Scina
← Personalized BCI

Research #1
BCI personalized

Protocol · 8 September 2026 · Source-backed discovery study

The answer we can honestly give

Not yet empirically answered. The primary outcome is frozen: anchored SAS change. Personalization succeeds only when it improves a later repeated session over a fixed population model and does not erase observed change from the person's initial SAS anchor.

This study has two deliberately separate questions. ds004148 can test whether personalization survives session drift. With only three sessions and limited SAS movement, it may be underpowered to establish whether personalization preserves genuine psychiatric symptom transitions. A prettier random-window score is not evidence for either claim.

1. Choose the right public dataset

Discovery dataset: OpenNeuro ds004148. The data paper documents 60 participants with three EEG sessions, including short-repeat and roughly one-month follow-up sessions. The official metadata define subject IDs and repeated Self-rating Anxiety Scale (SAS), Self-rating Depression Scale (SDS), Epworth Sleepiness Scale (ESS), Karolinska Sleepiness Scale (KSS), and positive/negative affect (PANAS) fields.

This is suitable for an honest stability-and-personalization discovery analysis, not a clinical monitoring claim. The target is not selected by intuition: anchored SAS change is frozen before modelling; SDS, PANAS, KSS, ESS, and quality/context measures are controls or secondary descriptions.

2. Measure the stability problem first

Train only on an earlier session, then predict the next later session from the same person. Separately, hold out whole people — all of their sessions — to measure new-person generalization. Fit preprocessing, tuning, and windowing inside each training split; no overlapping windows may cross a split.

Report the participant-level difference from a session-1 reference to the later repeated session and from familiar to unseen people. Participant/session resampling, rather than treating windows as independent patients, supplies uncertainty. This design follows the data's longitudinal structure and directly avoids the leakage warning demonstrated in translational EEG by Brookshire and colleagues.

3. Test personalization methods

Start simple: historical mean/last observation, spectral EEG + linear model, covariance/Riemannian model, and a small EEG neural network. Only after those establish later-session EEG signal should a foundation model enter the benchmark. Then compare the strongest fixed model with personal baseline, calibration head/adapters, and carefully budgeted fine-tuning.

Each method gets exactly the same prespecified earlier-session calibration labels and predicts a later untouched session. Success is frozen as: ≥0.10 training-SD reduction in later-session SAS MAE, a paired 95% interval excluding zero, no meaningful worsening of anchored-change error, and ≥60% of evaluable participants improved — all while surviving the controls below.

4. Check personalization is not cheating

Keep the first valid symptom/affect measurement as an immutable personal anchor. Evaluate both absolute target prediction and anchored within-person change. An update that makes all later predictions look normal by moving the baseline fails the question even if its distribution looks stable.

Research package and sources

The repository contains the full question, evaluation contract, assumption log, citations, official OpenNeuro metadata, and a local copy of the open-access data paper. Each methodological claim is marked as either sourced fact or proposed analysis. This prevents the research plan from quietly filling gaps with model-generated assumptions.

Wang et al. (2022), Scientific Data; OpenNeuro ds004148; Brookshire et al. (2024).