Detecting the sources of atrial fibrillation

Alexander Zolotarev, Nikita Kubrakov4 Jun 2022

Problem

Atrial fibrillation is treated by burning the area where an electrical vortex (a "driver") starts. A doctor finds that area by eye from electrode recordings, which is subjective: the burn may hit the driver or miss it. A model could help, but the doctors had no ready task statement and distrusted metrics they could not interpret. Objective labels exist only for lab hearts, while the target data are noisier clinical recordings with a subjective doctor annotation.

Example from the article: A catheter with 64 electrodes forms an 8×8 grid. The check is whether the brightest predicted electrode falls inside the driver area.

Solution in brief

Train on explanted hearts, where optical mapping of the same episode gives objective labels. Compute spectral features per electrode, then move from per-electrode classification (SVM) to segmentation of the whole 8×8 grid into a heat map (U-Net). Blend a clinic-only model with a lab-trained one. Judge quality by doctors' acceptance of the map and by stability across time windows.

Questions this case answers
  • How to localize a hidden source in a spatial grid of noisy sensor signals when the only objective ground truth exists in a lab setting and the target data comes from a different, noisier domain
  • How to transfer a model from lab data with objective labels to field data with subjective labels and show stakeholders that its predictions can be trusted
  • How to find the source of atrial fibrillation from the electrograms a doctor gets during a procedure?
  • How to use objectively labeled lab data to improve a model that must work on noisy clinical data?

Tools and methods

  • Spectral features — frequency amplitudes of each electrogram as model input
  • SVM — per-electrode driver classifier
  • Linear blend of two SVMs — one trained on clinical data, one on explanted hearts
  • U-Net with EfficientNet encoder — turns an 8×8×71 spectrum tensor into an 8×8 heat map
Continue reading — it's free

Google or a one-time email code. No password.