A frozen sleep foundation model gains a local time-frequency branch, reaching F1 0.808 for spindle localization at moderate overlap
Synopsis
Sleep foundation models compress multi-second EEG into compact tokens, leaving too little temporal detail to trace the onset and offset of sleep spindles. One approach keeps the SleepFM encoder frozen and adds a local time-frequency branch that preserves 11–16 Hz sigma-band energy and intra-token position, then uses residual cross-attention to bring in pretrained temporal context. On the MODA benchmark across six folds, event-level F1 is 0.8080±0.0040 at IoU 0.2 and 0.7766±0.0050 at IoU 0.5, with a faster decline at stricter boundary overlap.
Interpretation
Transient EEG events such as sleep spindles, which last 0.5–3 s, can now be localized sample by sample with a frozen sleep foundation model plus a local time-frequency branch, without training a dedicated spindle detector from scratch. Dense detectors such as SUMO and BLAST each learn task-specific spindle representations; foundation models like SleepFM transfer to many downstream tasks but compress multi-second EEG windows into compact tokens, a mismatch with the resolution dense localization needs. On MODA, the model reaches an event-level F1 of 0.8080±0.0040 at IoU 0.2, comparable to SUMO and BLAST, and 0.7766±0.0050 at IoU 0.5, against 0.7682±0.0103 for SUMO and 0.7554±0.0142 for BLAST.
Localization rests mainly on local time-frequency information, with the pretrained representation supplying complementary temporal context. In ablations with frame reinjection off, the frequency-query fusion reference reaches F1 0.8006±0.0051 at IoU 0.2, the frequency-only branch 0.7898±0.0030 and the SleepFM temporal-only representation 0.4024±0.0125; reversing the cross-attention direction drops F1 to 0.5137±0.0204. The same ablation set shows band pooling lowering F1 from 0.8006 to 0.6779 at IoU 0.2 and from 0.7662 to 0.5479 at IoU 0.5, while localized wavelet reconstruction beats linear interpolation, with the gain rising from the 0.7856 side at IoU 0.2 to the 0.7398 side at IoU 0.5.
The implementation uses a fixed Morlet wavelet transform for the time-frequency map and mean-pools over five physiological frequency bands and ordered temporal segments, keeping spectral position inside each token. Global pooling retains only whether relevant activity is present and discards where it sits within the 3-second token; the band-structured readout keeps that position explicitly. Fused tokens drive localized wavelet synthesis from a fixed atom bank, predicting coefficients and position probabilities for each temporal atom and combining overlapping windows by Hann-weighted overlap-add; retained intra-token features are reinjected by nearest-neighbor upsampling and refined by a five-block residual temporal convolutional network.
Perspective
Evaluation is confined to the MODA benchmark, using 749 N2 EEG excerpts from MASS, 180 subjects and 5,342 expert-consensus spindle events, with signals resampled to 128 Hz and inputs C3, F3 and O1, upper vertical EOG substituting when F3 is unavailable; 36 subjects form a fixed test set and the remaining 144 are split into 120/24 training and validation subjects across six folds.
At stricter overlap thresholds the model declines faster than SUMO and BLAST, and the authors state that very precise boundary alignment remains challenging; the specific values at those thresholds are not given.