← All research projects

When Consumer EEG Should Not Adapt: A Safety Audit of Source-Free Test-Time Adaptation

28th Annual Mersivity / WaterHCI Symposium 2026
Toronto and online, March 30 to May 10, 2026 · Published in the symposium proceedings (open access on Zenodo)

Figure 2 of the paper: rolling macro-F1 of two adapters over a Muse EEG stream, with a shaded corrupted segment in which both fall to zero and a row of dots marking the windows the gate froze

Figure 2 of the paper: the safety timeline under a severe corrupted stream (Subject B). The shaded segment is the corrupted part of the stream; inside it both adapters, CAFN and StreamAdapt, collapse to zero F1, and the gate (dots) freezes too few windows at the boundary to prevent it.

A model that adapts itself to each new user and session sounds like exactly what a four-electrode headband needs. We audited eight test-time adaptation methods on consumer EEG and found that most of them barely help, several make things worse, and the failures land on artefact-corrupted windows, where a BCI should be most careful. The useful finding is a gate: adapt only when confidence, consistency and margin all agree, and otherwise stay frozen.

Abstract

Source-free test-time adaptation (TTA) is appealing for consumer electroencephalography (EEG), where each user, session, and headset placement can shift the input distribution. However, consumer EEG streams also contain jaw electromyography, head motion, electrode-contact changes, packet irregularity, and other non-neural disturbances. In this setting, an online update rule can adapt to artifacts rather than to brain-relevant structure. We present a condensed safety audit of eight adaptation strategies on four-channel Muse-class EEG. We compare each adapter against a substrate-matched no-adaptation baseline, evaluate corrupted-stream behavior, and use an activation-patching diagnostic to test whether adaptation increases internal artifact-mediated loss. Across FFT, ACT, and FFT+ACT feature substrates, useful gains from source-free adaptation are very small, while harmful updates remain common. T3A fails consistently, with large negative transfer on all feature substrates. Under severe stream corruption, NoAdapt outperforms every adaptive method, and confidence-based gating fails to identify most corrupted windows. These results suggest that consumer EEG adaptation should be freeze-first and safety-audited before deployment rather than enabled by default.

Authors

Alexander Vicol, Angelo Wei, Khalil Chaghouri, Jeremy Liu, Ethan Yang, Steve Mann

Links

← All research projects