YOLO-1D ECG Delineation Weights (LUDB, single-lead, fold 0 / seed 2)

Pre-trained checkpoint of YOLO-1D, a 1-D anchor-free YOLO-style detector, for ECG wave delineation on the Lobachevsky University Electrocardiography Database (LUDB). The model takes one ECG lead (10 s at 500 Hz, z-scored) and returns the P, QRS and T waves as intervals [onset, offset] with a confidence score.

This is the single-lead model reported in the paper . The code, training configs and the full evaluation pipeline live in the source repository:

Source code: https://github.com/BystrovGladkov/YOLO_1d (branch EKG)

These weights correspond to the run experiments/final/fold0_seed2/train1/ of that repository.

Files

File	Description
`best.pt`	PyTorch checkpoint (`torch.save` dict with keys `model`, `ema`, `epoch`). Selected by the best mean F1 of the six fiducial points on the validation split (epoch 176 of 250). Inference uses the EMA weights. 1.08 M parameters, 8.5 MB.
`args.yaml`	Full resolved `TrainingSettings` of the run (architecture, augmentations, optimizer, etc.). Needed to build the model before loading `best.pt`.
`thresholds.json`	Per-class confidence thresholds (P = 0.30, QRS = 0.20, T = 0.25) tuned on the validation split and used for the test metrics below.
`metrics.json`	Complete test metrics of this run (overall, per rhythm group, with/without boundary points, detection mAP).
`splits/folds.csv`	Fold number of every LUDB record (all 5 folds).
`splits/fold_0/{train,val,test}.txt`	The exact example lists (`<record>_<lead>`) used for this checkpoint.
`LICENSE`, `NOTICE.txt`	Apache License 2.0 and attribution notice.

SHA-256 of best.pt: 31d9d1ffb32e159585cf56e5d90f0dc76ced54981c6249c814319cd177e403fb

LUDB split used for this checkpoint

LUDB contains 200 12-lead records (10 s, 500 Hz) with expert annotations of P, QRS and T boundaries. The split is patient-level (one record = one subject; all 12 leads of a record go to the same split) and was generated by data/make_splits.py of the source repository with the following protocol:

5-fold cross-validation, seed = 2026, folds stratified by rhythm group (sinus 175 records, af_flutter 15, paced 10). Every record is in the test set of exactly one fold.
For fold k: test = records of fold k; of the remaining records, ~10 % per rhythm group (at least one record per group) go to validation (early stopping, checkpoint selection, confidence-threshold tuning); the rest is train.
The unit of training is a single lead. 24 of the 2400 leads (7 records: 7, 8, 34, 95, 104, 111, 116) were excluded because of defective annotations inside the annotated zone (an unpaired wave boundary), leaving 2376 lead examples.

This checkpoint was trained on fold 0:

Split	Records	Lead examples	sinus / af_flutter / paced
train	144	1713	126 / 11 / 7
val	16	186	14 / 1 / 1
test	40	477	35 / 3 / 2

Record IDs:

test (40): 11, 16, 17, 19, 30, 34, 36, 40, 42, 44, 45, 50, 58, 78, 79, 88, 91, 100, 107, 113, 117, 123, 127, 129, 134, 135, 139, 142, 144, 146, 154, 167, 176, 181, 183, 184, 187, 191, 194, 196
val (16): 6, 7, 33, 35, 54, 55, 57, 70, 80, 85, 111, 118, 132, 138, 151, 169
train (144): all remaining records (see splits/fold_0/train.txt).

Do not evaluate this checkpoint on the records listed as train or val above; they were seen during training or used for model selection.

Model and training summary

Setting	Value
Architecture	YOLO-1D `nano`: DarkNet-style 1-D backbone (CSP blocks, SPP, PSA attention), FPN neck, anchor-free head with Distribution Focal Loss (16 bins)
Backbone strides	preset `B` → head strides 8 / 16 / 32
Input	1 channel × 5000 samples (10 s at 500 Hz), z-scored per lead, zero-padded to 5120 inside the transforms
Classes	`0 = P`, `1 = QRS`, `2 = T`
Loss weights	box 7.5, cls 0.5, DFL 1.5; task-aligned assigner
Optimizer	SGD, momentum 0.937, weight decay 5e-4, lr 0.01 → 1e-4, 3 warm-up epochs, batch 32
Epochs	250 (early-stopping patience 40); best epoch 176
EMA	decay 0.999, tau 200
Augmentations	16 ECG-specific augmentations (white / EMG / powerline / baseline-wander noise, real MIT-BIH NSTDB noise, time scaling, HRV warp, window warp, amplitude and per-wave scaling, inversion, step, time masking, mosaic, record splicing, band-pass filter); random edge masking of the unannotated record borders
Post-processing	1-D NMS with IoU 0.2, per-class confidence thresholds from `thresholds.json`
Training seed	2

Full details are in args.yaml.

Metrics

Evaluation protocol: per-lead detection → intervals; each predicted boundary is matched one-to-one to the nearest reference boundary within ±150 ms (ANSI/AAMI EC57). Se / PPV / F1 are computed per fiducial point; m ± σ is the mean ± standard deviation of the signed error (prediction − reference) in ms over matched pairs; FP/lead is the number of unmatched predictions per lead. Fiducials adjacent to the masked record edges (onset of the first and offset of the last wave of a lead) are excluded (exclude_boundary_points: true).

Test split of fold 0 (40 records, 477 leads) — this checkpoint

Fiducial	Se, %	PPV, %	F1, %	m ± σ, ms	FP/lead
P onset	98.55	97.20	97.87	1.3 ± 14.1	0.193
P offset	98.56	97.22	97.88	1.1 ± 11.9	0.193
QRS onset	100.00	99.89	99.95	0.2 ± 6.6	0.008
QRS offset	100.00	99.89	99.95	−0.1 ± 8.7	0.008
T onset	99.27	99.38	99.33	−1.8 ± 22.4	0.048
T offset	99.30	99.40	99.35	−1.8 ± 21.6	0.046
Mean over 6 points			99.05	σ = 14.2, \|m\| = 7.9

Usage

Installation, a ready-to-run inference example and the evaluation procedure are described in the source repository: https://github.com/BystrovGladkov/YOLO_1d (branch EKG, sections Installation and Quick Start → Run ECG Pipeline).

In short: clone the repository, put best.pt from this release at experiments/final/fold0_seed2/train1/best.pt (the matching args.yaml is already tracked there), and follow the Quick Start snippet. Inputs are single leads sampled at 500 Hz, at most 5000 samples long and z-scored; outputs are [onset, offset, confidence, class] in samples (1 sample = 2 ms), classes 0 = P, 1 = QRS, 2 = T.

The checkpoint can also be loaded without the repository's Trainer: build the model from args.yaml and load checkpoint["ema"]["ema_model_state_dict"] (or checkpoint["model"] for the raw, non-EMA weights).

Intended use and limitations

Research use: ECG delineation (P, QRS, T on- and offsets) on 500 Hz single-lead signals, trained and evaluated on LUDB only. The model was not validated on other databases, sampling rates or clinical populations.
Not a medical device. Not intended for clinical diagnosis.

Dataset attribution

LUDB — Kalyakulina A.I., Yusipov I.I., Moskalenko V.A., Nikolskiy A.V., Kosonogov K.A., Osipov G.V., Zolotykh N.Yu., Ivanchenko M.V. LUDB: A New Open-Access Validation Tool for Electrocardiogram Delineation Algorithms. IEEE Access, 2020, 8, 186181–186190. Distributed via PhysioNet under the Open Data Commons Attribution License (ODC-BY) v1.0. The data are not included in this release.
MIT-BIH Noise Stress Test Database (Moody G.B., Muldrow W.K., Mark R.G., 1984; PhysioNet, ODC-BY v1.0) — records bw, ma, em were used only as a training augmentation.

License

The weights and the accompanying files are released under the Apache License 2.0 (see LICENSE and NOTICE.txt), the same license as the source code.

Citation

If you use these weights, the architecture or the processing pipelines, please cite the paper (currently submitted to Trudy ISP RAN / Proc. ISP RAS):

@article{Gladkov2027,
  title   = {Neural Network YOLO-1D for Useful Signals Detection on Eddy Current Rail Defectograms},
  author  = {Gladkov, Artemy N. and Bystrov, Leonid Y. and Kuzmin, Egor V.},
  journal = {Trudy ISP RAN/Proc. ISP RAS},
  year    = {2027},
  note    = {Submitted}
}

and the source repository: https://github.com/BystrovGladkov/YOLO_1d

Downloads last month: -; Downloads are not tracked for this model. How to track