AnyECG: Evolved ECG Foundation Model for Holistic Health Profiling

Jun Li, Hongling Zhu, Yujie Xiao, Qinghao Zhao, Yalei Ke, Gongzheng Tang, Guangkun Nie, Deyun Zhang, Jin Li, Canqing Yu, Shenda Hong · 2026

Read AnyECG: Evolved ECG Foundation Model for Holistic Health Profiling by Jun Li, Hongling Zhu, Yujie Xiao, Qinghao Zhao, Yalei Ke, Gongzheng Tang, Guangkun Nie, Deyun Zhang, Jin Li, Canqing Yu, Shenda Hong (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-10748.txt.

Introduction

Jun Li and co-authors present AnyECG, an ECG foundation model for broad health profiling. AI methods that read electrocardiograms can spot many heart and non-heart conditions. But most research trains a separate model for each single disease. That split misses how diseases cluster in one patient. It also rarely forecasts future risk.

AnyECG is described as an evolved foundation model, trained to read ECGs as a whole. A single model that profiles many conditions could triage patients faster. It could also flag risks before symptoms grow. The work targets clinics that want one ECG tool with a wide view of health, not a shelf of narrow ones.

Catalog record

CREATORJun Li, Hongling Zhu, Yujie Xiao, Qinghao Zhao, Yalei Ke, Gongzheng Tang, Guangkun Nie, Deyun Zhang, Jin Li, Canqing Yu, Shenda Hong
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE80 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-10748

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

AnyECG: Evolved ECG Foundation Model for Holistic Health Profiling Jun Li, Hongling Zhu, Yujie Xiao, Qinghao Zhao, Yalei Ke, Gongzheng Tang, Guangkun Nie, Deyun Zhang, Jin Li, Canqing Yu, Shenda Hong Abstract ABSTRACT Background: Artificial intelligence-based electrocardiography (AI-ECG) has been proven to detect a variety of cardiac and non-cardiac pathological changes. However, current mainstream research remains limited to independent models for identifying single diseases. These models insufficiently address complex clinical comorbidities and rarely incorporate predictions of future disease risk. Although our previous work, ECGFounder, has achieved coverage of 150 cardiac diseases and significantly improved model performance through fine-tuning techniques, further breakthroughs and expansions are still needed to address key clinical needs and achieve holistic health profiling such as more comprehensive disease screening, long-term risk prediction, and in-depth comorbidity pattern recognition. Methods: In this study, we constructed a large-scale, multicenter electrocardiogram (ECG) dataset containing 13,348,593 ECG records and their corresponding ICD diagnostic codes from 2,984,209 patients. This dataset integrates retrospective and prospective data, providing a data foundation for multiple tasks including current disease diagnosis, future risk prediction, and comorbidity identification. Methodologically, we employed a transfer learning strategy to fine-tune the pre-trained ECGFounder, successfully developing an improved ECG foundation model, AnyECG, aimed at significantly enhancing its capabilities in holistic health profiling. Furthermore, we conducted extensive multicenter external validation and comprehensively evaluated the model’s robustness and accuracy in current diagnosis, future prediction, and comorbidity identification in a 10-year longitudinal cohort study. Results: We explored the systemic prediction ability of AnyECG across 1,172 ICD-coded conditions. AnyECG achieved an AUROC above 0.7 for 306 of these diseases. The AnyECG model reveals many novel discoveries that have never been reported, demonstrates robust detection of comorbidity patterns, and shows predictive capability for diseases that may emerge in the future. For example, AnyECG achieved surprising diagnostic abilities for hyperparathyroidism (AUROC: 0.941 [95% CI: 0.903–0.977]), type 2 diabetes mellitus (AUROC: 0.803 [95% CI: 0.798–0.807]), Crohn’s disease (AUROC: 0.817 [95% CI: 0.774–0.857]), lymphoid leukemia (AUROC: 0.856 [95% CI: 0.849–0.861]) and chronic obstructive pulmonary disease (AUROC: 0.773 [95% CI: 0.759–0.786]). Conclusion: This study introduces AnyECG, a new foundation model, and provides substantial empirical evidence for the clinical application of AI-ECG in both current disease detection and future disease prediction, highlighting its potential as a systemic diagnostic tool. KEYWORDS ECG, Foundation Model, Disease Profiling INTRODUCTION Electrocardiogram (ECG) is a non-invasive, cost-effective, and widely available tool traditionally used to assess cardiac electrical activity 47 , 31 . Abnormalities in electrical cardiac activity affect millions of people worldwide and are increasingly recognized as indicators and contributing factors for various health conditions 29 . It is worth noting that subtle changes in ECG signals often precede the clinical onset of diseases, including arrhythmias, heart failure, and ischemic heart disease 53 , 24 , 4 . These associations highlight the central role of cardiac electrophysiological homeostasis in maintaining overall health and underscore the potential of ECG for early prediction of a wide range of diseases 51 , 31 . However, most existing studies rely solely on isolated ECG indicators (such as QT interval or ST-segment deviations) to establish their associations with specific diseases 11 , 5 , while the complex physiological information embedded in ECG signals has yet to be fully explored and integrated for analysis. 7 In recent years, the integration of artificial intelligence (AI) with ECG—termed AI-enhanced ECG (AI-ECG)—has demonstrated significant potential in diagnosing and predicting various diseases beyond arrhythmias, including heart failure, chronic kidney disease, and diabetes 3 , 13 , 18 , 25 . While these efforts represent substantial progress, they face several critical limitations 9 , 43 . Most methods target specific diseases or limited disease categories and are constrained by relatively small datasets 35 . Furthermore, existing models lack large-scale multi-center external validation, limiting their generalizability to real-world clinical settings 34 , 23 . Additionally, most studies focus solely on diagnosis and do not conduct longitudinal cohort-based prediction and risk assessment. Moreover, prior research has not systematically explored the scientific discoveries underlying ECG’s ability to diagnose and predict diseases 37 . Foundation models offer a promising solution to address these challenges through transfer learning. In the biomedical field, foundation models have exhibited remarkable capabilities across multiple medical domains, advancing disease prediction, diagnosis, and therapeutic discovery 33 , 30 . Recently, several ECG foundation models have been proposed 32 , 44 , 28 . ECG-FM was pre-trained using self-supervised learning on two large ECG datasets 32 . CREMA was designed to learn generalizable representations through self-supervised pretraining on 1 million ECGs dataset. 44 Despite achieving commendable results, these studies on ECG foundation models narrowly focus on arrhythmia- and cardiac-specific outcomes, restricting the broader potential of foundation models in disease prediction. In our earlier research, we developed ECGFounder using a supervised learning paradigm, based on the largest current ECG dataset—the Harvard-Emory ECG Database (HEEDB) 28 , 26 . This dataset included 10,771,552 ECGs from 1,818,247 patients. While ECGFounder performed well on several downstream tasks, establishing the feasibility of our AI-ECG foundational model, its core capabilities remain primarily focused on arrhythmia detection. Further research is needed to explore how to leverage ECG for screening non-cardiac systemic diseases, predicting future health risks, and identifying comorbidity patterns—i.e., constructing holistic health profiling. In this study, we systematically explored the role of cardiac electrical activity in disease diagnosis and risk prediction both in cardiac or non-cardiac diseases. We fine-tuned a pre-trained ECGFounder model based on large-scale paired ECG–ICD data to enhance its ability to identify multiple disease spectrums, and performed external validation in multi-center independent cohorts. To further examine the model’s cross-population and cross-regional generalization performance, we also conducted additional validation in the China Kadoorie Biobank (CKB) 8 . Overall, this study constructed a large-scale, multicenter, multinational ECG cohort covering 2,984,209 patients and 13,348,593 ECGs, including both retrospective and prospective data, providing a data foundation for creating holistic health profiles of patients. This approach enables the model to align with various diseases and clinical events, ultimately developing AnyECG, a versatile model designed to predict over 800 disease types. Through extensive multi-center validation, we aim to demonstrate that AnyECG can (1) diagnose a wide spectrum of diseases, (2) analyze comorbidity relationships, and (3) perform effective disease screening (Figure 1b). Furthermore, we explored the correlations among different diseases within the ECG framework. By identifying association patterns across various disease systems, we hope to contribute to the understanding of disease mechanisms and potentially uncover new insights into disease interactions. Our objective was to further enhance ECGFounder using paired ECG and ICD diagnosis data, and to develop a unified model, AnyECG, covering the full spectrum of 1,172 diseases. With this model, we reported previously unrecognized disease associations detectable from ECGs, demonstrated its ability to reveal links between comorbidities, and showed its potential for predicting future disease onset. In this way, we extend the role of the ECG from a cardiac-centered tool to a comprehensive platform for current disease detection and future disease prediction. Figure 1 : Overview of the development and results of AnyECG. (a) The source and composition of the dataset. The dataset we used included large-scale, multicenter, multinational ECG cohorts: 13,348,593 ECGs, 2,984,209 patients. (b) Representative results of AnyECG across different systems are presented. Outcome indicators for different diseases

Read the complete text