ecg-algo

AFib detector v1 on 12-lead ECG

A five-class rhythm classifier reaches 0.921 F1 for atrial fibrillation on a held-out test set of 4,211 recordings.

2 min read

This is a sample page. It shows how a results write-up renders: tables, plots, formulas, code and downloadable data. The methods subpage explains how the data was prepared.

Setup

The model classifies 10-second, 500 Hz, 12-lead recordings into five rhythm classes:

  • NSR, normal sinus rhythm
  • AFib, atrial fibrillation
  • AFL, atrial flutter
  • PVC, premature ventricular contractions
  • Other, every remaining rhythm

Recordings were split by patient, so no patient appears in both training and test data.

Metrics

Sensitivity and positive predictive value are computed per class, one class against the rest:

Se=TPTP+FNPPV=TPTP+FPF1=2⋅Se⋅PPVSe+PPV\text{Se} = \frac{TP}{TP + FN} \qquad \text{PPV} = \frac{TP}{TP + FP} \qquad F_1 = \frac{2 \cdot \text{Se} \cdot \text{PPV}}{\text{Se} + \text{PPV}}
Class Support Sensitivity PPV F1 AUC
NSR 1980 0.966 0.962 0.964 0.987
AFib 910 0.925 0.916 0.921 0.962
AFL 216 0.815 0.834 0.824 0.931
PVC 411 0.944 0.926 0.935 0.978
Other 694 0.880 0.905 0.893 0.905

Download the per-class metrics as CSV.

Confusion matrix

Confusion matrix of true against predicted rhythm class.
Figure 1. Confusion matrix on the test set. Rows are true classes, columns are predictions.

Most AFL errors are predicted as AFib. The two rhythms share irregular atrial activity, and flutter waves are hard to see in short strips.

ROC curves

ROC curves for each class. NSR has the highest AUC at 0.987 and Other the lowest at 0.905.

Reproducing

from ecg_algo import load_model, evaluate

model = load_model("afib-detector", version="1.0.0")
report = evaluate(model, split="test", sample_rate=500)
print(report.per_class[["sensitivity", "ppv", "f1"]])

Next steps

  • Patient-level split
  • Add 1-lead wearable recordings
  • Calibrate thresholds per class1

Footnotes

  1. Thresholds are currently the argmax of the softmax output, with no per-class tuning. ↩

In this section

  • 011 subpage

    Methods

    Data preparation, preprocessing and the signal quality gate.

  • 021 subpage

    Detailed results

    Breakdowns of the headline numbers.