HEEDB: Harvard-Emory ECG Database

• HEEDB is a comprehensive open-access 12-lead ECG corpus containing over 10 million recordings used for model pretraining and external biometric validation.
• It underpins both supervised and self-supervised foundation-model training, achieving strong transfer performance and notable label efficiency on clinical tasks.
• HEEDB also functions as a validation platform for ECG biometrics, evaluating challenges like domain heterogeneity, gallery scaling, and temporal drift.

HEEDB, the Harvard-Emory ECG Database, is a large open-access 12-lead electrocardiographic corpus used in recent work as both a pretraining substrate for ECG foundation models and an external validation resource for ECG biometrics. In one benchmarking study it is described as containing 10.7 million ECG recordings and as being one of the largest available open-access 12-lead ECG datasets (Al-Masud et al., 29 Sep 2025). In a separate biometric evaluation, HEEDB v5.0 is reported as containing 11.67 million recordings from 2.17 million patients, with experiments conducted on a filtered Emory University Hospital subset (Scagnetto, 6 Apr 2026).

1. Corpus identity and reported scale

HEEDB is presented in the cited literature as a large-scale ECG dataset designed for representation learning and model development. One study characterizes it as a pretraining source for multiple ECG foundation models and gives its size as 10.7 million ECG recordings (Al-Masud et al., 29 Sep 2025). Another study, focused on biometric identification, refers specifically to HEEDB v5.0 and reports 11.67 million recordings and 2.17 million patients (Scagnetto, 6 Apr 2026). The latter study further states that its experiments used the i0006 subset, corresponding to Emory University Hospital, filtered to 385,079 ECGs from 118,756 patients, with only 500 Hz 12-lead 10-second recordings retained to match model training (Scagnetto, 6 Apr 2026).

Study	HEEDB description	Experimental role
(Al-Masud et al., 29 Sep 2025)	10.7 million ECG recordings	Foundation-model pretraining
(Scagnetto, 6 Apr 2026)	HEEDB v5.0: 11.67 million recordings, 2.17 million patients	External biometric validation

These descriptions establish HEEDB as a population-scale ECG resource rather than a task-specific benchmark. The corpus is used not merely for supervised classification, but also for self-supervised pretraining, transfer evaluation, gallery-based retrieval, and longitudinal stress testing.

2. Role in ECG foundation-model pretraining

HEEDB is used as the pretraining source for at least two foundation-model families in the cited benchmark: ECGFounder and ECG-CPC. ECGFounder uses a CNN backbone with supervised pretraining on HEEDB, whereas ECG-CPC uses a structured state-space model backbone with self-supervised pretraining via Contrastive Predictive Coding on the same corpus (Al-Masud et al., 29 Sep 2025). The benchmark emphasizes that HEEDB enables both supervised and self-supervised representation learning on population-scale ECG data. ECG-CPC is further described as lightweight, with approximately 3.8 million parameters, trained entirely on HEEDB using a CPC objective on a single GPU for 3 weeks (Al-Masud et al., 29 Sep 2025).

This usage situates HEEDB within the current foundation-model paradigm for biosignals. Rather than serving only as a downstream evaluation set, it functions as the large unlabeled or weakly labeled substrate from which transferable ECG representations are learned. The cited evidence also indicates that dataset choice and pretraining objective are consequential: the benchmark explicitly notes that scaling up compute and model size is not automatically beneficial, and that a small efficient backbone pretrained on HEEDB can outperform substantially larger alternatives (Al-Masud et al., 29 Sep 2025).

3. Transfer performance, scaling, and label efficiency

A broad benchmark evaluates eight ECG foundation models on 26 clinically relevant tasks using 12 public datasets comprising 1,650 regression and classification targets. Within that comparison, HEEDB-pretrained models show heterogeneous but often strong transfer behavior (Al-Masud et al., 29 Sep 2025). On adult ECG interpretation tasks, ECG-CPC is reported to achieve performance on par with the top models, with AUROCs up to approximately 0.95. Outside classical diagnosis, particularly for cardiac structure and function, outcome prediction, and patient characteristics, ECG-CPC is described as dominating or ranking among the top performers (Al-Masud et al., 29 Sep 2025).

The reported excerpt gives concrete examples. On EchoNext, ECG-CPC attains 0.831, compared with 0.817 for ECGFounder and 0.819 for a strong supervised S4 baseline. For cardiac outcomes, the corresponding values are 0.781 for ECG-CPC, 0.768 for ECGFounder, and 0.780 for S4; for non-cardiac outcomes, 0.719, 0.701, and 0.714, respectively (Al-Masud et al., 29 Sep 2025). The same study states that frozen-feature evaluation remains highly competitive outside classic diagnosis, which it interprets as evidence for generality of the representations learned from HEEDB under self-supervised objectives.

Label-efficiency experiments on EchoNext model performance as

$C N^{-\alpha} + L_0,$

where $N$ is sample size. For ECG-CPC pretraining, the reported label-efficiency ratio

$r = N^*/N$

is 0.21–0.40 for $N=250$–$1000$, meaning that only approximately 21–40% of the labeled data is needed to reach the performance of a model trained from scratch on $N$ samples (Al-Masud et al., 29 Sep 2025). The same benchmark states that ECG-CPC achieves up to a 5-fold reduction in required labeled data versus strong supervised training in smaller-data regimes. Reported scaling parameters are $$(C,\alpha,L_0,R^2)=(0.463,0.104,4.35\times10^{-7},0.946)$$ for ECG-CPC and $(0.462,0.109,0.018,0.933)$ for ECGFounder, compared with $(0.677,0.206,0.089,0.983)$ for supervised S4 (Al-Masud et al., 29 Sep 2025).

4. HEEDB in ECG biometric identification

HEEDB has also been used as an external validation platform for ECG biometrics under large-gallery, temporally separated, closed-set identification protocols. The cited study evaluates a 1D Inception-v1 model trained with ArcFace on an internal clinical corpus and tests it on HEEDB-derived cohorts (Scagnetto, 6 Apr 2026). Four task-specific datasets are defined from HEEDB: HEEDB-GC for general comparability, HEEDB-scale for varying gallery size and number of ECGs per patient, HEEDB-TST for temporal stress testing with fixed gallery size and 1- to 5-year enrollment-test gaps, and HEEDB-RR for post-hoc reranking on a 10,000-patient subset (Scagnetto, 6 Apr 2026).

Under general comparability, HEEDB-GC yields Rank@1 of 0.6884, Rank@5 of 0.7593, Rank@10 of 0.7807, TAR@FAR$=10^{-3}$ of 0.7210, and TAR@FAR$N$0 of 0.6002 (Scagnetto, 6 Apr 2026). Scale analysis on HEEDB shows monotonic degradation with gallery growth: with 2 ECGs per patient, Rank@1 decreases from 0.7517 at gallery size 500 to 0.6550 at gallery size 7,000. Conversely, increasing the number of ECGs per identity improves retrieval; averaged across gallery sizes, Rank@1 rises from 0.6864 with 2 ECGs per patient to 0.8928 with 7 ECGs per patient (Scagnetto, 6 Apr 2026).

Temporal stress testing quantifies longitudinal drift. On HEEDB-TST, Rank@1 declines from 0.6864 at a 1-year gap to 0.6452 at 2 years, 0.6047 at 3 years, 0.5776 at 4 years, and 0.5560 at 5 years (Scagnetto, 6 Apr 2026). Post-hoc reranking on HEEDB-RR shows that AS-norm provides the strongest improvement among the tested score-processing methods, increasing Rank@1 from a 0.7765 baseline to 0.8005 in the focused sweep (Scagnetto, 6 Apr 2026). The same study reports that ECG identity information remains measurable under externally validated large-scale closed-set conditions, but that operational quality is strongly affected by domain heterogeneity, longitudinal drift, gallery size, and second-stage score processing.

5. Methodological implications and limitations

The available evidence attributes several strengths to HEEDB. Its large scale and diversity are said to enable robust pretraining and to support generalizable, label-efficient models across a broad set of clinical ECG tasks (Al-Masud et al., 29 Sep 2025). Self-supervised CPC pretraining on HEEDB is described as especially effective for non-diagnostic targets such as structure, outcomes, and demographics. In the biometric setting, HEEDB provides a stringent external validation environment in which domain heterogeneity, gallery-size effects, and multi-year temporal gaps can be measured directly (Scagnetto, 6 Apr 2026).

The cited studies also state important limitations. The foundation-model benchmark evaluates only in-distribution tasks, and explicitly says that the benefits for out-of-distribution generalization remain unexplored (Al-Masud et al., 29 Sep 2025). It also notes that downstream label-task mismatch may limit transfer to rare or out-of-scope clinical tasks, and that token-level CPC pretraining may underperform for certain sequence-level prediction tasks and with purely linear heads (Al-Masud et al., 29 Sep 2025). In the biometric study, performance on HEEDB is described as lower than on internal or less heterogeneous external datasets because of multi-center data, acquisition-related variation, limited harmonization, and the absence of the filtering possible on proprietary data (Scagnetto, 6 Apr 2026). The same study further cautions that some reranking strategies, such as diffusion reranking, can degrade performance and confidence calibration.

These limitations are methodologically significant because they constrain what can be concluded from HEEDB-based experiments. Strong performance on HEEDB-pretrained models does not imply that larger models are intrinsically better, and measurable identity information in HEEDB does not imply deployment-ready ECG biometrics under domain shift or long temporal gaps.

6. Position within contemporary ECG machine learning

Across the cited literature, HEEDB occupies two distinct but complementary roles. First, it is a large pretraining corpus for ECG foundation models, supporting both supervised and self-supervised learning at scale (Al-Masud et al., 29 Sep 2025). Second, it is an external validation resource for large-scale biometric retrieval under realistic stressors including gallery growth, longitudinal drift, and score normalization effects (Scagnetto, 6 Apr 2026).

This suggests that HEEDB is valuable not only because of corpus size, but because it can anchor evaluations across multiple methodological regimes: representation learning, transfer learning, data-efficiency analysis, and identity retrieval. The combination of reported benchmark results and biometric stress tests positions HEEDB as a reference dataset for investigating how ECG models behave when scale, task diversity, domain heterogeneity, and temporal separation are all material to the outcome.