Try it yourself: pick a patient and switch between the full 21-electrode setup and the reduced 8-electrode one. Every chart below updates.
The headline numbers for the selected patient & setup:
1. What a seizure looks like in the brainwaves
👀 This is the actual brain signal. The orange band is where a real seizure happened — notice how the squiggle changes there.
One EEG channel () around the seizure. The shaded band is the clinician-marked seizure.
2. Did the model catch it?
🎯 Top row = when the seizure actually happened. Bottom row = when the model flagged one. Green means it got it right; red means a false alarm.
Each bar is a 4-second slice of the recording.
3. The main question: 21 electrodes vs 8
⭐ This is the heart of the project. If the two bars are about the same height, the cheaper 8-electrode setup works just as well.
Left: how many seizures each setup caught. Right: how often each cried wolf.
4. From detection to action — the alert tiers
🔔 A flag isn't an action. How sure the model is, and how long the activity lasts, decide what a real monitor would actually do.
Simulated decision only — this prototype never contacts anyone. Not a medical device.
5. How few electrodes can we use?
📉 We keep removing electrodes (always keeping the most useful ones) and re-test. If the lines stay flat, we can cut a lot of electrodes without losing much.
6. The honest test: would it work in real time?
⏱ We train the model only on the past, then ask it to catch a future seizure it has never seen — just like a live monitor. Watch the confidence line rise at the real seizure.
7. What the model actually pays attention to
🔬 These are the signals the model relies on most. They're the same things neurologists use — proof it learned real seizure markers, not noise.
Longer bar = more important to the model.
Honest limitations
- Small sample: a few patients, one seizure each. Promising, not proof.
- Results shown are per-patient (model calibrated to each person) — the realistic home-monitoring setup. Detecting seizures in a brand-new, unseen patient is much harder and needs far more data.
- This detects seizures as they happen; it does not predict them in advance.
- A research prototype, not a medical device.