Development and Multinational Validation of Artificial Intelligence-Enabled ASCVD Risk Stratification Using Electrocardiograms

This study developed and validated a scalable, AI-driven toolkit (ECG-ASCVD) that utilizes electrocardiogram data to predict atherosclerotic cardiovascular disease risk across diverse multinational cohorts, offering a solution for risk stratification in patients lacking traditional clinical data.

The Gist

Imagine you are trying to predict who might get a heart attack in the next few years. Right now, doctors use a "risk calculator" (called PREVENT) that works like a detailed financial audit. To get a score, you need to provide a lot of specific data: your blood pressure, your cholesterol levels, whether you smoke, your diabetes status, and more.

The Problem: Many people never get this "audit" done. They might not have seen a doctor recently, they might not have had their blood drawn, or their medical records might be missing these specific numbers. Without this data, the calculator can't run, and these people fly under the radar, missing out on life-saving treatments.

The Solution: This paper introduces a new, clever tool called ECG-ASCVD. Think of an ECG (electrocardiogram) as a "heartbeat snapshot." It's a quick, cheap, and common test where you stick a few stickers on your chest to see your heart's electrical rhythm.

The researchers asked a bold question: Can we look at this heartbeat snapshot and guess a person's heart risk, even if we don't know their cholesterol or blood pressure?

How They Did It (The "AI Chef")

The team didn't just look at the ECG with human eyes; they fed thousands of ECGs into a super-smart computer brain (Artificial Intelligence). They trained this AI in three different ways, like training a chef to recognize a dish:

1. The Master Chef (ECG-ASCVD-12): This AI looked at the full, detailed 12-lead ECG (the standard hospital test with many wires).
2. The Visual Artist (ECG-ASCVD-IMAGE): This AI looked at the ECG as a picture (like a photo of a graph on paper). This is huge because it means you could take a photo of a paper ECG with your phone, and the AI could still read it.
3. The Minimalist (ECG-ASCVD-1): This AI only looked at a single line of the heartbeat (Lead 1). This is the most exciting one because it mimics the tiny, wearable heart monitors (like smartwatches) that people wear every day.

The Results: The "Magic Crystal Ball"

The researchers tested these AI models on hundreds of thousands of people in the US, Brazil, and the UK. Here is what they found:

• It Works: The AI models were surprisingly good at predicting heart attacks. Their accuracy was almost as good as the traditional "financial audit" (PREVENT score), even though the AI only looked at the heartbeat and age/sex.
• It's Independent: Even when the researchers did have the full medical data (cholesterol, etc.) and used the traditional calculator, the AI ECG score still added extra value. It was like having a second opinion that caught risks the first calculator missed.
• It's Universal: The models worked well across different countries and populations, proving they aren't just a "US-only" trick.

The Real-World Impact: The "Safety Net"

The paper ran a simulation with 100,000 people. They found that 87% of people didn't have all the necessary data to use the traditional risk calculator. However, almost all of them had an ECG on file.

The Analogy:
Imagine you are trying to find people who need a life jacket before they jump into a stormy ocean.

• The Old Way: You ask everyone to fill out a complex form about their swimming ability, the weather, and their life jacket history. If they can't fill it out, you assume they are safe and let them jump.
• The New Way (ECG-ASCVD): You simply look at their pulse. If their pulse looks "stressed" or "risky," you immediately give them a life jacket and tell them to get a full check-up.

Why This Matters

This study suggests that the humble ECG, which is already done millions of times a year, could become a universal screening tool.

1. For the Uninsured or Underserved: In places where blood tests are expensive or unavailable, a simple ECG (or even a photo of one) could tell a doctor, "Hey, this person needs a full heart check-up."
2. For Wearables: Since the "Minimalist" model works on a single lead, your smartwatch could eventually warn you, "Your heart rhythm suggests you might be at higher risk; please see a doctor for a full check-up."
3. Filling the Gaps: It catches the people who fall through the cracks of the current system, ensuring that more people get the help they need before a heart attack happens.

In short: The researchers built an AI that turns a simple heartbeat graph into a powerful crystal ball for heart disease, offering a way to save lives even when we don't have all the usual medical data.

Technical Summary

1. Problem Statement

Current clinical risk prediction models for Atherosclerotic Cardiovascular Disease (ASCVD), such as the PREVENT equations, require comprehensive data including detailed medical history, laboratory tests (e.g., lipids, HbA1c), and biometric measurements. In real-world clinical practice, these data points are frequently missing, leading to a significant gap where many high-risk individuals never undergo formal risk assessment. Consequently, these patients are less likely to receive guideline-recommended preventive therapies. There is an urgent need for a scalable, accessible, and data-efficient method to identify high-risk individuals to trigger further comprehensive assessment.

2. Methodology

Data Sources and Cohorts:

• Development & Internal Validation: Conducted using Electronic Health Record (EHR) data from the Yale New Haven Health System (YNHHS) in the US, linked to a state death registry.
  • Development Set: 363,788 individuals (ages 30–79).
  • Held-out Test Set: 83,917 individuals without prior ASCVD.
• External Validation: Two independent prospective cohorts:
  • ELSA-Brasil: A community-based cohort of Brazilian civil servants (10,934 individuals).
  • UK Biobank (UKB): A UK volunteer cohort (54,166 individuals).

Model Architecture:
The authors developed a family of three deep learning models, collectively termed ECG-ASCVD, designed to predict time-to-first ASCVD event (fatal or non-fatal MI or stroke).

1. ECG-ASCVD-12: A 1D-Residual Convolutional Neural Network (1D-ResNet) trained on raw 12-lead ECG voltage signals.
2. ECG-ASCVD-IMAGE: A Vision Transformer (BEiT-based) model fine-tuned on ECG images (scanned paper ECGs). It was pre-trained using contrastive learning on ECG images and corresponding echocardiogram reports.
3. ECG-ASCVD-1: A 1D-ResNet trained on Lead-1 signals only, designed for compatibility with wearable/portable single-lead devices.

Training Strategy:

• Pre-training: Models were pre-trained as autoencoders to reconstruct ECG waveforms from Gaussian noise-augmented data to improve robustness and generalizability (especially for wearable data).
• Fine-tuning: Models were fine-tuned using a neural network adapted time-to-event learning objective.
• Fusion: Age and sex were incorporated via decision fusion, where the final risk score is a linear combination of the ECG-derived score and demographic factors.
• Outcome Definition: Time to first fatal/non-fatal MI or stroke, identified via ICD-10 codes, death registries, or adjudicated events depending on the cohort.

Comparators:
Performance was benchmarked against the PREVENT risk score (base form without optional predictors) and evaluated for additive value using multivariate penalized Cox proportional hazards models.

3. Key Contributions

• First Image-Based ASCVD Model: This study presents the first deep learning model capable of predicting ASCVD risk directly from ECG images, demonstrating performance comparable to signal-based models. This is crucial for scalability in settings where raw digital signals are unavailable.
• Wearable Compatibility: The development and validation of ECG-ASCVD-1 (single-lead) establishes the feasibility of using consumer-grade wearable devices for population-level ASCVD screening.
• Multinational Generalizability: The models were validated across diverse populations (US, Brazil, UK) with varying demographics and healthcare systems, proving robustness beyond the development cohort.
• Independence from Traditional Risk Factors: The study demonstrates that ECG-derived AI scores provide independent prognostic information beyond standard clinical risk factors (PREVENT score), even when those factors are available.

4. Results

Discriminative Performance (C-Index):

• Internal Validation (YNHHS): All three models showed strong discrimination.
  • ECG-ASCVD-12: 0.711 (95% CI: 0.699–0.722)
  • ECG-ASCVD-IMAGE: 0.714 (95% CI: 0.703–0.726)
  • ECG-ASCVD-1: 0.709 (95% CI: 0.698–0.721)
  • Comparison: In the subset with complete PREVENT data, ECG-ASCVD-12 (0.723) performed nearly identically to PREVENT (0.724).
• External Validation:
  • UK Biobank: ECG-ASCVD-12 achieved a C-index of 0.684, comparable to PREVENT (0.696).
  • ELSA-Brasil: ECG-ASCVD-12 achieved 0.746, slightly lower than PREVENT (0.782) but still robust.
  • Note: The image-based and single-lead models showed performance very close to the 12-lead signal model across all cohorts.

Additive Value:
In Cox models adjusting for age, sex, and the PREVENT score, the ECG-ASCVD-12 score remained a significant independent predictor of ASCVD:

• YNHHS: Adjusted Hazard Ratio (HR) 1.34 per SD increase.
• ELSA: Adjusted HR 1.23 per SD increase.
• UKB: Adjusted HR 1.33 per SD increase.

Simulated Deployment:
In a simulation of 100,000 adults in YNHHS:

• Only 12.5% of eligible individuals had all necessary data to calculate a PREVENT score.
• An additional 16.5% had a clinical ECG but insufficient data for PREVENT.
• Deploying ECG-ASCVD models could identify 51% of the population (at an 80% sensitivity threshold) who would experience 80% of future ASCVD events, specifically targeting those currently missed by traditional screening due to missing data.

5. Significance and Implications

• Scalable Screening: The study proposes a paradigm shift where a routine, low-cost ECG (available in billions of studies annually) serves as a "gatekeeper" or pre-screening tool. It can flag high-risk individuals who lack the data for traditional scores, prompting clinicians to order the necessary labs for a full risk assessment.
• Resource-Limited Settings: The ability to use ECG images (which can be photographed) or single-lead signals (from wearables) makes this approach viable for low-resource settings where digital ECG signals and laboratory infrastructure are scarce.
• Clinical Workflow Integration: The models can be integrated passively into EHRs to "nudge" clinicians or deployed via smartphone apps for community screening.
• Guideline Evolution: The findings challenge current guidelines that do not recommend resting ECGs for asymptomatic individuals, suggesting that AI-enhanced ECGs could become as ubiquitous and standard as blood pressure measurement for risk stratification.

Limitations Noted:

• Outcome definition in YNHHS relied on ICD codes rather than adjudicated events (though this mirrors PREVENT derivation).
• Slight performance drops in the UKB cohort suggest potential overfitting or genuine cohort differences, though similar trends were observed in other ECG prognostic studies.
• Targeted validation on actual wearable device data is required before clinical deployment of the single-lead model.