Upload an ECG
.hea + .dat of the same record (format 16, 500 Hz), or
CSV with a header row of lead names (I, II, III, aVR, aVL, aVF, V1–V6), one sample per row, mV, 500 Hz.
Result
The prediction will appear here once an ECG is analysed.
Read ECG from image
ExperimentalThe model was trained on 10-second digital signals. From an image, the signal has to be rebuilt from pixels, and a standard 3×4 print shows only 2.5 s of each lead. On synthetic paper ECGs, AUROC fell from 0.878 with the digital file to 0.787 for a clean 3×4 scan (0.752 for 6×2, 0.824 for 12×1), and at the image-mode threshold specificity was only 54% (3×4) versus 80% with the digital file, so many more false positives are expected. Real photographs (perspective, shadows, curved paper) may perform worse. Use the device’s digital export whenever possible. Details: Image-mode validation.
Drag on the image to select only the area containing the waveforms (leave out patient data and measurement text at the top). After analysis, the detected baselines (dashed) and the traced signal (green) are drawn over the image — if they do not follow the real ECG, do not use the result.
Preprocessed ECG (250 Hz, baseline-corrected)
How it works
Load an ECG
Choose a WFDB record or CSV export of a 10-second, 12-lead ECG — or, experimentally, a picture of a paper ECG.
Preprocess locally
The page reproduces the training pipeline in your browser: lead ordering, baseline removal, smoothing and resampling to 250 Hz.
Score on device
A 1-D residual network runs with WebAssembly and returns probabilities for LVEF ≤35%, ≤40% and <50%.
About the model
- 0.885
- AUROC, evaluation cohort
12,770 patients · 95% CI 0.879–0.892 - 84.7%
- Sensitivity at 0.0614
specificity 77.1% - 1.14 M
- Parameters
1-D residual CNN
- Trained with R
torchon MIMIC-IV-ECG waveforms labelled with MIMIC-IV-Echo transthoracic LVEF (44,567 ECG–TTE pairs; patients disjoint from the evaluation cohort). - Development-validation AUROC (LVEF ≤40%): 0.925. The evaluation cohort comes from the same institution.
- Preprocessing: lead ordering → baseline removal with a 1-s moving average → [1 2 1]/4 smoothing → 250 Hz → clipping at ±10 mV.
- Developed on single-centre data; performance and calibration may differ in other populations.
Image-mode validation
ExperimentalImage mode was tested on synthetic paper ECGs rendered from evaluation-cohort waveforms (red grid, calibration pulse, lead labels, header text, baseline wander), digitised by exactly the code this page runs, and scored by the model. Thresholds were fixed beforehand on 1,200 digitised development-validation images with mixed JPEG, low-resolution and rotation artefacts.
| Condition | n | AUROC digital file | AUROC image (95% CI) | Sens / spec, image | Sens / spec, digital | Same call as digital |
|---|---|---|---|---|---|---|
| Clean scan, 3×4+rhythm | 400 | 0.878 | 0.787 (0.739–0.837) | 87% / 54% | 81% / 80% | 68% |
| JPEG quality 60, 3×4 | 200 | 0.849 | 0.776 (0.699–0.850) | 82% / 56% | 80% / 79% | 72% |
| Low resolution (5 px/mm ≈ 127 dpi), 3×4 | 200 | 0.849 | 0.747 (0.667–0.829) | 92% / 38% | 80% / 79% | 58% |
| Rotated 1–2°, 3×4 | 200 | 0.849 | 0.773 (0.697–0.841) | 85% / 56% | 80% / 79% | 74% |
| Grey/black grid, 3×4 | 150 | 0.809 | 0.796 (0.691–0.884) | 90% / 58% | 74% / 78% | 74% |
| Clean scan, 6×2 | 300 | 0.873 | 0.752 (0.688–0.810) | 86% / 49% | 83% / 79% | 68% |
| Clean scan, 12×1 | 300 | 0.873 | 0.824 (0.766–0.877) | 76% / 75% | 83% / 79% | 86% |
Information ceiling. Even with perfect digitisation, a 3×4 print shows each lead for only 2.5 s (6×2: 5 s). In 3,000 evaluation ECGs this alone lowers AUROC from 0.883 to 0.818 (3×4) and 0.835 (6×2). To compensate, limb leads are completed with Einthoven/Goldberger identities and two reconstructions (segment in place, and whole R–R cycles tiled to 10 s) are averaged.
Image-mode thresholds (LVEF ≤40%, 85% / 90% sensitivity on development images): 3×4 0.123 / 0.098 · 6×2 0.114 / 0.081 · 12×1 0.120 / 0.104.
Sens / spec = sensitivity / specificity for LVEF ≤40% at the primary threshold (image: 85%-sensitivity threshold; digital: 0.0614). “Same call as digital” = share of ECGs with the same positive/negative result as the digital file. These are synthetic, flat, well-lit renderings from the same institution; real photographs (perspective, shadows, curved or creased paper, handwriting) are expected to perform worse and were not tested.