Inferring Systemic Biomarkers from Electrocardiogram Waveforms

32 concepts22 papers2 proposals

For over a century, the electrocardiogram (ECG) has been treated primarily as a localized diagnostic tool for the heart. Clinicians inspect its waves and intervals to detect arrhythmias, myocardial infarctions, or structural abnormalities. However, a growing body of research suggests that the heart acts as a highly sensitive, continuous sensor for the entire body's physiological state.

Because the cardiovascular system is intimately connected to metabolic, endocrine, and immune pathways, systemic disorders leave subtle, high-dimensional signatures in the heart's electrical activity. By leveraging deep learning, researchers are now unlocking these "invisible" digital biomarkers, transforming routine, non-invasive cardiac signals into windows for systemic disease screening.

The Intuition: The Heart as a Biochemical Sensor

Why should an electrical signal from the heart reveal chronic kidney disease, diabetes, or systemic inflammation? The answer lies in the cellular environment of the myocardium. The heart does not beat in a vacuum; its electrical conduction and repolarization processes depend heavily on the surrounding biochemical bath.

For instance, electrolyte imbalances directly alter the cardiac action potential. As noted by Nils Strodthoff et al. in EKG-Diagnostik mit Hilfe künstlicher Intelligenz: aktueller Stand und zukünftige Perspektiven – Teil 2, neural networks can accurately identify elevated potassium levels (hyperkalemia) directly from electrical heart patterns, offering a non-invasive proxy for metabolic blood tests.

Similarly, systemic conditions like chronic kidney disease (CKD) or type-2 diabetes (T2DM) alter blood chemistry and autonomic tone, which subtly reshape the ECG waveform. Xiayue Fan et al. demonstrated this by developing CRT-Net: A Generalized and Scalable Framework for the Computer-Aided Diagnosis of Electrocardiogram Signals, a hybrid architecture that detects CKD and T2DM with over 90% accuracy from standard ECGs.

Decoding High-Dimensional Morphological Shifts

Human eyes are poorly suited to detect these systemic signatures because they do not manifest as obvious, isolated abnormalities. Instead, they present as distributed, multi-variable shifts across the entire cardiac cycle.

Consider glucose regulation. Yao Mu developed IGRNet: A Deep Learning Model for Non-Invasive, Real-Time Diagnosis of Prediabetes through Electrocardiograms, demonstrating that even a brief 5-second ECG contains sufficient latent information to diagnose prediabetes. When metabolic disorders occur alongside conditions like non-alcoholic fatty liver disease (NAFLD), the systemic impact on the heart's electrical conduction becomes even more pronounced. Yaping Wang et al. showed in Association of UHR and ECG parameters with type 2 diabetes mellitus in non-alcoholic fatty liver disease that specific electrophysiological proxy markers—such as P-wave duration, QT intervals, and ST-T segment changes—serve as independent predictors of diabetic status, capturing structural and functional cardiac changes induced by systemic metabolic stress.

To systematically exploit these signatures, Juan Miguel Lopez Alcaraz et al. introduced CardioLab: Laboratory Values Estimation from Electrocardiogram Features - An Exploratory Study. By training machine learning models on engineered ECG features, the authors proved that cardiac electrical activity can directly estimate laboratory blood values, including renal function markers, immune system status, and oxygen transport capacity. In a companion study, Estimation of Cardiac and Non-cardiac Diagnosis from Electrocardiogram Features, the authors expanded this paradigm to show that standard ECG metrics can screen for 44 distinct conditions, including systemic lupus erythematosus and polycystic ovarian syndrome (PCOS), firmly establishing the heart as a proxy sensor for broader bodily health.

From Raw Waveforms to Latent Trajectories

The success of these diagnostic tools relies heavily on how the underlying models represent physiological signals. Traditional approaches analyze independent 1D lead signals, which can be sensitive to noise or electrode misplacement. Recent breakthroughs overcome this by mapping signals into physically grounded latent spaces.

For example, Bosong Huang et al. developed a self-supervised framework called LVCG, detailed in Learning Cardiac Latent Representations in Vectorcardiogram Space. By modeling the heart's electrical activity as a unified 3D trajectory in vectorcardiogram space rather than isolated 1D lines, the model learns latent representations that capture systemic metabolic and inflammatory states with high label efficiency.

Similarly, the development of large-scale open-source foundation models, such as DeepECG-SSL by Alexis Nolin-Lapalme et al. in Foundation models for electrocardiogram interpretation: clinical implications, highlights how self-supervised learning can extract robust digital biomarkers for conditions traditionally requiring complex imaging or genetic testing.

Cross-Domain and Cross-Organ Inference

This paradigm of treating localized physiological metrics as proxies for systemic health extends beyond the ECG. The core mechanism—that localized biological outputs reflect systemic states—applies across multiple modalities:

By shifting the diagnostic lens from localized organ pathology to systemic pattern recognition, this research direction redefines the utility of everyday physiological signals. Whether through a 5-second ECG or peripheral vascular patterns, deep learning models exploit the interconnected nature of human physiology, turning non-invasive sensors into powerful, comprehensive screening tools for systemic health.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_systemic_physiological_sensing_a4ddf0f8,
  title = {systemic physiological sensing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/systemic-physiological-sensing/rdr_a4ddf0f8cdb544c7bc249ca2bb72ad53},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_systemic_physiological_sensing_a4ddf0f8_v1,
  title = {systemic physiological sensing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/systemic-physiological-sensing/rdr_a4ddf0f8cdb544c7bc249ca2bb72ad53/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake