⚠️ RESEARCH TOOL | NOT FOR CLINICAL USE | CC BY-NC 4.0

AF isn't just
"there" or "not there"
it has patterns

We trained a deep learning model on 51,000+ hours of ECG data. It doesn't just find AF—it tells you when it happens. Night owl? Morning person? Turns out your heart rhythm might be too.

Try the DemoRead the PapersPublished in Nature NPJ
51,386
hours of ECG analyzed
from 2,147 patients
0.95
F1 score
across 4 countries
5
AF timing patterns
nobody knew existed
90%
sensitivity
patient-level detection

Here's what we actually did

1Built a model that actually generalizes

Most AF detection models fall apart when you test them on data from a different hospital. Ours doesn't. Same F1 score whether the ECG is from Virginia, Israel, Japan, or China. We tested it—trust us, that almost never happens.

2Discovered AF has a clock

Some people get AF at 3am. Others at 3pm. We found 5 distinct patterns. Why does this matter? Maybe your medication should target when your AF actually happens. Wild concept, right?

3Made it open source

All our code is on GitHub. The model weights are public. Use it for research, modify it, break it, fix it—whatever. Just don't sell it (CC BY-NC 4.0 license).

The numbers (no BS edition)

From Table 3 of our Nature paper. Real test set, 1,825 consecutive patients at Rambam Hospital.

90%
Sensitivity
Catches 9 out of 10 patients with AF
96%
Specificity
Rarely cries wolf on healthy hearts
99%
NPV
If it says no AF, it's almost certainly right
74%
PPV
If it flags AF, correct 3/4 times
What we won't tell you: That this is ready for clinical use. It's not FDA-cleared, not CE-marked. It struggles with atrial flutter (69% misclassification rate). Use it for research, not for patients.

Want to try it?

Load up our demo with sample data, or dig into the research papers to see how we built this thing. Or clone the repo and break stuff—that's cool too.