Position: Evaluation of ECG Representations Must Be Fixed

Zachary Berger; Daniel Prakah-Asante; John Guttag; Collin M. Stultz

cs.LGFeb 19, 2026

Position: Evaluation of ECG Representations Must Be Fixed

Authors: Zachary Berger, Daniel Prakah-Asante, John Guttag, Collin M. Stultz

Organizations: Massachusetts Institute of Technology, Cambridge, MA, USA · Massachusetts General Hospital, Boston, MA, USA

Abstract

This position paper argues that current benchmarking practice in 12-lead ECG representation learning must be fixed to ensure progress is reliable and aligned with clinically meaningful objectives. The field has largely converged on three public multi-label benchmarks (PTB-XL, CPSC2018, CSN) dominated by arrhythmia and waveform-morphology labels, even though the ECG is known to encode substantially broader clinical information. We argue that downstream evaluation should expand to include an assessment of structural heart disease and patient-level forecasting, in addition to other evolving ECG-related endpoints, as relevant clinical targets. Next, we outline evaluation best practices for multi-label, imbalanced settings, and show that when they are applied, the literature's current conclusion about which representations perform best is altered. Furthermore, we demonstrate the surprising result that a randomly initialized encoder with linear evaluation matches state-of-the-art pre-training on many tasks. This motivates the use of a random encoder as a reasonable baseline model. We substantiate our observations with an empirical evaluation of five representative ECG pre-training approaches across six evaluation settings: the three standard benchmarks, a structural disease dataset, hemodynamic inference, and patient forecasting.

Figures & tables

Figure 1: Overview of the evaluation pipeline for 12-lead ECG representations. (a) Current practice focuses on arrhythmia/waveform tasks and macro-AUROC point estimates, which can produce misleading method rankings. (b) We propose a broader set of clinically relevant tasks and evaluation best-practices that more reliably assess methods. We find that no method consistently prevails and for many tasks, many methods overlap with the baseline of a randomly initialized encoder.

Dataset		Random	CLOCS	KED	HeartLang	MERL	D-BETA
PTB-XL	1%	0.770 0.759–0.780	0.752 0.740–0.764	0.861 0.852–0.869	0.645 0.632–0.658	0.823 0.813–0.832	0.867 0.858–0.875
Super	10%	0.832 0.822–0.842	0.807 0.796–0.817	0.894 0.886–0.901	0.790 0.780–0.800	0.884 0.876–0.892	0.887 0.879–0.895
	100%	0.861 0.851–0.870	0.821 0.810–0.830	0.909 0.902–0.915	0.824 0.815–0.834	0.901 0.893–0.908	0.893 0.885–0.901
PTB-XL	1%	0.689 0.673–0.704	0.657 0.626–0.688	0.783 0.770–0.795	0.574 0.548–0.599	0.757 0.740–0.772	0.791 0.781–0.802
Sub	10%	0.763 0.739–0.786	0.771 0.752–0.790	0.878 0.857–0.898	0.708 0.688–0.729	0.857 0.837–0.875	0.862 0.834–0.888
	100%	0.834 0.808–0.859	0.803 0.784–0.819	0.920 0.909–0.929	0.836 0.819–0.853	0.905 0.890–0.918	0.899 0.882–0.914

Table 1: On the standard arrhythmia/waveform benchmarks, KED and D-BETA appear to dominate according to macro-AUROC. However, once uncertainty is quantified, neither ECG pre-training method consistently prevails. A randomly initialized encoder is often competitive, often beating the ECG-specific pre-training methods CLOCS and HeartLang. Entries report macro-AUROC with 95% confidence intervals in the subscript under linear probing at varying levels of training data. Green indicates no significant difference from the top method.

Method	Norm	IMI	AMI	STTC	LVH	CLBBB
Random	0.891 0.877–0.903 / 0.839 0.813–0.863	0.864 0.840–0.884 / 0.556 0.506–0.609	0.894 0.874–0.915 / 0.646 0.593–0.700	0.827 0.801–0.851 / 0.319 0.275–0.368	0.917 0.896–0.935 / 0.648 0.593–0.704	0.973 0.934–0.999 / 0.880 0.784–0.955
CLOCS	0.859 0.843–0.872 / 0.788 0.763–0.812	0.674 0.644–0.702 / 0.299 0.262–0.341	0.857 0.832–0.883 / 0.590 0.537–0.646	0.823 0.796–0.849 / 0.319 0.279–0.366	0.900 0.875–0.921 / 0.593 0.528–0.650	0.984 0.974–0.993 / 0.676 0.551–0.794
KED	0.932 0.921–0.942 / 0.900 0.883–0.916	0.881 0.862–0.899 / 0.597 0.549–0.647	0.951 0.939–0.962 / 0.808 0.770–0.845	0.893 0.873–0.911 / 0.507 0.444–0.570	0.949 0.935–0.961 / 0.745 0.695–0.794	0.998 0.997–1.000 / 0.949 0.906–0.983
HeartLang	0.875 0.860–0.890 / 0.822 0.798–0.846	0.750 0.721–0.776 / 0.353 0.310–0.399	0.876 0.857–0.895 / 0.548 0.496–0.601	0.799 0.772–0.825 / 0.295 0.250–0.346	0.823 0.794–0.853 / 0.405 0.343–0.472	0.993 0.984–0.998 / 0.867 0.767–0.944
MERL	0.926 0.915–0.937 / 0.885 0.862–0.904	0.843 0.819–0.864 / 0.538 0.488–0.589	0.959 0.949–0.969 / 0.841 0.809–0.873	0.885 0.863–0.905 / 0.480 0.421–0.539	0.929 0.912–0.945 / 0.680 0.623–0.734	0.999 0.997–1.000 / 0.947 0.889–0.988
D-BETA	0.929 0.919–0.939 / 0.894 0.875–0.912	0.897 0.880–0.914 / 0.691 0.649–0.734	0.956 0.943–0.967 / 0.834 0.793–0.869	0.862 0.837–0.887 / 0.429 0.370–0.490	0.862 0.837–0.885 / 0.489 0.426–0.552	0.999 0.997–1.000 / 0.970 0.937–0.995

Table 2: Task-level results reveal heterogeneous behavior. Shown are the five highest-prevalence PTB-XL Sub labels plus CLBBB. Method rankings vary by endpoint. AUPRC can help capture differences masked by AUROC, as in the case of CLBBB. Each cell reports AUROC (top) and AUPRC (bottom) with 95% confidence intervals; methods not statistically different from the best are highlighted in green. Results on the remaining tasks are in Table 20 .

Method	PRC(S)	STE_	TAB_
Random	0.92 0.91–0.94 / 0.02 0.01–0.02	0.64 0.41–0.78 / 0.01 0.01–0.01	0.56 0.39–0.84 / 0.01 0.00–0.02
CLOCS	0.85 0.83–0.87 / 0.01 0.01–0.01	0.56 0.10–0.91 / 0.01 0.00–0.04	0.70 0.59–0.80 / 0.01 0.01–0.02
KED	0.97 0.96–0.98 / 0.03 0.03–0.05	0.59 0.30–0.94 / 0.01 0.00–0.05	0.74 0.52–0.98 / 0.03 0.01–0.13
HeartLang	0.87 0.85–0.89 / 0.01 0.01–0.01	0.61 0.20–0.93 / 0.01 0.00–0.05	0.27 0.14–0.50 / 0.00 0.00–0.01
MERL	0.76 0.73–0.79 / 0.01 0.00–0.01	0.92 0.84–0.99 / 0.11 0.02–0.43	0.87 0.80–0.97 / 0.04 0.01–0.12
D-BETA	0.97 0.96–0.98 / 0.04 0.03–0.06	0.71 0.51–0.95 / 0.02 0.01–0.06	0.66 0.23–0.95 / 0.02 0.00–0.07

Table 3: Tasks with very few positives can yield high-variance estimates that distort benchmark summaries. Confidence intervals are wide for two of the three tasks with lowest prevalence in PTB-XL Form . Each cell reports AUROC (top) and AUPRC (bottom) with 95% confidence intervals.

Method	ORIG	CLEAN	$\Delta$
Random	0.756 0.734–0.780	0.754 0.733–0.774	-0.002
CLOCS	0.715 0.687–0.741	0.710 0.690–0.731	-0.005
KED	0.851 0.827–0.875	0.862 0.846–0.876	+0.011
HeartLang	0.707 0.679–0.736	0.723 0.699–0.745	+0.016
MERL	0.853 0.839–0.866	0.847 0.833–0.860	-0.005
D-BETA	0.845 0.821–0.868	0.855 0.841–0.868	+0.009

Table 4: Removing low-support labels can materially change macro-AUROC and alter which method appears most performant. ORIG uses the standard PTB-XL Form test set; CLEAN excludes labels with fewer than 10 positive test examples. Values are macro-AUROC with 95% confidence intervals.

Method	SHD	LVEF $\leq 45$	TR
Random	0.81 0.79–0.82 / 0.77 0.76–0.79	0.86 0.84–0.87 / 0.62 0.59–0.65	0.78 0.76–0.81 / 0.23 0.20–0.27
CLOCS	0.72 0.71–0.74 / 0.69 0.67–0.70	0.78 0.76–0.79 / 0.48 0.45–0.51	0.71 0.69–0.74 / 0.14 0.12–0.17
KED	0.80 0.79–0.81 / 0.76 0.75–0.78	0.86 0.85–0.87 / 0.62 0.59–0.65	0.78 0.75–0.80 / 0.23 0.19–0.26
HeartLang	0.77 0.76–0.79 / 0.72 0.70–0.74	0.83 0.81–0.84 / 0.55 0.52–0.58	0.75 0.73–0.78 / 0.18 0.16–0.21
MERL	0.81 0.80–0.82 / 0.78 0.76–0.80	0.88 0.87–0.89 / 0.67 0.64–0.70	0.82 0.79–0.84 / 0.27 0.23–0.30
D-BETA	0.80 0.79–0.81 / 0.76 0.75–0.78	0.86 0.85–0.87 / 0.61 0.58–0.64	0.79 0.77–0.82 / 0.24 0.20–0.27

Table 5: On structural disease endpoints, CLOCS and HeartLang underperform while the other methods cluster closely. Each cell reports AUROC/AUPRC with 95% confidence intervals; methods not significantly different from the top-performing method are highlighted in green.

Dataset	All	ResNet	ViT
PTB-XL Super	$0.82{\scriptstyle\,\pm\,0.07}$	$0.87{\scriptstyle\,\pm\,0.01}$	$0.76{\scriptstyle\,\pm\,0.07}$
PTB-XL Sub	$0.79{\scriptstyle\,\pm\,0.09}$	$0.86{\scriptstyle\,\pm\,0.02}$	$0.71{\scriptstyle\,\pm\,0.08}$
PTB-XL Form	$0.70{\scriptstyle\,\pm\,0.07}$	$0.76{\scriptstyle\,\pm\,0.01}$	$0.64{\scriptstyle\,\pm\,0.05}$
PTB-XL Rhythm	$0.75{\scriptstyle\,\pm\,0.09}$	$0.81{\scriptstyle\,\pm\,0.03}$	$0.68{\scriptstyle\,\pm\,0.07}$
CPSC2018	$0.79{\scriptstyle\,\pm\,0.09}$	$0.86{\scriptstyle\,\pm\,0.01}$	$0.73{\scriptstyle\,\pm\,0.08}$
CSN	$0.76{\scriptstyle\,\pm\,0.09}$	$0.82{\scriptstyle\,\pm\,0.03}$	$0.69{\scriptstyle\,\pm\,0.07}$

Table 6: The competitiveness of the random baseline in this paper is not an artifact of a single configuration. Values are macro-AUROC with standard deviation across grid of randomly initialized encoders. All summarizes all 36 random encoder configurations; ResNet and ViT summarize each backbone subset.

Method	m PCWP	m PA	1yr-HF
Random	0.70 0.67–0.73 / 0.71 0.68–0.75	0.72 0.68–0.76 / 0.86 0.83–0.88	0.78 0.78–0.79 / 0.58 0.58–0.59
CLOCS	0.68 0.64–0.71 / 0.71 0.68–0.74	0.66 0.63–0.70 / 0.84 0.81–0.86	0.75 0.74–0.75 / 0.53 0.53–0.54
KED	0.72 0.69–0.75 / 0.75 0.71–0.78	0.76 0.73–0.79 / 0.89 0.87–0.91	0.83 0.82–0.83 / 0.66 0.65–0.66
HeartLang	0.68 0.65–0.71 / 0.70 0.66–0.73	0.71 0.67–0.74 / 0.86 0.84–0.88	0.79 0.78–0.79 / 0.58 0.58–0.59
MERL	0.74 0.71–0.77 / 0.76 0.73–0.79	0.76 0.73–0.80 / 0.88 0.86–0.90	0.83 0.83–0.83 / 0.66 0.66–0.67
D-BETA	0.71 0.68–0.74 / 0.73 0.70–0.77	0.74 0.71–0.78 / 0.87 0.85–0.90	0.82 0.82–0.82 / 0.64 0.63–0.64

Table 7: Performance on m PCWP , m PA , and 1yr-HF (AUROC/AUPRC; 95% CIs). Most methods are statistically similar on the hemodynamic tasks; MERL and KED prevail on 1yr-HF .

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Year	Venue	Pretraining Dataset	Evaluation Dataset
CLOCS ( Kiyasseh et al., 2021 )	2021	ICML	PhysioNet 2020, Chapman	PhysioNet 2020, Chapman, Cardiology, PhysioNet 2017
3KG ( Gopal et al., 2021 )	2021	ML4H	PhysioNet 2020	PhysioNet 2020
ISL ( Lan et al., 2022 )	2022	AAAI	PTB-XL, Chapman, CPSC2018	PTB-XL, Chapman, CPSC2018
( Oh et al., 2022 )	2022	CHIL	PTB-XL, CPSC2018, G12EC, CSN	PTB-XL, CPSC2018, G12EC
CRT ( Zhang et al., 2024 )	2022	TNNLS	PTB-XL, HAR, Sleep-EDF	PTB-XL, HAR, Sleep-EDF
CPC ( Mehari and Strodthoff, 2022 )	2022	CIBM	PhysioNet 2020, Chapman, Ribeiro	PTB-XL

Appendix

Table 8: Selected survey set of 12-lead ECG representation learning methods. For each method we indicate the year and venue of publication, which datasets were used for pre-training, and which datasets were used for downstream evaluation. PhysioNet 2020 includes PTB-XL, CPSC2018, INCART, and G12EC. PhysioNet 2021 includes PTB-XL, CPSC2018, INCART, and G12EC, CSN, and UMich.

Method	Metrics Reported	Per-label?	UQ?	Rand?
CLOCS ( Kiyasseh et al., 2021 )	AUROC	N	Y	Y
3KG ( Gopal et al., 2021 )	AUROC, $F_{1}$	Y	Y	N
ISL ( Lan et al., 2022 )	AUROC	N	Y	Y
( Oh et al., 2022 )	Acc.	N	Y	Y
CRT ( Zhang et al., 2024 )	AUROC, Acc., $F_{1}$	Y	Y	N
CPC ( Mehari and Strodthoff, 2022 )	AUROC	Y	Y	N

Appendix

Table 9: Evaluation and reporting choices in our surveyed set of 12-lead ECG representation learning papers. “Y” indicates the practice is clearly reported, and “N” otherwise. The column per-label indicates whether the given method reports its results per individual task, or aggregates them together. UQ indicates whether the paper uses uncertainty quantification when reporting their results. Rand indicates whether the paper compares to a random encoder as a baseline. AUROC denotes “area under the receiver operating curve”, AUPRC denotes “area under the precision-recall curve”, and Acc. denotes “accuracy”.

Task	Description	Prevalence (%)	N Train	N Val	N Test
CD	Conduction Disturbance	22.90%	3907	495	496
HYP	Hypertrophy	12.39%	2119	268	262
MI	Myocardial Infarction	25.57%	4379	540	550
NORM	Normal ECG	44.48%	7596	955	963
STTC	ST/T Change	24.48%	4186	528	521

Appendix

Table 10: Downstream tasks with their definition in PTB-XL Super . For each label, we report the prevalence of the positive class in the whole dataset. We then report the total number of positive examples in the train, validation, and test set after splitting.

Task	Description	Prevalence (%)	N Train	N Val	N Test
AMI	Anterior myocardial infarction	14.39%	2466	306	306
CLBBB	Complete left bundle branch block	2.51%	428	54	54
CRBBB	Complete right bundle branch block	2.53%	432	55	54
ILBBB	Incomplete left bundle branch block	0.36%	62	7	8
IMI	Inferior myocardial infarction	15.29%	2618	326	327
IRBBB	Incomplete right bundle branch block	5.23%	894	112	112

Appendix

Table 11: Downstream tasks with their definition in PTB-XL Sub . For each label, we report the prevalence of the positive class in the whole dataset. We then report the total number of positive examples in the train, validation, and test set after splitting.

Task	Description	Prevalence (%)	N Train	N Val	N Test
AFIB	Atrial fibrillation	7.20%	1211	151	152
AFLT	Atrial flutter	0.35%	59	7	7
BIGU	Bigeminal pattern (unknown origin, SV/Ventricular)	0.39%	66	8	8
PACE	Normal functioning artificial pacemaker	1.40%	237	29	28
PSVT	Paroxysmal supraventricular tachycardia	0.11%	19	3	2
SARRH	Sinus arrhythmia	3.67%	618	77	77

Appendix

Table 12: Downstream tasks with their definition in PTB-XL Rhythm . For each label, we report the prevalence of the positive class in the whole dataset. We then report the total number of positive examples in the train, validation, and test set after splitting.

Task	Description	Prevalence (%)	N Train	N Val	N Test
ABQRS	Abnormal QRS	37.06%	2683	322	322
DIG	Digitalis-effect	2.02%	145	18	18
HVOLT	High QRS voltage	0.69%	49	7	6
INVT	Inverted T-waves	3.27%	235	30	29
LNGQT	Long QT-interval	1.30%	94	12	11
LOWT	Low amplitude T-waves	4.88%	350	44	44

Appendix

Table 13: Downstream tasks with their definition in PTB-XL Form . For each label, we report the prevalence of the positive class in the whole dataset. We then report the total number of positive examples in the train, validation, and test set after splitting.

Task	Description	Prevalence (%)	N Train	N Val	N Test
AF	Atrial fibrillation	17.77%	845	124	251
IAVB	1st degree AV block	10.50%	509	80	132
LBBB	Left bundle branch block	3.42%	175	23	37
NSR	Sinus rhythm	13.37%	653	77	188
PAC	Premature atrial contraction	8.94%	420	60	134
PVC	Premature ventricular contractions	10.18%	498	67	134

Appendix

Table 14: Downstream tasks with their definition in CPSC2018. For each label, we report the prevalence of the positive class in the whole dataset. We then report the total number of positive examples in the train, validation, and test set after splitting.

Task	Description	Prevalence (%)	N Train	N Val	N Test
1AVB	1 degree atrioventricular block	2.19%	476	80	144
2AVB	2 degree atrioventricular block	0.07%	14	2	6
2AVB1	2 degree atrioventricular block(Type one)	0.05%	10	1	5
AF	Atrial Flutter	10.46%	2388	324	624
AFIB	Atrial Fibrillation	4.43%	975	147	291
ALS	Axis left shift	2.89%	651	88	182

Appendix

Table 15: Downstream tasks with their definition in CSN. For each label, we report the prevalence of the positive class in the whole dataset. We then report the total number of positive examples in the train, validation, and test set after splitting.

Task	Description	Prevalence (%)	N Train	N Val	N Test
AR	Moderate or severe aortic regurgitation	1.22	878	62	66
AS	Moderate or severe aortic stenosis	4.19	2919	252	286
LVEF $\leq 45$	Left ventricular ejection fraction is $\leq 45\%$	22.76	16962	866	962
LVWT $\geq 13$	Max of interventricular septum/posterior wall $\geq 1.3$ cm	23.75	17667	877	1061
MR	Moderate or severe mitral regurgitation	8.18	6137	282	337
PASP $\geq 45$	Pulmonary artery systolic pressure is $\geq 45$ mmHg	18.18	13727	581	699

Appendix

Table 16: Downstream tasks with their definition in EchoNext. For each label, we report the prevalence of the positive class in the whole dataset. We then report the total number of positive examples in the train, validation, and test set after splitting.

Task	Description	Prevalence (%)	N Train	N Val	N Test
m PA	Mean pulmonary arterial pressure $\geq 20$ mmHg measured by right heart catheterization	68.16	4,297	393	751
m PCWP	Mean pulmonary capillary wedge pressure $\geq 15$ mmHg measured by right heart catheterization	49.30	3,066	306	561

Appendix

Table 17: Downstream hemodynamic inference tasks. For each binary label, we report the prevalence of the positive class in the full dataset, as well as the number of positive examples in the train, validation, and test sets.

Task	Description	Prevalence (%)	N Train	N Val	N Test
1yr-HF	Development of heart failure within one year, defined as left ventricular ejection fraction $<40\%$ on an echocardiogram	29.49	93,960	12,929	18,774

Appendix

Table 18: Patient forecasting task. For the binary outcome, we report the prevalence of the positive class in the full dataset, as well as the number of positive examples in the train, validation, and test sets.

Task	Random	CLOCS	KED	HeartLang	MERL	D-BETA
CD	0.843 0.821–0.863 / 0.724 0.689–0.755	0.794 0.767–0.818 / 0.630 0.590–0.666	0.904 0.887–0.919 / 0.805 0.776–0.831	0.802 0.779–0.824 / 0.637 0.600–0.673	0.893 0.876–0.909 / 0.796 0.765–0.822	0.902 0.885–0.917 / 0.815 0.788–0.841
HYP	0.847 0.819–0.873 / 0.565 0.510–0.620	0.840 0.811–0.867 / 0.547 0.486–0.601	0.896 0.875–0.915 / 0.669 0.614–0.714	0.782 0.752–0.813 / 0.386 0.336–0.441	0.873 0.848–0.896 / 0.611 0.556–0.661	0.813 0.786–0.839 / 0.456 0.395–0.510
MI	0.841 0.821–0.861 / 0.655 0.616–0.698	0.744 0.719–0.767 / 0.563 0.524–0.597	0.885 0.868–0.900 / 0.764 0.732–0.792	0.801 0.780–0.822 / 0.585 0.546–0.628	0.887 0.871–0.902 / 0.777 0.747–0.802	0.905 0.892–0.919 / 0.815 0.791–0.838
NORM	0.891 0.877–0.904 / 0.839 0.814–0.863	0.859 0.843–0.873 / 0.789 0.764–0.815	0.932 0.922–0.942 / 0.901 0.884–0.917	0.875 0.860–0.889 / 0.822 0.797–0.846	0.927 0.916–0.938 / 0.885 0.862–0.904	0.930 0.919–0.939 / 0.895 0.875–0.913
STTC	0.881 0.864–0.896 / 0.701 0.660–0.737	0.866 0.848–0.882 / 0.688 0.647–0.726	0.927 0.915–0.938 / 0.808 0.776–0.838	0.860 0.844–0.877 / 0.637 0.598–0.677	0.924 0.911–0.936 / 0.799 0.764–0.830	0.916 0.901–0.928 / 0.786 0.753–0.817

Appendix

Table 19: Performance for every task in PTB-XL Super . Each cell reports AUROC (top line) and AUPRC (bottom line) with 95% confidence intervals.

Task	Random	CLOCS	KED	HeartLang	MERL	D-BETA
AMI	0.894 0.874–0.915 / 0.646 0.593–0.700	0.857 0.832–0.883 / 0.590 0.537–0.646	0.951 0.939–0.962 / 0.808 0.770–0.845	0.876 0.857–0.895 / 0.548 0.496–0.601	0.959 0.949–0.969 / 0.841 0.809–0.873	0.956 0.943–0.967 / 0.834 0.793–0.869
CLBBB	0.973 0.934–0.999 / 0.880 0.784–0.955	0.984 0.974–0.993 / 0.676 0.551–0.794	0.998 0.997–1.000 / 0.949 0.906–0.983	0.993 0.984–0.998 / 0.867 0.767–0.944	0.999 0.997–1.000 / 0.947 0.889–0.988	0.999 0.997–1.000 / 0.970 0.937–0.995
CRBBB	0.989 0.979–0.996 / 0.789 0.691–0.875	0.982 0.970–0.992 / 0.749 0.647–0.843	0.998 0.996–0.999 / 0.910 0.838–0.966	0.964 0.936–0.983 / 0.632 0.490–0.753	0.998 0.997–0.999 / 0.891 0.788–0.979	0.997 0.995–0.999 / 0.834 0.733–0.928
ILBBB	0.899 0.733–0.988 / 0.197 0.045–0.499	0.916 0.812–0.979 / 0.064 0.025–0.119	0.913 0.758–0.994 / 0.206 0.071–0.472	0.876 0.769–0.975 / 0.066 0.019–0.140	0.886 0.683–0.987 / 0.094 0.044–0.160	0.935 0.818–0.993 / 0.211 0.071–0.470
IMI	0.864 0.840–0.884 / 0.556 0.506–0.609	0.674 0.644–0.702 / 0.299 0.262–0.341	0.881 0.862–0.899 / 0.597 0.549–0.647	0.750 0.721–0.776 / 0.353 0.310–0.399	0.843 0.819–0.864 / 0.538 0.488–0.589	0.897 0.880–0.914 / 0.691 0.649–0.734
IRBBB	0.856 0.815–0.891 / 0.341 0.262–0.433	0.810 0.770–0.850 / 0.227 0.173–0.292	0.955 0.935–0.970 / 0.617 0.530–0.697	0.770 0.726–0.813 / 0.187 0.137–0.246	0.971 0.961–0.980 / 0.683 0.602–0.757	0.929 0.905–0.949 / 0.517 0.431–0.602

Appendix

Table 20: Performance for every task in PTB-XL Sub . Each cell reports AUROC (top line) and AUPRC (bottom line) with 95% confidence intervals.

Task	Random	CLOCS	KED	HeartLang	MERL	D-BETA
AFIB	0.858 0.827–0.888 / 0.384 0.318–0.463	0.836 0.803–0.867 / 0.337 0.277–0.410	0.984 0.967–0.995 / 0.936 0.903–0.964	0.928 0.908–0.947 / 0.577 0.500–0.651	0.981 0.965–0.992 / 0.891 0.839–0.937	0.984 0.967–0.997 / 0.955 0.924–0.979
AFLT	0.948 0.897–0.991 / 0.231 0.046–0.531	0.911 0.748–0.997 / 0.302 0.090–0.632	0.966 0.906–0.999 / 0.508 0.178–0.824	0.795 0.527–0.999 / 0.451 0.117–0.801	0.969 0.914–0.999 / 0.611 0.277–0.901	0.950 0.859–1.000 / 0.582 0.241–0.883
BIGU	0.636 0.405–0.871 / 0.102 0.011–0.286	0.703 0.498–0.896 / 0.147 0.006–0.408	0.955 0.893–0.996 / 0.445 0.139–0.751	0.884 0.790–0.969 / 0.106 0.020–0.309	0.744 0.541–0.926 / 0.148 0.015–0.427	0.975 0.960–0.989 / 0.350 0.071–0.669
PACE	0.905 0.830–0.967 / 0.589 0.410–0.759	0.889 0.801–0.957 / 0.428 0.252–0.597	0.971 0.937–0.994 / 0.733 0.579–0.864	0.905 0.810–0.979 / 0.551 0.371–0.721	0.962 0.900–0.997 / 0.796 0.657–0.916	0.993 0.985–0.999 / 0.887 0.776–0.973
PSVT	0.921 0.834–0.998 / 0.096 0.006–0.333	0.998 0.994–1.000 / 0.613 0.143–1.000	0.999 0.996–1.000 / 0.664 0.200–1.000	0.998 0.995–1.000 / 0.626 0.166–1.000	0.998 0.995–1.000 / 0.477 0.167–1.000	0.999 0.996–1.000 / 0.664 0.200–1.000
SARRH	0.644 0.581–0.708 / 0.079 0.053–0.114	0.611 0.552–0.676 / 0.063 0.046–0.092	0.873 0.837–0.907 / 0.233 0.170–0.309	0.625 0.557–0.683 / 0.059 0.045–0.079	0.690 0.631–0.743 / 0.099 0.065–0.146	0.950 0.917–0.970 / 0.515 0.414–0.618

Appendix

Table 21: Performance for every task in PTB-XL Rhythm . Each cell reports AUROC (top line) and AUPRC (bottom line) with 95% confidence intervals.

Task	Random	CLOCS	KED	HeartLang	MERL	D-BETA
ABQRS	0.788 0.757–0.819 / 0.687 0.639–0.736	0.684 0.649–0.717 / 0.538 0.491–0.583	0.829 0.798–0.858 / 0.741 0.697–0.785	0.731 0.696–0.765 / 0.615 0.568–0.661	0.811 0.780–0.839 / 0.720 0.675–0.766	0.785 0.757–0.814 / 0.648 0.602–0.698
DIG	0.755 0.643–0.852 / 0.070 0.039–0.119	0.628 0.491–0.756 / 0.045 0.023–0.087	0.907 0.861–0.949 / 0.326 0.151–0.534	0.714 0.583–0.816 / 0.061 0.032–0.126	0.878 0.819–0.929 / 0.149 0.076–0.269	0.882 0.806–0.943 / 0.239 0.111–0.431
HVOLT	0.907 0.839–0.968 / 0.078 0.030–0.184	0.829 0.742–0.911 / 0.046 0.017–0.131	0.943 0.916–0.967 / 0.080 0.049–0.143	0.808 0.622–0.926 / 0.035 0.015–0.062	0.936 0.888–0.973 / 0.089 0.043–0.169	0.907 0.847–0.948 / 0.052 0.030–0.083
INVT	0.864 0.819–0.908 / 0.161 0.104–0.250	0.864 0.811–0.910 / 0.161 0.103–0.240	0.899 0.854–0.940 / 0.421 0.258–0.595	0.765 0.679–0.845 / 0.145 0.078–0.244	0.910 0.872–0.945 / 0.317 0.183–0.471	0.915 0.880–0.948 / 0.353 0.213–0.519
LNGQT	0.679 0.495–0.847 / 0.046 0.018–0.105	0.690 0.535–0.818 / 0.037 0.017–0.082	0.804 0.638–0.941 / 0.218 0.051–0.494	0.701 0.544–0.858 / 0.058 0.020–0.151	0.850 0.716–0.972 / 0.348 0.125–0.620	0.898 0.818–0.965 / 0.229 0.067–0.455
LOWT	0.745 0.666–0.823 / 0.145 0.100–0.202	0.726 0.654–0.794 / 0.112 0.081–0.155	0.831 0.775–0.881 / 0.240 0.157–0.344	0.687 0.616–0.757 / 0.122 0.075–0.191	0.833 0.789–0.874 / 0.214 0.139–0.320	0.843 0.795–0.885 / 0.186 0.136–0.260

Appendix

Table 22: Performance for every task in PTB-XL Form . Each cell reports AUROC (top line) and AUPRC (bottom line) with 95% confidence intervals.

Method	ORIG	CLEAN	$\Delta\text{Macro-AUROC}$
Random	0.834 0.808–0.859	0.847 0.834–0.860	+0.013
CLOCS	0.803 0.784–0.819	0.803 0.787–0.820	+0.000
KED	0.920 0.909–0.929	0.913 0.905–0.921	-0.007
HeartLang	0.836 0.819–0.853	0.830 0.819–0.841	-0.006
MERL	0.905 0.890–0.918	0.897 0.883–0.909	-0.007
D-BETA	0.899 0.882–0.914	0.906 0.899–0.914	+0.007

Appendix

Table 23: ORIG uses the standard PTB-XL Sub test set; CLEAN excludes labels with fewer than 10 positive test examples. Values are macro-AUROC with 95% confidence intervals.

Method	ORIG	CLEAN	$\Delta\text{Macro-AUROC}$
Random	0.787 0.737–0.833	0.802 0.781–0.823	+0.015
CLOCS	0.810 0.762–0.854	0.798 0.775–0.819	-0.013
KED	0.959 0.948–0.969	0.949 0.938–0.959	-0.009
HeartLang	0.854 0.827–0.879	0.861 0.841–0.879	+0.008
MERL	0.903 0.870–0.933	0.914 0.898–0.927	+0.011
D-BETA	0.968 0.956–0.978	0.958 0.941–0.971	-0.010

Appendix

Table 24: ORIG uses the standard PTB-XL Rhythm test set; CLEAN excludes labels with fewer than 10 positive test examples. Values are macro-AUROC with 95% confidence intervals.

Method	ORIG	CLEAN	$\Delta\text{Macro-AUROC}$
Random	0.756 0.734–0.780	0.754 0.733–0.774	-0.002
CLOCS	0.715 0.687–0.741	0.710 0.690–0.731	-0.005
KED	0.851 0.827–0.875	0.862 0.846–0.876	+0.011
HeartLang	0.707 0.679–0.736	0.723 0.699–0.745	+0.016
MERL	0.853 0.839–0.866	0.847 0.833–0.860	-0.005
D-BETA	0.845 0.821–0.868	0.855 0.841–0.868	+0.009

Appendix

Table 25: ORIG uses the standard PTB-XL Form test set; CLEAN excludes labels with fewer than 10 positive test examples. Values are macro-AUROC with 95% confidence intervals.

Task	Random	CLOCS	KED	HeartLang	MERL	D-BETA
AF	0.873 0.848–0.897 / 0.660 0.605–0.714	0.837 0.807–0.862 / 0.617 0.565–0.669	0.987 0.979–0.993 / 0.957 0.938–0.973	0.914 0.897–0.930 / 0.700 0.650–0.750	0.979 0.967–0.989 / 0.947 0.923–0.968	0.995 0.993–0.998 / 0.983 0.974–0.991
IAVB	0.774 0.732–0.814 / 0.296 0.235–0.366	0.706 0.660–0.751 / 0.211 0.167–0.265	0.987 0.980–0.993 / 0.922 0.889–0.950	0.818 0.780–0.859 / 0.437 0.356–0.521	0.984 0.974–0.992 / 0.916 0.880–0.946	0.993 0.986–0.997 / 0.954 0.927–0.975
LBBB	0.946 0.868–0.999 / 0.839 0.693–0.965	0.976 0.960–0.991 / 0.758 0.638–0.877	0.996 0.990–0.999 / 0.931 0.861–0.983	0.962 0.914–0.997 / 0.855 0.751–0.944	0.987 0.967–0.998 / 0.898 0.820–0.964	0.998 0.996–1.000 / 0.957 0.910–0.991
NSR	0.893 0.871–0.913 / 0.593 0.532–0.658	0.846 0.820–0.872 / 0.461 0.400–0.523	0.958 0.946–0.969 / 0.804 0.752–0.850	0.893 0.870–0.916 / 0.585 0.516–0.658	0.951 0.938–0.963 / 0.760 0.699–0.820	0.955 0.943–0.965 / 0.762 0.702–0.816
PAC	0.706 0.659–0.749 / 0.196 0.160–0.240	0.597 0.554–0.641 / 0.121 0.106–0.141	0.861 0.828–0.892 / 0.470 0.390–0.556	0.659 0.610–0.709 / 0.178 0.144–0.222	0.779 0.734–0.821 / 0.334 0.263–0.407	0.927 0.898–0.952 / 0.703 0.622–0.784
PVC	0.748 0.700–0.796 / 0.370 0.296–0.444	0.699 0.649–0.745 / 0.220 0.170–0.280	0.881 0.849–0.910 / 0.603 0.528–0.674	0.840 0.802–0.874 / 0.468 0.390–0.552	0.815 0.778–0.852 / 0.460 0.379–0.545	0.910 0.876–0.939 / 0.781 0.717–0.840

Appendix

Table 26: Performance on CPSC2018 sub-tasks. Each cell reports AUROC (top line) and AUPRC (bottom line), each with a 95% confidence interval.

Task	Random	CLOCS	KED	HeartLang	MERL	D-BETA
1AVB	0.773 0.735–0.811 / 0.093 0.071–0.122	0.705 0.663–0.746 / 0.057 0.042–0.081	0.975 0.953–0.991 / 0.757 0.688–0.817	0.865 0.832–0.896 / 0.228 0.173–0.291	0.981 0.968–0.990 / 0.719 0.647–0.789	0.990 0.986–0.994 / 0.803 0.742–0.853
2AVB	0.470 0.266–0.728 / 0.007 0.001–0.033	0.936 0.880–0.990 / 0.217 0.011–0.553	0.981 0.951–0.997 / 0.103 0.040–0.187	0.880 0.718–0.968 / 0.011 0.004–0.020	0.867 0.643–0.988 / 0.034 0.008–0.101	0.983 0.961–0.995 / 0.066 0.027–0.125
2AVB1	0.593 0.422–0.795 / 0.067 0.001–0.268	0.796 0.651–0.931 / 0.015 0.002–0.050	0.940 0.897–0.983 / 0.019 0.005–0.055	0.575 0.291–0.834 / 0.057 0.001–0.258	0.792 0.550–0.954 / 0.008 0.002–0.024	0.978 0.952–0.992 / 0.039 0.016–0.067
AF	0.867 0.852–0.881 / 0.410 0.376–0.444	0.844 0.829–0.859 / 0.340 0.309–0.371	0.973 0.969–0.976 / 0.712 0.674–0.748	0.911 0.902–0.921 / 0.475 0.440–0.510	0.975 0.971–0.979 / 0.733 0.697–0.772	0.977 0.974–0.981 / 0.757 0.723–0.790
AFIB	0.856 0.834–0.877 / 0.243 0.207–0.280	0.856 0.835–0.874 / 0.221 0.186–0.261	0.966 0.961–0.971 / 0.510 0.460–0.566	0.909 0.897–0.922 / 0.282 0.246–0.322	0.962 0.957–0.966 / 0.473 0.423–0.526	0.969 0.964–0.973 / 0.527 0.478–0.581
ALS	0.974 0.964–0.981 / 0.541 0.472–0.612	0.856 0.828–0.882 / 0.180 0.143–0.224	0.981 0.975–0.986 / 0.597 0.523–0.668	0.893 0.869–0.914 / 0.248 0.199–0.302	0.938 0.922–0.952 / 0.398 0.331–0.469	0.984 0.979–0.988 / 0.673 0.607–0.739

Appendix

Table 27: Performance on CSN sub-tasks whose labels begin with letter A through M. Each cell reports AUROC (top line) and AUPRC (bottom line), each with a 95% confidence interval.

Task	Random	CLOCS	KED	HeartLang	MERL	D-BETA
PRIE	0.717 0.229–0.983 / 0.039 0.001–0.148	0.681 0.141–0.977 / 0.032 0.001–0.130	0.843 0.566–0.994 / 0.074 0.001–0.277	0.783 0.640–0.940 / 0.003 0.001–0.008	0.935 0.841–0.991 / 0.232 0.003–0.672	0.964 0.931–0.986 / 0.014 0.007–0.028
PWC	0.772 0.622–0.896 / 0.020 0.007–0.041	0.773 0.662–0.875 / 0.015 0.005–0.037	0.902 0.803–0.967 / 0.062 0.020–0.137	0.854 0.749–0.930 / 0.019 0.009–0.038	0.927 0.869–0.974 / 0.099 0.032–0.228	0.879 0.751–0.959 / 0.033 0.015–0.062
QTIE	0.656 0.576–0.740 / 0.043 0.018–0.094	0.764 0.695–0.828 / 0.036 0.022–0.056	0.931 0.899–0.960 / 0.248 0.144–0.369	0.761 0.700–0.821 / 0.037 0.021–0.061	0.940 0.914–0.964 / 0.225 0.133–0.325	0.936 0.900–0.965 / 0.243 0.155–0.346
RAH	0.905 0.898–0.912 / 0.002 0.002–0.002	0.989 0.986–0.991 / 0.014 0.011–0.018	0.994 0.991–0.995 / 0.024 0.018–0.033	1.000 0.999–1.000 / 0.317 0.125–1.000	0.978 0.974–0.981 / 0.007 0.006–0.008	0.992 0.989–0.994 / 0.018 0.014–0.024
RBBB	0.954 0.921–0.983 / 0.792 0.710–0.869	0.957 0.928–0.980 / 0.556 0.463–0.648	0.995 0.988–0.999 / 0.945 0.908–0.972	0.921 0.895–0.944 / 0.407 0.316–0.496	0.991 0.976–0.999 / 0.931 0.881–0.971	0.998 0.997–0.999 / 0.957 0.929–0.981
RVH	0.826 0.482–0.998 / 0.172 0.030–0.467	0.824 0.519–0.999 / 0.165 0.012–0.468	0.990 0.973–0.999 / 0.260 0.066–0.632	0.920 0.766–0.999 / 0.334 0.027–0.706	0.993 0.984–0.998 / 0.174 0.044–0.434	0.969 0.916–0.999 / 0.281 0.028–0.655

Appendix

Table 28: Performance on CSN sub-tasks whose labels begin with letter P through Z. Each cell reports AUROC (top line) and AUPRC (bottom line), each with a 95% confidence interval.

Task	Random	CLOCS	KED	HeartLang	MERL	D-BETA
AR	0.698 0.629–0.764 / 0.037 0.025–0.054	0.620 0.553–0.693 / 0.040 0.018–0.080	0.626 0.553–0.692 / 0.034 0.019–0.062	0.627 0.567–0.687 / 0.030 0.016–0.059	0.696 0.626–0.764 / 0.061 0.028–0.110	0.681 0.629–0.731 / 0.022 0.017–0.027
AS	0.753 0.723–0.784 / 0.187 0.152–0.226	0.695 0.661–0.729 / 0.140 0.112–0.169	0.763 0.735–0.792 / 0.171 0.143–0.205	0.710 0.677–0.744 / 0.133 0.110–0.161	0.802 0.776–0.827 / 0.228 0.189–0.269	0.776 0.751–0.800 / 0.159 0.135–0.190
LVEF $\leq 45$	0.856 0.842–0.870 / 0.619 0.589–0.648	0.777 0.760–0.793 / 0.483 0.453–0.512	0.858 0.846–0.873 / 0.622 0.593–0.653	0.826 0.810–0.840 / 0.548 0.521–0.580	0.878 0.865–0.892 / 0.667 0.638–0.697	0.862 0.849–0.874 / 0.613 0.584–0.641
LVWT $\geq 13$	0.742 0.726–0.758 / 0.416 0.391–0.442	0.670 0.651–0.688 / 0.342 0.319–0.367	0.744 0.728–0.760 / 0.417 0.392–0.443	0.715 0.699–0.731 / 0.361 0.340–0.384	0.744 0.728–0.760 / 0.413 0.388–0.441	0.718 0.702–0.734 / 0.354 0.332–0.378
MR	0.789 0.762–0.814 / 0.246 0.210–0.287	0.723 0.696–0.752 / 0.169 0.143–0.198	0.796 0.772–0.819 / 0.234 0.200–0.272	0.744 0.717–0.773 / 0.161 0.140–0.188	0.804 0.780–0.827 / 0.236 0.204–0.274	0.796 0.771–0.820 / 0.222 0.190–0.256
PASP $\geq 45$	0.741 0.721–0.760 / 0.312 0.283–0.341	0.696 0.676–0.715 / 0.241 0.221–0.263	0.737 0.715–0.755 / 0.300 0.272–0.328	0.712 0.693–0.730 / 0.261 0.240–0.285	0.752 0.733–0.771 / 0.335 0.306–0.367	0.730 0.710–0.750 / 0.293 0.266–0.321

Appendix

Table 29: Performance on EchoNext sub-tasks. Each cell reports AUROC (top line) and AUPRC (bottom line), each with a 95% confidence interval.

Explore similar work

Information-theoretic Multimodal Representation Learning for Electrocardiogram Signals
May 26, 2026Phu X. Nguyen, Konstantinos Kontras, Wei Dai +5Representation Learning Cross-Modal Representation Learning
1KU Leuven · 3MIT · 4DFKI +1
Pretraining Strategies and Scaling for ECG Foundation Models: A Systematic Study
May 12, 2026M A Al-Masud, Nils StrodthoffElectrocardiography Self-Supervised Time Series Representation Learning
AI4Health Division Carl von Ossietzky Universität Oldenburg Oldenburg, Lower Saxony, Germany
Foundation Model for Cardiac Time Series via Masked Latent Attention
Mar 27, 2026Moritz Vandenhirtz, Samuel Ruipérez-Campillo, Simon Böhi +6Electrocardiogram Foundation Models Self-Supervised Time Series Representation Learning
Department of Computer Science, ETH Zurich, Switzerland · Department of Biomedical Engineering, University of Basel, Switzerland

May 26, 2026J/K move · Enter open · S save

Information-theoretic Multimodal Representation Learning for Electrocardiogram Signals

Phu X. Nguyen, Konstantinos Kontras, Wei Dai, Huy Phan+4

1KU Leuven · 3MIT · 4DFKI+1

Representation Learning Cross-Modal Representation Learning Self-Supervised Time Series Representation Learning Multimodal Pretraining Electrocardiogram Classification Multimodal Contrastive Learning

Open paper AskPDF arXiv

Electrocardiograms (ECGs) are widely used non-invasive measurements of cardiac activity and play a central role in clinical diagnosis. Recent multimodal approaches align ECG signals with clinical reports to incorporate diagnostic semantics, but clinical reports often fail to preserve the rich physiological structure of ECG waveforms, particularly across multiple levels of abstraction ranging from coarse diagnostic categories to fine-grained morphology. To address this limitation, we formulate ECG representation learning from an information-theoretic perspective and derive a tractable objective that jointly preserves signal structure and integrates clinical semantics. Based on this principle, we propose MERIT (Multimodal ECG Representation via Information Theory), a dual-branch pretraining framework combining masked ECG modeling with ECG--text contrastive alignment. Extensive experiments on PTB-XL and additional benchmarks demonstrate consistent improvements over prior methods, including gains exceeding

3%

F1 on PTB-XL All and

5%

F1 on SubClass classification. In zero-shot evaluation, MERIT further improves performance by up to

+2.66\%

AUC and

+2.11\%

F1 on PTB-XL SubClass, while also demonstrating robustness under multiple distribution-shift settings. Moreover, leveraging the learned ECG representations for ECG-conditioned clinical text generation with large language models improves text quality across several metrics, including ROUGE and METEOR. Together, these results demonstrate that MERIT learns more informative and clinically meaningful ECG representations, particularly for fine-grained clinical applications.

Abstract

Explore similar work

Information-theoretic Multimodal Representation Learning for Electrocardiogram Signals

Pretraining Strategies and Scaling for ECG Foundation Models: A Systematic Study

Foundation Model for Cardiac Time Series via Masked Latent Attention