eeg-test-time-adaptation-benchmark

name	eeg-test-time-adaptation-benchmark
description	NeuroAdapt-Bench: Systematic benchmark for test-time adaptation (TTA) on EEG foundation models under real-world distribution shifts. Evaluates TTA methods across multiple FMs, tasks, and datasets including extreme modality shifts (Ear-EEG). Finds gradient-based TTA degrades, optimization-free methods more stable.

EEG Test-Time Adaptation Benchmark (NeuroAdapt-Bench)

Paper: Test-Time Adaptation for EEG Foundation Models: A Systematic Study under Real-World Distribution Shifts
arXiv: 2604.16926 (April 2026)
Authors: Gabriel Jason Lee, Jathurshan Pradeepkumar, Jimeng Sun
Categories: cs.LG, cs.AI, eess.SP

Core Contribution

NeuroAdapt-Bench is the first systematic benchmark for evaluating test-time adaptation (TTA) methods on EEG foundation models under realistic distribution shifts. It reveals that standard TTA methods from other domains are unreliable for EEG.

Problem

EEG foundation models face distribution shifts across:

• Clinical settings: Different hospitals, protocols
• Devices: Different amplifier hardware, electrode types
• Populations: Different age groups, conditions
• Modalities: Scalp EEG vs. Ear-EEG

TTA enables adaptation to unlabeled target data during inference without source data access — critical for healthcare privacy.

Benchmark Design (NeuroAdapt-Bench)

Distribution Shift Types

1. In-distribution: Same domain as training
2. Out-of-distribution: Different but related domain
3. Extreme modality shift: e.g., Scalp EEG → Ear-EEG

Evaluation Dimensions

• Multiple pretrained foundation models
• Diverse downstream tasks
• Heterogeneous datasets
• Representative TTA approaches from other domains

Key Findings

1. Standard TTA Methods Are Unreliable for EEG

• Standard TTA approaches yield inconsistent gains
• Often degrade performance compared to no adaptation
• Results don't transfer across tasks or datasets

2. Gradient-Based TTA Fails

• Gradient-based approaches particularly prone to heavy degradation
• EEG signal characteristics make gradient estimation unstable
• Distribution shifts in EEG are fundamentally different from image domain shifts

3. Optimization-Free Methods Are More Stable

• Methods that don't require gradient computation show greater stability
• More reliable improvements across settings
• Suggests EEG requires fundamentally different adaptation strategies

Implications

For Practitioners

1. Avoid naive TTA: Don't apply standard TTA methods directly to EEG FMs
2. Prefer optimization-free: Use methods that don't rely on gradients
3. Validate per-task: TTA effectiveness varies by task — test before deployment
4. Domain-specific needed: EEG requires custom adaptation strategies

For Researchers

1. EEG ≠ images: Distribution shifts in EEG have different characteristics
2. Gradient instability: EEG signal properties make gradient-based TTA unreliable
3. Need domain-specific TTA: Current TTA literature is vision-focused
4. Ear-EEG challenge: Extreme modality shifts remain largely unsolved

TTA Methods Evaluated

Gradient-Based (Found to Degrade)

• Tent (entropy minimization)
• EATA (entropy minimization with sample selection)
• SAR (sharpness-aware regularization)
• MEMO (multi-expansion for test-time adaptation)

Optimization-Free (Found More Stable)

• Feature alignment methods
• Statistical normalization approaches
• Non-parametric adaptation

Comparison with TTA in Other Domains

Domain	Gradient-Based TTA	Optimization-Free TTA
Vision (ImageNet-C)	Strong improvements	Moderate improvements
EEG	Heavy degradation	Stable, modest gains

Application Scenarios

• Clinical EEG deployment across hospitals
• Cross-device model transfer
• Consumer EEG headset adaptation
• Ear-EEG and alternative modality deployment
• Privacy-preserving model adaptation (no source data needed)

Trigger Keywords

• neuroadapt-bench, eeg test-time adaptation, tta eeg, test-time adaptation foundation model, optimization-free tta, EEG测试时自适应

Related Skills

• eeg-channel-adaptation-benchmark
• eeg-foundation-model-adapters
• tta-eeg-foundation-models
• laya-eeg-foundation