PaperlayerView exact revision on arXiv

Source-linked AI summary

Automated Artifact Removal in EEG Age Prediction: A systematic comparison

Davoud Hajhassani, Paul-Adrien Graignic, Bruno Aristimunha, Apolline Mellot, Tom Mariani, Clément Nober, Bruna J. Lopes, Léo Burgund, Lionel Kusch, Thomas Semah, Arnault H. Caillet

arXiv:2609.31195v1eess.SP

TL;DR

Automated EEG artifact removal can change information relevant to downstream prediction, yet its effects on deep-learning performance remain comparatively underexplored. The paper benchmarks nine methods for cross-dataset age prediction and introduces SQI-guided GEDAI, which selectively corrects channel–epoch pairs. GEDAI and its SQI-guided version were the only methods that consistently improved prediction across datasets and architectures; SQI guidance reduced modification while retaining closely matched performance.

  • Problem

    The effects of automated artifact removal on deep-learning performance remain comparatively underexplored, despite method differences and the need to balance artifact suppression with preserving useful predictive information.

    Source: §1 · p. 1

  • Method

    The study compares nine artifact-removal methods with no removal for cross-dataset EEG age prediction and introduces SQI-guided GEDAI to restrict correction to contaminated channel–epoch pairs.

    Source: §1 · p. 1, §2.2 · p. 2

  • Results

    GEDAI and SQI-guided GEDAI were the only methods that consistently improved median MAE and R2 across datasets and architectures; the other seven were generally neutral or detrimental.

    Source: §4 · p. 4

  • Takeaways & Limitations

    Automated EEG artifact removal should be validated for the intended downstream task, since gains differed by method and varied across datasets.

    Source: §4 · p. 4

    The study covers one source corpus, one downstream task, and a fixed SQI threshold selected on source validation data; modification ratio does not measure correction magnitude, artifact attenuation, or neural preservation.

    Source: §4 · p. 4

Abstract

from arXiv · show

Automated EEG artifact removal may improve downstream analysis but can also alter predictive information. We benchmarked nine automated artifact removal methods against a common no-artifact-removal baseline for cross-dataset EEG age prediction. We introduce Signal Quality Index (SQI)-guided GEDAI, which leverages local signal-quality assessment to restrict correction to the channel--epoch pairs requiring intervention. Three deep neural architectures were trained on TUEG and evaluated without target-domain fitting on ds005385, LEMON, and TDBRAIN. Across this setting, GEDAI and SQI-guided GEDAI were the only methods with consistent gains over baseline in age prediction performance across all datasets and architectures ($\Delta$MAE $=-0.77/-0.64$ years, $\Delta R^2=+0.083/+0.072$, respectively). The remaining methods were neutral or detrimental on average ($\Delta$MAE $=+0.22\pm0.16$ years, $\Delta R^2=-0.023\pm0.016$ across methods). The two GEDAI-based methods achieved closely matched performance, while SQI guidance reduced the median modification ratio from $74.78\%$ to $42.80\%$. These findings show that curation benefits are method-dependent and establish SQI guidance as a more selective operating point, leaving more of the original EEG unchanged and limiting the potential loss of neural activity while retaining most of GEDAI's predictive benefit.

1. INTRODUCTION

Automated artifact-removal methods differ in how they handle contamination, and their effects on deep-learning prediction remain comparatively underexplored. This study benchmarks those methods for cross-dataset EEG age prediction, where chronological age enables comparison across cohorts without matched tasks or participants.

1. INTRODUCTION

  • AAR methods differ substantially in identifying, rejecting, or reconstructing contaminated EEG, motivating comparison under common downstream conditions.Selecting an appropriate method remains difficult as EEG datasets grow.Source: §1 · p. 1
  • Artifact handling must balance suppressing contamination against preserving predictive information, because non-neural activity may also predict age and extensive correction can attenuate neural signals.Rejection can discard usable EEG and limit online applicability.Source: §1 · p. 1
  • Chronological age supports comparison across independent resting-state cohorts, while cross-dataset evaluation tests transfer across cohorts, recording systems, and acquisition conditions.The study benchmarks nine methods against no artifact removal using three architectures and four public EEG datasets, and introduces SQI-guided GEDAI to correct only contaminated channel–epoch pairs.Source: §1 · p. 1

2. METHODS

The study preprocesses resting-state EEG from four public datasets, compares automated artifact-removal methods, and evaluates age prediction on held-out cohorts without target-domain fitting. SQI-guided GEDAI uses local quality scores to limit which channel–epoch pairs are corrected.

2.1. Datasets and Preprocessing

  • Recordings came from healthy participants aged 20–90 years in four public datasets and were filtered, resampled, referenced, epoched, spatially harmonized, and normalized before prediction.At least 60 seconds of resting-state data were required; up to the first 240 seconds were retained, then segmented into nonoverlapping 5-second epochs.Source: §2.1 · pp. 1–2

2.2. Automated artifact removal methods

  • The benchmark included nine automated methods, while SQI-guided GEDAI restricted correction to channel–epoch pairs identified as contaminated by local signal-quality assessment.Its SQI combines seven signal descriptors, maps them to severity levels, and uses a threshold of 0.05 selected on source validation data; GEDAI estimates correction from the complete multichannel epoch before masking.Source: §2.2 · p. 2

2.3. Benchmark protocol and evaluation

  • Models trained on TUEG were tested on ds005385, LEMON, and TDBRAIN without further tuning, using three neural architectures and participant-level age-regression metrics.Training and splitting were repeated over 50 seeds; paired method comparisons used identical splits, and the modification ratio counted altered or rejected channel–epoch pairs.Source: §2.3 · pp. 2–3

3. RESULTS

GEDAI and SQI-guided GEDAI performed best overall across datasets and architectures, while the other methods were generally neutral or detrimental. SQI guidance achieved closely matched prediction performance with a lower median modification ratio.

3. RESULTS

  • 9.27 MAE/0.559 R2 for GEDAI and 9.40 MAE/0.548 R2 for SQI-guided GEDAI, versus 10.04 MAE/0.476 R2 for no-AAR; both GEDAI methods ranked first or second in all nine combinations.These averages are across cell-wise seed medians.Source: §3 · p. 3
  • 17/18 GEDAI and 18/18 SQI-GEDAI comparisons significantly improved over no-AAR, versus only 8/63 MAE and 9/63 R2 comparisons for the remaining methods.Without target-domain fitting, the best configurations achieved median-across-seed MAE/R2 of 7.69/0.589 on ds005385, 9.91/0.632 on LEMON, and 8.36/0.592 on TDBRAIN.Source: §3 · p. 3
  • SQI-guided GEDAI had 0.11-year higher MAE and 0.008 lower R2 than GEDAI on average, with only 7 of 18 cell-wise paired tests significant after Holm correction.Both GEDAI methods had lower across-seed variability than no-AAR in most dataset–architecture combinations.Source: §3 · p. 3
  • −1.12 and −1.05 years were the mean ∆P90 reductions for GEDAI and SQI-guided GEDAI, respectively; they were the only methods with negative ∆P90 across all nine combinations.Panel A summarizes ∆P90 by dataset–architecture combination and its overall average.Source: §3 · p. 3, Figure 3 · p. 4
  • The SQI-guided method’s modification ratio was lower than those of all other AAR methods except Autoreject and FAAR.Autoreject and FAAR had median ratios of 19.97% and 8.33%, respectively.Source: §3 · p. 4

4. DISCUSSION

GEDAI and SQI-guided GEDAI were the only methods that consistently improved median MAE and R2 across datasets and architectures; SQI guidance retained closely matched predictive benefit with fewer modifications. The discussion supports downstream validation while noting that predictive performance cannot identify whether altered information was neural or artifactual.

  • GEDAI and SQI-guided GEDAI alone consistently improved median MAE and R2 across datasets and architectures, and reduced upper-tail error in all nine dataset–architecture combinations.Their gains appeared across all three deep-learning architectures, though their magnitude varied across datasets.Source: §4 · p. 4
  • The other seven methods produced significant gains in only 17 of 126 baseline comparisons and were generally neutral or detrimental on average.Source: §4 · p. 4
  • Matched participant splits, random seeds, and target cohorts isolate AAR effects under a common protocol, but downstream performance cannot identify whether altered information was neural or artifactual.Source: §4 · p. 4
  • Best GEDAI-based models reached R2 = 0.632 on LEMON and R2 = 0.592 on TDBRAIN; differing training data and evaluation protocols prevent direct comparison with a previous study.Source: §4 · p. 4
  • 42.80% versus 74.78% median modification ratio accompanied closely matched predictive performance for SQI-guided GEDAI and GEDAI, respectively.The ratio reflects how often EEG is altered, not correction magnitude or neural preservation.Source: §4 · p. 4
  • 8.33% was FAAR’s modification ratio, achieved by discarding contaminated epochs; it did not match GEDAI-based predictive gains and reduced downstream data.Source: §4 · p. 4
  • The study considered one source corpus, one downstream task, and a fixed SQI threshold selected on source validation data.Future work should evaluate recording-adaptive thresholding and other SQI-guided correction methods.Source: §4 · p. 4

Related papers

Loading 2609.31195v1…