Learning from VAE Errors to support ECG-based Differential Diagnosis of Myocardial Scar Shayan Sharifi1 , Riccardo Treu2 , Ilaria Gandin3 , Federico Garoia2 , Marco Merlo2 , Giulia Cisotto1 Department of Mathematics, Informatics, and Geosciences, University of Trieste, Italy [email protected], [email protected] 2 Centre of Diagnosis and Management of Cardiomyopathies; Azienda Sanitaria Universitaria Giuliano Isontina; University of Trieste, Italy; Member of Ern Guard-Heart [email protected], [email protected] [email protected] 3 Department of Medical, Surgical and Health Sciences, University of Trieste, Italy [email protected]
arXiv:2609.05294v1 [cs.LG] 4 Sep 2026
1
Abstract. Late Gadolinium Enhancement (LGE) on cardiac magnetic resonance is a key marker of myocardial scar, but its limited accessibility motivates routine ECG-based screening. We evaluated whether β-variational autoencoder (VAE)-derived ECG representations can discriminate LGE+ from LGE- cardiomyopathic patients in a local cohort of 300 subjects. We compared 32-dimensional features from the foundation ECGx.AI model with those from a shallower β-VAE trained on normal PTB-XL ECGs, evaluating downstream classification and Dynamic Time Warping (DTW)-based reconstruction errors. ECGx.AI reached an area under ROC of 0.686 with Random Forest, while the proposed β-VAE reached 0.577 with sensitivity of 0.775 with Gradient Boosting. Notably, DTW-reconstruction errors significantly differed between classes in 10 out of 12 leads according to Mann-Whitney U test and help in classification, leading to an area under ROC of 0.643 with Logistic Regression, supporting their potential as markers of scar-related ECG alterations. Keywords: ECG · cardiomyopathy · β-VAE · DTW · machine learning.
1
Introduction and state of the art
Myocardial fibrosis and scar tissue, detectable as Late Gadolinium Enhancement (LGE) on cardiac magnetic resonance imaging (CMR), are key biomarkers for diagnosis and risk stratification in cardiomyopathies [1, 3]. In dilated cardiomyopathy (DCM) and non-dilated left ventricular cardiomyopathy (NDLVC), LGE supports disease characterization and clinical decision-making, but CMR is costly, time-consuming, and not uniformly available [4]. This motivates electrocardiography (ECG)-based screening strategies to identify patients most likely to benefit from CMR evaluation [11]. Standard 12-lead ECG is inexpensive, routinely acquired, and reflects cardiac electrical activity. Structural alterations,
2
Sharifi et al.
including fibrosis and focal scar tissue, may perturb ventricular depolarization and repolarization, inducing subtle changes in QRS complexes, ST segments, and T waves [10]. Since these changes are weak and spatially distributed, LGE prediction from ECG is well suited to machine learning and representationlearning methods. Recent ECG-AI studies have shown that deep learning can infer imaging-defined cardiac phenotypes from ECG. A supervised model detected left ventricular systolic dysfunction with an Area Under the ROC Curve (AUROC) of 0.93 [2], but such methods require large disease-specific labelled datasets. To mitigate label scarcity, unsupervised and self-supervised approaches have been explored, including convolutional VAEs for ECG compression and clustering [6], deep feature extraction for genetic discovery [14], compact latent representations preserving ECG morphology [7], and ECGx.AI, a β-VAE trained on approximately 1.1 million ECGs that detected reduced left ventricular ejection fraction (LVEF) with an AUROC of 0.89 [8]. Other recent works support ECG-based scar detection: In [5], the authors achieved an AUROC of 0.80 using fully supervised convolutional neural networks trained solely on raw ECGs to detect ischaemic myocardial scars (improving to 0.89 when clinical parameters were added), the XplainScar model [9] achieved an F1-score of 0.89 and sensitivity of 0.90 for left ventricular scar localization, ECGWiz [15] reached 0.74 accuracy using only 5 training ECGs, and in [19], the authors proposed a 34-layer neural network that achieved an AUROC of 0.80 for myocardial scar prediction from paired ECG/MRI data. Despite these advances, the use of unsupervised ECG representations for LGE prediction in DCM and NDLVC remains underexplored, especially in small local cohorts with limited CMR annotations. In this work, we assess whether the pretrained ECGx.AI encoder provides informative latent features for LGE+/LGE- classification, and compare it with those extracted by an alternative, shallower, β-VAE trained on a reduced set of normal public ECGs. Furthermore, we quantify its reconstruction errors via DTW and propose a way to classify LGE+/LGE- based on their different statistical distributions, following prior works on EEG signals [18].
2
Materials and Methods
Datasets and preprocessing We used two datasets: a local DCM/NDLVC cohort and the public PTB-XL ECG dataset. The local cohort included 300 patients, each with a 10 s 12-lead ECG; 174 are LGE+ at CMR, confirming myocardial scar, and 126 are LGE-. CMR annotations were given by expert cardiologists, data were fully anonymized, and all patients provided informed consent before LGE-CMR. From PTB-XL, available on PhysioNet [16], we extracted 3000 normal ECGs to train the proposed model. Preprocessing was kept minimal: data quality was assessed through expert visual inspection and exploratory analysis, and the median beat was extracted from each 10 s recording. As the sampling frequency is 500 Hz, each median beat contains 520 time samples. A set of 32 latent features was extracted for each patient included in the local DCM/NDLVC cohort using two different models: the ECGx.AI [8] and
VAE to support ECG-based Myocardial Scar Diagnosis
3
our proposed β-VAE model. They are both based on β-VAE, but differ in their encoder and decoder architectures, as well as in the training they underwent. Baseline foundation model. The ECGx.AI model was proposed by [8] as a foundation model to extract meaningful latent features, called AI Factors, from any given ECG signal. Its architecture is a β-VAE with 7 hidden 1D convolutional layers in the encoder, kernel size of 5, a 32-dimensional latent space, and 7 hidden 1D transposed convolutional layers in the decoder (not all details were available, so reproducibility could not be fully achieved). Training was conducted using the UMC Utrecht Dataset, including 1.1 million ECGs (median beats), from an unselected clinical population encompassing normal sinus rhythm and patients suffering different pathologies (e.g., myocardial infarction, atrial fibrillation, left bundle branch block, and reduced ejection fraction). A probabilistic Gaussian decoder was employed in this model, leading to the following loss function: " # N (xi − µθ,i (z))2 1X 2 log 2πσθ,i (z) + + βDKL [qϕ (z|x)||p(z)]. (1) Lβ-VAE = 2 (z) 2 i=1 σθ,i The hyperparameter β was a-priori selected by the authors of [8] in the set {8, 16, 32, 64, 128} with 32 giving the best results. The ECGx.AI model showed high reconstruction quality, with mean Pearson correlations of 0.90 and 0.88 in internal and external validation (on the UK Biobank cohort), respectively. When broken down by diagnostic subgroup, reconstruction fidelity peaked for conditions like sinus rhythm and pericarditis (mean r = 0.91 to 0.92), but dropped significantly for rarer abnormalities such as ventricular tachycardia and ST elevation suspected of myocardial infarction (mean r = 0.62 to 0.70). As Pearson correlation measures the fidelity of the overall shape of a time-series, Mean Squared Error (MSE) could have been useful to evaluate the point-to-point reconstruction fidelity. However, no MSE values were reported. In the task of reduced LVEF detection (a different clinical objective than the scar prediction targeted in our study), evaluated on an independent internal test set of 5, 669 patients, the Extreme Gradient Boosting (XGBoost) decision trees achieved satisfactory performance, with an AUROC of 0.89, slightly below the competitor deep neural network (fully black box) model developed by the authors for direct comparison on the same task (AUROC of 0.91). This performance was proved robust under external validation on the population-based UK Biobank cohort (4, 855 subjects), where the explainable pipeline achieved an AUROC of 0.89 compared to 0.86 for the black-box model. Interestingly, by computing Pearson correlation coefficients between conventional ECG clinical features and ECG factor values over all samples in the training dataset, the authors reported significant correlations: ventricular rate is mostly correlated to Factor 10 (r = 0.96), the PR interval to Factor 8 (r = 0.62), QRS duration to Factor 25 (r = −0.47), and the QT interval to Factor 30 (r = −0.52). This makes ECGx.AI a gray-box model, offering a partial - but very promising - degree of explainability. While external validation confirmed generalizability for reconstruction fidelity and LVEF classification, the specific correlations between individual latent factors (such as Factors 8, 10, 25, and 30) and standard clinical measurements were established
4
Sharifi et al.
on the development cohort and were not explicitly re-evaluated on external data. In the present study, this model was used to encode the local dataset using 32 AI Factors. Later, the 32-d representation was fed to 5 machine learning models for the classification of LGE+ versus LGE- patients. Our proposed β-VAE Model. We propose a shallower β-VAE model with 4 layers, both in the encoder and in the decoder. The decoder is deterministic, leading to the loss function: N
Lβ-VAE (θ, ϕ; x) =
1 X 2 (xi − x̂i ) + βDKL (qϕ (z|x)||p(z)) N i=1
(2)
with N the total number of samples in an ECG signal. The training used 3000 normal ECGs of the public PTB-XL Dataset [16] using a 90/10 train/validation split over 88 epochs, Adam optimizer, learning rate of 0.001, and batch size of 16. Signals were Z-score normalized, making the reconstruction error unitless (standard deviations squared). We evaluated 4 distinct β scheduling strategies [13], including cyclic, fixed, gradually descending and gradually ascending schedules, with β = 0.0001 yielding the best reconstruction, corresponding to an average validation MSE of 0.1. To evaluate reconstruction fidelity and address comparisons with ECGx.AI, the Pearson correlation coefficient (r) was calculated across the time dimension (520 time samples) between original and reconstructed waveforms for each lead, and averaged across all 12 leads. On the normal PTB-XL validation split, the proposed model achieved high reconstruction fidelity with an overall mean Pearson correlation of r = 0.94 ± 0.09, demonstrating comparable reconstructive capacity to the foundation model on its respective training distribution. When evaluated on the local cardiomyopathy cohort, the mean correlation dropped to r = 0.36±0.04. As expected for out-of-distribution pathological signatures, reconstruction alignment dropped more severely in scar-positive cases (r = 0.35 ± 0.24) compared to scar-negative cases (r = 0.38 ± 0.26). The trained model was used to reconstruct the ECGs of patients of the local DCM/NDLVC cohort. We quantified the similarity between any ECG in the cohort and its reconstruction via block-wise DTW [17], and collected all leadwise DTW-based errors. This provided us with a 12-dimensional feature vector per subject. Fig. 1 illustrates representative reconstructions of a normal PTB-XL sample alongside LGE- and LGE+ samples, demonstrating that MSE and DTW reconstruction errors increase in the presence of pathology. Downstream classification task. The aim of the study is to classify LGE+/LGE- patients based on their ECG data. Four different factors were varied in our experiments and fairly compared to find the best configuration to maximize classification performance. Factors are: (1) input representation, (2) input normalization, (3) machine learning model, and (4) cross-validation strategy. We used three input representations: two of them were obtained by encoding the ECG signals using the ECGx.AI and our shallow β-VAE models, trained as explained above. Both representations gave 32 latent features, separately, for
VAE to support ECG-based Myocardial Scar Diagnosis
(a) Normal ECG (PTB-XL)
(b) LGE-
5
(c) LGE+
Fig. 1: Comparison of original (black) and reconstructed (red) ECG signals, showing a median beat extracted from Lead II. (a) A normal ECG sample from the PTB-XL validation set (MSE = 0.0007, DTW = 0.2872, r = 0.9846). (b) A LGE- sample (MSE = 0.5280, DTW = 1.4314, r = 0.8734) and (c) a LGE+ sample (MSE = 0.6374, DTW = 11.6183, r = 0.6194).
each patient. Additionally, we explored the 12-d representation given by the 12 lead-wise DTW-based reconstruction errors obtained by attempting reconstruction of LGE+/LGE- patients’ ECGs using our shallow β-VAE. To normalize the input, we tested z-score normalization either lead-wise, patient-wise or feature-wise (strictly applied within each data partition to eliminate data leakage). Five common machine learning models were selected: Extra Trees (ET), Random Forest (RF), Gradient Boosting (GB), Logistic Regression (LR), and Support Vector Machine (SVM). Cross-validation was performed using leave-one-subject-out (LOSO) with metrics derived via bootstrap iterations, 20 repeated stratified 5-fold, or 200 repeated stratified 80/20 train/test. Hyperparameters for the evaluated classifiers were tuned separately within each outer validation split using an inner 3-fold grid search. Algorithm-specific parameter grids were explored, including regularization strength (C) for logistic regression and SVM models, the kernel coefficient (γ) for the RBF SVM, the number of estimators and maximum tree depth for tree-based ensembles, and the learning rate for gradient boosting. For classifiers supporting class weighting, both a default unweighted approach and a balanced weighting option were evaluated dynamically during the grid search. The balanced option assigns weights N inversely proportional to class frequencies according to Wj = 2N (where N is j the total number of training samples and Nj is the number of samples in class j), ensuring that the class with fewer samples inherently receives a proportionally higher weight. The configuration achieving the highest mean cross-validated performance across the inner folds was selected and refitted on the complete outer training set before evaluation on the held-out test set. Finally, a two-tailed non-parametric Mann-Whitney U test was applied leadwise to assess whether the distribution of the DTW-based reconstruction errors was significantly different between the two classes (LGE+/LGE-). The experimental conditions were varied one at a time, in order to fairly compare the classification results. Furthermore, data splits for cross-validation were kept the same for different input representations and machine learning models. Fig. 2 shows the entire processing pipeline used in this study.
6
Sharifi et al.
Fig. 2: Study processing pipeline overview All pipelines were implemented in Python 3 using the scikit-learn library, with fixed random seeds to ensure full computational reproducibility.
3
Results and Discussion
Table 1 and Table 2 report the classification performance when using ECGx.AI and the proposed shallower β-VAE encoders, respectively, as feature extractors. To identify the upper-bound performance achievable by each feature representation, these tables present the top-performing pipeline configurations resulting from a grid search across classifiers, normalizations, and validation schemes (with full comparative grids omitted due to space constraints). From Table 1, Random Forest (RF) yielded the highest performance for ECGx.AI with an AUROC of 0.686, an accuracy of 0.657, and a high sensitivity of 0.852 (alongside a moderate specificity of 0.389), which aligns with the clinical priority of maximizing recall to screen candidate scar-positive patients for CMR evaluation. When evaluated on identical validation splits (e.g., Leave-One-Subject-Out with latent features), the ECGx.AI foundation encoder demonstrates higher discriminative capacity than the proposed shallow β-VAE latents (AUROC 0.657 vs. 0.577 with Gradient Boosting). However, using DTW-based reconstruction errors as input features substantially improves performance for the shallow architecture, reaching an AUROC of 0.643 with Logistic Regression (80/20 split) and 0.638 with Extra Trees. Among all results (varying the 4 experimental factors mentioned in Section 2), Tables 1 and 2 report the best AUROC performance for the ECGx.AI model and the best ones for our proposed shallower model. It is worth noting that the best performance can be achieved with different classification models, if input representation is obtained with the ECGx.AI or the proposed β-VAE, respectively.
VAE to support ECG-based Myocardial Scar Diagnosis
7
Table 1: ECGx.AI: best classification models (ranked by AUROC). Model (CV) Normalization Sensitivity Specificity Accuracy AUROC Random Forest (5-Fold CV) Factor-wise 0.852 ± 0.049 0.389 ± 0.074 0.657 ± 0.036 0.686 ± 0.053 Random Forest (80/20 Split) Factor-wise 0.858 ± 0.048 0.394 ± 0.086 0.664 ± 0.042 0.676 ± 0.047 Random Forest (LOSO) None 0.873 ± 0.024 0.380 ± 0.044 0.666 ± 0.028 0.672 ± 0.032 Gradient Boosting (LOSO) Patient-wise 0.868 ± 0.026 0.383 ± 0.046 0.664 ± 0.029 0.657 ± 0.033 Logistic Regression (LOSO) Patient-wise 0.880 ± 0.025 0.340 ± 0.042 0.653 ± 0.028 0.644 ± 0.032
Table 2: Shallow β-VAE: best classification models (ranked by AUROC). Model (CV, input) Normalization Sensitivity Specificity Accuracy AUROC Logistic Regression (80/20, DTW errors) Lead-wise 0.885 ± 0.078 0.191 ± 0.115 0.596 ± 0.030 0.643 ± 0.068 Extra Trees (80/20, DTW errors) Lead-wise/None 0.815 ± 0.067 0.383 ± 0.089 0.635 ± 0.052 0.638 ± 0.066 Logistic Regression (5-Fold CV, DTW errors) Lead-wise 0.895 ± 0.078 0.188 ± 0.108 0.597 ± 0.038 0.637 ± 0.081 Extra Trees (5-Fold CV, DTW errors) Lead-wise/None 0.801 ± 0.067 0.373 ± 0.100 0.621 ± 0.057 0.626 ± 0.088 Logistic Regression (LOSO, DTW errors) None 0.753 ± 0.032 0.431 ± 0.042 0.618 ± 0.027 0.624 ± 0.032 Gradient Boosting (LOSO, latents) Patient-wise 0.775 ± 0.033 0.308 ± 0.041 0.579 ± 0.029 0.577 ± 0.034 Logistic Regression (LOSO, latents) Feature-wise 0.823 ± 0.029 0.216 ± 0.038 0.569 ± 0.029 0.574 ± 0.034 Support Vector Machine (LOSO, latents) Feature-wise 0.846 ± 0.028 0.151 ± 0.033 0.554 ± 0.029 0.573 ± 0.034
On the other hand, the best classification performance using the proposed β-VAE as a features extractor is achieved by Gradient Boosting with AUROC of 0.577 and sensitivity of 0.775. Interestingly, using DTW-based reconstruction errors as input, a notable performance boost was observed: Logistic Regression reached a top AUROC of 0.643, while Extra Trees preserved a more robustly balanced outcome, reaching an accuracy of 0.635 and a specificity of 0.383. Overall, this shows that a shallower pipeline (including architecture and training strategy) can yield comparable results with respect to the foundation model in a very specific differential diagnosis, such as distinguishing LGE+ from LGE-. Our results are also consistent with previous findings showing a trade-off between reconstruction and classification objectives [12], questioning the discriminative informativeness of latent features extracted from β-VAEs (both ECGx.AI and the proposed β-VAE), trained for high-fidelity reconstruction, for classification purposes. Reconstruction quality of the proposed β-VAE was then quantified by DTWbased similarity metric, yielding an average of 5.557 ± 2.705 (over the entire dataset). For the sake of completeness, we also evaluated errors in terms of MSE and obtained an average of 1.195 ± 0.463 (significantly higher than the value of 0.1 reported during training). The two-sided non-parametric Mann-Whitney U test revealed significant distributional differences between LGE+ and LGEpatients in all but two leads, as shown in Fig. 3.
4
Conclusions
This work investigated ECG-derived signatures of LGE-defined myocardial scar in DCM/NDLVC patients, aiming to support CMR prioritization through routine ECGs. We compared LGE+/LGE- classification based on latent features extracted from the pretrained ECGx.AI foundation model and from the proposed shallower β-VAE trained on normal public ECGs. Additionally, we computed DTW-based reconstruction errors from the β-VAE and evaluated their
8
Sharifi et al.
Fig. 3: Lead-wise DTW-based reconstruction errors in LGE- (blue bars) and LGE+ (red bars) classes, after averaging over patients. Statistical significance between classes was evaluated via two-tailed Mann-Whitney U test (* p < 0.05, ** p < 0.01, *** p < 0.001, n.s.: not significant). Error bars indicate standard error of the mean. ability to discriminate LGE+ from LGE- patients. Overall, our results suggest that while deep foundation representations offer stronger latent discriminative capacity, DTW-based reconstruction errors from a significantly lighter β-VAE trained exclusively on normal ECGs capture meaningful scar-related anomalies. While end-to-end black-box models achieve higher raw performance metrics, error-based representation learning provides a compact and interpretable screening alternative to assist in CMR prioritization. External multi-center validation, integration of clinical variables, and improved strategies to balance reconstruction and classification objectives remain necessary to assess and improve clinical utility.
References 1. Arbelo, E., Protonotarios, A., Gimeno, J.R., Arbustini, E., Barriales-Villa, R., Basso, C., Bezzina, C.R., Biagini, E., Blom, N.A., De Boer, R.A., et al.: 2023 esc guidelines for the management of cardiomyopathies: developed by the task force on the management of cardiomyopathies of the european society of cardiology (esc). European heart journal 44(37), 3503–3626 (2023) 2. Attia, Z.I., Kapa, S., Lopez-Jimenez, F., McKie, P.M., et al.: Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nature Medicine 25(1), 70–74 (2019) 3. Bacigalupi, E., Merlo, M., Barbati, G., Villatore, A., Molinari, L., Dal Ferro, M., Paldino, A., Radesich, C., Bassetto, G., Manzi, L., et al.: Predictors of disease progression in patients with left ventricular nondilated cardiomyopathy. Canadian Journal of Cardiology 42(6), 1099–1110 (2026) 4. Garoia, F., Capovilla, T.M., Reginato, A., Rubbo, F.M., Paldino, A., Gava, C.P.L.D., Bassetto, G., Dal Ferro, M., Merlo, M., Sinagra, G.: Genetic testing in cardiomyopathies: updates and future perspectives. Journal of Cardiovascular Medicine 26(7), 386–397 (2025)
VAE to support ECG-based Myocardial Scar Diagnosis
9
5. Gumpfer, N., Grün, D., Hannig, J., Keller, T., Guckert, M.: Detecting myocardial scar using electrocardiogram data and deep neural networks. Biological Chemistry 402(8), 911–923 (2021) 6. Jang, J.H., Kim, T.Y., Lim, H.S., Yoon, D.: Unsupervised feature learning for electrocardiogram data using the convolutional variational autoencoder. PLoS One 16(12), e0260612 (2021) 7. Kuznetsov, V.V., Moskalenko, V.A., Gribanov, D.V., Zolotykh, N.Y.: Interpretable feature generation in ECG using a variational autoencoder. Frontiers in Genetics 12, 638191 (2021) 8. van de Leur, R.R., Bos, M.N., Taha, K., Sammani, A., Yeung, M.W., van Duijvenboden, S., Lambiase, P.D., Hassink, R.J., van der Harst, P., Doevendans, P.A.: Improving explainability of deep neural network-based electrocardiogram interpretation using variational autoencoders. European Heart Journal - Digital Health 3(3), 415–424 (2022). https://doi.org/10.1093/ehjdh/ztac038 9. Nezamabadi, K., Sivalokanathan, S., Lee, J., Tanriverdi, T., Chen, M., Lu, D., Abraham, J., Sardaripour, N., Li, P., Mousavi, P., et al.: Explainable artificial intelligence identifies and localizes left ventricular scar in hypertrophic cardiomyopathy using 12-lead electrocardiogram. Scientific Reports 15(1), 33918 (2025) 10. Nijveldt, R., van der Vleuten, P.A., Hirsch, A., Beek, A.M., Hoogslag, G.E., van der Vleuten, H., Tio, R.A., Zijlstra, F., van Rossum, A.C.: Early electrocardiographic findings and MR imaging-verified microvascular injury and myocardial infarct size. JACC: Cardiovascular Imaging 2(10), 1187–1194 (2009) 11. Perotto, M., Pio Loco Detto Gava, C., Garoia, F., Folgheraiter, A., Allegro, V., Grilli, G., Porcari, A., Dal Ferro, M., Merlo, M., Sinagra, G.: Classification of cardiomyopathies: bringing order to complexity. European Heart Journal Supplements 28(Supplement_5), v159–v163 (2026) 12. Senellart, A., Solal, M., Allassonnière, S., Burgos, N.: Mitigating the reconstruction-detection trade-off in vae-based unsupervised anomaly detection. In: 2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI). pp. 1–5. IEEE (2026) 13. Shao, H., Yao, S., Sun, D., Zhang, A., Liu, S., Liu, D., Wang, J.W., Abdelzaher, T.: Controlvae: Controllable variational autoencoder. In: International Conference on Machine Learning. vol. 119, pp. 8655–8664 (2020) 14. Sieliwonczyk, E., Sau, A., Patlatzoglou, K., McGurk, K.A., Pastika, L., Thami, P.K., Mangino, M., et al.: Unsupervised feature extraction using deep learning empowers discovery of genetic determinants of the electrocardiogram. Genome Medicine 17(1), 1–15 (2025) 15. Vatsaraj, I., Unde, I., Loeffler, S., Mohsen, Y., Aronis, K., Carrick, R., Wu, K., Chrispin, J., Trayanova, N.: Ecgwiz: a unified and generalizable ai model for comprehensive ecg diagnosis. European Heart Journal 45(Supplement_1), ehae666– 3433 (2024) 16. Wagner, P., Strodthoff, N., Bousseljot, R.D., Kreiseler, D., Lunze, F.I., Samek, W., Schaeffter, T.: Ptb-xl, a large publicly available electrocardiography dataset. Scientific Data 7(1), 154 (2020). https://doi.org/10.1038/s41597-020-0495-6 17. Zancanaro, A., Badia, L., Cisotto, G.: Blockdtw: Efficient and scalable similarity search algorithm for healthcare-focused time-series. In: 2025 IEEE Conference on Standards for Communications and Networking (CSCN). pp. 1–6. IEEE (2025) 18. Zancanaro, A., Cisotto, G.: The impact of mis-labeled artefacts on deep learning models for eeg analysis: a case study. In: International Meeting on Computational Intelligence Methods for Bioinformatics and Biostatistics. pp. 248–255. Springer (2024)
10
Sharifi et al.
19. Zhang, J., Zhang, H., Jethani, N., Chinitz, L.A., Aphinyanaphongs, Y., Jankelson, L.: Abstract 17893: Deep learning for the prediction of myocardial scarring through electrocardiograms. Circulation 148(Suppl_1), A17893 (2023). https://doi.org/10.1161/circ.148.suppl\_1.17893