Adapting Sleep Foundation Models for Dense Sleep Spindle Localization
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "Adapting Sleep Foundation Models for Dense Sleep Spindle Localization".
Dev: Here's the gist of the paper. Adapting sleep foundation models to dense electroencephalography event localization by preserving sub-token information is presented in this work The proposed architecture adapts a frozen SleepFM encoder with a local time-frequency pathway that preserves spindle-relevant spectral content and intra-token position This adaptation results in an event-level F1 of 0.8080±0.0040 at IoU 0.2 across six folds on the Massive Online Data Annotation (MODA) benchmark Pruning redundant reasoning tokens The core challenge addressed is the temporal-resolution mismatch between pretrained sleep representations and dense spindle localization, where increasing output resolution alone cannot recover precise location The proposed method preserves local information in a parallel branch and reuses it during sample-level reconstruction Specifically, the architecture keeps a pretrained SleepFM encoder frozen and adds a local time-frequency branch that captures sigma-band oscillations and the transient morphology of sleep spindles How the training signal is built The training objective jointly supervises spindle presence, event boundaries, and temporal structure using several loss functions The segmentation objective combines focal loss and Dice loss to supervise spindle presence Boundary regression is optimized using Smooth L1 loss on onset and offset distances normalized by the sampling frequency An auxiliary objective regularizes the task-specific frequency representation toward the pretrained SleepFM representation via a DKL divergence constraint Furthermore, a frequency loss encourages spindle-related spectral structure by comparing energy in
11, 16: Hz with surrounding context Where the gains come from The localization primarily depends on local spectral and sub-token temporal information, while the pretrained representation provides complementary context The key contributions include dense sample-level adaptation of a frozen SleepFM encoder and a physiologically structured time-frequency representation preserving local spectral content and intra-token position Localized wavelet synthesis outperforms linear interpolation for reconstruction, showing an improvement from 0.0150 at IoU 0.2 to 0.0264 at IoU 0.5 Frame reinjection yields an additional gain, increasing F1 from 80.75±42% to 80.89±4% at IoU 0.
Title and authors: Rosa: So, to recap what we're looking at here, they took these big sleep foundation models and modified them by adding this local pathway that keeps track of the specific spectral details and where each tiny piece of data sits within a sleep cycle.
Dev: Right. It’s not just about getting a general idea of the sleep state; it’s about getting down to the exact moment a spindle pops up, which is much harder with dense recordings.
Rosa: Exactly. The whole point is they needed something more than just the standard model to pinpoint those events precisely when you have lots of data points packed together.
Taro: So, what's the actual result they’re showing? They claim an event level F1 score of zero point eight zero eight at an intersection over union of zero point two on this massive online data annotation benchmark.
Dev: That number is pretty solid for localization, but the catch is that it only works because they adapted this frozen SleepFM encoder to handle those dense sample-level details.
Rosa: It means that if you can capture that specific local spectral content and the sub-token temporal position effectively, you get better than just relying on what the foundation model already knows about sleep generally.
Taro: What’s the mechanism they use for that refinement? I want to know how they actually connect those local details to a usable spindle map.
Dev: They use a localized wavelet synthesis method for reconstruction. Instead of just smoothly connecting things, it predicts the temporal patterns and their position within the window, constrained by a fixed atom bank.
Rosa: That’s what helps them recover that detail; it outperforms simple linear interpolation when you look at their reconstruction accuracy metrics. They see improvement going from zero point oh one five zero at an intersection over union of zero point two to zero point two six four at an intersection over union of zero point five.
Taro: So, the system can actually predict where things are in time and space within a window, even when the data is dense. Does that mean it can handle messy real-world scenarios better than older methods?
Dev: It means it’s getting closer to being able to handle the noise and complexity of real sleep data because that local information acts like a very precise guide for where the true event boundaries are.
Rosa: But there is a limitation they mention; this whole reconstruction process still relies on that fixed atom bank, which constrains how much temporal detail it can recover beyond what’s already encoded in those atoms.
Taro: So, the system isn't perfectly free to predict any arbitrary pattern outside of that pre-set structure. What about the training side? How do they get this local branch to actually learn the right things?
Dev: They use a multi-objective loss function setup. They supervise spindle presence with focal loss and dice loss, they optimize boundary regression using smooth l one loss on onset and offset distances, normalized by the sampling frequency.
Rosa: And there's also this auxiliary objective that regularizes the task specific frequency representation toward the pretrained SleepFM representation by representing tokens as diagonal gaussian distributions in a shared latent space.
Taro: That sounds like they’re using some kind of soft constraint to make sure their new local information aligns with what the main model already understands about sleep physics.
Dev: It’s an attempt to guide the learning process so it doesn't just create noise, but actually creates something physiologically structured, which is what drives that localization pathway we talked about.
Rosa: So, in a nutshell, they show that by adapting the foundation model densely and adding this structured local branch for spectral content and position, you can achieve better event localization even when you are dealing with high-density EEG data.
Taro: It suggests that future work needs to focus on making that fixed atom bank more dynamic or improving how they combine this local refinement if we want to push past those current reconstruction limits.
The paper's summary: Taro: So, we've seen how they built that localized pathway and used wavelet synthesis for reconstruction to get those better numbers on the benchmark. What else are they suggesting as improvements to this whole approach?
Rosa: They also highlight how their context fusion is asymmetric, which means frequency tokens query the pretrained sequence through residual cross-attention.
Dev: That’s a big deal because it lets the local time-frequency representation maintain its own localization pathway while still pulling in that broader SleepFM temporal context.
Taro: So they aren't just doing parallel processing; they’re making it interactive, where the fine detail can ask the general model for help.
Rosa: Right, and on the training side, they introduced an auxiliary objective to regularize that task specific frequency representation toward what the pretrained SleepFM already knows about sleep.
Dev: They do this by representing those tokens as diagonal gaussian distributions in a shared latent space. It’s a way to encourage that spindle related spectral structure from the very start of the training.
Taro: So, it's like they're using a gentle nudge during training to make sure the model learns to look for spindles in a way that matches its existing knowledge base about sleep.
Rosa: Exactly, and they also show how frame reinjection adds another layer of gain when you’re looking at the F1 score, boosting it from eighty point seven five plus or minus forty two percent up to eighty point eight nine plus or minus four percent at an intersection over union of zero point zero.
Dev: That small increase in F1 when the intersection over union is near zero shows that even these minor additions really help refine the final prediction under tight constraints.
Taro: But what’s the actual limitation they flag? What is this system *not* doing well, or where does it stop working gracefully?
Rosa: They point out that while frame reinjection helps, the overall architecture still relies on a fixed atom bank during reconstruction. This constrains how much temporal detail it can recover beyond what’s already encoded in those atoms.
Dev: That means the system isn't totally unconstrained at the sample level; it’s constrained by those fixed atoms, which limits its ability to predict truly novel temporal patterns outside of that structure.
Taro: So, it’s good enough for dense data localization, but if you wanted to go beyond that, you have to rethink how that fixed atom bank interacts with the local synthesis step.
The paper's improvements: Rosa: So we're wrapping up this look at "Adapting Sleep Foundation Models for Dense Sleep Spindle Localization." The main thing is that they showed how adding this local pathway and structured training objectives lets you get a very precise event level F1 score of zero point eight zero eight when you have dense data.
Dev: It means that if you can capture those local spectral and sub-token temporal details, you don't need to rely solely on the general context the foundation model has about sleep.
Taro: I think what this changes for people in real-world robotics or autonomy is that we can get much more accurate localization even when things are noisy and dense because the system is looking at the immediate local environment, not just a broad overview.
Rosa: Exactly. And they proved that frame reinjection adds another layer of gain, pushing that F1 up to eighty point eight nine plus or minus four percent at an intersection over union of zero point zero.
Dev: The safety case here is that the performance relies primarily on those local spectral and sub-token temporal information, so if you can get that right locally, you get good results regardless of how complex the global context is.
Taro: If we want to push past this current reconstruction limit, we'll need to look at making that fixed atom bank more dynamic or figuring out a better way to combine that local refinement with the global information.
Rosa: Right, so they found a way to adapt these powerful foundation models by injecting fine-grained local spectral content and using specialized reconstruction techniques for better localization on dense EEG.
Dev: It’s a solid result for getting those precise boundaries down below the standard token scale while keeping the core model frozen.
Taro: We’re definitely watching how they handle that constraint with the fixed atom bank, because that's where we'll go next to see if we can get even finer temporal control.
Conclusion: Rosa: So we're wrapping up our look at "Adapting Sleep Foundation Models for Dense Sleep Spindle Localization." Basically, they showed how adding this local pathway and structured training objectives lets you get a very precise event level F1 score of zero point eight zero eight when you have dense data.
Dev: It means that if you can capture those local spectral and sub-token temporal details, you don't need to rely solely on the general context the foundation model has about sleep.
Taro: I think what this changes for people in real-world robotics or autonomy is that we can get much more accurate localization even when things are noisy and dense because the system is looking at the immediate local environment, not just a broad overview.
Rosa: Exactly. And they proved that frame reinjection adds another layer of gain, pushing that F1 up to eighty point eight nine plus or minus four percent at an intersection over union of zero point zero.
Dev: The safety case here is that the performance relies primarily on those local spectral and sub-token temporal information, so if you can get that right locally, you get good results regardless of how complex the global context is.
Taro: If we want to push past this current reconstruction limit, we'll need to look at making that fixed atom bank more dynamic or figuring out a better way to combine that local refinement with the global information.
Rosa: Right, so they found a way to adapt these powerful foundation models by injecting fine-grained local spectral content and using specialized reconstruction techniques for better localization on dense EEG.
Dev: It’s a solid result for getting those precise boundaries down below the standard token scale while keeping the core model frozen.
Taro: We’re definitely watching how they handle that constraint with the fixed atom bank, because that's where we'll go next to see if we can get even finer temporal control.
Rosa: Yeah, it shows how preserving those sub-token temporal details through a local branch can complement the global context from the SleepFM encoder for better event localization.
Dev: That’s the core finding of "Adapting Sleep Foundation Models for Dense Sleep Spindle Localization." It’s a solid result for getting those precise boundaries down below the standard token scale while keeping the core model frozen.
Taro: I think what this implies is that future work needs to focus on making that fixed atom bank more dynamic or improving how that temporal refinement is combined if we want to push past those current reconstruction limits.
Rosa: Right, so they found a way to adapt these powerful foundation models by injecting fine-grained local spectral content and using specialized reconstruction techniques for better localization on dense EEG.
Dev: It’s a solid result for getting those precise boundaries down below the standard token scale while keeping the core model frozen.
Taro: We’re definitely watching how they handle that constraint with the fixed atom bank, because that's where we'll go next to see if we can get even finer temporal control.
Alexis Bollengier, Abel Díaz Berenguer, Olivier Pallanca, Andre Anjos ´, Hichem Sahli
Idiap Research Institute · Department of Electronics and Informatics (ETRO) Vrije Universite Brussel (VUB) · Interuniversity Microelectronics Centre (IMEC)
eess.SP, cs.LG
Submitted: 2026-11-04
Updated: 2026-11-04
Comments: 5 pages, 3 figures, 1 table. Submitted to IEEE ICASSP 2027 (under review)
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 76/100
The gist: Adapting sleep foundation models to dense electroencephalography event localization by preserving sub-token information is presented in this work The proposed architecture adapts a frozen SleepFM…
Key concepts
- SleepFM Encoder
- This is a pre-trained model used as the foundation. It provides rich, general sleep representations that are kept frozen during adaptation. It serves as a powerful starting point to understand the overall structure of sleep data before adding specific localization features.
- Local Time-Frequency Pathway
- A new branch added to the model that focuses specifically on spindle characteristics. This pathway captures important spectral content, like sigma-band oscillations, and tracks the precise position of these events within individual tokens, which is crucial for fine localization.
- Dense Sample-Level Adaptation
- The core technique involves adapting the frozen SleepFM encoder at a fine level for each sample. This allows the model to recover very precise boundaries below its original token scale, significantly improving localization accuracy compared to standard methods.
Terminology
Summary
Adapting sleep foundation models to dense electroencephalography event localization by preserving sub-token information is presented in this work The proposed architecture adapts a frozen SleepFM encoder with a local time-frequency pathway that preserves spindle-relevant spectral content and intra-token position This adaptation results in an event-level F1 of 0. p. 08080±0.0040 at IoU 0. p. 02 across six folds on the Massive Online Data Annotation (MODA) benchmark
Pruning redundant reasoning tokens
The core challenge addressed is the temporal-resolution mismatch between pretrained sleep representations and dense spindle localization, where increasing output resolution alone cannot recover precise location The proposed method preserves local information in a parallel branch and reuses it during sample-level reconstruction Specifically, the architecture keeps a pretrained SleepFM encoder frozen and adds a local time-frequency branch that captures sigma-band oscillations and the transient morphology of sleep spindles
How the training signal is built
The training objective jointly supervises spindle presence, event boundaries, and temporal structure using several loss functions The segmentation objective combines focal loss and Dice loss to supervise spindle presence Boundary regression is optimized using Smooth L1 loss on onset and offset distances normalized by the sampling frequency An auxiliary objective regularizes the task-specific frequency representation toward the pretrained SleepFM representation via a DKL divergence constraint Furthermore, a frequency loss encourages spindle-related spectral structure by comparing energy in [11, 16] Hz with surrounding context
Where the gains come from
The localization primarily depends on local spectral and sub-token temporal information, while the pretrained representation provides complementary context The key contributions include dense sample-level adaptation of a frozen SleepFM encoder and a physiologically structured time-frequency representation preserving local spectral content and intra-token position Localized wavelet synthesis outperforms linear interpolation for reconstruction, showing an improvement from 0.0150 at IoU 0. p. 32 to 0.0264 at IoU 0. p. 35 Frame reinjection yields an additional gain, increasing F1 from 80. p. 375±42% to 80.89±4% at IoU 0. p. 0
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Dense sample-level adaptation of a frozen SleepFM encoder allows for precise boundary recovery below the pretrained token scale, achieving an event-level F1 of 0.8080 at IoU 0.2 on the MODA benchmark.
-
A physiologically structured time-frequency representation is used to preserve local spectral content and intra-token position, which drives localization by preserving
local spectral and sub-token temporal information.
-
The system can perform dense spindle detection by using localized wavelet synthesis to predict both temporal patterns and their position within the window, constrained by a fixed atom bank.
-
Context fusion is asymmetric, where frequency tokens query the pretrained sequence via residual cross-attention, allowing the local time-frequency representation to maintain its primary localization pathway while retrieving complementary context from SleepFM.
-
The system can estimate spindle probability, onset and offset distances, and class (background or boundary) at each sample level by using a Temporal Convolutional Network with five residual blocks on the reconstructed sequence.
-
An auxiliary objective regularizes the task-specific frequency representation toward the pretrained SleepFM representation by representing tokens as diagonal Gaussian distributions in a shared latent space, encouraging spindlerelated spectral structure.
Sources
Related papers
- Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative
- Scalable Attention for 5G NR Channel Estimation
- Low-Altitude Wireless Networks: A Comprehensive Survey
- Sensing-Assisted OFDM with Index Modulation: A Waveform for Optical Wireless Integrated Communication and Localisation
- Iteration-Free Candidate Pruning and an Index-Domain VLSI Architecture for Sort-Free MIMO Box Decoding up to 1024-QAM
- A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization