Semi-Supervised ECG Delineation

Updated 7 July 2026

Semi-supervised ECG delineation comprises techniques that combine scarce expert annotations with large amounts of unlabeled data to accurately identify P waves, QRS complexes, and T waves.
It employs varied representations, including 1D semantic segmentation, interval-wise detection, and beat-wise keypoint estimation, to capture detailed cardiac waveform structures.
The approach integrates self-training, teacher–student consistency, and physiology-aware data fusion to improve robustness and cross-domain performance.

Semi-supervised ECG delineation is the problem of recovering clinically relevant ECG waveform structure when only a limited subset of recordings carries expert delineation labels and a larger corpus is available only as unlabeled ECG, weak labels, or automatically generated pseudo-labels. In current literature, the task is formulated as 1D semantic segmentation, interval-wise detection with boundary regression, or beat-wise keypoint estimation, but the common objective is the same: identify P waves, QRS complexes, and T waves with sufficient temporal precision to derive fiducial points and intervals such as PR, QRS duration, and QT. Recent work has moved from purely supervised delineators toward self-training, teacher–student consistency, mixed-quality supervision, and physiology-constrained data fusion, while simultaneously exposing how strongly performance depends on dataset diversity, arrhythmia coverage, and evaluation protocol (Avetisyan et al., 2024, Park et al., 24 Jul 2025, Lim et al., 15 May 2026).

1. Task definition and representational variants

In the contemporary literature, ECG delineation denotes more than R-peak detection. One formulation divides the ECG into sample-wise classes such as background, P wave, QRS complex, and T wave; another predicts, for each temporal interval, whether a wave is present together with its start and end coordinates; a third predicts wave-specific keypoint probability maps. These formulations differ in output geometry but all target wave-level structure rather than mere beat timing (Park et al., 24 Jul 2025, Avetisyan et al., 2024, Atienza et al., 2024).

The interval-wise formulation in ECG-CODE is explicit: for each non-overlapping interval and for each class $c \in \{\mathrm{P,QRS,T}\}$ , the model outputs a confidence $PC_{k,c}$ , a start $PS_{k,c}$ , and an end $PE_{k,c}$ , trained with a tolerant confidence loss and a start-end loss,

$\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$

where $CL$ has a dead zone for $|PC-TC|<0.25$ , and $SEL$ has a tolerance region for $SS=(PS-TS)^2+(PE-TE)^2<0.15$ and is gated by object presence $TC$ (Avetisyan et al., 2024). By contrast, the segmentation formulation used in the SemiSegECG benchmark treats delineation as a 1D semantic segmentation problem over a 10 s single-lead signal resampled to 250 Hz, with sample-wise labels for waveform components and clinical interval extraction delegated to post hoc onset/offset transitions (Park et al., 24 Jul 2025).

KEED introduces a different representation: beat-wise keypoint estimation inspired by human pose estimation. For each R–R interval resampled to a fixed length $PC_{k,c}$ 0, the model outputs $PC_{k,c}$ 1 keypoint probability maps $PC_{k,c}$ 2, with presence decided by a threshold $PC_{k,c}$ 3 on $PC_{k,c}$ 4 and localization derived by $PC_{k,c}$ 5. In the reported experiments, KEED is trained and evaluated only for P-wave presence/absence and P-peak localization, but the architecture is presented as generic for onset, peak, and offset of P, QRS, and T without architectural changes (Atienza et al., 2024). This suggests that semi-supervised delineation is not tied to a single output representation; rather, semi-supervision attaches to whichever representation admits confidence estimation, structural regularization, or pseudo-label selection.

2. Why semi-supervision emerged

The central driver is label scarcity. Public delineation datasets are small, heterogeneous in lead configuration and sampling rate, and often limited in pathological diversity. DENS-ECG, for example, is fully supervised and trained on 105 QTDB records with dense sample-wise labels, but that setting already illustrates the cost of manual delineation and the dependence of deep models on carefully curated annotations (Peimankar et al., 2020). ECG-DelNet makes the scarcity even more explicit by distinguishing approximately 30 manually delineated beats per QTDB recording from a much larger set of automatic DSP-generated labels, resulting in $PC_{k,c}$ 6 beats with both high- and low-quality labels and $PC_{k,c}$ 7 beats with only low-quality labels (Jimenez-Perez et al., 2020).

Several adjacent regimes appeared before explicit semi-supervised delineation. ECG-DelNet uses noisy-label pre-training on low-quality labels followed by fine-tuning on expert labels; the paper describes this as mixed-quality labeling rather than semi-supervision, but functionally it is weak supervision with a clean refinement stage. The pseudo-synthetic augmentation study of 2021 addresses the same bottleneck by generating labeled ECG traces from pools of real P, PQ, QRS, ST, T, and TP segments under rhythm rules for AF, VT, AV block, sinus arrest, ST elevation/depression, and beat-level merging events; synthetic-only training outperforms real-only training, and real plus synthetic training performs best (Jimenez-Perez et al., 2021). A different precursor is KEED, which is fully supervised on only $PC_{k,c}$ 8 labeled beats after removing overlap with MIT-PWave; its label-efficient behavior and direct probabilistic outputs make it naturally compatible with later pseudo-labeling schemes, although the paper itself does not use unlabeled data (Atienza et al., 2024).

Another motivation is domain mismatch. Deep learning based ECG segmentation for diverse arrhythmias shows that models trained on standard public delineation datasets can retain high aggregate benchmark scores while failing substantially on underrepresented arrhythmias such as tachycardias and VT. In that study, a model trained on LUDB performs much worse on ST and VT than one trained on a more arrhythmia-diverse internal dataset, despite both using the same U-Net-like segmentation paradigm (Joung et al., 2023). This directly motivates semi-supervised use of large unlabeled arrhythmia-rich corpora: the problem is not only the number of labels, but the coverage of morphologies represented by those labels.

3. Principal methodological families

Semi-supervised ECG delineation currently comprises several distinct but overlapping families. The most direct is self-training with pseudo-labels; the most systematic evaluation has been done with teacher–student semantic segmentation algorithms; and physiology-aware data fusion has recently become a major refinement. Adjacent weakly supervised families include mixed-quality labeling and pseudo-synthetic supervision.

Paradigm	Core mechanism	Representative work
Self-training / pseudo-labeling	Train teacher on labeled delineations, pseudo-label unlabeled ECG, retrain from scratch, fine-tune on gold labels	ECG-CODE (Avetisyan et al., 2024)
Semi-supervised semantic segmentation	Apply MT, FixMatch, CPS, ReCo, ST++ to 1D ECG segmentation	SemiSegECG (Park et al., 24 Jul 2025)
Pattern-guided bidirectional fusion	L2U and U2L CutMix guided by segmentation-space similarity and confidence gating	CardioMix (Lim et al., 15 May 2026)
Mixed-quality weak supervision	Pre-train on low-quality automatic labels, fine-tune on expert labels	ECG-DelNet (Jimenez-Perez et al., 2020)
Pseudo-synthetic supervision	Generate rule-based labeled ECGs and train with instance- and boundary-aware losses	(Jimenez-Perez et al., 2021)

ECG-CODE is the clearest self-training pipeline. A teacher is first trained on the expert-annotated ISP training set. That teacher is then applied to 50% of PTB-XL, and a delineation score is computed as

$PC_{k,c}$ 9

so that confident presence and confident absence are treated symmetrically. The top $PS_{k,c}$ 0 of predictions for each object are retained as pseudo-labels, a new model is trained from scratch on those pseudo-labels, and the resulting model is fine-tuned on ISP. No joint supervised-plus-unsupervised loss is used; the procedure is curriculum-style and staged (Avetisyan et al., 2024).

The benchmarked semantic-segmentation family instead adopts the standard semi-supervised objective

$PS_{k,c}$ 1

or its method-specific equivalent, where $PS_{k,c}$ 2 is sample-wise cross-entropy on labeled signals and $PS_{k,c}$ 3 is supplied by Mean Teacher, FixMatch, CPS, ReCo, or ST++. In this setting, ECG-specific weak augmentation is random resized cropping, strong augmentation is RandAugment over powerline noise, sine-wave noise, amplitude scaling, and white noise, and all methods are adapted to 1D encoder–decoder architectures with either ResNet-18 or ViT-Tiny backbones (Park et al., 24 Jul 2025).

CardioMix refines the augmentation side of semi-supervision. It is not a standalone learner but a plug-and-play module for consistency-based SemiSeg. Its defining feature is that mixing is performed in segmentation space rather than at random: for two segments $PS_{k,c}$ 4 and $PS_{k,c}$ 5, similarity is

$PS_{k,c}$ 6

This supports labeled-to-unlabeled fusion and unlabeled-to-labeled fusion, with the latter gated by a segment-confidence threshold, typically $PS_{k,c}$ 7, computed from teacher softmax outputs. The motivation is specifically ECG-structural: random CutMix can create physiologically implausible P–T–P or T–T sequences, whereas pattern-guided fusion preserves plausible P–QRS–T ordering (Lim et al., 15 May 2026).

Mixed-quality and pseudo-synthetic supervision remain adjacent rather than fully semi-supervised. ECG-DelNet pre-trains on noisy automatic delineations, then fine-tunes on expert labels (Jimenez-Perez et al., 2020). The pseudo-synthetic generator of 2021 uses expert rules to produce “perfectly” labeled synthetic masks and pairs these with two additional losses: BoundaryLoss, which applies Dice on edge maps near boundaries, and F1InstanceLoss, which penalizes disagreement in the number of contiguous wave instances (Jimenez-Perez et al., 2021). These regimes do not use unlabeled real ECG, but they address the same label-efficiency problem and often function as practical precursors to semi-supervised learning.

4. Datasets, preprocessing, and evaluation frameworks

Three data ecosystems currently anchor the field. The first is the ECG-CODE setting: ISP as the main expert-labeled 12-lead delineation dataset, LUDB as an external labeled source and test set, and PTB-XL as unlabeled data for pseudo-labeling. In this pipeline, all signals are resampled to 1000 Hz, z-score normalized, and optionally perturbed during training with bandpass and notch filtering. Input is a stack of 12 mel spectrograms, one per lead, and post-processing merges gaps shorter than 300 samples and removes segments shorter than 50 samples (Avetisyan et al., 2024).

The second ecosystem is SemiSegECG, the first systematic benchmark for semi-supervised semantic segmentation in ECG delineation. It unifies LUDB, QTDB, ISP, Zhejiang, PTB-XL, and mECGDB. All labeled ECGs are cropped or zero-padded to 10 s, resampled to 250 Hz, bandpass filtered to $PS_{k,c}$ 8 Hz, z-score normalized, and treated per lead as independent training samples. The benchmark defines in-domain labeled ratios of $PS_{k,c}$ 9, $PE_{k,c}$ 0, $PE_{k,c}$ 1, and $PE_{k,c}$ 2, and a cross-domain setting in which labeled training data are merged from LUDB, QTDB, ISP, and Zhejiang while PTB-XL supplies unlabeled data; mECGDB is used only for external interval evaluation (Park et al., 24 Jul 2025).

The third ecosystem is CardioMix, which adopts the SemiSegECG preprocessing and focuses on single-lead dense segmentation. Window size $PE_{k,c}$ 3 for pattern-guided mixing is sampled from $PE_{k,c}$ 4, corresponding to 1–5 s at 250 Hz, with stride $PE_{k,c}$ 5. This scale is chosen to preserve beat-level or multi-beat cardiac pattern continuity during fusion (Lim et al., 15 May 2026).

Evaluation varies with task representation. ECG-CODE reports sensitivity, PPV, F1, and $PE_{k,c}$ 6 timing error for each fiducial class under a 150 ms matching tolerance, excluding the first and last second of LUDB due to missing annotations (Avetisyan et al., 2024). SemiSegECG and CardioMix instead use mean Intersection over Union,

$PE_{k,c}$ 7

and mean absolute error of clinically derived PR, QRS, and QT intervals (Park et al., 24 Jul 2025, Lim et al., 15 May 2026). KEED, because it is evaluated only for P-wave presence and peak localization, reports accuracy, sensitivity, specificity, peak error in samples, and runtime on MIT-PWave and BUT PDB (Atienza et al., 2024). The multiplicity of metrics is not incidental: sample-wise overlap, fiducial-point accuracy, interval accuracy, and runtime stress different aspects of delineation quality.

5. Empirical findings

The current empirical picture is consistent on three points: unlabeled or weakly labeled data help most when labels are scarce, gains are largest for subtle structures such as P and T waves, and physiology-aware augmentation matters.

In ECG-CODE, supervised training on ISP generalizes better to LUDB than the reverse, indicating that dataset breadth and abnormality coverage matter as much as raw performance on a familiar benchmark. On the ISP test set, self-training with PTB-XL improves P onset F1 from $PE_{k,c}$ 8 to $PE_{k,c}$ 9, P offset F1 from $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 0 to $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 1, T onset F1 from $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 2 to $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 3, and T offset F1 from $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 4 to $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 5, while QRS F1 declines slightly from $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 6 to $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 7. Error dispersion also falls for several fiducials. The reported interpretation is that abundant unlabeled data help more for subtle, variable waves than for QRS, which is already comparatively easy (Avetisyan et al., 2024).

The SemiSegECG benchmark shows that standard semi-supervised segmentation methods transfer meaningfully to ECG, but with nontrivial modality-specific behavior. In-domain gains are strongest at $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 8 and $\mathcal{L}=\sum_{k,c}\left(CL_{k,c}+SEL_{k,c}\right),$ 9 label ratios. ViT-Tiny almost always outperforms ResNet-18 for the same method and labeled ratio, and Mean Teacher is frequently among the strongest methods. In the cross-domain setting with PTB-XL as unlabeled data, ViT-Tiny improves from Scratch mIoU $CL$ 0 to $CL$ 1 with MT and $CL$ 2 with ST++, while average interval MAE drops from $CL$ 3 ms to $CL$ 4 ms with MT or FixMatch. By contrast, the ResNet-18 backbone gains little or can degrade slightly under out-of-domain unlabeled data (Park et al., 24 Jul 2025). A common misconception is therefore that more unlabeled data are automatically beneficial; the benchmark shows that the effect depends strongly on representation capacity and domain shift.

CardioMix strengthens those results by replacing random or uncertainty-driven CutMix with cardiac pattern-guided fusion. On LUDB with MT as the base method, mIoU at $CL$ 5 labels rises from $CL$ 6 with plain MT to $CL$ 7 with CardioMix; on QTDB at $CL$ 8, the gain is from $CL$ 9 to $|PC-TC|<0.25$ 0. On LUDB, MT plus CardioMix at $|PC-TC|<0.25$ 1 labels surpasses vanilla MT at $|PC-TC|<0.25$ 2 labels in both mIoU and interval MAE. In the cross-domain QTDB-labeled / ISP-unlabeled setting, the average improvement over the base method is $|PC-TC|<0.25$ 3 mIoU for CardioMix versus $|PC-TC|<0.25$ 4 for UPC. Ablations show a graded pattern: random L2U helps, signal-similarity matching helps slightly more, segmentation-pattern similarity helps more again, adding U2L improves further, and confidence gating yields the best result. The paper also reports a cardiac pattern consistency ratio of approximately $|PC-TC|<0.25$ 5 for CardioMix, higher than random or signal-based selection (Lim et al., 15 May 2026).

Weakly supervised and synthetic approaches show the same label-efficiency trend. ECG-DelNet reports F1 gains from low-quality-label pre-training of $|PC-TC|<0.25$ 6 for P, $|PC-TC|<0.25$ 7 for QRS, and $|PC-TC|<0.25$ 8 for T, relative to training on high-quality labels alone (Jimenez-Perez et al., 2020). The pseudo-synthetic augmentation study finds synthetic-only training to outperform real-only by about $|PC-TC|<0.25$ 9 percentage points in $SEL$ 0, and real plus synthetic to improve $SEL$ 1 by about $SEL$ 2 points over real-only (Jimenez-Perez et al., 2021). These results do not constitute semi-supervised learning in the strict sense, but they show that additional supervision proxies can substitute for scarce expert delineations.

Efficiency and structure of the output representation also matter. KEED, though not semi-supervised, attains competitive P-wave performance with only $SEL$ 3 labeled beats and reduces inference time by factors ranging from $SEL$ 4 to $SEL$ 5 relative to the wavelet baselines reported in that paper (Atienza et al., 2024). This suggests that keypoint-probability outputs may be attractive teachers or pseudo-label sources in future semi-supervised pipelines, because they yield both localization and a directly usable confidence measure.

6. Limitations, misconceptions, and research directions

Semi-supervised ECG delineation is sometimes treated as synonymous with self-training. The literature no longer supports that simplification. Self-training with pseudo-labels is one branch, but benchmarked ECG SemiSeg also includes EMA consistency, thresholded pseudo-label consistency, cross pseudo supervision, regional contrastive learning, and multi-stage self-training (Park et al., 24 Jul 2025). CardioMix adds a further distinction: the main contribution is not a new unsupervised loss but a physiology-aware way to exchange information between labeled and unlabeled sets (Lim et al., 15 May 2026).

A second misconception is that vision augmentations transfer unchanged to ECG. The benchmark explicitly finds horizontal flipping harmful because it reverses P–QRS–T order, and baseline shift harmful as a strong augmentation in that setting (Park et al., 24 Jul 2025). CardioMix shows that random temporal CutMix can create implausible cardiac sequences, which is why it replaces random region selection with segmentation-pattern matching (Lim et al., 15 May 2026). The plausible implication is that semi-supervised ECG delineation is unusually sensitive to structural priors: invariance should be enforced only under transformations that preserve cardiac semantics.

Several unresolved problems remain. Multi-lead structure is still underused in the main benchmark, which treats each lead as an independent sample, whereas ECG-CODE operates jointly on 12-lead mel spectrograms (Park et al., 24 Jul 2025, Avetisyan et al., 2024). Arrhythmia coverage remains inadequate: supervised work on diverse arrhythmias shows that standard benchmark performance can mask poor behavior on tachycardia, AFL, or VT (Joung et al., 2023). Pseudo-label quality is still only partially studied: ECG-CODE does not ablate different $SEL$ 6 thresholds or PTB-XL fractions, and CardioMix remains dependent on early pseudo-label quality even with confidence gating (Avetisyan et al., 2024, Lim et al., 15 May 2026). Cross-domain unlabeled data are helpful for transformers but not consistently for CNNs, so unlabeled data selection and adaptation remain open problems (Park et al., 24 Jul 2025).

The most technically promising directions already appear in adjacent work. ECG-CODE explicitly leaves consistency regularization and teacher–student methods as future extensions on top of pseudo-labeling (Avetisyan et al., 2024). CardioMix points toward combining pattern-guided fusion with self-supervised representation learning on large unlabeled ECG pools (Lim et al., 15 May 2026). Structured signal models suggest another route: graph-constrained changepoint detection encodes physiological transition structure and is described as amenable to semi-supervised graph search with mixed supervised and unsupervised criteria, while variable-projection beat models produce low-dimensional, interpretable parameters that can be fitted without dense labels and then used as structured latent features (Fotoohinasab et al., 2021, Böck et al., 2021). Sampling-rate-independent wavelet delineation similarly suggests a route to consistent pseudo-label generation across heterogeneous databases by recalculating scales as $SEL$ 7 rather than retuning filters per dataset (Chereda et al., 2016).

Taken together, current research defines semi-supervised ECG delineation as a convergence of three agendas: exploiting unlabeled ECG at scale, preserving physiological validity in both model outputs and augmentations, and reducing dependence on narrow benchmark distributions. The field has moved beyond the question of whether unlabeled data can help. The remaining question is how to exploit them without breaking cardiac structure, degrading cross-domain robustness, or overfitting to pseudo-label artifacts.