🧬 biology

What Does EEG Preprocessing Contribute to Cross-Subject Decoding? A Subject-Level Ablation Study of High-Density EEG

In a high-density, subject-independent EEG regression setting, this ablation study demonstrates that per-sample temporal z-score normalization provides the vast majority of the predictive performance gains, rendering additional preprocessing steps like filtering, clipping, and masking statistically redundant.

Original authors: Arvind Gyandatt Mishra, Alice Wong

Published 2026-09-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Arvind Gyandatt Mishra, Alice Wong

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The human brain is a restless, noisy engine, constantly firing electrical signals that ripple across the scalp. To hear the specific thoughts or reactions hidden within that static, scientists use electroencephalography, a technique that places sensors on the head to record these electrical whispers. The challenge has always been that the raw signal is messy, filled with interference from muscle movement, eye blinks, and the varying thickness of skin and bone between the brain and the sensor. For decades, researchers have relied on a long list of cleaning steps—filtering out certain frequencies, cutting off extreme spikes, and adjusting for volume differences—to make the data readable. The assumption was that more cleaning meant better results, especially when trying to build computer models that could understand the brain of a person they had never met before. But in the complex world of brain decoding, it has been difficult to know which cleaning steps actually help and which are just busy work.

A team of researchers recently set out to test this assumption with a focused experiment using data from a large-scale challenge involving high-density brain recordings. They asked a simple but profound question: when trying to predict a person's reaction time based on their brain activity, how much of the success comes from the complex cleaning process, and how much comes from just one specific step? They took a standard, multi-stage pipeline used for these tasks and systematically stripped it away, testing six different versions of data preparation. They kept everything else exactly the same: the same computer model, the same group of people, and the same task, which involved detecting changes in visual contrast. The goal was to see if adding more filters or correction steps actually improved the model's ability to guess the reaction times of new, unseen participants, or if a simpler approach was just as good.

The results were striking in their clarity. The researchers found that the single most powerful step was a simple process of normalizing the data for each individual recording window. This step essentially adjusted the volume and center of the signal for every single channel of data, removing the average background hum and scaling the fluctuations to a standard range. When they applied this single step, the model's error rate dropped dramatically, falling from a high of 0.7267 down to 0.3717. In practical terms, this means the model's predictions became nearly twice as accurate. More importantly, this improvement was consistent across the vast majority of the people tested; about 77 percent of the unseen participants showed better results with this normalization alone. The researchers then added the other standard cleaning steps back in one by one: band-pass filtering to remove very low and very high frequencies, clipping extreme values, and masking out bad sensors. None of these additions provided any measurable extra benefit. The model performed just as well with the simple normalization as it did with the full, complex suite of cleaning tools.

The study also looked at whether these methods made the system more robust against noise, simulating what happens when the signal gets distorted by artificial interference. Here, the full system did show it was far less sensitive to these disturbances than the raw, unprocessed data. However, the researchers were careful to note that they could not pinpoint exactly which part of the system provided this stability, as they did not test the normalization step in isolation against the perturbations. What they could say with certainty was that for the specific goal of predicting reaction times in new people, the extra complexity did not translate into better accuracy. The additional steps of filtering, clipping, and masking did not lower the error rate any further once the normalization was in place. In fact, for some participants, adding these extra steps made the predictions slightly worse, though not enough to be statistically significant.

This finding challenges the common practice of stacking every available cleaning tool onto a dataset in the hope that it will help. The researchers demonstrated that in this specific high-density setting, the predictive power came almost entirely from that single normalization step. The remaining operations, while perhaps useful for other types of analysis or different tasks, did not contribute to the model's ability to generalize to new people. The study suggests that in the pursuit of decoding the brain, simplicity can be more effective than complexity. By focusing on the one transformation that truly mattered, the researchers were able to achieve the same level of performance with a much leaner pipeline. The lesson is not that cleaning the data is unimportant, but that the value of each cleaning step must be weighed against the specific goal, and that sometimes, the most powerful tool is the one that simply levels the playing field for every new person the model encounters.

Technical Summary: "What Does EEG Preprocessing Contribute to Cross-Subject Decoding? A Subject-Level Ablation Study of High-Density EEG"

Problem Statement

Electroencephalography (EEG) decoding pipelines rely heavily on preprocessing steps (filtering, artifact handling, normalization, clipping, augmentation) to shape the signal representation for machine learning models. However, the specific predictive value of individual preprocessing operations is often difficult to disentangle from the model architecture and dataset characteristics. This is particularly critical in cross-subject decoding, where the goal is to generalize to unseen participants. Differences in scalp impedance, physiology, and recording environments create substantial inter-individual variation. While recent "multiverse" studies (e.g., Kessler et al., 2025; Huang et al., 2025) have established that preprocessing choices materially affect downstream results, they have not determined which specific operations are essential for large-scale, subject-independent regression tasks.

Methodology

The authors conducted a controlled ablation study using data from the NeurIPS 2025 EEG Foundation Challenge (Contrast Change Detection task). The study utilized a fixed high-density EEG representation (129 channels, 100 Hz, 2-second windows) and the EEGNeX convolutional architecture.

Experimental Design:

Dataset: 10,383 training windows, 1,717 within-subject held-out windows, and 1,470 cross-subject test windows involving 26 held-out subjects.
Protocol: Six progressively composed preprocessing conditions were evaluated. Crucially, the input windows, subject-level partitions, model architecture, optimization procedure, and evaluation metrics were identical across all conditions. Only the per-window transformation varied.
Conditions:
1. Raw Minimal: No preprocessing.
2. Per-sample Normalization: Temporal z-score normalization (mean and standard deviation calculated per window, per channel).
3. Filtering + Normalization: 0.5–35 Hz Butterworth filter + normalization.
4. Filtering + Normalization + Clipping: Added amplitude clipping to $[-10, 10]$ .
5. Full Preprocessing: Added channel masking (zeroing flat or noisy channels).
6. Full + Augmentation: Added training-time augmentations (temporal shifts, noise, scaling, channel dropout).
Evaluation: The primary metric was Root Mean Square Error (RMSE) on the cross-subject test set. Analysis was performed at two levels: pooled RMSE (all windows) and subject-level RMSE (treating the participant as the unit of inference). Statistical significance was assessed using paired Wilcoxon signed-rank tests and bootstrap confidence intervals (10,000 resamples). Repeatability was checked using three fixed training seeds. Robustness was tested via evaluation-time perturbations.

Key Contributions

Subject-Level Ablation: Unlike previous studies that varied pipelines across different datasets or tasks, this study isolates the contribution of preprocessing operations within a single, fixed regression problem with a strict subject-held-out split.
Quantification of Normalization: The study provides empirical evidence that per-sample temporal z-score normalization is the dominant factor driving performance gains in this specific high-density, subject-independent setting.
Diminishing Returns of Complexity: The paper demonstrates that adding standard filtering, clipping, channel masking, and training augmentation to a normalized pipeline yields no statistically significant improvement in clean-data cross-subject prediction.
Robustness vs. Accuracy Distinction: The study distinguishes between clean-data accuracy and robustness to synthetic perturbations, showing that while the full pipeline (including augmentation) is more robust than raw data, the specific contribution of augmentation to clean-data accuracy is negligible.

Results

Normalization Impact: Per-sample normalization reduced the pooled cross-subject RMSE from 0.7267 (Raw) to 0.3717, a reduction of approximately 48.9%.
Subject-Level Consistency: Normalization improved performance for 76.9% (20/26) of the held-out subjects. The mean subject-level improvement was 0.1274 RMSE (Wilcoxon $p = 0.0014$ ).
Lack of Incremental Gain:
- Adding filtering to normalization resulted in a slight, non-significant increase in error (Mean $\Delta$ RMSE = -0.0097, $p = 0.0669$ ).
- Adding clipping and channel masking (Full Preprocessing) resulted in an RMSE of 0.3736, statistically indistinguishable from normalization alone (0.3717).
- Adding training augmentation resulted in an RMSE of 0.3722, with no significant difference from the full preprocessing condition ( $p = 0.7265$ ).
Repeatability: The normalization condition showed high stability across three training seeds (CV = 0.20%), whereas the raw condition was highly unstable (CV = 21.4%).
Robustness: The combined "Full + Augmentation" system was significantly more robust to synthetic perturbations (e.g., high-frequency contamination, low-frequency drift) than the raw baseline, though the study does not isolate which component (normalization vs. augmentation) drives this robustness.
Error Distribution: Normalization drastically reduced the upper tail of subject-level errors. Under raw input, one subject contributed 70.3% of the total squared error; after normalization, the error distribution became tightly clustered with no extreme outliers.

Significance and Claims

The paper concludes with a restrained methodological conclusion: in this specific high-density, subject-independent EEG regression setting, per-sample normalization captures the vast majority of the predictive benefit obtained from the tested preprocessing pipeline.

The authors emphasize that:

Additional preprocessing operations (filtering, clipping, masking) did not provide measurable incremental predictive value once normalization was applied.
This does not imply that filtering or artifact handling are undesirable, but rather that in this specific context, they did not yield further clean-data gains.
The results suggest that preprocessing pipelines should be evaluated incrementally rather than accepted as additive packages based solely on intuitive signal-processing justifications.
The findings are specific to the CCD task, the HBN-EEG dataset, and the EEGNeX model; they do not claim universal optimality for all EEG tasks or model families.

The study serves as an empirical characterization of a concrete pipeline, highlighting that in subject-independent decoding, a simple normalization step may be the most critical component for generalization, while complexity does not necessarily equate to predictive benefit.