Development and validation of an ECG-based 10-year risk prediction model for Major Adverse Cardiac and Cerebrovascular Events in UK Biobank

This study demonstrates that while ECG-derived risk scores independently predict 10-year major adverse cardiac and cerebrovascular events in UK Biobank participants, their addition to traditional QRISK3 clinical factors yields only a marginal improvement in overall risk prediction, with the most notable benefit being enhanced risk stratification for women.

The Gist

The Big Picture: Can a Heart Test Predict the Future?

Imagine you are trying to predict who might get a flat tire on a long road trip. You have a list of known risk factors: the age of the car, how many miles it's driven, and whether the driver is known for speeding. This is like the standard medical risk scores doctors use today (called QRISK3). They are pretty good, but they aren't perfect. Sometimes, a "new" car with a "good" driver still gets a flat tire because of a hidden defect in the rubber.

This study asked a simple question: Can we get a better prediction by actually listening to the engine while it's revving?

In medical terms, the researchers wanted to know if adding an Electrocardiogram (ECG)—a test that measures the heart's electrical activity—taken while a person is resting, exercising, and recovering, could help predict heart attacks and strokes better than just looking at age, blood pressure, and cholesterol.

The Experiment: The "Heart Gym" Test

The researchers used data from the UK Biobank, a massive database of half a million people. They focused on 41,000 people who were healthy enough to go to a "heart gym" (a cycle station) but had never had a heart attack or stroke before.

Here is what they did:

1. The Warm-up: They took a 15-second reading of the heart while the person sat still (Rest).
2. The Workout: They had the person pedal a stationary bike at a moderate pace for 6 minutes (Exercise).
3. The Cool-down: They took a reading for 1 minute after the person stopped (Recovery).

They didn't just look at the obvious numbers (like heart rate). They used two different "detectives" to analyze the data:

• Detective A (The Traditionalist): Looked at standard, old-school measurements (like the length of the heart's electrical waves).
• Detective B (The AI Super-Scanner): A powerful Artificial Intelligence (Deep Learning) that looked at the entire raw electrical signal, searching for tiny, invisible patterns that human eyes or standard machines would miss.

The Findings: The Good, The Bad, and The "Meh"

After tracking these people for over 12 years, here is what they found:

1. The Heartbeat is a Crystal Ball (Sort of)
Both the traditional measurements and the AI scanner found that the way the heart behaves during exercise does contain clues about future heart trouble. If the heart's electrical signal looked "weird" during the workout, that person was more likely to have a heart event later.

• Analogy: It's like hearing a specific, subtle rattle in a car engine while driving uphill. Even if the car looks fine, that rattle suggests trouble ahead.

2. The AI was a Sharp Detective
The AI model (ECGAI) was surprisingly good at spotting these risks on its own, even without knowing the person's age or blood pressure. It performed better than the traditional measurements.

3. The "Extra Credit" Was Small
This is the most important part. When the researchers added the heart test results to the standard risk scores (the list of age, blood pressure, etc.), the improvement was very small.

• Analogy: Imagine you are guessing the winner of a horse race. You already know the horse's age, jockey, and past wins (the standard risk score). You then add a new piece of info: "The horse's ears were twitching." While that might tell you something, it doesn't change your prediction much. The standard score was already doing 95% of the heavy lifting.

4. A Glimmer of Hope for Women
There was one interesting exception. The heart test helped reclassify women slightly better than men. It helped identify a few more women who were at higher risk than the standard score thought they were.

• Analogy: The standard risk score is like a generic raincoat. It fits most people okay. But for women, adding the heart test is like adding a hood to that raincoat—it doesn't make it waterproof, but it keeps the rain off your head a little better.

The Conclusion: Should We All Exercise-Test Our Hearts?

The researchers concluded that while the exercise heart test is scientifically interesting and does show a link to future heart trouble, it probably shouldn't be used as a routine screening tool for everyone right now.

Why? Because the standard risk scores (age, blood pressure, cholesterol) are already so good that adding the heart test doesn't change the doctor's decision enough to justify the cost and effort of the test for the average person.

The Takeaway:
Think of the standard risk score as a very accurate weather forecast. The heart test is like a fancy, high-tech barometer. While the barometer gives you some extra data, it doesn't tell you to bring an umbrella if the forecast already said "sunny." However, for specific groups (like certain women), that extra data might just be the nudge needed to take extra precautions.

In short: The heart test is a cool tool that works, but for now, it's not a game-changer for the general population. The old-school risk factors are still the MVPs of prediction.

Technical Summary

1. Problem Statement

Cardiovascular disease (CVD) remains a leading cause of mortality, and early risk stratification is critical for primary prevention. Current clinical risk scores, such as QRISK3 (widely used in UK primary care), rely heavily on traditional risk factors (age, sex, blood pressure, lipids, etc.) but often suffer from limited accuracy and heterogeneity in individual predictions. While Electrocardiograms (ECGs) provide real-time insights into myocardial structure and electrical activity, their utility for long-term risk prediction—particularly when combined with exercise stress testing and deep learning—remains unclear. Previous studies have been limited by small sample sizes, short follow-up durations, or a reliance on conventional, hand-crafted ECG features rather than raw signal analysis.

The core research question is: Do ECG measurements (resting, exercise, and recovery), analyzed via both conventional parameters and deep learning, improve the 10-year prediction of Major Adverse Cardiovascular and Cerebrovascular Events (MACCE) beyond established clinical risk models like QRISK3?

2. Methodology

Study Population and Data Source

• Cohort: 41,076 participants from the UK Biobank (aged 40–69) who underwent a submaximal cycle ergometry test between 2009 and 2013.
• Inclusion Criteria: No prior history of MACCE, no use of calcium channel blockers or antiarrhythmics prior to testing.
• Exclusion: Participants with missing or corrupted ECG data.
• Follow-up: Median of 12.5 years. Primary outcome: Incident MACCE (hospitalization or death due to ischemic heart disease, heart failure, or cerebrovascular disease).

ECG Data Acquisition & Processing

• Device: Single-lead ECG (CAMUSB 6.5) at 500 Hz sampling rate.
• Phases: Recordings captured during:
  1. Rest (15 seconds).
  2. Submaximal Exercise (6 minutes).
  3. Recovery (1 minute).
• Total Duration: 7 minutes 15 seconds per participant.
• Preprocessing: Signal cleaning to remove baseline wander and high-frequency noise.

Model Development

The study developed and compared two distinct ECG-based risk scores:

1. C-ECG (Conventional ECG): A risk score derived from 12 hand-crafted features measured at rest, peak exercise, and recovery.
   • Morphological: PR, QRS, QTc intervals (corrected via Bazett's at rest, Fridericia's during exercise).
   • Rhythm: Premature ventricular contractions (PVCs), Coefficient of Sample Entropy (CosEn), and six Poincaré plot-derived features (irregularity, regularity, density, anisotropy, AF evidence, Atrial Tachycardia evidence).
2. ECGAI (AI-based ECG): A deep learning risk score derived from the raw 7-minute 15-second ECG signal.
   • Architecture: An adapted ResNet-V2 (50 layers) with 1D convolutions combined with a Long Short-Term Memory (LSTM) network.
   • Framework: Deep Survival Analysis (DeepSurv) to explicitly model time-to-event and censoring, rather than simple classification.
   • Output: A continuous 10-year risk probability.

Baseline Clinical Models

To assess incremental value, the ECG scores were added to three baseline models:

1. Age + Sex.
2. Cox Clinical: A re-fitted Cox proportional hazards model using QRISK3 covariates (adjusted for the UK Biobank population structure).
3. Clinical Deep Survival Network (DSN): A deep learning model trained only on QRISK3 risk factors.

Validation and Metrics

• Internal Validation: Five-fold stratified cross-validation with 1,000 bootstrap iterations.
• Performance Metrics:
  • Harrell's C-index: Discrimination ability.
  • Net Reclassification Index (NRI): Continuous and categorical (at 10% risk threshold).
  • Net Benefit: Clinical utility via decision curve analysis.
  • Calibration: Integrated Calibration Index (ICI).

3. Key Results

Independent Associations

• Both C-ECG and ECGAI scores were independently associated with incident MACCE after adjusting for traditional risk factors.
  • ECGAI: Hazard Ratio (HR) of 1.18 (95% CI: 1.15–1.21) per standard deviation (SD) increase.
  • C-ECG: HR of 1.76 (95% CI: 1.63–1.91) per SD increase.
• ECGAI showed a linear association with risk, with the highest quintiles showing significantly elevated risk compared to the lowest.

Predictive Performance (Discrimination)

• Univariate Performance: ECGAI derived from exercise phases alone yielded the strongest univariate C-index (0.591), outperforming resting and recovery phases.
• Added Value to Clinical Models:
  • Adding ECGAI to the Cox Clinical model resulted in a modest improvement in discrimination ($\Delta$C-index = 0.002, p=0.002).
  • Adding ECGAI to the Clinical DSN (Deep Learning Clinical) model showed a slightly larger but still modest improvement ($\Delta$C-index = 0.007 for women).
  • Sex Differences: The improvement in discrimination was statistically significant only in women when adding ECGAI to the Clinical DSN. Men showed no significant C-index improvement.

Reclassification and Clinical Utility

• Net Reclassification Index (NRI):
  • Continuous NRI: Significant improvements were observed across all models (e.g., +0.097 for Cox Clinical).
  • Categorical NRI (at 10% threshold): Generally minimal. However, a significant improvement was observed in women ($NRI = 0.09$) when adding ECGAI to the Clinical DSN.
• Net Benefit: No significant improvement in net benefit was observed at the 10% risk threshold for the overall population, suggesting limited impact on clinical decision-making for statin initiation in this low-risk cohort.

Subgroup Analysis

• The greatest reclassification benefits were found in younger men (<45 years)** and **older women (>55 years).
• Sensitivity analyses excluding events within the first 2 and 5 years confirmed that reverse causation bias did not drive the results.

4. Key Contributions

1. First Large-Scale Deep Survival Study on Exercise ECG: This is the largest population-based study (n=41,076) to apply deep learning to submaximal exercise ECGs for 10-year MACCE prediction, moving beyond the standard resting 12-lead ECGs used in prior AI studies.
2. Comprehensive Phase Analysis: Unlike previous studies, this work explicitly evaluated the prognostic value of rest, exercise, and recovery phases separately and combined.
3. Rigorous Benchmarking: The study compared AI models not just against simple baselines but against a re-calibrated QRISK3 model and a Deep Survival Clinical Network, providing a realistic assessment of incremental value in a clinical context.
4. Sex-Specific Insights: The study identified that deep learning ECG features may offer unique predictive value specifically for women, a subgroup often under-represented in CVD risk modeling.

5. Significance and Limitations

Significance

• Validation of Independent Association: Confirms that ECG signals (especially during exercise) contain independent prognostic information regarding future cardiovascular events, even in asymptomatic individuals.
• Limited Incremental Utility in Low-Risk Populations: The study aligns with USPSTF conclusions that while ECGs improve statistical discrimination, they offer minimal clinical utility for reclassifying low-risk individuals across standard treatment thresholds (e.g., 10% 10-year risk). This suggests routine exercise ECG screening for primary prevention in low-risk populations may not be cost-effective.
• Potential for High-Risk Subgroups: The findings suggest that deep learning ECG models might be more valuable in specific subgroups (e.g., older women, younger men) or higher-risk populations where the baseline risk is higher.

Limitations

• Healthy Volunteer Bias: The UK Biobank cohort is healthier than the general population, leading to a lower event rate (7.3% vs. ~18% in general populations), which may attenuate the predictive power of the models.
• Submaximal Exercise: Participants only reached ~50% of their predicted maximum workload, potentially failing to unmask latent ischemia that a maximal stress test would reveal.
• Signal Quality: Single-lead exercise ECGs are prone to motion artifacts, leading to the exclusion of some participants and potentially obscuring subtle morphological features.
• Lack of External Validation: Due to the uniqueness of the UK Biobank exercise protocol, external validation in a comparable dataset was not possible.

Conclusion

The study demonstrates that deep learning-derived ECG risk scores (ECGAI) are independently associated with 10-year MACCE risk and can modestly improve the discrimination of established clinical risk models, particularly in women. However, the incremental clinical value (reclassification and net benefit) at standard treatment thresholds is limited. The authors conclude that while ECGs are powerful diagnostic tools, their routine use for primary prevention risk stratification in low-risk populations adds little value over current clinical guidelines. Future research should focus on applying these deep learning models to higher-risk populations where the potential for improved risk stratification is greater.