ICTRABiosignal ResearchResearch code ↗
WAICY 2026 · AI Showcase · Vancouver

A seizure detector that listens to brain, heart and muscle at once.

Two electrodes behind the ear, one on the chest, one on the shoulder, and a model developed on the public SeizeIT2 dataset: 11,626 hours from 125 patients. Built by four students to count the seizures nobody sees.

See the resultsWhy it mattersHow the AI works
Locked test opened 28 October 2026 · run once

29 patients no model had ever seen. Three of the four pre-registered hypotheses held; one did not, and one of our own earlier findings failed to replicate. The locked-test results →

Concept render: a volunteer seen from behind wearing three small white sensor patches — behind the left ear, on the shoulder and on the upper arm
Concept render (AI-generated) How ICTRA would be worn: EEG behind the ear, ECG at the chest, EMG at the shoulder, motion on the upper back. AI-generated illustration of the design intent (18 Oct 2026), not a photograph of the prototype or of a study participant. About the renders
Watch concept animation (AI-generated)

Concept render (AI-generated), not a verified hardware assembly or a functioning device.

Three signals. Shared time.

Synthetic signal demonstration

Scrolling EEG, ECG and EMG demonstrationThree synthetic waveforms scroll together, cycling between baseline patterns and increased activity. Not patient recordings, live device measurements or model predictions.EEGBrainECGHeartEMGMuscle−6 s−3 sNow
Static preview6-second window
Synthetic waveforms, adapted from the original demo. Illustrative amplitude; no patient data or model predictions.
11,626 hof wearable recordings from 125 patients, 883 seizures — the only open dataset with ear-EEG, ECG and EMG togetherSeizeIT2, Sci Data 2025
33 %of the average patient’s seizures caught on the 29 locked-test patients, at 0.23 false alarms a day — and 77 % of convulsive ones. Counted over all 181 seizures together instead of per patient, 16 %: a few patients with many brief seizures pull it down. Seizures under 15 s: none caught.Locked test, opened once on 28 Oct 2026 · model C, four signals · the threshold frozen before the test · per-patient average, the pre-registered primary aggregation
0.04 vs 0.86false alarms a day at the same sensitivity, four signals against EEG alone. This was the project’s primary hypothesis, written down and hashed before the locked test was opened. It held on every bootstrap sample.H1 · locked test · matched sensitivity 0.38 · ΔFA/24h −0.82 [−1.29, −0.36], P(improvement) = 1.000
40 %of non-motor focal seizures leave no trace on any wearable signal — a limit we measured and publishExp R, 695 seizures
29patients locked away from day one, opened once on 28 October 2026 — twelve seconds after the pre-registration was pushed to a public repositoryTier 2 of 3 · the test cannot be run again

Explore the research

The question, the method, the engineering and the evidence behind ICTRA.

Problem statement

The seizures nobody counts

50 million people; BC has 0.83 monitoring beds per 1,000 refractory patients; diaries miss half of all seizures; SUDEP follows unwitnessed convulsions.

Why this project → Artificial intelligence

Measure the timing, then fuse

Signal timing measured across 695 seizures before any model was built; three models that differ by one signal each; a decision rule anyone can read.

AI & method → Technical skills

126 GB, verified, every hour processed

Open Arduino acquisition, SHA-256 downloads, one signal-processing code path, GPU training, a live lab dashboard.

Engineering → Evidence

Numbers from patients we never trained on

Three tiers of test, patient-bootstrap intervals, the dataset authors' own algorithm as the reference, and every dead end kept.

Results → AI ethics

Silence must never mean safety

Missed seizures, false alarms, brain data that stays home, who the model was not trained on, and why this is not a medical device.

Ethics → Presentation

Four branches, four owners

Who presents what, the 30-second version of each result, and the questions each member is ready to take.

Team →

What is new here

Three contributions, with links to the supporting work.

Relationships before models

We measured when the heart, muscle and movement react relative to the brain in every seizure type before designing the fusion — and rewrote a hypothesis when the data disagreed.

Exhibit 1 →
Four signals, one model, every seizure type

The published ear-EEG detector uses EEG + ECG and needs a human reviewer; consumer watches see only convulsions. ICTRA models all four signals and reports every focal seizure type, including the ones it misses.

Where we stand →
Honesty as a design feature

Patient-wise folds, a locked test set, hypotheses fixed in advance, a stated accept rule and a public record of dead ends. The pre-registration was hashed and published before the locked patients were scored — and we report the hypothesis that failed and the finding that did not replicate.

Dead ends →
The story so far

From a noisy home-built board toward a frozen model and a blind test: the timeline, the version that was rejected, and the plan. Read the journey →

Signal coverage comparison: ICTRA records EEG behind the ear, ECG on the chest, EMG on the left shoulder, and movement on the upper back. Wrist sensing provides heart rate and movement, without EEG or shoulder EMG.
What a wrist device cannot reach. A seizure watch reads heart rate and movement. It is not blind — in a convulsion it often sees something. What it has no access to is the brain signal, and on the 29 locked-test patients every learned model beat every single-signal rule by 0.12 to 0.28 event F1, on every bootstrap sample. We had also found, on the development set, that the brain signal changes first; that one did not replicate on the locked patients, and we say so on the evidence page rather than quietly dropping it.

Status · 28 October 2026

Data and training

Seizure recordings
1,163 files · 24.8 GB · complete, SHA-256 verified
Seizure-free hours
all 6,894 files downloaded and featured (stage 2 complete)
Versions
v02g (offline) and v11 (causal) frozen and hashed on 26 Oct · v03 and v03b rejected under the pre-set rule · nothing changed after the freeze — how it evolved
Locked test
opened 28 Oct 2026, run once. sub-097…125 are development data from now on — results
Real ICTRA Lab capture, 18 October UTC: v02 all-hours training in progress

Real training record · screenshots explained →

Next

This week
The locked test is spent. Writing up what it showed, including the hypothesis that did not hold and the one that failed to replicate
November
External evaluation: a request to the SeizeIT2 authors to run the frozen model on data we have never had access to, and a submission of the EEG-only model to the EPFL SzCORE benchmark. SzCORE accepts only 19-channel scalp EEG, so that one measures how far a behind-the-ear model transfers to scalp recordings — not the four-signal system. Both are submitted once. They are the only untouched tests we have left
Hardware
Wearable prototype and a real-time demo video by 10 Nov
Submission
29 November 2026

Recorded pipeline: ICTRA Lab screenshots and explanations →