acceptodds
Under review as a conference paper at ICLR 2027

SPAN: Set-Prediction Adaptation for Continuous Biosignal Event Decoding

Abstract

Continuous biosignal applications, including seizure monitoring, sleep analysis and motor-imagery decoding, require predictions with an accurate class, onset and offset. Yet pretrained biosignal foundation models are almost universally adapted through per-window classifiers, whose outputs must be thresholded, smoothed and merged into events after the fact. We show that this adaptation interface, and not only the pretrained representation, is a decisive factor in event-level quality. We introduce SPAN, a temporal set-prediction framework that maps tokens from heterogeneous biosignal backbones into a shared physical-time representation and directly predicts event intervals via multiscale temporal features, anchor-conditioned queries and duration-dependent attention. Using a paired protocol that fixes the pretrained initialization, training data and event scorer, we compare SPAN against dense per-token fine-tuning (Dense-FT) across six public EEG foundation-model backbones, four tasks and three seeds. SPAN improves test event F1 at IoU in all 72 seed-matched comparisons, with average gains of 4.5–8.3 percentage points per task, and reduces onset error on matched seizure, motor-imagery and arousal events. A further comparison against a plain set decoder on identical backbone features shows that these gains come from SPAN's combined design rather than set prediction alone. On densely overlapping artifact annotations, boundary localization and fragmentation remain difficult, pointing to a concrete direction for future work. Together, these results establish the adaptation interface as a substantial, previously underexamined lever for transferring pretrained biosignal representations to temporally localized event decoding, and position SPAN as a strong, broadly validated default for this setting.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG foundation models, continuous event detection, temporal set prediction, transfer learning, seizure detection, motor imagery, self-supervised representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_eienOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks how much the adaptation head matters when a pretrained EEG foundation model is used for event detection. SPAN maps backbone tokens to physical time and decodes a set of intervals. It uses a multiscale memory, anchor-conditioned queries, a duration-scaled Gaussian attention bias, denoising queries and an auxiliary dense head. The baseline is dense per-token fine-tuning (DENSE-FT: a two-layer MLP per token plus tuned post-processing). Pairs share checkpoint, data, adapter and scorer. The study covers six backbones, four tasks (seizures, motor imagery, artifacts, arousals) and three seeds. SPAN wins all 72 pairs on event F1 at IoU>0.5, by 4.5-8.3 points per task (Table 1). It adds a DANCE-style plain set decoder (Table 2), a component ablation (Table 3) and boundary and fragmentation analyses (Section 5.2).

Critical strengths
  1. The paired design is careful, and I enjoyed reading it. Each pair shares checkpoint, data, adapter, training caps and scorer (L062-064, L730-731), and all 72 differences stay positive when the dense checkpoint is picked by event F1 instead (Table 15). A useful warning for per-second evaluation: the dense head has higher per-second AUPRC in 21 of 24 backbone-task means, but finds fewer events (L352-354, Table 10).

  2. The paper shows where SPAN fails and why. On TUAR, onset error rises for every backbone and fragmentation reaches 70.5-77.8% (Tables 6-7). Appendix D traces this to duplicates (best-aligned candidate 0.43 vs. 0.46 s, Table 13), and Appendix H tests suppression. L482-484 also say frozen encoders and random initialisation were not tested.

Critical weaknesses
  1. Dense baseline without temporal context. Because of this, the comparison doesn't separate out set prediction. Each backbone encodes 10-s windows that are joined in time (L122-123), and DENSE-FT applies a two-layer MLP to each token (L130). Apart from post-hoc smoothing and merging, nothing in it works across windows. SPAN's decoder has four Transformer layers of width 256 that attend over the whole 60-s or 300-s chunk (L199-203, L747, Table 5). So the gain in Table 1 mixes two changes. One is predicting intervals as a set. The other is a learned temporal head with chunk-level context and more parameters. This matters most when events are longer than a window. TUSZ references have a median of 34.9 s (Table 11), and on TUSZ the dense output has a median signed onset error of +7.50 s (L895-896). The paper frames the contrast as set prediction versus "per-window classifiers" (abstract L014-016). But dense segmentation with temporal context, like the U-Time family it cites (L094), is also a standard way to decode events. A dense baseline with the same multiscale memory and a Transformer of similar size, with per-token outputs and the same post-processing, would separate the two. Head parameter counts for both methods would help too. I could only find the 0.2 M of the pyramid (L1078).

  2. Attribution. The claims go further than the controls. The abstract says the gains "come from SPAN's combined design rather than set prediction alone" (L026-027) and that the ablation "traces the gain to duration-dependent attention locality and query denoising" (L027-028). The plain set decoder is the authors' own DANCE-style implementation (L971-972). It has half the width and half the heads (128-d and 4 heads vs. 256-d and 8, Table 17). On arousal it was also trained under a shorter cap than the primary pairs (L1024-1025). There it gets 5.7 and 3.9 F1, far below even DENSE-FT (24.9 and 24.0, Table 2). This looks like a training failure, not a limit of set prediction, so the +24.4 gain on that task says little. The ablation (Table 3) tests only two tasks, and on arousal every interval includes zero. Nothing in Table 3 explains the 5.5-point arousal gain. The anchor-conditioned queries, the iterative refinement and the auxiliary dense head are not ablated. The last one gives SPAN the same dense supervision DENSE-FT gets. I'd ask for a width- and budget-matched plain decoder, an ablation of the auxiliary dense head, and an abstract limited to what Tables 2-3 show.

A few smaller points. The training caps don't match. Appendix B gives 12 h for TUSZ and arousal (L730), but Appendix F refers to "the 23 h cap of the primary LaBraM and REVE pairs" (L1025). On TUAR, SPAN's precision falls from 38.8% to 25.1%, and false positives go up from 2,238 to 8,293 per 24 h (L822-825). I couldn't find a false-alarm rate for TUSZ, which seizure-detection work usually reports. The text says Table 8 adds a false-positive rate, but the table doesn't show one (L808-809). So "a strong, broadly validated default" (L032-033) is stronger than the TUAR results support. Typos: "We transfers" (L106), "sill" (L431).

Questions for authors
  1. How does SPAN compare on TUSZ and arousal with a dense head that has chunk-level temporal context and a size similar to SPAN's decoder? What are the head parameter counts of DENSE-FT, the plain set decoder and SPAN?
  2. What happens to SPAN without the auxiliary dense head?
  3. Which training cap did the primary arousal pairs use, 12 h or 23 h? With the same cap and width as SPAN, is the plain set decoder on arousal still below DENSE-FT?
  4. What are the false alarms per 24 h on TUSZ for both methods at the selected operating points?
2 comments

Dear Reviewer eien,

Thank you for your thorough review, recognizing both our paired design and our diagnostic analyses of duplicate predictions on TUAR, and we really appreciate that you enjoy reading it.

We have uploaded a revised manuscript (changes highlighted in blue) and run additional experiments to address your concerns directly below. In brief:

  • New chunk-level Transformer dense baseline (Dense-Ctx) and head parameter counts (W1, Q1): reported below.
  • Abstract claims scoped to the experimental evidence and auxiliary dense head ablated (W2, Q2): abstract revised and reported below.
  • Training budgets corrected and Plain set decoder retrained at SPAN's width and training budget (W2, Q3): Appendix B and reported below.
  • False-alarm results (Q4): Section 5.2 and Appendix E, Table 18 and Figure 4.
  • Typos fixed.

W1 / Q1: Dense baseline with temporal context

Response: To separate the effect of chunk-level temporal context from set prediction, we implemented Dense-Ctx. Dense-Ctx inserts SPAN's multiscale pyramid and four Transformer decoder layers (width 256, 8 heads, feed-forward 512) between the adapter and Dense-FT's per-token MLP. Each token attends to the same chunk as SPAN and to the multiscale memory. Targets, loss, checkpoint selection, training caps and the post-processing search are Dense-FT's. We also evaluated Dense-Ctx (threshold only), which sweeps only the confidence threshold like SPAN, without tuning smoothing windows, merge gaps, or minimum durations. Due to limited computation resources, we report event F1 at IoU >> 0.5 for LaBraM and REVE with one seed 37 below:

TaskBackbone (Seed 37)Dense-FTDense-CtxDense-Ctx, threshold onlySPAN
TUSZLaBraM45.757.748.959.5
TUSZREVE47.957.240.258.3
ArousalLaBraM25.029.628.231.6
ArousalREVE24.226.625.327.5

Adding chunk-level multiscale temporal context improves Dense-FT, while SPAN remains ahead in all four comparisons (+0.9 to +2.0 points). Moreover, without post-processing, Dense-Ctx falls 14.4 points below SPAN on TUSZ, because SPAN outputs intervals directly while a dense head depends on the smoothing, merge-gap and minimum-duration search to convert probabilities into events.

Below are head parameters, excluding the adapter projection that all heads share (0.05--0.53 M by backbone).

HeadDense-FTDense-CtxPlain set decoder (paper)Plain set decoder, matched widthSPAN
Parameters0.07 M3.50 M1.64 M4.36 M3.60--3.66 M

Dense-Ctx (3.50 M) closely matches SPAN (3.60--3.66 M), and the width-matched 256-d plain set decoder (4.36 M) is slightly larger than SPAN.


W2 / Q2: Attribution and the auxiliary dense head

Response: We agree and have tightened the abstract to reflect what Tables 2 and 3 show.

To evaluate the auxiliary dense head, we trained SPAN without the auxiliary dense head on LaBraM and REVE across all four tasks on one seed 37 due to limited computation resources. The dense head reads the adapter tokens and feeds nothing into the decoder, so this changes only the training. Table below reports Event F1 at IoU >> 0.5:

Backbone (Seed 37)MethodTUSZMITUARArousal
LaBraMDense-FT45.718.214.825.0
LaBraMFull SPAN →\rightarrow No dense head (Δ\Delta)59.5 →\rightarrow 55.3 ($-$4.2)22.6 →\rightarrow 22.3 ($-$0.3)19.8 →\rightarrow 19.9 (+0.1)31.6 →\rightarrow 28.3 ($-$3.3)
REVEDense-FT47.926.613.724.2
REVEFull SPAN →\rightarrow No dense head (Δ\Delta)58.3 →\rightarrow 51.2 ($-$7.1)34.0 →\rightarrow 25.6 ($-$8.4)18.7 →\rightarrow 19.3 (+0.6)27.5 →\rightarrow 30.7 (+3.2)

We use the auxiliary dense head to strengthen the features the decoder reads. One-to-one matching supervises only the few matched queries, while the dense head gives every adapter token a per-second target, these tokens form the decoder's memory and condition each query's initial content. From the results above, the auxiliary dense head is beneficial on TUSZ and MI, while having a neutral or mixed effect on TUAR and arousal, where set-level supervision alone is already dense. Moreover, even without the auxiliary dense head, SPAN outperforms Dense-FT in 7/8 comparisons.

Q3: Training caps and the matched plain set decoder

Response: Dense-FT and SPAN share the same cap in all 72 primary pairs. On arousal task, Appendix F was correct, Appendix B misstated the caps, and we have corrected it (Initial runs were scheduled with a conservative 23 h wall-clock limit, this was reduced to 12 h for left runs after observing consistent early stopping within 12 h). Based on the log, table below reports completed Dense-FT and SPAN runs on arousal for LaBraM and REVE:

RunsCapStopped byOptimizer stepsTraining time
Dense-FT and SPAN (12)23 hearly stopping, 12/12123k--232kall within 12 h (max 11.4 h)
Component ablation (18)23 hearly stopping, 18/18137k--260k17/18 within 12 h, except REVE −- denoising seed 7: 12.7 h
Plain set decoder (6)12 hearly stopping, 6/6109k--410kall within 12 h (max 10.8 h)

All 12 Dense-FT and SPAN runs reached within 12 h, after 123--232k optimizer steps. 17/18 ablation runs early stopped within 12 h, after 137--260k optimizer steps, only REVE −- denoising seed 7 early-stopped at 12.7 h.

To discover the gap between SPAN and plain set decoder, we retrained the plain set decoder at SPAN's width (256-d, 8 heads) and the same training caps, for LaBraM and REVE with three seeds each. Table below reports event F1 at IoU >> 0.5:

TaskDense-FTPlain set (paper)Plain set, matched width and capSPANSPAN −- matched plain [95% CI]
TUSZ45.249.150.455.7+5.3 [+1.4, +9.0], 5/6 pairs
MI23.216.719.227.1+7.9 [+5.6, +9.7], 6/6 pairs
TUAR13.510.410.218.9+8.7 [+8.1, +9.4], 6/6 pairs
Arousal24.54.84.729.2+24.6 [+22.0, +26.8], 6/6 pairs

With SPAN's width and cap, the plain set decoder is still far below Dense-FT on arousal. From the results, matching width and budget changes the plain decoder by at most 2.5 points, so neither explains SPAN's advantage. On arousal, where the gap is largest, when we also replace DANCE's matching loss with SPAN's one in both matching and training, the SPAN remains 2.6 points higher [+0.0, +4.8], 5/6 pairs. Because pure IoU loss from DANCE has zero gradient when predicted and reference intervals do not overlap, the plain set decoder fails to localize short, sparse arousals. Equipping the plain decoder with SPAN's L1 + dIoU objective resolves this optimization bottleneck.


Q4: False alarms on TUSZ

Response: We have updated Table 8 to include false positives per 24 h at both overlap thresholds across all tasks. On TUSZ, at the default operating points, both methods produce comparable false-alarm rates:

TUSZ, false positives per 24 hDense-FTSPAN
IoU >> 0.33945
IoU >> 0.54749

Because the default operating point maximizes validation event F1 rather than targeting a clinical false-alarm burden, we also reselected the threshold of every run under fixed false-alarm limits (Section 5.2; Appendix E, Table 18 and Figure 4). Under ≤1\le 1 FP per hour on TUSZ, SPAN achieves 31.8% recall and 42.0% event F1 at IoU>0.5\text{IoU} > 0.5 compared to 15.1% recall and 21.2% event F1 for Dense-FT. On TUAR, held to Dense-FT's own FP rate, SPAN is 1.0 point lower there. From practitioner's perspective, SPAN is the better choice for events that rarely overlap within a class like seizures, motor imagery and arousals, while Dense-FT is the safer choice for densely overlapping events like artifacts under a tight false-alarm restriction.

We hope these new results and revisions fully address your concerns. Please let us know if any further clarification would be helpful during the discussion period!

Reviewer_Kmj8OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how pretrained EEG foundation models should be adapted for continuous event decoding, where each prediction requires a class, onset, and offset. It proposes SPAN, a temporal set-prediction interface that converts heterogeneous backbone tokens into a common physical-time representation, constructs a multiscale temporal memory, and decodes event intervals using anchor-conditioned queries, duration-dependent Gaussian attention, iterative boundary refinement, and denoising supervision. A dense auxiliary head is used during training and for per-second evaluation. The main experiment fully fine-tunes six pretrained backbones on seizure, motor-imagery, artifact, and arousal detection. Under a paired protocol, SPAN improves event F1 at IoU > 0.5 over dense per-token fine-tuning in all 72 backbone-task-seed comparisons, with task-level average gains of 4.5--8.3 percentage points. Additional experiments compare SPAN with a plain set decoder, ablate three components, and analyze onset error, recall, precision, fragmentation, checkpoint selection, and duplicate suppression. SPAN improves onset localization on three tasks, but produces substantially more fragmented predictions, especially for overlapping artifact annotations.

Critical strengths
  1. The paper is well motivated to go beyond the pretrained representation and study the effect of the adaptation interface. The paired evaluation then supports this motivation convincingly: each comparison shares the pretrained initialization, data, adapter, and event scorer, while covering six substantially different EEG backbones, four tasks, and three seeds. SPAN improves event F1 at IoU > 0.5 in all 72 seed-matched comparisons, with average gains of 4.5--8.3 percentage points across tasks
  2. Beyond a single aggregate score, the onset, precision/recall, per-class, per-second, fragmentation, matched-subset, and suppression analyses distinguish better detection from better probability ranking and event coherence.
Critical weaknesses
  1. On TUAR, reference classes may overlap, but DENSE-FT uses a single (K+1)-way softmax target and assigns only one class to each token. It is therefore structurally unable to express simultaneous events that SPAN retains in its set targets. A multi-label dense baseline would be needed to separate the benefit of set prediction from this output-space restriction.
  2. The component ablations cover only two backbones and two tasks. Moreover, on arousal, the confidence interval for the effect of removing each component includes zero, so the current three-seed experiment does not establish that any individual component reliably improves event F1.
  3. SPAN lowers per-second AUPRC in 21 of 24 backbone-task means and increases fragmentation in every backbone-task combination. On TUAR, it raises predictions per reference from 0.37 to 1.14 and increases false positives per 24 hours from 2,238 to 8,293. The appendix further shows that overlap-based suppression cannot remove this fragmentation without reducing F1 at IoU > 0.5. These trade-offs are consequential in continuous monitoring, where false-alarm burden, coherent event counts, and computational cost can matter as much as thresholded event F1.
Questions for authors
  1. How much of SPAN's gain remains under more closely matched controls: (i) a multi-label sigmoid dense head that can represent overlapping TUAR classes, and (ii) a plain set decoder matched to SPAN in memory resolution, width, interval loss, and intermediate supervision, with anchors, Gaussian locality, and denoising then added incrementally?
  2. How should a practitioner choose among the methods when false-alarm burden or coherent event counting is the deployment constraint? Some discussion, along with analysis such as operating-point curves, would be helpful.
3 comments

Dear Reviewer Kmj8,

Thank you for your detailed review and for recognizing the value of adaptation interface study and the analysis beyond the aggregate score useful.

We have uploaded a revised manuscript (changes highlighted in blue) and run experiments to address your concerns directly below. In brief:

  • Multi-label dense baseline on TUAR (W1, Q1(i)): Section 5.1, Appendix B and Appendix D, Table 14.
  • Thresholds under deployment constraints: matched false-positive rate, fixed false-alarm limits and at most one prediction per reference (W3, Q2): Section 5.2 and Appendix E, Table 18 and Figure 4.
  • Plain set decoder retrained at SPAN's width and training budget, and the component ablation (W2, Q1(ii)): reported below.

W1 / Q1(i): TUAR against a multi-label dense baseline

Response: We trained a dense baseline with independent per-class sigmoid outputs and multi-hot targets, extracted per class with the same post-processing search as Dense-FT, on all six backbones and three seeds (Section 5.1, Appendix B, Appendix D, Table 14):

TUAR, event F1Dense-FTSigmoid denseSPANSigmoid −- Dense-FT [95% CI]SPAN −- Sigmoid [95% CI]
IoU >> 0.513.516.618.0+3.1 [+2.6, +3.6], 18/18 pairs+1.4 [+0.7, +2.0], 14/18 pairs
IoU >> 0.320.625.226.5+4.6 [+4.0, +5.4], 18/18 pairs+1.2 [+0.4, +2.0], 15/18 pairs

The sigmoid head recovers about 70% of SPAN's TUAR gain, so part of the original 4.5-point gain reflects the single-label dense output. SPAN remains ahead by +1.4 points [+0.7, +2.0], with positive gains in 14/18 pairs and on five of the six backbones.


W2 / Q1(ii): Matched set-decoder controls and component ablation

Response: We agree that a broader ablation would strengthen the attribution to individual components. Each ablation arm requires full fine-tuning of the backbone, so our computational budget limited the ablation to two backbones and two tasks. We also acknowledge that the current ablation supports task-specific conclusions only, so we do not claim a general effect for any single component. Future study would be extending it to more backbones and tasks, and identifying components that are consistent across different task categories (long/short event, overlap/non-overlap event).

To discover the gap between SPAN and plain set decoder, we retrained the plain set decoder at SPAN's width (256-d, 8 heads) and the same training caps, for LaBraM and REVE with three seeds each. Table below reports event F1 at IoU >> 0.5:

TaskDense-FTPlain set (paper)Plain set, matched width and capSPANSPAN −- matched plain [95% CI]
TUSZ45.249.150.455.7+5.3 [+1.4, +9.0], 5/6 pairs
MI23.216.719.227.1+7.9 [+5.6, +9.7], 6/6 pairs
TUAR13.510.410.218.9+8.7 [+8.1, +9.4], 6/6 pairs
Arousal24.54.84.729.2+24.6 [+22.0, +26.8], 6/6 pairs

From the results, the gain persists across four tasks, matching width and budget changes the plain decoder by at most 2.5 points, so neither explains SPAN's advantage. On arousal, where the gap is largest, when we also replace DANCE's (1−IoU)(1-\mathrm{IoU}) matching cost and with SPAN's L1 + (1−dIoU)(1-\mathrm{dIoU}) in both matching and training, the SPAN remains 2.6 points higher [+0.0, +4.8], 5/6 pairs. Because pure IoU loss from DANCE has zero gradient when predicted and reference intervals do not overlap, the plain set decoder fails to localize short, sparse arousals. Equipping the plain decoder with SPAN's L1 + dIoU objective resolves this optimization bottleneck.

W3 / Q2: False alarms, event counting and choosing a method in deployment

Response: Thank you for this suggestion. The default threshold maximizes validation event F1 at IoU >> 0.3, which is not a deployment criterion. We therefore reselected the threshold of every run on the validation split under three deployment constraints and scored it once on test (Section 5.2; Appendix E, Table 18), and we added the operating-point curves that helps for illustration (Figure 4). A false positive is a prediction not matched at IoU >> 0.5. The table below reports Δ\DeltaF1 at IoU >> 0.5, SPAN −- Dense-FT across six backbones ×\times three seeds with 95% CI:

TaskFP/h ≤\le Dense-FT's own rateFixed FP limit≤\le 1 prediction per reference
TUSZ+7.2 [+5.2, +9.3]≤\le 1/h: +20.8 [+17.3, +24.3]+8.6 [+7.0, +10.4]
MI+7.6 [+5.5, +9.5]≤\le 100/h: +5.4 [+3.6, +7.0]+7.9 [+6.1, +9.7]
TUAR$-1.0[1.0 [-$1.7, $-$0.4]≤\le 50/h: $-2.2[2.2 [-$2.7, $-$1.8]+3.4 [+2.7, +4.1]
Arousal+5.3 [+3.6, +7.1]≤\le 5/h: +5.6 [+3.7, +7.5]+6.0 [+4.2, +7.8]

Under the same false-alarm limit, SPAN stays better on TUSZ, MI and arousal. The curves in Figure 4 show the same picture over the whole range, where SPAN's recall is higher at every FPR on those tasks. At one FP per hour on TUSZ, SPAN recalls 31.8% of seizures against 15.1% for Dense-FT. TUAR is the exception, Dense-FT's recall saturates near 10%, while SPAN's rises only as the FP budget grows. Under the same event counting, SPAN is better on all four tasks, in 18 of 18 pairs each, including TUAR.

Practical guidance. Our practical guidance is to select the threshold on the validation split under the deployment constraint, a false-alarm budget or a count constraint. SPAN is then the better choice for events that rarely overlap within a class like seizures, motor imagery and arousals, while Dense-FT is the safer choice for densely overlapping events like artifacts under a tight false-alarm restriction.

We hope these new results and revisions fully address your concerns. Please let us know if any further clarification would be helpful during the discussion period!

Thank you for the additional experiments and clarification. Most of my concerns have been addressed. However, for the component-attribution study on arousal, matching the localization objective reduces SPAN’s advantage from 24.6 to 2.6 points, with the reported confidence interval touching zero. This could potentially indicate that the original large gap is mainly due to the localization objective rather than SPAN’s architectural components. The revised paper should report and discuss this result explicitly, and limit claims about individual architectural components (particularly for the arousal task studied here).

Beyond this, I have no further questions and would like to maintain my positive score.

Reviewer_6B69OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how the downstream adaptation interface affects continuous event decoding with pretrained EEG foundation models. SPAN maps heterogeneous backbone tokens into a common physical-time representation and predicts event classes and intervals using multiscale temporal memory, anchor-conditioned queries, duration-dependent attention, iterative refinement, and denoising supervision. The evaluation compares SPAN with dense per-token fine-tuning across six backbones, four tasks, and three seeds, with both encoders fully fine-tuned from the same pretrained initialization. The paper reports improvements in event F1 at IoU > 0.5 in all 72 paired comparisons, with task-average gains of 4.5–8.3 percentage points. Additional experiments examine checkpoint selection, a plain set decoder, component ablations, boundary accuracy, and duplicate suppression. The central contribution is a broadly evaluated adaptation framework and evidence that event-oriented adaptation substantially affects the downstream utility of pretrained EEG representations.

Critical strengths
  1. Broad and consistent evidence for a practically important adaptation choice. The paired evaluation across six pretrained backbones and four distinct event-decoding tasks is the strongest aspect of the paper. Within each primary comparison, the pretrained initialization, data, adapter architecture, and event scorer are shared. The improvements at IoU > 0.5 occur in all 72 seed-matched comparisons, and Appendix E shows that they persist when the dense baseline also selects checkpoints using event F1. This makes the central finding substantially more convincing than an improvement demonstrated on a single backbone or dataset.

  2. Useful analysis beyond the headline metric. The paper carefully distinguishes event detection from sample-level discrimination and prediction coherence. In particular, it reports that better event F1 can coexist with lower auxiliary per-second AUPRC and greater fragmentation. The shared-reference onset analysis addresses the concern that boundary improvements could simply reflect differences in which events are detected. The TUAR diagnostics and suppression experiments also make the remaining limitations explicit. Together with the common physical-time interface and the component ablations, these analyses provide useful guidance for researchers adapting EEG foundation models. I view this combination of consistent empirical gains and informative diagnostics as sufficient to support a weak acceptance.

Critical weaknesses
  1. The plain-set comparison does not fully isolate the source of the improvement. Table 17 shows that SPAN and the plain decoder differ in decoder width (256 versus 128), memory construction, interval parameterization, and several training losses. Appendix F additionally states that the arousal plain-set runs used an 11-hour cap, whereas the primary LaBraM and REVE pairs used 23 hours; this also needs reconciliation with the 12-hour description in Appendix B. The component ablations help, particularly on TUAR, but none of the individual removals produces a clear event-F1 effect on arousal. These limitations do not invalidate the primary comparison with Dense-FT, but they weaken the attribution of the large arousal gain over the plain decoder to SPAN's specific design.

  2. The overlapping-artifact setting remains less conclusive. Appendix B assigns only one class to each dense target cell even when artifact classes overlap, whereas set targets preserve the individual intervals. Consequently, the TUAR comparison partly reflects a difference in target expressiveness, and an overlap-aware multilabel dense baseline would make the conclusion stronger. SPAN also produces substantial same-class duplication: Table 13 reports approximately 2.9 candidates per matched reference, and onset error remains worse on references detected by both methods. Appendix H already investigates suppression, but it does not resolve this tradeoff without reducing F1 at IoU > 0.5 on TUAR. The authors acknowledge these issues appropriately; nevertheless, they limit the claim that SPAN is a broadly suitable default for densely overlapping events.

Questions for authors
  1. Clarification: Please reconcile the training budgets in Appendices B and F and specify the actual caps used for the arousal Dense-FT, SPAN, plain-set, and component-ablation runs. Were Dense-FT and SPAN matched within every primary pair? Learning curves or completed optimization-step counts would help assess the impact of the shorter plain-set budget.

  2. Targeted comparison: Is there evidence that the large gain over the plain set decoder persists when decoder width and training budget are matched? A representative comparison would materially strengthen the attribution to SPAN's design.

  3. Targeted comparison: How much of the TUAR improvement remains against a dense baseline with independent per-class sigmoid outputs and multilabel targets? If such an experiment is unavailable, a breakdown between overlapping and non-overlapping reference events would help assess the scope of the current evidence.

1 comment

Dear Reviewer 6B69,

Thank you for your constructive feedback and for recognizing the breadth of our paired evaluation and the value of our diagnostic analysis beyond the headline metric.

We have uploaded a revised manuscript (changes highlighted in blue) and run experiments to address all of your questions directly below. In brief:

  • Multi-label dense baseline on TUAR (W2, Q3): Section 5.1, Appendix B and Appendix D, Table 14.
  • Operation-point under deployment constraints (W2): Section 5.2 and Appendix E, Table 18 and Figure 4.
  • Training budgets corrected (Q1): Appendix B.
  • Plain set decoder retrained at SPAN's width and training budget (W1, Q2): reported below.

Q1: Training budgets in Appendices B and F

Response: Dense-FT and SPAN share the same cap in all 72 primary pairs. Appendix F was correct, Appendix B misstated the arousal caps, and we have corrected it (Initial runs were scheduled with a conservative 23 h wall-clock limit, this was reduced to 12 h for left runs after observing consistent early stopping within 12 h). From the training log, table below reports completed Dense-FT and SPAN runs on arousal for LaBraM and REVE:

RunsCapStopped byOptimizer stepsTraining time
Dense-FT and SPAN (12)23 hearly stopping, 12/12123k--232kall within 12 h (max 11.4 h)
Component ablation (18)23 hearly stopping, 18/18137k--260k17/18 within 12 h, except REVE −- denoising seed 7: 12.7 h
Plain set decoder (6)12 hearly stopping, 6/6109k--410kall within 12 h (max 10.8 h)

All 12 Dense-FT and SPAN runs reached within 12 h, after 123--232k optimizer steps. 17/18 ablation runs early stopped within 12 h, after 137--260k optimizer steps, only REVE −- denoising seed 7 early-stopped at 12.7 h, so the cap barely affects any results.


W1 / Q2: Plain set decoder at matched width and budget

Response: The gain persists across all four tasks. We retrained the plain set decoder at SPAN's width (256-d, 8 heads) and the same training caps, for LaBraM and REVE with three seeds each. Table below reports event F1 at IoU >> 0.5, mean over the two backbones:

TaskDense-FTPlain set (paper)Plain set, matched width and capSPANSPAN −- matched plain [95% CI]
TUSZ45.249.150.455.7+5.3 [+1.4, +9.0], 5/6 pairs
MI23.216.719.227.1+7.9 [+5.6, +9.7], 6/6 pairs
TUAR13.510.410.218.9+8.7 [+8.1, +9.4], 6/6 pairs
Arousal24.54.84.729.2+24.6 [+22.0, +26.8], 6/6 pairs

Matching width and budget changes the plain decoder by at most 2.5 points, so neither explains SPAN's advantage. On arousal, where the gap is largest, when we also replace DANCE's (1−IoU)(1-\mathrm{IoU}) matching cost and with SPAN's L1 + (1−dIoU)(1-\mathrm{dIoU}) in both matching and training, the SPAN remains 2.6 points higher [+0.0, +4.8], 5/6 pairs. Because pure IoU loss from DANCE has zero gradient when predicted and reference intervals do not overlap, the plain set decoder fails to localize short, sparse arousals. Equipping the plain decoder with SPAN's L1 + dIoU objective resolves this optimization bottleneck.

W2 / Q3: TUAR against a multi-label dense baseline

Response: We trained a dense baseline with independent per-class sigmoid outputs and multi-hot targets, extracted per class with the same post-processing search as Dense-FT, on all six backbones and three seeds (Section 5.1, Appendix B, Appendix D, Table 14):

TUAR, event F1Dense-FTSigmoid denseSPANSigmoid −- Dense-FT [95% CI]SPAN −- Sigmoid [95% CI]
IoU >> 0.513.516.618.0+3.1 [+2.6, +3.6], 18/18 pairs+1.4 [+0.7, +2.0], 14/18 pairs
IoU >> 0.320.625.226.5+4.6 [+4.0, +5.4], 18/18 pairs+1.2 [+0.4, +2.0], 15/18 pairs

The sigmoid head recovers about 70% of SPAN's TUAR gain, so part of the original 4.5-point gain reflects the single-label dense output. SPAN remains ahead by +1.4 points [+0.7, +2.0], with positive gains in 14/18 pairs and on five of the six backbones.

We hope these new results and revisions fully address your concerns. Please let us know if any further clarification would be helpful during the discussion period!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.