, Belgium
Abstract
Adapting sleep foundation models to dense electroencephalography (EEG) event localization is challenging because multi-second representations may not preserve sufficient temporal detail for precise boundary recovery. We address sleep spindle localization by adapting a frozen SleepFM encoder with a local time-frequency pathway that preserves spindle-relevant spectral content and intra-token position. The local representation drives localization, while residual cross-attention incorporates SleepFM temporal context. Dense predictions are obtained through localized token-to-sample reconstruction and temporal refinement. Tested on the Massive Online Data Annotation (MODA), the proposed model achieves an event-level of at IoU 0.2 across six folds. Ablations show that localization primarily depends on local spectral and sub-token temporal information, while the pretrained representation provides complementary context. These results indicate that dense adaptation benefits from preserving information below the pretrained token scale. Code will be released upon acceptance.
Figures & tables
| Frame reinjection | Fusion | Representation | Readout | Reconstruction | |||
|---|---|---|---|---|---|---|---|
| Metric | On | Off (ref.) | Temp. query | Freq. only | Temp. only | Band pooling | Linear interp. |
| [email protected] | 0.8080 0.0040 | 0.8006 0.0051 | 0.5137 0.0204 | 0.7898 0.0030 | 0.4024 0.0125 | 0.6779 0.0119 | 0.7856 0.0089 |
| [email protected] | 0.8075 0.0042 | 0.7989 0.0054 | 0.4921 0.0230 | 0.7892 0.0044 | 0.3707 0.0148 | 0.6490 0.0132 | 0.7828 0.0097 |
| [email protected] | 0.7766 0.0050 | 0.7662 0.0058 | 0.3909 0.0298 | 0.7442 0.0213 | 0.2606 0.0112 | 0.5479 0.0142 | 0.7398 0.0170 |
Electroencephalography
Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.
See all 198 papers