AI-driven ECG age prediction transforms early disease detection

By Vijay Kumar MalesuReviewed by Susha Cheriyedath, M.Sc.Feb 9 2025

A simple ECG scan could now predict your risk for heart disease, Alzheimer’s, and cancer before symptoms appear—thanks to AI-powered biological age tracking.

Study: Reclassification of the conventional risk assessment for aging-related diseases by electrocardiogram-enabled biological age. Image Credit: totojang1977 / Shutterstock

In a recent study published in the journal npj Aging, researchers evaluated whether artificial intelligence (AI)-enabled electrocardiogram (ECG)-estimated biological age (ECG-BA) improves risk classification for aging-related diseases beyond chronological age (CA).

Background

Did you know that two people of the same age can have drastically different health outcomes? Aging affects individuals differently, with some staying active and disease-free while others develop serious conditions.

Aging is a universal process that leads to physiological decline, increasing the risk for neurodegenerative, cardiovascular (CV), metabolic, musculoskeletal, and immune disorders. CA is commonly used in disease prediction models, yet it fails to capture the variability in biological aging across individuals. The study excluded individuals with pre-existing conditions such as hypertension, diabetes, and heart failure to focus on a "healthy" population.

ECG-BA, derived from physiological biomarkers, provides a more personalized measure of health status. AI now enables real-time analysis of ECG signals to estimate ECG-BA, improving risk stratification. Five-fold cross-validation was applied to optimize the model’s performance, ensuring robust results. Further research is needed to validate its predictive value for diverse populations.

About the Study

The study utilized ECG recordings collected from Taipei Veterans General Hospital between 2006 and 2017. Initially, 51,061 valid ECGs were recorded, but after applying exclusion criteria, 48,783 healthy individuals aged 20-80 years were analyzed.

A deep learning model integrating a residual network (ResNet), squeeze-and-excitation network (SENet), and multitask learning was developed to estimate ECG-BA from 12-lead ECGs. The model was optimized using the Adam optimizer, which fine-tuned network weights for enhanced accuracy. CA and medical records were linked using International Classification of Disease (ICD) codes to categorize participants into aging-related disease and control groups.

Model training involved five-fold cross-validation to optimize performance. The primary evaluation metric was the correlation between ECG-BA and CA in a healthy population.

Diagnostic performance for CV and non-CV diseases was assessed using area under the receiver operating characteristic (ROC) curves. Net reclassification improvement (NRI) was calculated to measure the improvement in risk classification after incorporating ECG-BA.

Statistical analysis included conditional logistic regression to evaluate the predictive utility of the model in disease classification. The model’s mean absolute error (MAE) was 6.25 years, with a mean absolute percentage error (MAPE) of 15.35%, indicating strong predictive accuracy compared to previous models.

Data processing and model implementation were conducted using PyTorch, with results validated against established clinical benchmarks.

Study Results

Imagine being able to predict future health risks with a simple ECG, just like how a smartwatch monitors daily heart activity. This study reveals that ECG-BA is a powerful tool for identifying aging-related diseases earlier and more accurately than CA alone.

The model showed a strong correlation between ECG-BA and CA (R² = 0.70, p < 0.01). The model's predictive accuracy was higher than previous AI-based ECG models, which had greater error margins. However, the real value of this technology is its ability to identify people at risk of developing serious diseases before traditional symptoms appear.

Compared to using only CA, incorporating ECG-BA significantly improved risk prediction for conditions like coronary artery disease (CAD), stroke, and myocardial infarction (MI).

For example, the net reclassification improvement (NRI) for peripheral arterial occlusive disease (PAOD) was 1.1% (from 0.8632 to 0.8653, p < 0.01), meaning ECG-BA refined risk classification beyond CA alone. Cancer risk classification improved by 29% in terms of NRI, demonstrating that this technology can refine medical assessments and target high-risk individuals more effectively.

For real-world impact, consider cancer detection. Early diagnosis can be the difference between life and death. The study demonstrated that ECG-BA corrected 21% of misclassifications made by CA alone, reducing the number of incorrectly classified patients. This means that more people at high risk could be identified earlier, potentially allowing for timely interventions that save lives.

The most significant improvements were seen in individuals aged 40 and older, reinforcing the idea that biological aging—not just the number of years lived—should be considered in healthcare assessments.

Despite its success in refining disease prediction, the model had limitations in predicting arrhythmia-related conditions such as atrial fibrillation (AF) and sick sinus syndrome (SSS). The study suggests that arrhythmias are influenced by factors beyond aging, such as hyperthyroidism, smoking, and lifestyle habits, which may explain why ECG-BA is less effective for these conditions.

However, for conditions driven by biological aging, such as Alzheimer’s disease (AD) and osteoarthritis (OA), this tool presents a groundbreaking opportunity to enhance early detection and preventive healthcare strategies.

With the increasing accessibility of ECG monitoring via wearable devices, these findings have far-reaching implications. However, the study notes that ECG-BA models need further validation across different ECG machines, such as Philips and GE Healthcare, as variations in device settings could impact predictions.

Imagine a future where routine ECGs not only detect immediate heart problems but also provide a personalized aging risk score, helping individuals take proactive steps to maintain long-term health.

This study marks a significant step toward that future, demonstrating that ECG-BA can reshape preventive medicine and risk assessment, ultimately improving health outcomes globally.

Conclusions

To summarize, ECG-BA provides additional value in risk classification for aging-related diseases beyond CA. The deep learning-based model demonstrated significant improvements in predictive accuracy, particularly for CV conditions, AD, OA, and cancers.

Net reclassification improvement (NRI) analysis indicated that incorporating ECG-BA could correct misclassifications in 21% of cases, with the highest improvement (29%) observed in cancer risk prediction. The results highlight the potential of ECGs as a non-invasive, cost-effective biomarker for systemic aging.

However, the study also emphasizes the need for multi-center validation to confirm generalizability across diverse populations and device platforms.