User Linkage and Replay Utility in BCI

Research question

Similarity between archived recordings can reflect user identity, temporal structure, or information useful for updating a decoder. This study measures these properties separately: identifying a similar recording and finding useful training history are distinct tasks.

Study design

The study uses 62 users, with 20 for development and 42 for confirmation. Three fixed-volume replay contrasts compare a user’s own history with matched other-user history, newer with older history, and similarity-ranked with matched other-user replay. Configurations, contrasts, and decision rules were locked before the confirmation run using internal time-stamped records. The data had been used in earlier analyses. Later exploratory analyses are reported separately.

A record-linkage audit measures what retrieval descriptors reveal about user identity and session indices, alongside the predictive value of the selected history.

Main findings

• On the foundation-model decoder, own-history replay improved next-session balanced accuracy by 0.48 percentage points over matched other-user history, with a 95% confidence interval of 0.18–0.78.
• Replay at lag 2 and at lags 3–5 was equivalent within ±0.5 percentage points when the newest session entered every update.
• Similarity-ranked versus matched other-user replay was inconclusive. Retrieval descriptors carried same-user information.
• In an explanatory comparison, uniform replay scored 0.99 percentage points lower than fine-tuning on the newest session alone.

Interpretation

The controlled comparisons estimate the value of replay composition. The linkage audit identifies information captured by retrieval scores. Together, they help assess whether a retrieval rule selects useful history under the tested update protocol. Results cover one dataset and the evaluated systems; they do not establish a general advantage for similarity-based replay.