Data, not model, is the bottleneck: a leakage-controlled study of Wolff-Parkinson-White detection from the 12-lead electrocardiogram at 471:1 class imbalance

Wolff-Parkinson-White (WPW) syndrome is a congenital cardiac pre-excitation that is clinically important and frequently under-recognized on the resting 12-lead electrocardiogram, so automatic detection is useful; it is also statistically hard, because the surface signature is subtle and the condition is rare. We study WPW detection on two pooled public corpora, PTB-XL and Chapman-Shaoxing-Ningbo, comprising 66,951 recordings of which 142 are WPW, a prevalence of 0.21% (about 471:1).

The contribution is methodological rather than a state-of-the-art detection result: under a single pre-specified, leakage-controlled protocol with a held-out test fold contacted exactly once, we build and compare seven representations of the signal (beat-delineation morphology, global-statistical summaries, wavelet time-frequency localization, median-beat morphology, vectorcardiographic geometry, an on-machine commercial measurement set, and a one-dimensional convolutional neural network), changing only the representation.

Scoped to models trained or pretrained only on these two corpora at modest compute, added diversity and capacity do not raise the ceiling: adding an orthogonal detector to the two strongest significantly hurts, a feature-union model and a deep network are statistically indistinguishable from a simple two-member vote, and self-supervised pretraining fails a pre-specified transfer gate. A leak-free learning curve that re-runs feature selection at every training size is still rising at the full 115 training positives for the strongest deployed detector (paired 90-to-100% difference +0.027, 95% CI [0.019, 0.033]); we therefore claim only that the deployed system has not been shown to have saturated and that its strongest component demonstrably has not, not that the quantity of positives has been proven the sole ceiling.

An accompanying error and label analysis, conducted against independent evidence rather than against its own plausibility, finds that the cases the system misses are those of minimal pre-excitation, a narrower QRS confirmed by an independent on-machine measurement whose sign we show is delineator-dependent; that uncertain labels are not enriched among the misses; and that a share of the apparent false positives are recordings the source corpus itself codes as pre-excited, placing part of the label problem in the negative class and making our reported precision conservative.

The deployed system is a two-member percentile-rank vote, whose output we deliberately express as a rank within a frozen reference distribution rather than as a probability. On the held-out fold it attains an area under the receiver operating characteristic curve of 0.950 (95% CI [0.89, 1.00]) at a false-positive rate of 0.045%; this rests on 14 positives, so we read it as a floor on discriminative ability rather than a resolved operating performance. We present it not as a diagnostic tool, which its stage of validation does not support, but as a screening pre-filter whose sensitivity is validated only in-distribution, and we argue that its most consequential use is to make affordable the larger, better-labeled corpus that would raise its own ceiling.