acceptodds
Withdrawn from ICLR 2027

Decision-Identifiable Control of Partially Observed Epilepsy Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Selecting a suppressive target in an epileptic network requires accounting for pathways that intracranial electrodes cannot observe. PULSE connects local probe responses to suppression benefits and uses the uncertainty in those benefits to decide where to measure next. In a stable nonnegative Hawkes network with calibrated probes, integrated observed responses and baseline activity determine stationary outcomes under local gain suppression, even when hidden structure is not uniquely identified. A target's baseline rate and two response summaries suffice; correcting for recurrent returns explains why the strongest evoked response need not identify the best target. In 24 network families whose members differ in hidden structure and transient response, all 2,880 intervention conditions share the benefits predicted by this formula, and predictions from finite measurements err by 2.38 percentage points of baseline observed burden on average across 3,072 interventions. The same summaries also direct measurement: in 96 benchmark networks, decision-focused acquisition reaches 2.35% selection regret at 2,560 trial blocks against 5.33% for uniform allocation, and at 5,120 blocks it keeps regret at 1.72% against 3.11% for an acquisition rule that reconstructs the responses more accurately. Independent post-action simulations of the intervened networks confirm the predicted benefits, and experiments on hidden coverage, probe calibration, and response windows delineate the required measurement conditions, including a 250-ms benchmark built on empirical human response timing. The information a suppression decision needs can thus be read from local probes, without reconstructing the hidden network.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Epilepsy Networks, Neurostimulation, EEG

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zxgMOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper appears to propose a method for target selection / seizure focus identification for responsive neurostimulation based on Hawkes network models.

Critical strengths

None.

Critical weaknesses
  • The problem setting is not sufficiently outlined. It is unclear what the goal of the paper is.
  • The paper is hard to follow. Concepts are not introduced. The writing is lacking structure and a logical progression. For instance, in L048 a statement about nonnegative Hawkes networks is made without without any explanation of what they are, why they would be a relevant model for the considered problem, or citation.
  • It is unclear what the machine learning contribution of this work is. ICLR may not be the right venue for this work.
Questions for authors
  • The paper would benefit from a better description of the general problem setting. That responsive neurostimulation is a treatment approach for drug-resistant epilepsy, briefly what it is, why and how the seizure focus is currently identified, and what the paper tries to improve in this setting.
  • While the paper appears to positions itself as a method for seizure focus identification, the bulk of the paper appears to focus on theoretical network models without justifying any connection or relevance to the former problem. If the point is to analyse or prove any properties of these network models, it might be better to frame the paper as such and perhaps note seizure focus identification as a potential application.

  • Figures should be vector graphics
Reviewer_3cPsOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to select a suppressive target when part of a network is unobserved. Under a stable nonnegative Hawkes model and calibrated probes, PULSE derives stationary benefits of local gain suppression from integrated observed responses and baseline rates, then allocates probes according to uncertainty in the estimated benefits. Validation is primarily simulation-based; human recordings provide propagation-timing priors and a supplementary analysis of response representations.

Critical strengths
  1. Decision identifiability provides a clear and useful theoretical contribution. Under explicit assumptions about the model, probes, and intervention, suppression benefits can be compared without recovering the full hidden structure. The recurrent-return correction also explains why the strongest evoked response need not identify the best target, making the connection between measurements and decisions concrete and testable.

  2. The experiments provide useful checks of the identification result within the assumed model. Response-equivalent networks test whether different hidden structures yield the same predicted benefits, and independent post-intervention simulations compare those predictions with simulated event-rate reductions. I also appreciate the separate evaluation of response-reconstruction accuracy and target-selection regret. These experiments support the identification result independently of the covariance ablation discussed below.

Critical weaknesses
  1. The evidence for the covariance component is internally inconsistent. Figure 5a, interpreted using the legend in Figure 5b, shows lower regret for PULSE without covariance than for full PULSE at 640, 1,280, and 2,560 blocks. At 1,280 blocks, the values read from the plot are approximately 4.3% for PULSE without covariance and 5.5% for full PULSE. Table 3 instead reports 6.18% for PULSE without covariance and 4.15% for full PULSE, and Appendix F.3 states that covariance reduces regret at every tested budget. If Table 3 is correct, correcting the figure could resolve this concern. If the plotted ordering is correct, the claimed consistent benefit of covariance, particularly at intermediate budgets, is not supported. I cannot currently assess how much of the acquisition improvement should be attributed to this component.

  2. The decision consequences of probe miscalibration remain unclear. Assumption 1 excludes direct recruitment of hidden populations. In the diagnostic, ε is the probability that a probe additionally activates one randomly selected hidden population. At ε = 0.3, value NMAE remains 14.2% even with exact expected responses (Section 6.3; Appendix H.1). This establishes persistent bias in predicted benefits, but the diagnostic does not report whether the selected target changes or how much selection regret increases. That distinction is particularly important for a paper that separates estimation accuracy from decision quality. Appendix B.2 handles known input spread across observed populations, but no procedure is given to detect or bound unknown hidden recruitment. The human-timing benchmark still generates suppression outcomes from the model, while the supplementary human-response analyses do not validate the selected targets against observed suppression outcomes. I therefore find the evidence insufficient to assess target-selection robustness when this probe assumption fails.

  3. Additional reporting and presentation comments

  • Figures 6b–c contain further numerical discrepancies. At ε = 0.6, the estimated value error is approximately 26% read from Figure 6b, versus 21.8% in Section 6.3 and Appendix H.1. At the 0.5-second window, Figure 6c gives approximately 18% value error and 46% response error, both read from the plot, versus 15.2% and 38.5% in Appendix H.2. Please reconcile the plotted and written values. These are reporting concerns distinct from the reversed covariance ordering above.

  • A compact protocol table for Section 6 would make the experiments easier to compare. In particular, total versus per-target trial budgets and known versus estimated baseline rates differ across experiments and currently require tracking the text and captions.

-Please label Figure 7b explicitly as a simulation benchmark using empirical human timing. This would make the scope of the human-data evidence clear directly from the figure.

Questions for authors
  1. After reconciling Figure 5, Table 3, and Appendix F.3, what are the paired regret differences between full PULSE and the no-covariance variant at each budget, with uncertainty? Do the corrected results support a consistent covariance benefit or a benefit restricted to particular budgets?

  2. For the existing hidden-recruitment experiment, could you report selection regret and the fraction of networks selecting an optimal target across ε, using both estimated and exact expected responses? This would show whether the reported value bias changes decisions or mainly shifts the predicted benefits.

  3. What can be established about direct hidden recruitment from the existing pulse/sham recordings, and what additional calibration information would be needed to detect or bound it? Could any available bounds certify that the selected target remains optimal or bound its regret?

  4. Section 3 interprets nonnegativity at the level of pathological events rather than synaptic sign. What specific event definition is intended for the proposed application, and what empirical evidence supports nonnegative interactions for that event stream?

Reviewer_SnFyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In clinical practice, patients with drug-resistant epilepsy may be treated by suppressing activity in specific regions of the brain, for example, through responsive neurostimulation. Selecting which regions to suppress is challenging because implanted electrodes record activity from only part of the brain network, while unrecorded regions may still relay neural activity.

This paper addresses the problem at a theoretical and simulation level. It proposes PULSE, a method that predicts the benefit of suppressing each recorded site based on its responses to brief stimulation “probes” and adaptively selects where to probe next to identify the best target under a limited measurement budget. Brain activity is modeled as a Hawkes process, with the assumption that each probe adds exactly one known event at the stimulated site without directly activating hidden regions. The intervention is modeled as reducing the activity of a single recorded site by a fixed fraction. Under these assumptions, the authors prove that probe responses and baseline activity at each site are sufficient to predict the benefit of suppressing that site, without requiring knowledge of the hidden network.

The experiments use a new synthetic benchmark consisting of 96 simulated networks, each with 8 recorded populations and up to 12 hidden populations, along with a variant whose event timing is derived from published human recordings involving 74 patients. PULSE is compared with alternative approaches that select targets based on firing rate, the strongest evoked response, or a Hawkes model fitted to passive recordings.

Performance is measured using selection regret, defined as the proportion of the maximum achievable benefit lost by selecting a suboptimal site. PULSE achieves the lowest regret of 1.72% with 5,120 trial blocks, compared with 2.85% for the best passive Hawkes model, 11.29% for firing-rate-based selection, and 19.65% for strongest-evoked-response selection. With a budget of 2,560 blocks, adaptive probing reduces regret by more than half compared with uniform probing (2.35% vs. 5.33%). The predicted benefits differ from the true benefits by an average of 2.38 percentage points of baseline activity. On the human-timing benchmark, PULSE achieves 1.52% regret, compared with 2.45% for uniform probing.

However, all experimental data, including the ground-truth outcomes, are generated using the same model assumed by the proposed method.

Critical strengths

The paper presents a novel formulation of an existing clinical problem: selecting a suppression target for epilepsy when part of the brain network is unobserved. Rather than reconstructing the hidden network, as prior network-modeling approaches implicitly require, the authors show that two summary measures of each site’s probe response, together with its baseline activity, are sufficient to predict the benefit of suppression (Theorem 2). The method also uses this result to guide subsequent measurements toward identifying the optimal suppression target.

Critical weaknesses
  1. Lack of validation on real patient data. All experiments use synthetic event streams generated from the same Hawkes model assumed by the proposed method. Real patient data are used only to derive response-latency distributions for setting simulated delays (Section 6.4). Therefore, the paper demonstrates that PULSE works when its modeling assumptions hold, but does not establish whether these assumptions accurately reflect real epileptic activity.

  2. Limited comparability with prior work. Performance is reported only in terms of selection regret relative to an exact simulated ground truth. This metric can only be computed in simulation because the true benefit of suppressing every candidate site cannot be observed in patients. Prior work instead evaluates agreement with clinically identified seizure-onset zones, overlap with resected tissue, or the prediction of seizure freedom. Since PULSE is not evaluated using these metrics or compared with established target-selection methods, it is difficult to assess whether it offers improvements over existing approaches. The baselines in Table 1 consist primarily of simple heuristics, generic Hawkes model fits, and variants of PULSE itself.

  3. Insufficient justification for the Hawkes model. The paper models epileptic activity using a linear, excitatory-only, stationary Hawkes process but does not adequately explain why this model is appropriate for epilepsy. Section 3 acknowledges that excluding inhibition is “a modeling choice about the event stream rather than about synaptic sign.” However, the paper does not sufficiently justify this assumption or examine how violations of it might affect the proposed method.

Questions for authors
  1. Why was PULSE evaluated only on synthetic data generated from the same Hawkes model assumed by the proposed method? Could it be evaluated on a publicly available intracranial recording dataset with clinical labels? Such an evaluation would help determine whether the improvements reported in Table 1 generalize to real patients and how PULSE compares with existing methods evaluated using clinical metrics rather than simulated regret.

  2. What motivates the use of a linear, excitatory-only Hawkes process to model epileptic activity, and what does an “event” represent in real intracranial recordings? If the underlying dynamics involve inhibition or nonlinear responses, as considered in the neural-mass and dynamical-systems models cited in the paper, would Theorem 2 and the resulting target-selection approach still hold, at least approximately?

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

Related papers

Discussion (0)

Sign in to comment.