Rare-ECG: Framework for Rare Cardiac Anomalies
- Rare-ECG is a regime defined by scarce, heterogeneous ECG patterns and underrepresented cardiac conditions.
- The framework employs self-supervised anomaly detection pretraining and supervised multi-label classification to enhance diagnostic accuracy.
- Benchmark datasets like ECG-LT and PTB-XL-Image-17K support digitization, synthetic augmentation, and rigorous evaluation of rare cardiac anomalies.
Searching arXiv for papers on Rare-ECG-related methods and benchmarks to ground the article. Rare-ECG is both a named framework for rare cardiac anomaly detection and a broader research designation for ECG problems characterized by long-tailed label distributions, extreme rarity, heterogeneity, and modality scarcity. In the framework sense, Rare-ECG denotes a demographic-aware, two-stage system that first performs self-supervised anomaly detection pretraining on ECGs and then fine-tunes a multi-label diagnostic model over 116 ECG types (Huang et al., 20 Mar 2026). In the broader literature, the term is used for settings such as cardiac amyloidosis screening with few labeled cases, congenital heart disease classification with small heterogeneous cohorts, contactless radar-to-ECG recovery under limited paired data, and recovery of rare paper ECGs through digitization (Speziale et al., 13 Jan 2026, Alkan et al., 2023, Zhang et al., 24 Jun 2025, Mehdi, 7 Feb 2026).
1. Conceptual scope
In its explicit formulation, Rare-ECG refers to ECG problems in which the patterns of interest are rare, underrepresented, and heterogeneous. The cardiac amyloidosis study describes this regime through rarity, severe class imbalance, heterogeneity across phenotypes, and instability of unsupervised analysis when only a few confirmed cases are available (Speziale et al., 13 Jan 2026). The congenital heart disease study provides a structurally similar small-data setting: 436 adults, 4,153 ECGs, and five mutually exclusive anatomical diagnoses, with substantial heterogeneity due to repairs, hemodynamics, and long-term remodeling (Alkan et al., 2023). The radar-to-ECG recovery study defines a related limited-data regime in which only about 80 minutes of radar-ECG data from 5 subjects in 2 indoor scenarios are available, and labeled pairs are further reduced to 80%, 60%, or 40% of training data to simulate new scenarios (Zhang et al., 24 Jun 2025).
Across these works, Rare-ECG is not restricted to a single disease. It includes rare arrhythmias, rare cardiomyopathies, uncommon conduction patterns, localized myocardial infarction phenotypes, congenital structural diagnoses, and historically archived paper ECGs that must first be digitized before any modern model can use them (Yu et al., 25 May 2026, Mehdi, 7 Feb 2026). This suggests that Rare-ECG is best understood as a regime of scarce, imbalanced, or partially inaccessible electrocardiographic evidence rather than as a single benchmark or ontology.
2. Data resources and benchmark ecology
Several datasets and benchmark constructions define the current Rare-ECG landscape.
| Resource | Composition | Rare-ECG role |
|---|---|---|
| ECG-LT / Rare-ECG | 1,089,367 clinical ECGs, 116 anomaly types, 43 rare types with fewer than 400 occurrences | Long-tail diagnosis, anomaly detection, and localization (Huang et al., 20 Mar 2026) |
| PTB-XL-Image-17K | 17,271 synthetic 12-lead ECG images with masks, signals, YOLO boxes, and metadata | Digitization of legacy paper ECGs and rare archives (Mehdi, 7 Feb 2026) |
| ECG-Expert-QA | 47,211 bilingual multimodal QA pairs with multi-turn dialogues | Reasoning benchmark including rare cardiac conditions and temporal progression (Wang et al., 16 Feb 2025) |
| ECGCLIP pretraining corpus | 2,837,962 ECG-report pairs from 1,324,856 patients | Foundation representations for low-prevalence and rare diseases (Yu et al., 25 May 2026) |
The Rare-ECG framework’s internal cohort is notable for both scale and tail severity. In ECG-LT, 43 conditions occur fewer than 400 times over nine years, and some important entities have prevalence below 0.002%, including second-degree type 2 AV block (Huang et al., 20 Mar 2026). The anomaly-detection benchmark introduced in the multi-scale cross-restoration work complements this by providing 8,167 normal PTB-XL ECGs for training, 912 normal plus 1,248 abnormal ECGs for testing, and 400 records with signal point-level annotations covering 22 abnormal types (Jiang et al., 2023).
PTB-XL-Image-17K addresses a different bottleneck: the lack of image-signal paired corpora for ECG digitization. It supplies 17,271 A4 portrait 12-lead ECG images at 300 DPI, pixel-level waveform masks, ground-truth signals, YOLO-format annotations, and metadata, thereby supporting the full digitization pipeline needed before rare paper ECGs can be reintroduced into computational workflows (Mehdi, 7 Feb 2026). ECG-Expert-QA, by contrast, is not a waveform benchmark but a QA-style evaluation corpus; it converts ECG evidence into textual descriptions and uses single-turn and multi-turn tasks to probe complex diagnosis, dynamic context reasoning, counterfactual reasoning, and ethics, including rare cardiac conditions and temporal progression patterns (Wang et al., 16 Feb 2025).
3. Core methodological paradigms
The central Rare-ECG architecture is a two-stage system: self-supervised anomaly detection pretraining followed by supervised multi-label classification (Huang et al., 20 Mar 2026). In pretraining, a convolutional autoencoder reconstructs masked global 10-second ECG signals and masked local heartbeat segments, models signal trends, and predicts demographic and physiologic attributes. The total pretraining objective is
with . At inference, anomaly localization is obtained from the score map , and the scalar anomaly score is
This design explicitly couples global rhythm, local morphology, trend reconstruction, and attribute-aware representation learning (Huang et al., 20 Mar 2026).
A closely related precursor is the multi-scale cross-restoration framework for ECG anomaly detection, which also uses a global branch, a heartbeat-level local branch, multi-scale cross-attention, uncertainty-aware restoration, and a trend generation module (Jiang et al., 2023). Its motivation is that anomalies may be global, such as rhythm irregularity, or local, such as beat-level morphological changes. That decomposition is now a recurrent design principle in Rare-ECG work.
Synthetic generation constitutes a second major paradigm. ECGAN uses self-supervised autoencoder pretraining followed by conditional WGAN training to synthesize class-conditioned ECG beats and is explicitly motivated by rare arrhythmias and privacy-constrained data sharing (Simone et al., 2023). The cardiac amyloidosis generation tool uses separate class-specific bidirectional-LSTM GANs, a command-line workflow, and a Streamlit GUI so that minority ECG phenotypes can be trained and generated per class, although its cardiac amyloidosis-specific evaluation protocol remains planned rather than completed (Speziale et al., 13 Jan 2026). WearECG addresses a different sparse-observation problem: it uses a variational autoencoder to reconstruct 12-lead ECGs from three leads, specifically II, V1, and V5, so that spatially localized disease signals are not lost in wearable settings (Guan et al., 13 Oct 2025).
A third paradigm emphasizes geometry and transfer. For congenital heart disease, covariance matrices of 12-lead ECG segments are modeled on the symmetric positive definite manifold, augmented along Riemannian geodesics, and projected into class-specific tangent spaces to improve classification under small-sample, heterogeneous conditions (Alkan et al., 2023). For radar-based ECG recovery, RFcardi combines cardio-focusing and -tracking with a sparsity-aware transfer-learning pretext task so that only a small number of radar-ECG pairs are needed for fine-tuning in new scenarios (Zhang et al., 24 Jun 2025).
Signal-language and multimodal reasoning models add a fourth paradigm. ECGCLIP aligns ECG waveforms with expert reports through contrastive learning at scale, while ECG-R1 uses protocol-guided interpretation, modality-decoupled encoders, interleaved modality dropout, and evidence-grounded reinforcement learning to reduce hallucinations and improve robustness when either signal or image is missing (Yu et al., 25 May 2026, Jin et al., 4 Feb 2026).
4. Diagnostic performance on rare and long-tail targets
On the ECG-LT rare subset, the Rare-ECG joint model achieves an AUROC of 0.947, sensitivity of 0.922, and specificity of 0.925, compared with 0.858, 0.876, and 0.762 for training from scratch (Huang et al., 20 Mar 2026). The common-versus-rare AUROC gap shrinks from 8.2 percentage points in the scratch model to 2.2 percentage points in the joint model, which the paper describes as about a 73% reduction in the common-rare performance gap (Huang et al., 20 Mar 2026). These gains are not restricted to retrospective evaluation: in a simulated emergency-department setting, AI-assisted cardiologists reached 84.0% comprehensive diagnosis accuracy, reduced diagnosis time by 36 seconds on average, and improved detailed signal interpretation in 11.8% of cases. For long-tail anomalies specifically, sensitivity increased from 0.469 without AI assistance to 0.714 with AI assistance, while specificity remained 0.997 (Huang et al., 20 Mar 2026).
Foundation pretraining also improves low-prevalence disease detection. ECGCLIP reports internal PRAUC values of 0.253 for Ebstein anomaly, 0.175 for constrictive pericarditis, 0.121 for dextrocardia, and 0.201 for cardiac amyloidosis, and the average rare-disease tier PRAUC rises to 0.045 for ECGCLIP-R34 versus 0.011 for Merl-R18 (Yu et al., 25 May 2026). Although these absolute values remain modest, they are substantially above near-chance baselines for several rare structural or infiltrative diseases.
In small specialized cohorts, geometric approaches remain competitive. In congenital heart disease, the best model—multiple tangent spaces with covariance augmentation and an MLP—reaches accuracy 0.71, macro-AUC 0.84, and macro-F1 0.69 on a five-way anatomical classification task derived from 436 adults and 4,153 ECGs (Alkan et al., 2023). This is a Rare-ECG result in the strict sense of limited, heterogeneous, and structurally specific data.
Sparse-observation reconstruction also preserves downstream rare-pattern information. WearECG yields macro-AUROC 0.8333 for downstream multi-label classification on reconstructed 12-lead ECGs, versus 0.8465 on original 12-lead ECGs and 0.7837 on direct 3-lead input (Guan et al., 13 Oct 2025). For regional myocardial infarction, the reconstructed-signal macro-AUC is 0.8764, compared with 0.8817 for real 12-lead ECG and 0.8233 for 3-lead only. Septal myocardial infarction, a strongly spatially localized target, improves from 0.7839 on 3-lead input to 0.8794 after reconstruction (Guan et al., 13 Oct 2025).
5. Digitization, multimodality, and reasoning infrastructure
Rare-ECG increasingly depends on infrastructure that precedes or complements diagnosis. PTB-XL-Image-17K is foundational for legacy archives because it explicitly supports lead detection, waveform segmentation, and calibrated signal extraction from synthetic paper-like ECG images (Mehdi, 7 Feb 2026). Its round-trip validation reports mean squared error , Pearson correlation , and maximum absolute error 0.018 mV for signal image signal mapping, indicating that the rendering and inverse calibration are essentially lossless under ideal conditions (Mehdi, 7 Feb 2026). This matters for rare archived ECGs because many rare morphologies remain trapped in paper repositories.
Reasoning-oriented Rare-ECG research uses separate resources. ECG-Expert-QA provides 47,211 expert-validated QA pairs and explicitly includes rare cardiac conditions, subtle ischemic changes, complex arrhythmias, dynamic context reasoning, and multi-turn dialogue, but it remains a text-centric benchmark rather than a raw waveform corpus (Wang et al., 16 Feb 2025). ECG-R1 addresses the reliability problem directly: it reports that severe hallucinations are widespread in general-purpose and medical multimodal LLMs for ECG interpretation, and its interleaved modality dropout raises image-missing diagnosis accuracy from about 36.8 without IMD to 77.91 with IMD, while full-modality accuracy changes from 81.99 to 80.29 (Jin et al., 4 Feb 2026). This suggests that modality robustness is crucial when rare cases exist only as raw signal, only as image, or in both forms with partial corruption.
Contactless and cross-modal sensing further extends the scope of Rare-ECG. In radar-based recovery, the cardio-focusing and -tracking algorithm reduces median peak-timing error to 0.022 s and missed detection rate to 14%, outperforming De-ViMo and MMECG; under the hardest label-scarcity condition with 40% labeled pairs, transfer learning with sparsity yields MSE 0.0093, PCC 78.72%, R-peak error 8.70 ms, and MDR 9.02%, compared with 0.0098, 75.89%, 11.15 ms, and 12.15% for supervised training from scratch (Zhang et al., 24 Jun 2025). A plausible implication is that Rare-ECG should be viewed as a multimodal reconstruction and sensing problem as much as a classification problem.
6. Limitations, misconceptions, and future directions
A common misconception is that Rare-ECG denotes a single standardized dataset. The literature instead supports two concurrent meanings: a named demographic-aware framework for equitable rare cardiac diagnosis, and a broader problem family spanning long-tail classification, synthetic augmentation, digitization, contactless sensing, and reasoning benchmarks (Huang et al., 20 Mar 2026, Speziale et al., 13 Jan 2026). Another misconception is that all “Rare-ECG” resources are waveform datasets. ECG-Expert-QA is explicitly a QA benchmark, not a raw signal repository, and its multimodality is mediated through textualized ECG descriptions rather than direct waveform supervision (Wang et al., 16 Feb 2025).
Current systems also have important limitations. The cardiac amyloidosis GAN tool is demonstrated on MIT-BIH arrhythmia data, while its cardiac amyloidosis-specific evaluation protocol is prospective rather than reported (Speziale et al., 13 Jan 2026). PTB-XL-Image-17K does not yet simulate perspective distortions, scanner noise, blur, folds, stains, or lighting artifacts, all of which are deferred as future work (Mehdi, 7 Feb 2026). ECGCLIP’s rare-disease results, while improved, remain modest in absolute terms for several targets and are best interpreted as opportunistic screening rather than definitive diagnosis (Yu et al., 25 May 2026). In Rare-ECG’s own fairness analysis, AUROC remains above 90% for ages 10–90, but the under-10 group drops to about 88% AUROC with specificity 78%, and the over-90 group shows specificity about 85% (Huang et al., 20 Mar 2026). These results indicate that age-extreme robustness remains incomplete.
Future work in this area is converging on several directions already named in the cited papers: richer real-world digitization distortions for legacy ECG images, multi-lead and longer-segment synthetic generation, multimodal augmentation with imaging or biomarkers, domain-shift evaluation across hospitals and devices, and protocol-guided rare-pattern reasoning for multimodal LLMs (Mehdi, 7 Feb 2026, Simone et al., 2023, Jin et al., 4 Feb 2026, Yu et al., 25 May 2026). A plausible synthesis is that Rare-ECG will continue to evolve from isolated rare-disease classifiers into a layered ecosystem: anomaly pretraining to define normality, generative models to address scarcity, digitization to unlock historical archives, and grounded multimodal reasoning to integrate signals, images, and narrative evidence in clinically verifiable form.