WPW Detector

Machine learning · cardiology research

Detecting Wolff-Parkinson-White from a 12-lead ECG

A machine learning research project, trained on 53,540 real hospital recordings. Tested on 6,713 ECGs it had never seen.

Try the demoRead the paper
The challenge

A needle in a haystack

Machine learning needs examples to learn from, and the more the better. But WPW is rare. In the two public hospital datasets this project uses, there are 142 confirmed cases among 66,951 ECGs. That is one in every 471. In the heart, one dot in 471 is orange.

1 in 471the ratio of people living with WPW

The 99.8% trap

With one positive case for every 471 ECGs, a model that simply answers "no" every time is right 99.8% of the time. Accuracy is meaningless here. The real challenge is catching the rare case without drowning in false alarms.

Easy to miss, even for the eye

The signature of WPW is a delta wave, a subtle slurred start to the heartbeat's main spike. On a busy ECG, with other conditions layered on top, it can be faint enough to slip past a trained reader.

Where a machine can help

A model never gets tired and reads every ECG the same way. It will not replace a cardiologist. But it can flag the tracings that deserve a second, closer human look, and some of these models read the signal in ways no eye can, through wavelets, statistics, or a neural network.

The tool

See what it does

Upload an ECG. The tool returns a suspicion level from 1 to 100, shows which features pushed the score, and says plainly what the number means.

Open the demo

Runs the same model validated in the paper. A research tool, not a diagnosis.

Result
Suspicion level94/100
High suspicion
What drove the score
Delta wave / pre-excitation
44%
QRS-localized patterns
33%
Repolarization (ST/T)
18%
Other features
5%
The method

Seven ways to read the same heartbeat

There is no single right way to turn an ECG into numbers a model can learn from. So this project builds seven, each looking at the signal through a different lens, then combines the ones that see best.

M1, QRS onset morphology

The shape of the QRS onset, where the delta wave lives.

M2, Global statistics

The shape of the whole signal, summarized.

M3, Wavelet localization

Focusing in on the exact millisecond that matters.

M4, Median-beat morphology

One clean, representative heartbeat.

M5, Spatial VCG

How the leads relate to each other geometrically.

M6, Commercial baseline

A benchmark from an industry system.

M7, 1D-CNN

A neural network, letting a deep model find its own patterns.

The surprising result: the bottleneck was never the algorithm. It was the data.

How we proved it, in the paper

The proof

Tested where it had never looked

The honest test of a model is data it has never seen. The deployed detector, an equal-weight vote between M3 and M4, was scored once, at the very end, on a held-out fold kept untouched throughout. It reached an AUC of 0.95: show it one real WPW tracing and one healthy one at random, and it ranks the WPW higher about 95 times out of 100.

0ECGs it had never seen, used for the final test
The full held-out results
0.00AUC
The why

Why I built this

I was diagnosed with Wolff-Parkinson-White as a child. In 2024, a catheter ablation at the Montreal Children's Hospital removed that extra pathway, and I am healthy today because of it.Years later, I took a machine learning class and fell for it. This project sits exactly where the two met: the condition I had lived with, and the thing I had just discovered I loved doing. I wanted to know whether a machine could learn to detect WPW on an ECG, and help doctors catch a condition that is genuinely hard to see...

Read the full story
Where to go next