acceptodds
Under review as a conference paper at ICLR 2027

When Does EEG Decoding Survive Electromagnetic Interference?

Abstract

EEG decoders are trained in electromagnetically clean laboratories but deployed in hospitals, cockpits, and wearable settings. Structured electromagnetic interference collides with neural frequency bands, driving decoders to chance. Removing those frequencies can also erase activity that the task needs. We ask when decoding survives this shift. We introduce the Interference–Neural Confusion benchmark with four physically motivated interference families and a compact sensing-gating-calibration front end. It locates occupied frequencies in a fixed short-time Fourier representation, suppresses those bins, and withholds predictions when sensing is unreliable. For a fixed mask, the output is unchanged by interference whose transform support the mask covers; an operator-norm bound quantifies residual leakage. In subject-disjoint experiments on STEW, narrowband interference at -10 dB drops mean accuracy from 0.858 to 0.504, while gating restores 0.747, with gains on every held-out subject. For left- versus right-hand imagery in BCI Competition IV 2a, accuracy remains 0.556; band ablation points to the task information lost by gating. Prediction-set coverage on STEW rises from 0.478 to 0.854 after gating (clean: 0.915), but broadband sweeps remain largely undetected. These results connect recovery to detectable interference and the preservation of task-relevant rhythms, identifying when spectral gating can support reliable EEG decoding.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG Decoding, Electromagnetic Interference in Biosignals, Time Series and Biosignals, Reliable Machine Learning, Out-of-Distribution Robustness

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9mMFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces the Interference–Neural Confusion by adding four families of simulated electromagnetic interference to two public EEG datasets. It proposes the Sense–Gate–Calibrate (SGC) framework, combining spectral detection, STFT frequency masking, and conformal calibration. Experiments on workload and motor imagery examine how interference detectability and the loss of task information affect decoding performance.

Critical strengths
  1. The research problem has practical value.
  • Electromagnetic interference can overlap with neural frequency bands needed by a task, so removing interference may also remove useful information. The paper examines when the method succeeds and fails.
  1. The controlled evaluation is reasonably systematic.
  • Experiments cover different interference types, multiple interference levels, and evaluation across subjects. The paper also includes the complete four-class motor imagery task and a replication covering all 48 STEW participants.
  1. Both recovery and failure cases are reported.
  • The narrowband interference experiments on STEW show substantial recovery. The motor imagery results, frequency-band ablations, and broadband detection failures also document cases where the method has limited effectiveness or fails.
Critical weaknesses
  1. The added value over conventional filtering needs clarification.
  • How does detecting interference and masking the corresponding frequency bands differ substantively from conventional band-stop or notch filtering? A direct comparison under suitable narrowband conditions would help clarify this.
  1. The realism of the simulated interference has not been validated with measurements.
  • Interference is directly added to previously recorded EEG, assuming that the original signal remains unchanged and only additive noise is introduced. The authors should explain whether the interference considered in real environments only contaminates the recordings or may also affect the underlying brain activity. If the latter is outside the scope of this work, this should be stated, together with the real conditions represented by this noise addition.

  • The simulation also adds the same waveform to all channels, changing only its amplitude. Measurements are lacking to establish which channels and frequency bands are affected by real interference, whether channels are affected differently, and how these differences influence task decoding. Such evidence is needed to support the representativeness of the simulated interference types, parameters, and distribution across channels.

  1. Task coverage is too narrow to generalize the reported gains to other EEG tasks.
  • Only workload and motor imagery are evaluated. Substantial recovery is mainly observed on STEW. Under F1 at SIR −10 dB, gating provides no mean gain in the main left/right motor imagery experiment. The complete four-class task shows improvements of approximately 1.8–3.5 percentage points, but these gains are limited. Increasing the number of participants and extending the motor imagery classes do not broaden the types of tasks evaluated. The evaluation should cover more representative EEG tasks to establish whether other tasks also benefit.
  1. Effectiveness across EEG frequency bands and tasks needs broader validation.
  • EEG activity is commonly analyzed in frequency bands such as alpha, beta, and gamma, and different tasks use these bands and their associated information differently. The authors should explain which bands are affected by real interference, whether multiple bands are affected simultaneously, and what information is lost from different tasks after masking. The paper includes motor imagery band ablations and analysis of failures under broadband interference, but these results are insufficient to establish performance on other bands and tasks. Further validation should examine the overlap between interference bands and bands needed by each task, identifying when the method restores performance, provides limited gains, or removes useful information.
  1. Validation in real recording settings is missing.
  • The authors should clarify whether STEW and BCI Competition IV 2a were collected in electromagnetically shielded environments and what filtering was applied during acquisition. Although the underlying EEG was recorded from real participants, the interference is added digitally afterward. This cannot replace validation with interference occurring during actual acquisition. At minimum, the method should be evaluated on EEG collected in an environment without electromagnetic shielding, using either public data or data collected by the authors. The authors should explain how real EMI is identified and report the decoding gains and their stability. This limitation is acknowledged, but the current simulations alone still make it difficult to assess the method’s effectiveness in real settings.
Questions for authors
  1. Are there EEG recordings from real settings, particularly unshielded environments, that can be used to evaluate the method? Either public data or data collected by the authors could be used. Does the method provide stable gains on these data?

  2. Does the interference in real environments mainly contaminate the recordings, or may it also affect the underlying brain activity? If the latter is outside the scope of this work, please state this. Under what real recording conditions is directly adding simulated noise a suitable model?

  3. Which channels and frequency bands are affected by real interference? Is there measurement evidence supporting the assumption that interference has the same waveform across all channels and differs only in amplitude?

  4. For EEG tasks beyond workload and motor imagery that use frequency bands differently, what task information is affected by the interference? When interference overlaps with bands needed by a task, under what conditions does the method work, provide limited gains, or fail?

Reviewer_Z6f8OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the impact of four different types of electromagnetic interference (slow drifts, bursts, chirps) on EEG decoding. To mitigate these sources of interference, a noise handling “frontend” (no learned parameters) is added before the decoder to gate contaminated frequency bins in the spectral domain, before transforming the data back in the time domain. Experiments on two decoding tasks (mental workload and motor imagery classification; using an EEGNet architecture) show that the proposed approach improves robustness on mental workload, but not on motor imagery, under different simulated interference conditions.

Critical strengths
  1. The paper studies a relevant question for EEG decoding, i.e. test-time noise robustness, which is necessary to resolve to make decoders practical in real-world applications and therefore deserves more research attention.
Critical weaknesses
  1. The paper’s potential significance is reduced by its limited scope: experiments are run on two datasets only, with simulated EMI only (no datasets with real EMI examples to show efficacy of method in real case scenarios); only one decoding pipeline family and model (EEGNet) is evaluated and no handcrafted features baselines or EEG foundation models were tested, while they may react differently to EMI than task-specific models; baselines omit key EEG denoising baselines, especially those relying on spatial filtering (ICA, SSP) which should be particularly relevant when targeting components such as the studied EMI.

  2. The paper conflates two questions which makes its goal unclear. The title and abstract frame it as (1) a study of when EEG decoding survives EMI, while the rest of the paper is focused around (2) the benchmarking of a method to mitigate EMI using a model “frontend”. For (1), I would have expected a more complete study of when EEG decoding fails, encompassing more decoding tasks (including other BCI protocols, cognitive decoding tasks, etc.), and more decoder families (handcrafted baselines, task-specific models, EEG foundation models, etc.). In contrast, the conclusion states that “EEG decoding survives when the interference is visible and the occupied band is expendable”, which is a bit vague and could have been explored in more detail. For (2), I would have expected more baselines, including spatial filtering ones (ICA, SSP), to showcase the particular usefulness of the proposed frontend approach over existing ones.

Questions for authors
  1. If the objective of the paper is about characterising the behaviour of EEG decoders under EMI, does the conclusion change under typical spatial removal noise techniques (ICA, SSP), or over other types of experimental paradigms (SSVEP, ERP, cognitive decoding, etc.)? Does it work similarly on data that contains real (non-simulated) sources of noise?
Reviewer_mayrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates EEG decoding under structured electromagnetic interference (EMI), a practically important distribution shift largely overlooked in existing EEG robustness benchmarks. The authors introduce the Interference‑Neural Confusion (INC) benchmark with four physically‑motivated interference families. They propose Sense‑Gate‑Calibrate (SGC), a fixed spectral preprocessing pipeline combining CFAR‑based interference sensing, STFT spectral gating, and split‑conformal adaptive prediction sets for abstention. The paper provides formal invariance and leakage bounds for spectral masking, and validates on two public EEG datasets (STEW workload classification, BCIC‑LR motor imagery). Empirically, SGC recovers classification accuracy on STEW when narrow‑band interference is detectable and the corrupted frequency bands carry no critical task information. However, gains vanish for broadband chirp interference and for motor‑imagery tasks where discriminative neural information overlaps heavily with interference frequency ranges.

Critical strengths

The critical strengths of the paper are as follows

  1. Most EEG domain generalization work focuses on cross‑subject or cross‑site shifts. This work systematically studies structured electromagnetic interference, which is ubiquitous for wearable, cockpit and hospital BCI deployments.
  2. The paper derives exact invariance theorem for fixed STFT masking together with operator‑norm leakage bounds. It further connects gating performance to Bayes risk lower bound when task‑relevant frequency bands are removed. The theory clearly articulates when spectral gating can and cannot work.
Critical weaknesses

The critical strengths of the paper are as follows

  1. All interference is synthetic additive common‑mode simulation. Real‑world EMI involves non‑linear coupling, electrode‑impedance variations, muscle artifacts mixed together. The simulator cannot fully replicate real hardware recording conditions.
  2. The baselines include augmentation and TENT, but miss comparisons against classical artifact subspace reconstruction (ASR), notch‑filter banks, or other widely‑used EEG preprocessing pipelines. Adding comparisons against established domain preprocessing methods would better position SGC.
Questions for authors
  1. All interference used in the paper is synthetic additive common‑mode simulation. What is the main gaps between the synthetic interference and real‑world interference?
  2. How is the hyperparameter sensitivity in the sensor thresholds and mask dilation? Practical guidance for hyper‑parameter transfer across datasets is missing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.