Measured, not labelled: un-analyzable input versus misclassification in single-lead wearable ECG

Proposes an unsupervised morphological analyzability criterion for single-lead ECG windows, calibrated against three expert readers on 15 free-living subjects. At threshold τ = 0.50, criterion-positive (un-analyzable) windows account for 8.1 % (Holter), 56.7 % (handheld) and 69.6 % (patch) of the strict performance loss. In other words, much of what looks like AF-classifier misclassification on wearable ECG is really unreadable input.

• Form factor: mixed (Holter, handheld, patch)

Objective. Wearable single-lead ECG has expanded arrhythmia screening, but many recordings are inconclusive, and each corpus defines that word differently. We ask how much deployment error is un-analyzable input rather than misclassification when analyzability is measured with one instrument. Approach. We evaluate three diagnostic databases in-domain and leave-dataset-out under a strict loss counting windows carrying each corpus's inconclusive label as errors unless deferred. We score analyzability with an unsupervised morphological criterion calibrated against three experts on free-living ECG. Reference unreadability is sparse and imbalanced: only seven of 15 subjects contain an unreadable window and one supplies 88% of all unreadable windows. Thresholded sensitivity and every prevalence or true-share correction are therefore exploratory. Main results. The criterion flags almost all expert-unreadable windows but only subsets of the windows labelled inconclusive by their corpora: an inconclusive rate measures convention as much as signal. At tau_a = 0.50, criterion-positive windows are 8.1%, 56.7% and 69.6% of the strict proxy loss on Holter, handheld and patch. The true-share point intervals lie below 50% on handheld and above it on patch, but joint calibration-and-sampling envelopes span 50% on both and still do at the Youden threshold 0.60. On handheld, the calibration-only envelope is more than twice as wide as the corpus-sampling-only envelope. Gate-versus-confidence ordering changes with the accounting and corpus; no general selective-risk advantage of the gate is established. Significance. Corpus labels, criterion flags and latent expert unreadability must not be conflated. A gate can win by predicting a labelling convention, and this retrospective study validates neither transfer to smartwatch dry electrodes nor improvement in patient outcomes. Measure the inconclusive rate with a stated criterion on the raw signal; do not inherit it from a corpus label.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://publishingsupport.iopscience.iop.org/iop-standard/v1

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.