Window-level AF detector: feasibility results

A 0.5 MB gradient-boosting classifier on RR-interval and beat-morphology features reaches 98.7 % sensitivity and 98.6 % specificity per 30-second sensor strip on patients never seen in training.

3 min read

Feasibility study for an atrial-fibrillation (AF) decision stage built on top of the existing beat detector. The beat detector and the per-beat measurements are unchanged; the AF decision is a new window-level classifier. All results below are on data the model never trained on; operating points were chosen on separate tuning records, never on a test set. Status: feasibility, not validation. Last update: 14 September 2026, 23:00.

Headline results

Sensor strips (30 s), 47 held-out patients, 755 strips, no expert artefact flags used. Product rule: 16-beat windows; AF if the highest window probability is at least 0.8, not AF if at most 0.5, otherwise inconclusive; strips without plausible beats are inconclusive.

Metric Value
Inconclusive 31.8 % (28 % of this set is artefact by the expert’s flags)
Classified strips 515
Sensitivity 98.7 %
Specificity 98.6 %
PPV / NPV 96.9 % / 99.4 %
Accuracy 98.6 %
Errors 5 false positives, 2 misses

Public benchmark, MIT-BIH Atrial Fibrillation Database (23 records, 10 h each), never used in training, RR features only.

Metric Value
Window sensitivity / specificity 95.6 % / 96.7 %
EC57 duration sensitivity / positive predictivity 96.5 % / 92.0 % (94.0 % with the two-scale cascade)
Episodes of 30 s or longer detected 100 %
Episodes of 10-30 s detected 89-100 % depending on window scale
EC57 episode positive predictivity, all lengths 47-56 %: short false episodes in ectopy-rich stretches, see results

Beat detector (unchanged), MIT-BIH Arrhythmia Database, bxb rule: sensitivity 98.1 %, positive predictivity 98.7 %.

What this shows

  • Per-strip performance on unseen sensor patients is at the level reported by cleared consumer ECG apps (about 98 % / 99.6 % with roughly 12 % unclassifiable), on a set with far more artefact.
  • The long-recording benchmark confirms the algorithm is sound on cardiologist-annotated public data: duration metrics above 92 % and every AF episode of clinical length detected.
  • The remaining weakness, short false episodes in long recordings, does not apply to a 30-second strip claim; the pending morphology test targets it for a future continuous-monitoring claim.

Caveats

  • Sensor labels are whole-strip annotations made by the original developer; cardiologist re-annotation is the next step and may move the numbers either way.
  • 165 AF-positive strips give a lower 95 % confidence bound near 95 % for sensitivity; more annotated strips are needed for tighter bounds.
  • The public benchmark was used during exploration before the protocol was frozen; the validation of record will be a single pre-specified run.
  • The EC57 scorer is a Python implementation of the published definitions; cross-check with the WFDB tools is pending.

Pages: the algorithm, article-style, methods, data and audit, detailed results, plan to certification, gap analysis and to-do list. Downloads: feasibility report, draft (PDF), full experiment log (Markdown).

In this section

  • 01

    The algorithm: a window-level AF detector for 30-second single-lead ECG

    How the detector works, from the cardiac cycle to the strip decision: the signal, the beat measurements it reuses, the rhythm features it computes, the model, the safeguards, and what the data shows.

  • 02

    Methods

    Window features on RR intervals and engine beat measurements, gradient boosting with monotonic constraints, a physiological guard, hysteresis episodes and an inconclusive output.

  • 03

    Data and audit

    LTAFDB, MIT-BIH, a clinical AF set and sensor strips for training; AFDB and held-out sensor patients for testing; a label audit found and fixed a loader bug and two classes of bad windows.

  • 04

    Detailed results

    End-to-end sensor results, EC57 benchmark tables, sensitivity by episode length, false-episode analysis and the comparison with the previous model.

  • 05

    Plan to certification

    Certify the 30-second strip claim first; what is met, what is missing, and the order of work.

  • 06

    Gap analysis and to-do list

    FDA 510(k) evidence for the 30-second AF strip classifier: what exists, what is missing, who owns it, and in what order.