EEG-Jev← Back to the replay

Data & methods

How the replay is assembled

Recorded EEG and the answers one checkpoint already gave for it. The page itself runs no model.

What is on the page

Loading the material list…

QuestionRecording played in the replayWindowsWhy this oneSources
Loading…

Every file is listed with its SHA-256 in material/SHA256SUMS; each replayed window with its identifier, label, question and saved output is in manifest.json and the case files it names. The build checks are in BUILD_REPORT.json.

Sources and protocols

Loading…

All sources by application and protocol
ApplicationSourcesProtocol

Model-input channels of the four sources shown

Names and order as the model reads them. The complete check of what the page holds is in CHANNEL_COVERAGE.json.

Channel names in input order
SourceInput channelsShown by defaultIn order

Recordings, answers and labels

Real recordings with the answers one checkpoint gave for them, each bound to the question it was asked and to the dataset’s own label.

The model

One checkpoint answers all four questions: a frozen REVE-L encoder, a trainable MLP connector and Qwen3.5-2B with frozen base weights and LoRA, trained once on 35 public sources for sleep staging, motor imagery, seizure detection and pain rating. The EEG, the question and the answer specification are read in one forward pass; the readout selected by the answer type returns the probabilities, and no text is generated. The manuscript link will be added to the paper section of the main page once posted.

Choice · sleep stage

Sleep-EDF Expanded, leads Fpz–Cz and Pz–Oz, 30-second epochs. The model returns one probability for each of W, N1, N2, N3 and REM; the stage shown is the most probable one. The label is the hypnogram stage of the epoch, by one scorer (the dataset’s stages 3 and 4 are both N3).

Noul · seizure activity

CHB-MIT in its BIDS conversion, 18 longitudinal bipolar leads, 5-second windows that follow each other without overlap. The model input of this source has 24 channel slots: these 18 leads and six alternative lead names (F7–T3, T3–T5, T5–O1, F8–T4, T4–T6, T6–O2) that do not occur in the shown recordings and are zero in the model input. The number, Model P(seizure overlap), is the model’s probability that an annotated seizure overlaps the window. The dataset label is “yes” when the window and an annotated seizure interval share any time at all, and “no” otherwise; it is not the share of the window that is seizure. The amber colouring is the dataset’s time annotation. Where the page states how many of the 5 s are covered, that is a separate statistic of the annotation, not what the model estimates.

The marked threshold was fixed on development data before any test evaluation: the lowest value that keeps 99% of non-seizure development windows below it.

The threshold is a reading aid of this page. The balanced accuracy reported in the paper uses a threshold of 0.5 on the raw probability.

Choice · motor imagery

Stieger et al. 2021, a cursor-control experiment with 60 scalp leads: imagining the left or the right hand moves a cursor left or right, both hands move it up, and resting moves it down. Each trial has a 2-s pause, 2 s with the target on screen, then up to 6 s of cursor control with feedback. The model reads the 4 s after the target appears and returns a probability for each of the four options listed in the question: left hand, right hand, both hands, rest. The label is the cued target, not a measured movement. The benchmark split keeps only part of the trials of a session (177 of 450 in the replayed one; trials flagged as artefacts and trials that end early are among those left out), so the replay plays every trial of two runs end to end, 42 trials, and the model’s saved answer appears only for the 24 of them that the split kept; the others are played and marked as having no saved answer. The session file stores one array per trial, each with its own time vector counted from the appearance of the target, and no clock time between trials or runs. The time axis of the replay is therefore the trial arrays laid end to end in file order (the same axis the benchmark split lists use), not clock time; trial boundaries are drawn, the phases of each trial (pause, target shown, cursor control, the second after the result) are shown as a strip, and the two runs are separate segments because the time between them was not recorded.

Score · pain rating

OpenNeuro ds005292, 63 scalp leads. Each trial is one laser stimulus; the window runs from 1 s before the stimulus to 2 s after it. The model returns a probability for each rating from 0 to 10. The expected value shown is the sum over the eleven ratings of rating × probability; the error is that value minus the rating the participant reported for the stimulus a few seconds later (0 no sensation, 4 onset of pain, 6 moderate, 8 severe, 10 unbearable). The target is the reported rating, not the laser energy. The replay plays the three recorded blocks of one participant in full, on the time axis of each block’s recording, so the EEG between stimuli is shown as recorded. Each rated laser stimulus is marked at its real time with the 3-s window the model read; the answer appears when that window is complete and stays on the card until the next one, and the line under the answer says whether the playhead is inside a model window or between stimuli. The blocks are separate recordings and stay separate segments: the time between them was not recorded. Only stimuli of the benchmark split list are marked.

The numbers are raw outputs

Probabilities are shown as saved, to seven decimals, without temperature scaling or prior correction. Training draws the classes more evenly than they occur, so raw seizure probabilities sit higher than the rate of seizures in a recording; the paper reports what prior correction and temperature scaling change. A window’s answer appears only after the whole window has been drawn. The drawing speed is a playback setting and says nothing about inference time.

The waveform

Exported from the original recording files in microvolts, with the 0.5 Hz high-pass of the model’s preprocessing (the pain recording also gets a 40 Hz display low-pass). Each model window inside the export was compared with the window the model actually read: the two agree lead by lead (correlations in BUILD_REPORT.json). The model’s own input is additionally standardised per recording and lead. The replays are stored at 100 Hz (sleep, pain) or 200 Hz (seizure, motor imagery); the pain and motor-imagery replays show 9 and 13 leads by default and can show every input lead (63 and 60). The leads outside the default set are stored in time chunks (60 s for pain, 30 s for motor imagery), with the same rate, unit and filters, and are fetched when they are on screen.

Channels. “Input channels” are the channels the model reads for a source, in the order of its features. “Default” is a display subset chosen for readability; “All” shows every input channel; “Custom” is the viewer’s own choice. The scalp map uses the electrode positions stored with the encoder features (REVE position bank); a bipolar lead is drawn between its two electrodes and enters the encoder at the position of the first. The map, the lead list and the waveform rows share one selection: pointing at one highlights the others, a click pins it.

Amplitude. Files hold 16-bit integers with one scale per file; the browser multiplies them back to microvolts, and the step is far below what a pixel can show (peak amplitude ÷ 32,767). The gain is stated in µV per row, and the bracket at the right edge of a plot spans one row. Every lead is drawn inside its own row only: where a trace would leave the display range of its row it is cut, and a thin grey line with a small outward arrow (▲▼) at the row edge marks the cut, so a cut cannot be mistaken for a flat signal and no lead runs into its neighbour. The mark says “outside display range” and nothing else: it depends on the gain, not on the recording. In the single windows the default gain is one value per task, widened for a window whose amplitude needs it (at most 5% of its samples outside their rows); the gain can be set by hand, and Fit amplitude chooses the value at which nothing is cut. The replay has the same controls: an automatic scale per recording, manual steps and Fit amplitude for the span on screen. Three limits are separate things: the display range (the gain, in the viewer only); the 16-bit encoding of the files, where only the peak sample of a file reaches full scale, so nothing is cut; and the ±15 clip of the model’s preprocessing, which exists only in the standardised model input (0.016% of the samples of the replayed sleep windows, 0.015% of the pain windows, none in the seizure and motor-imagery windows) and is not drawn. Each pixel column shows the minimum and maximum of its samples, so short peaks are kept. Gain changes the drawing only, never a saved probability.

Chosen examples, measurements and scope

The examples are curated, not sampled. Results for whole test sets are in the paper.

How the examples were chosen

Choosing examples decides only what is displayed. The checkpoint, the threshold and every reported result were fixed by the evaluation protocol beforehand.

Results and checkpoint

Task results on held-out test data, probability quality and inference timing are reported in the paper, linked from the main page. This page shows examples only and quotes no result.

Signal analysis

Sleep signal cues

Seizure lead view

The lead selector and the paired non-seizure window are for inspection. The non-seizure window is the last window of the same record that ends at least 300 s before the annotated onset, or the first one that starts at least 300 s after the offset. The dataset gives seizure start and end times, not onset channels or labels for individual waves.

Limits