ECGomics: A Multidimensional ECG Framework
- ECGomics is a multidimensional framework that treats ECG signals as high-dimensional biological data, organizing digital biomarkers across structural, intensity, functional, and comparative dimensions.
- It employs deep learning, interpretable feature extraction, and multimodal alignment to enhance phenotype discovery and disease detection from standard 12-lead recordings.
- The approach integrates platformization and generative models to support cross-center validation, synthetic signal generation, and scalable clinical application.
ECGomics is an omics-style view of electrocardiography in which the ECG is treated as a structured, high-dimensional biological data resource rather than only a narrow test for rhythm interpretation and a small set of heuristic intervals. It has been proposed as both a conceptual framework and a deployable software platform for turning ECG signals into a structured, multidimensional set of digital biomarkers (Zhang et al., 19 Jan 2026). Earlier work can already be read as an ECGomics study when it constructs an automated and interpretable 725-element patient ECG profile from a full 10-second, 12-lead recording for disease detection, tracking, and discovery (Tison et al., 2018). Taken together, subsequent work suggests a broad program of ECG-derived phenotyping that spans interpretable feature construction, latent representation learning, multimodal alignment, longitudinal drift analysis, and synthetic signal generation.
1. Conceptual definition and taxonomic structure
ECGomics is explicitly framed as a four-dimensional taxonomy inspired by genomics. In this formulation, Structural ECGomics captures waveform shapes, durations, amplitudes, and intervals; Intensity ECGomics captures energy distribution, spectrum, signal complexity, and nonlinear descriptors such as ECG spectrum and R-R interval skewness; Functional ECGomics targets physiological performance and autonomic regulation through quantities such as heart rate variability and the degree of QRS fragmentation; and Comparative ECGomics uses deep representations and association features to derive quantities such as ECG age, individual trait prediction, cardiovascular risk stratification, and cross-modal prediction of laboratory indices or imaging features (Zhang et al., 19 Jan 2026).
The same framework also introduces a three-layer biomarker hierarchy. Engineered biomarkers correspond mainly to the Structural, Intensity, and Functional dimensions; predictive biomarkers correspond mainly to Comparative ECGomics; and deep biomarkers are hidden-layer embeddings or latent features extracted by supervised, unsupervised, and generative deep models (Zhang et al., 19 Jan 2026). The analytical workflow is described as three stages—multi-dimensional extraction, systematic biomedical mapping, and clinical translation—which positions ECGomics less as a single algorithm than as a systems architecture for extracting, structuring, and deploying ECG-derived digital biomarkers.
An important feature of this taxonomy is that it is described as conceptual and systems-oriented, not as a mathematically formalized theory. The framework does not provide explicit equations defining the four dimensions, and its novelty lies in organizing ECG biomarkers into a formal multidimensional taxonomy while combining expert-defined morphology, physiological descriptors, and deep latent signals within one deployable ecosystem (Zhang et al., 19 Jan 2026).
2. Interpretable phenotype construction and morphology-native representations
A foundational ECGomics pattern is the construction of patient-level representations that preserve direct physiological meaning. A prominent example is the UCSF study of automated and interpretable patient ECG profiles, which restricted analysis to 36,186 normal-sinus-rhythm ECGs and built a 725-element vector per ECG study: five global interval summaries—heart rate, PR interval, P-wave duration, QRS duration, and QT interval—plus 720 lead- and segment-resolved waveform features obtained from three waveform regions, 12 leads, and 20 interpolated bins per region. The upstream segmentation model combined a convolutional neural network with a Hidden Markov Model, recognized six primitive waveform labels at millisecond resolution, and produced interval estimates that agreed closely with clinical workflow measurements, including median absolute deviations of 0.6% for heart rate, 3.0% for PR interval, 5.6% for QRS duration, and 4.4% for QT interval. On downstream tasks, the same profile supported disease detection with AUROCs of 0.94 for pulmonary arterial hypertension, 0.91 for hypertrophic cardiomyopathy, 0.86 for cardiac amyloidosis, and 0.77 for mitral valve prolapse, while preserving variable-level interpretability through named intervals and lead-specific waveform bins (Tison et al., 2018).
This representation is characteristic of ECGomics because it occupies a middle position between hand-crafted rules and opaque embeddings. Its coordinates remain semantically legible—such as “mid-QRS in V1” or “P-wave duration”—yet the feature space is much richer than standard clinical ECG summaries. The same study also demonstrated primitive longitudinal tracking by showing that pulmonary arterial hypertension scores rose and fell with visually appreciable changes such as more prominent in V1, increasing T-wave inversion in V1, rightward QRS axis in lead I, and increasing P-wave amplitude (Tison et al., 2018).
A second interpretable strand replaces fixed feature vectors with representative morphology signatures. A motif-based framework defines a recording-level or window-level ECG motif as a beat-aligned representative cardiac cycle selected by minimizing Dynamic Time Warping distance within fixed windows. From these motifs it derives three continuous biomarkers: deviation from a population normal sinus rhythm template, deviation from a personalized baseline, and a motif instability index. In PTB-XL, deviation from normal sinus rhythm separated normal from abnormal ECGs across major diagnostic subtypes with and Cliff’s up to 0.93; in MIT-BIH, the personalized drift and instability measures separated predominantly normal from arrhythmic recordings with . This suggests that ECGomics can be implemented not only as a fixed patient profile but also as a longitudinal morphology trajectory built from directly inspectable waveform exemplars (Bijlani et al., 26 May 2026).
3. Platformization and software ecosystems
The 2026 ECGomics platform operationalizes the taxonomy as a software ecosystem. On the web side, it supports .npy upload, preloaded sample data, calibration through Sampling Rate, ADC Gain, and Zero-point Voltage, high-fidelity 12-lead visualization, expert-defined feature extraction through FeatureDB and ENCASE, latent representation mining through ECGFounder based on the Net1D architecture family, downstream modeling with XGBoost, and export of structured outputs in CSV. On the mobile side, it is embedded in a WeChat Mini-Program workflow with portable sensors, Bluetooth pairing, cloud-side ECGomics analysis, and near-instant report delivery. The reported outputs include biomarkers grouped under the four ECGomics dimensions, clinically interpretable morphology and physiology measures, deep hidden-layer embeddings, heart age, laboratory-test correlations, disease probability scores, rhythm parameters such as heart rate, PR interval, and QRS duration, and cardiovascular risk stratification (Zhang et al., 19 Jan 2026).
The platform paper is supported by representative application studies rather than a single unified benchmark. These include atrial fibrillation detection on the PhysioNet 2017 Challenge with F1 score = 0.825 using ENCASE; prediction of atrial fibrillation recurrence after cryoablation in 201 patients with AUC = 0.872 and Accuracy = 0.902 using 12-lead ECG representation differences and XGBoost; severe coronary stenosis detection in 392 patients, where sensitivity improved from 0.545 under standard ECG interpretation to 0.848 with ECGomics/deep-feature integration and AUC = 0.847; and portable maternal cardiac monitoring in 99 pregnant women, with diagnostic consistency > 0.900, heart rate correlation , QT interval correlation , arrhythmia sensitivity 0.842, and specificity 0.975 (Zhang et al., 19 Jan 2026).
A parallel software-and-benchmark ecosystem is provided by Heartcare Suite, which comprises Heartcare-220K, Heartcare-Bench, and HeartcareGPT. Heartcare-220K is built from 21,799 PTB-XL 12-lead ECG recordings and 12,170 PDF-format hospital ECG reports; Heartcare-Bench contains approximately 18,000 curated test samples spanning diagnosis, waveform, and rhythm tasks; and HeartcareGPT uses the Bidirectional ECG Abstract Tokenization (Beat) tokenizer to convert raw multi-lead signals into discrete tokens for multimodal language modeling. In its reported evaluations, HeartcareGPT reached a Closed-QA average accuracy of 43.6% and Report Generation scores of 55.8 F1-Rad and 58.0 ROUGE-L on the signal subset, which positions it as an ECG-specific multimodal understanding stack rather than a conventional classifier (Xie et al., 6 Jun 2025).
4. Foundation models, multimodal alignment, and broad phenotyping
One major ECGomics trajectory treats the ECG as a broad phenotyping substrate rather than a disease-specific detector. A proof-of-concept study using a population-based dataset of >250,000 patients, >1000 medical conditions, and >2 million ECGs showed that deep learning on the first in-hospital ECG could uncover 128 diseases and 68 disease categories with strong discriminative performance. The work is important because it extends ECG-derived phenotyping beyond classical cardiovascular labels to an ICD-wide exploration of latent disease signatures encoded in routine 12-lead recordings (Sun et al., 2022).
A second trajectory concerns reusable public representation-learning infrastructure. OpenECG introduced a benchmark of 1,233,337 public 12-lead ECG recordings from nine centers and used leave-one-dataset-out experiments to compare self-supervised learning strategies. Across this benchmark, BYOL and MAE generally outperformed SimCLR, and the data-scaling experiments showed that performance for BYOL and MAE saturated at approximately 60%–70% of total data, whereas SimCLR required more data (Wan et al., 2 Mar 2025). This suggests that ECGomics increasingly depends on shared foundation-model infrastructure and multi-center heterogeneity rather than on single-institution supervised pipelines alone.
Multimodal self-supervised learning extends this logic by forcing ECG representations to absorb non-electrical phenotypes. Echo2ECG aligns ECGs with multi-view echocardiographic studies rather than with single views, on the argument that single-view alignment creates a representational mismatch between a global electrical signal and a spatially restricted anatomical snapshot. In downstream evaluation, its ECG feature extractor consistently outperformed unimodal and multimodal baselines on structural phenotype classification and echo-retrieval tasks, despite being 18x smaller than the largest baseline (Liman et al., 9 Mar 2026). A complementary signal-language strategy is ECGCLIP, which aligns ECG waveforms with expert diagnostic reports. ECGCLIP was pretrained on 2,837,962 ECG studies from 1,324,856 patients, evaluated on 89 downstream tasks across nine independent external cohorts, and improved low-prevalence or diagnostically elusive phenotypes including Ebstein anomaly, constrictive pericarditis, dextrocardia, and cardiac amyloidosis, with internal PRAUC values of 0.253, 0.175, 0.121, and 0.201, respectively (Yu et al., 25 May 2026). Taken together, these studies suggest that ECGomics increasingly treats the ECG as a latent phenotype carrier that can be enriched by language, imaging, and large-scale pretraining.
5. Generative, synthetic, and mechanistic ECGomics
Synthetic ECG generation occupies a distinct ECGomics role because it can support class balancing, phenotype coverage, robustness analysis, and mechanistic labeling. A graph-based synthesis framework illustrates the mechanistic end of this spectrum. It uses a unified heart graph with two interchangeable backends—an eikonal-template backend and a pseudo-diffusion reaction–eikonal backend—and introduces activation-consistency certification through the Bellman residual of graph eikonal activation times. On the cardiac graph, reaction–eikonal activation times showed near-millisecond agreement with the eikonal backbone and achieved after causal predecessor filtering. In final balanced multi-lead curation, the reaction–eikonal backend accepted 658/2000 samples versus 578/2000 for the eikonal-template backend and increased per-model morphology coverage from 0.09248 to 0.09888 (Yoo, 25 Jun 2026). For ECGomics, this matters because the synthetic signals come with controllable activation and recovery variables rather than only visual realism.
A more translational but less mechanistically grounded strand is represented by a class-specific GAN tool proposed for cardiac amyloidosis research. The tool is motivated by the rarity, heterogeneity, and class imbalance of cardiac amyloidosis ECG cohorts, but the implemented experiments are trained on the MIT-BIH Arrhythmia Database, not on a cardiac amyloidosis dataset. Its generator produces single ECG beats, one channel, and 187 samples per beat, and the paper explicitly presents cardiac amyloidosis evaluation as prospective rather than completed. The work is therefore best interpreted as a clinician-facing prototype for augmentation studies rather than as a validated cardiac-amyloidosis ECGomics resource (Speziale et al., 13 Jan 2026).
Generative ECGomics also includes attempts to unify interpretation and waveform synthesis within one model. UniECG is presented as the first unified ECG model that supports both evidence-based ECG interpretation and text-conditioned ECG generation, using a two-stage scheme in which the model first learns ECG-to-text interpretation and then injects text-to-ECG generation through latent alignment to a pretrained diffusion generator (Jin et al., 23 Sep 2025). A very different mechanistic strand is provided by the oscillator-and-genetic-algorithm model that reproduces a normal sinus rhythm ECG using three coupled modified van der Pol oscillators for the pacemaker complex and four FitzHugh–Nagumo equations for P, Ta, QRS, and T waves, then fits clinical ECGs that do not deviate much from normal sinus rhythm by optimizing delay-differential-equation parameters with a genetic algorithm (Chowdhury et al., 2024). These strands suggest that ECGomics generation is not a single paradigm: it includes data-driven augmentation, signal-language synthesis, and low-dimensional mechanistic modeling.
6. Interpretability, evaluation, and unresolved questions
Interpretability is not incidental to ECGomics; it is one of its defining tensions. The ExECG framework responds to this by formalizing ECG explainability as a three-stage pipeline—Wrapper, Explainer, and Visualizer—that standardizes heterogeneous ECG formats and model interfaces, unifies attribution-, counterfactual-, and concept-based explanation methods under a common execution protocol, and renders explanations in clinically aligned 12-lead chart formats (Jang et al., 19 May 2026). This infrastructure is significant because ECGomics increasingly relies on comparing not only predictions but also explanatory artifacts across models, tasks, and institutions.
A recurrent misconception is that ECGomics is simply another name for molecular multi-omics. One prognostic study is explicit that a VCG/GEH pipeline built from routine 12-lead ECGs should be understood as ECGomics / electrocardiographic phenomics, not as multi-omics in the molecular sense, even though it derives a richer phenotype space through quantities such as QRST angle and Spatial Ventricular Gradient (Costa et al., 2024). ECGomics can certainly be integrated with genomics, transcriptomics, proteomics, metabolomics, imaging, and laboratory data, but its primary object remains the systematic extraction of latent structure from ECG signals themselves.
The principal limitations identified across the literature are methodological rather than conceptual. The early UCSF profile study used data from a single center, restricted modeling to ECGs in normal sinus rhythm, and did not include an external validation cohort (Tison et al., 2018). The 2026 ECGomics platform explicitly notes challenges in data scale, annotation quality, and the need for rigorous multi-center validation (Zhang et al., 19 Jan 2026). A large public mortality benchmark further showed that cross-dataset external evaluation between Code-15 and MIMIC-IV produced substantial AUROC and concordance loss, indicating that ECG-derived prognostic signal is real but strongly shaped by site, care setting, and cohort composition (Lukyanenko et al., 2024). These patterns suggest that the unresolved questions of ECGomics are less about whether the ECG contains latent phenotypic information than about how to preserve interpretability while scaling across institutions, rhythms, devices, and populations.
A plausible implication is that the field will continue to differentiate into at least three interacting layers: interpretable phenotype construction, foundation-model representation learning, and mechanism-aware or policy-curated generation. The central unresolved comparison is not merely performance versus performance, but whether a given ECGomics representation is scientifically useful—meaning transferable, semantically grounded, and compatible with external validation—rather than only predictive on a local benchmark.