Before a model sees an electrocardiogram (ECG), the signal is usually rescaled in height (to zero mean and unit variance, or to [0, 1]) and often resampled so each heartbeat has a fixed number of samples. Separately, the boundary labels the model learns from, such as where a T-wave ends, are usually placed by one cardiologist, and each cardiologist has consistent habits. This proposal argues that these two facts are linked.
A rater's habit is a rule about where to put a mark. It might be "a fixed number of milliseconds after the peak," "a fixed fraction of the beat," or "where the trace visibly flattens at this voltage." Each rule is simple, but only in its own units. Normalizing the waveform changes the units the model works in. So the choice of normalization decides whether one annotator's bias looks to the model like a single constant offset or like a tangle that depends on heart rate and amplitude.
We propose to measure which units make each annotator's bias simplest. We will then build a delineation model that works in those units and represents each annotator with a few interpretable parameters. The practical payoff is the ability to train on labels from one annotator and then adapt to a different clinician's conventions, or to a new hospital's, from a handful of labeled beats. That adaptation should also keep working at heart rates and signal amplitudes the new annotator never labeled. A second payoff is diagnostic: the fitted parameters can separate how much of a cross-dataset accuracy drop comes from annotation conventions rather than from recording hardware.
Core Claim
Write $t^*$ for the boundary location a model would predict from the waveform alone. Write $t_a$ for where annotator $a$ actually marks it. Write $\phi$ for a coordinate system: raw milliseconds, fraction of the RR interval (the time between consecutive beats), or an amplitude-threshold coordinate in raw millivolts or in normalized units.
Our hypothesis is that, for each annotator and boundary type, there is a coordinate system in which bias is nearly constant:
$$\phi(t_a) - \phi(t^*) \approx b_a,$$
where $b_a$ is a scalar offset that does not depend on heart rate or amplitude. In the wrong coordinate system, the same bias becomes a function of heart rate and gain, and a model trained on one annotator memorizes it entangled with the waveform.
This matters for normalization in a concrete way. Resampling beats to fixed length turns a constant millisecond bias into one that grows with heart rate. Z-scoring turns a threshold set in millivolts into one that moves with signal energy. The claim is testable and falsifiable: if a constant millisecond offset explains annotator differences as well as anything else does, normalization and bias really are separable. That would be a useful negative result.
Relation to Prior Work
The motivation comes from two separate observations. Transformer-based Annotation Bias-aware Medical Image Segmentation (TAB) learns one query vector per annotator and separates systematic preference from random error in 2D fundus segmentation. It shows that annotator style is structured rather than noise. However, TAB's preference is an opaque embedding in image space, and the question of what coordinates the preference lives in never arises for images of fixed scale.
SemiSegECG benchmarks 1D ECG delineation across LUDB, QTDB, ISP and Zhejiang, and it treats all four label sets as a single homogeneous truth. Its finding that time-reversal augmentation hurts delineation constrains our design. Transforms that ignore the physiology of the signal are harmful, so our candidate coordinates are limited to physiologically meaningful ones: absolute time, cycle fraction, and voltage thresholds.
The normalization side supplies the transforms themselves:
- fixed-length pulse resizing in Regression analysis for peak designation in pulsatile pressure signals;
- dual-axis [0, 1] normalization in Clustered photoplethysmogram pulse wave shapes and their associations with clinical data, which also notes that manual shape labels are inconsistent;
- log and z-score harmonization in SPINDLE.
All three treat normalization as a fixed preprocessing step. SPINDLE additionally collapses its five raters into one consensus. We instead treat normalization as the object of study, judged by how simply it expresses rater disagreement.
Influence of Inter-Annotator Variability on Automatic Mitotic Figure Assessment constrains the data budget. It finds that three annotators capture most of the benefit of a panel (the IQR of single-annotator models is 0.038, versus 0.016 for three-annotator consensus). This justifies a small three-rater re-annotation study rather than a large panel.
Towards Domain Generalization for ECG and EEG Classification attributes cross-center drops entirely to physical shift. Our decomposition experiment tests part of that assumption for delineation.
Research Program
The work proceeds in three parts.
Part 1: measure bias structure in existing labels. We use the subset of QTDB that carries annotations from more than one expert, and we treat whole datasets (LUDB, QTDB, ISP, Zhejiang) as distinct annotating groups. For each pair of annotators and each boundary (P onset and offset, QRS onset and offset, T offset), we compute boundary differences on shared beats. We then regress these differences on heart rate and on local amplitude and slope under each candidate coordinate system. The coordinate system in which the residual dependence is smallest is the annotator's "native" coordinate.
Part 2: collect controlled human data. Three cardiologists re-annotate a stratified subset of LUDB. Each record is shown to them rendered at different display gains and sweep speeds, in randomized order. This directly tests whether human boundary placement shifts with rendering scale. Rendering scale is the human analogue of waveform normalization, and no existing dataset measures its effect. To check that the fitting procedure itself works, we also simulate raters with known rules (a millisecond offset, a cycle fraction, or a millivolt slope threshold) on real LUDB beats. We then confirm that the procedure recovers each simulated rater's rule.
Part 3: build the model. A shared 1D delineation network is trained on waveforms, and each annotator is attached through an explicit head. The head is a small number of parameters, $b_a$, applied in that annotator's selected coordinate system, with a separate variance term for random placement error in the spirit of TAB's covariance head. Adapting to a new annotator means fitting only these parameters from $k$ labeled beats.
Baselines
Naive reference. The simplest reference is a ResNet-18 1D segmenter trained on a single annotator's labels with z-score input and no annotator modeling. This is the standard pipeline evaluated in SemiSegECG. It is the "do nothing" point: it shows how badly a single-annotator model misses another annotator's marks.
Strongest delineation reference. The strongest delineation reference is the arrhythmia-guided segmenter of An Arrhythmia Classification-Guided Segmentation Model for Electrocardiogram Delineation, which reports F1 "exceeding 99%" for QRS and T waves and "over 97%" for P waves on LUDB. We also report the Mean Teacher mIoU of 73.6% with 1/16 of the labels from SemiSegECG. For QTDB, no reported result was found on Lacuna for millisecond boundary error. These tolerance-based F1 scores are near ceiling and hide millisecond-level annotator offsets, so our primary metric is signed and absolute boundary error in milliseconds against a held-out annotator.
Strongest annotator-aware competitor. The strongest annotator-aware competitor is a 1D port of TAB. It uses learnable per-annotator queries plus a low-rank Gaussian output, trained on raw and on z-scored input, and it adapts to a new annotator by fitting a new query from $k$ beats. This is the most flexible competitor, and it directly tests whether explicit coordinates add anything beyond a learned embedding.
Per-annotator constant offset. A simpler competitor is a per-annotator constant offset in raw milliseconds, fitted after training the naive model.
Ablation. The ablation that isolates our key ingredient keeps our full architecture but forces every annotator into one fixed coordinate system: raw milliseconds for all, or cycle fraction for all.
Fairness and success criterion. All methods share the same backbone, training data, augmentation (only the physiological noise augmentations from SemiSegECG), hyperparameter search budget and seeds. Beating the baselines means lower millisecond error on a held-out annotator at small $k$ (for example $k \le 50$ beats). It also means an error that stays flat when the adaptation beats and test beats come from disjoint heart-rate or amplitude ranges.
Why not just…?
Why not just average several annotators and z-score the input? Consensus removes idiosyncrasy only when a panel exists, and for most ECG biobanks it does not. Consensus also cannot tell a deployed model which convention a given hospital follows. TAB outperformed consensus-style baselines on RIGA (average soft Dice of 96.70% for the optic disc and 85.52% for the optic cup), which indicates that modeling individual styles helps. We still include a consensus-trained model and report its error against each individual annotator.
Why not just learn an annotator embedding, as TAB does, and let the network sort out coordinates? This is the real threat. With abundant per-annotator data spanning all heart rates, a learned embedding can represent any bias. Our claim is narrower: when adaptation data is scarce or narrow, a bias expressed in the correct coordinates extrapolates, while a free embedding does not. The disjoint-range adaptation experiment decides this. If the TAB port matches us there, the coordinate idea adds nothing beyond interpretability, and we will report that.
Why not just subtract a constant millisecond offset per annotator? If annotators' native coordinate is raw milliseconds, this is exactly the right model, and our method reduces to it. Part 1 tests this directly. We expect cycle-fraction behavior for T offset, since QT intervals scale with heart rate, and we expect amplitude-threshold behavior for low-slope P boundaries. However, this is an empirical question, and a result showing that constant offsets suffice is still informative about whether normalization and bias can be treated separately.
Why not just augment with time stretching and random gain? Augmentation makes the model insensitive to scale. It does not model a clinician who systematically marks later than peers. SemiSegECG also shows that augmentations that ignore physiology can hurt. We include a stretch-and-gain augmented baseline to measure whether it helps or hurts cross-annotator error.
Proposed Experiments
Decisive check 1: human rendering study. If the three cardiologists' boundary marks do not move with display gain or sweep speed beyond their test-retest noise, the amplitude-threshold hypothesis for humans is false. We would then restrict the claim to temporal coordinates.
Decisive check 2: coordinate recovery on simulated raters. The procedure must recover the known rule for each synthetic annotator. If it cannot, the Part 1 fits on real annotators are not trustworthy.
Decisive check 3: disjoint-range adaptation. We adapt to a held-out annotator, or to a held-out dataset treated as an annotator, using $k$ beats from the low-heart-rate tertile, and test on the high tertile. We then repeat with the tertiles swapped, and repeat again for amplitude. We compare the coordinate-based head against the TAB port, the constant offset, and the fixed-coordinate ablation.
Decisive check 4: cross-dataset error decomposition. We train on LUDB and test on QTDB, ISP and Zhejiang. We measure the fraction of boundary error removed by fitting only the few annotator-convention parameters, with the network frozen. A large removable fraction would show that part of what BioDG treats as physical domain shift is, for delineation, annotation convention.
References
- Transformer-Based Annotation Bias-Aware Medical Image Segmentation
- SemiSegECG: A Multi-Dataset Benchmark for Semi-Supervised Semantic Segmentation in ECG Delineation
- Regression analysis for peak designation in pulsatile pressure signals
- Clustered photoplethysmogram pulse wave shapes and their associations with clinical data
- SPINDLE: End-to-end learning from EEG/EMG to extrapolate animal sleep scoring across experimental settings, labs and species
- Influence of Inter-Annotator Variability on Automatic Mitotic Figure Assessment
- Towards Domain Generalization for ECG and EEG Classification: Algorithms and Benchmarks
- An Arrhythmia Classification-Guided Segmentation Model for Electrocardiogram Delineation