EDSS
Research prototype · not a diagnostic device
Step 1 · Choose an ECG

Demo cases (verified against the cardiologists' labels)

Test patient (MIT-BIH)

Patients the models never saw in training, with every beat labelled by cardiologists.

Hospital ECG

WFDB (.hea + .dat), CSV, EDF, GE MUSE XML, or a PDF / photo of the printout.

or drop files here

Step 2 · Watch EDSS read the heart
0–10 s
♥Listenspikes sense change
→
⚡Decidegate opens on activity
→
◎ThinkMamba reads the rhythm
→
🔔Alerta clear, explained flag
Standard AIevery step processed
0steps computed
EDSSquiet steps skipped
0steps computed
Step 3 · The alert, and why
EDSS flags
–

Why: traced back to the beats

Try it: click any beat above to ask "what if this beat were normal?", every model is rerun on the edited ECG.

Each beat is replaced by this strip's typical beat and EDSS is run again: the bar is how much of the evidence disappears. The model is re-run, not guessed from gradients; it shows what the decision depends on, and it is not a clinical diagnosis.

Step 4 · Every model, the same ECG

Test AUROC = accuracy on all held-out test patients (95% interval); energy is estimated from operation counts (45 nm figures, Horowitz 2014), not measured. One training run per model shown here.

Rhythm-change timeline

When does the rhythm change, and when does it revert?

The whole recording, 10 seconds at a time: cardiologists' rhythm notes and beat labels against EDSS and Mamba, with the share of steps EDSS chose to compute. Click any window to open it above.

Results on held-out patients

Less compute. Honest numbers.