ECG-derived age deviation predicts cardiovascular diseases across lead configurations and cohorts

The Heart’s Hidden Clock

Cardiovascular diseases (CVDs) remain the primary global health burden. They account for approximately one-third of all global mortality. While clinicians traditionally rely on chronological age—the literal number of years since birth—to assess risk, this metric is a blunt instrument. Two people of the same age can possess vastly different physiological states. One may have the cardiovascular profile of a thirty-year-old. Another may mirror someone twice that age. Scientists have long sought a way to measure this "biological age" (BA)—a proxy for physiological status. Finding non-invasive, scalable biomarkers has proven difficult.

A new study proposes that the electrocardiogram (ECG), the standard tool for measuring the heart's electrical activity, contains a hidden clock. By leveraging a massive foundation model pretrained on over 10 million recordings, the researchers demonstrate that the discrepancy between a person's actual age and their ECG-derived biological age is a potent predictor of cardiovascular disease and mortality. This discrepancy is termed "age deviation." Crucially, they find that when a heart "looks" older than it should, the risk of death increases significantly. This holds true even in relatively young patients.

Beyond the Calendar: The Gap in Risk Assessment

Current cardiovascular risk assessments often struggle with a "hidden window" of physiological variation. Chronological age is a strong predictor of disease. However, it fails to capture the cumulative impact of metabolic stress, inflammation, and vascular dysfunction. Until now, efforts to use deep learning to estimate biological age from ECGs have faced two main issues. First, most models are institution-specific and task-specific. They are trained on narrow datasets. This means they often fail to generalize to different hospitals or patient populations. Second, the electrical signals themselves can be deceptive. Certain conduction abnormalities (disruptions in the heart's electrical impulse) can fundamentally alter the ECG waveform. These changes can confuse an AI. They might make a healthy heart look artificially "old" or "young."

The authors aim to fill these gaps by moving toward a foundation model approach. They seek to determine if age deviation can serve as a universal biomarker. This biomarker must work across different types of ECG equipment. This includes standard 12-lead hospital machines and single-lead sensors found in consumer wearables. They also aim to see if it can reliably predict long-term survival.

Decoding the Rhythm: The Foundation Model Pipeline

The researchers' approach rests on a sophisticated multi-stage pipeline. It is designed to isolate biological aging from simple chronological passage. The process follows these logical steps:

1. Feature Extraction via Foundation Models: Instead of training a model from scratch, the authors use ECGFounder. This is a model pretrained on 10 million ECGs. It converts raw electrical signals into 1,024-dimensional "embeddings" (mathematical vectors representing complex features).
2. Establishing a Normative Baseline: To ensure the model understands "normal" aging, the researchers trained an ElasticNetCV regression model exclusively on healthy subjects from the PTB-XL dataset. This model uses regularization (a technique to prevent overfitting) to select the most important features. This creates a benchmark for a healthy heart at any given age.
3. Calculating Age Deviation and Acceleration: For any subject, the model predicts a biological age ($\hat{y}$). The deviation is the raw difference between this prediction and chronological age ($y$). To remove systematic biases, they calculate "age acceleration." This is the residual (the leftover error) after regressing predicted age against chronological age.
4. Cross-Cohort Validation: The framework was then applied to the massive MIMIC-IV-ECG dataset. This is a hospital-based cohort of over 160,000 patients. The goal was to see if the "heart clock" held true in a real-world clinical setting.

The workflow for this estimation process is summarized in .

Evidence of an Accelerated Heart

The results demonstrate that the ECG-derived biological age is a clinically relevant signal. In the PTB-XL cohort, the model achieved an $R^2$ of 0.611. It had a mean absolute error (MAE) of 8.31 years .

This means the model could predict a healthy person's age within roughly eight years.

When the researchers looked at diseased subjects, the associations became even clearer. Non-conduction diseases were associated with significant "age acceleration." Specifically, subjects with ischemic conditions (reduced blood flow to the heart) and ST/T changes (electrical patterns indicating heart stress) showed the largest increases in perceived age [Figure 4d]. On average, diseased subjects exhibited an age acceleration of +3.33 years compared to healthy controls [Figure 4a].

Most importantly, the study validates the clinical utility of this metric through mortality prediction. In the MIMIC-IV-ECG cohort, the authors found that every one-year increase in age acceleration was associated with a 0.5% higher risk of all-cause mortality (Hazard Ratio = 1.005) [Table 1]. This prognostic power was most pronounced in patients under 65 years old. In this group, each year of acceleration increased mortality risk by 1.4% (HR = 1.014) [Table 1]. Furthermore, the authors demonstrated that this signal is remarkably resilient. Even when reducing the ECG from a standard 12-lead configuration to a single-lead setup, the ability to discriminate between healthy and diseased patients was largely preserved [Figure 5c].

Morphological Confounds and Technical Limits

The study is notably transparent about the "traps" inherent in using AI to interpret biological signals. The most striking issue discovered was a morphological confound involving Complete Left Bundle Branch Block (CLBBB). While most diseases are associated with age acceleration, CLBBB showed a massive negative age deviation of approximately -15.18 years [Figure 4c]. This happens because the specific electrical shape of a CLBBB waveform is interpreted by the foundation model as a pattern typically seen in much younger individuals.

This discovery led the authors to a vital methodological pivot. They propose using "absolute age deviation" ($|\text{predicted age} - \text{chronological age}|$) as a more robust marker. By taking the absolute value, the metric can flag both the "prematurely old" heart and the "morphologically unusual" heart. This covers both directions of abnormality under a single risk metric.

There are also practical limitations to consider. While single-lead ECGs are effective for general mortality prediction, the authors found they could produce paradoxical results for specific conditions. For example, single-lead recordings showed negative deviation for hypertension and diabetes in the MIMIC-IV-ECG cohort. Consequently, they warn that while a smartwatch might indicate if you are "aging fast," it may not be reliable for specific disease interpretation. Finally, the study notes that ethnic disparities exist in the model's accuracy. Asian patients showed a larger deviation (-2.84 years). This highlights the need for more diverse training data to ensure algorithmic equity.

The Verdict: A Scalable Sentinel

Is ECG-derived age deviation a viable clinical tool? The evidence suggests it is a strong candidate for preventative screening. The signal is robust across different populations. Crucially, it maintains predictive power even on the low-power, single-lead hardware common in wearables.

By identifying individuals whose hearts are aging faster than their birthdays suggest, clinicians can intervene earlier. The researchers have made their framework accessible via a web application at https://bioinformatics.mdc-berlin.de/ECGage. However, the transition to clinical reality will depend on handling morphological outliers like CLBBB. It will also require calibrating the "clock" across all ethnicities. For now, it stands as a powerful proof-of-concept for the next generation of non-invasive cardiovascular monitoring.

Figures from the paper