Multi-Lead ECG: The Diagnosis Lives in the Disagreement

8 min read

Companion note to Advanced Neural Network Architecture for Enhanced Multi-Lead ECG Arrhythmia Detection, IJFMR 2024.

Here is the thing that made me want to write this paper. A cardiologist reading an ECG doesn’t stare at one waveform. They read twelve at once — different electrodes, different vantage points on the same heart — and they look at how the leads disagree. That disagreement is where a lot of arrhythmias give themselves away.

The vast majority of published ECG classifiers are single-lead. They throw twelve leads’ worth of signal away and read the shortest one.

Twelve leads, twelve angles on the same heart

Every electrode watches the heart from a different direction. Lead II sees the standard rhythm-strip view. Lead V1 sits on the right side of the chest and picks up the right ventricle. V5 sits on the left. aVF looks bottom-up. Same electrical event, four (or twelve) different projections.

A normal heartbeat looks pretty similar across all of them, so single-lead does fine on healthy patients. But diseased hearts break the symmetry. In left bundle branch block, the QRS complex widens on the leads that view the left ventricle and stays narrow on the ones that don’t. You literally cannot see that from one lead.

Click through the conditions and watch what happens to the leads — and where a fusion model actually looks.

Interactive · Reading the same heartbeat from four angles

Four ECG leads, same heart, same instant. Each lead is a different vantage point — and each disagrees a little. Click a diagnosis to see how the leads shift, and where the model looks.

How this demo works

Each track shows a stylized ECG waveform for the selected condition, from one of four leads (II, V1, V5, aVF). The colored spans mark which regions the paper's per-lead attention block weights most heavily for the diagnosis. Single-lead classifiers see just one row; the fusion block reads all four and their disagreements.

Look at what the fusion attention does on LBBB. It ignores Lead II (the one every single-lead paper uses) and locks onto V1 and V5, because the disagreement between them is the diagnosis. That’s the entire pitch of the paper in one screenshot.

What the architecture actually does

Two changes. Neither is fancy on its own; together they do most of the work.

The alternative most people default to is channel-stacking — treat the 12 leads as 12 input channels of a single CNN. It works, but the model can’t ever look at leads independently; every filter is a joint function of all leads. The disagreement signal gets mashed at the earliest possible layer, where the model has the least information to know what to preserve.

The auxiliary head is not scored at test time. It only exists to shape the encoder during training. Turning it off costs us about 4 F1 points on the harder classes and barely moves the aggregate number, which is the exact shape of “this actually mattered.”

What actually moved the numbers

Class-conditional recall. Not aggregate F1. Aggregate F1 on ECG datasets is dominated by the majority class (normal sinus rhythm), so a model that just says “normal” 90% of the time already looks 90% good.

The interesting question is what happens on SVT, LBBB, and PVC — the classes where single-lead classifiers tend to collapse. That’s where multi-lead pays off, and where our recall improvements were.

The bit I still lose sleep over

I never did the clean ablation that separated the architectural gain from the auxiliary-loss gain. Both changes were in the final model, both changes were in the ablation-off baseline. I have suspicions about which one is doing more of the work, but suspicions aren’t a paper.

Also — and this is bigger — everything in the paper is trained and tested on public datasets (MIT-BIH, PTB-XL). Real hospital ECG data has noise, lead misplacement, patient motion, and 12 leads that don’t always agree with the public-dataset conventions. I have no idea how the model holds up on that. Nobody who publishes on public ECG data does, really.

What I’d do next

  • Learn a per-lead noise gate. In practice, leads drop out — an electrode falls off, a patient moves. The fusion block should be able to say “lead V5 is garbage this window, downweight it.” We didn’t have that.
  • Domain-shift-aware training. Mix hospital data (even a little) into training, or explicitly augment for common real-world noise.
  • Report ROC-AUC per class alongside F1. F1 hides a lot. Per-class ROC is where you actually see how the model behaves at different operating points, which is what a cardiologist actually cares about.

Full paper: arXiv:2404.15347.

Research referenced in this post