Repurposing cardiovascular disease prediction models for cancer

This study demonstrates that widely implemented cardiovascular disease risk models, when recalibrated, perform comparably to dedicated cancer risk models in predicting 10-year cancer risk, suggesting they can be effectively repurposed to guide cancer prevention and risk-stratified monitoring.

The Gist

The Big Idea: Reusing an Old Map for a New Journey

Imagine you have a very reliable, well-worn map that helps drivers avoid traffic jams and car accidents. This map is called a Cardiovascular Disease (CVD) model. Doctors use it all the time to predict who is likely to have a heart attack or stroke in the next 10 years.

Now, imagine scientists asked a bold question: "Could this same map also help us predict who is likely to get cancer?"

Usually, doctors use different maps for different problems. They have one map for heart health and a separate, less commonly used map for cancer. But this study suggests that the "Heart Map" is actually so good at spotting risk factors (like smoking, high blood pressure, and age) that it can also spot the danger signs for cancer.

The Experiment: A Test Drive on Two Roads

The researchers took four of the most popular "Heart Maps" (QRISK3, PCE, SCORE2, and SCORE2-OP) and put them to the test.

1. The Training Ground: They first taught these models to look for cancer using data from the UK Biobank (a massive database of half a million people). They adjusted the models slightly, like tuning a radio to get a clearer signal.
2. The Real-World Test: Then, they drove these models on a completely different road: the Clinical Practice Research Datalink (CPRD), which contains records from millions of real patients in GP clinics across England.

The Result? The "Heart Maps" didn't just work; they worked surprisingly well.

What Did They Find?

1. The "Jack-of-All-Trades" Performance
For many types of cancer—especially those linked to lifestyle, like lung, liver, and kidney cancer—the heart models performed just as well as the specialized cancer models.

• The Analogy: Think of it like a Swiss Army Knife. You usually buy a specific tool for every job (a screwdriver for screws, a knife for cutting). But this study found that the "Heart Screwdriver" is actually sharp enough to cut through "Cancer Wood" just as effectively as a dedicated "Cancer Knife."

2. The Smoking Connection
The study found that the most important thing these models looked at was smoking.

• The Analogy: If the models were detectives, "Smoking" was the biggest fingerprint at the crime scene. Since smoking causes both heart attacks and many cancers, the models that were already good at spotting smoking-related heart risks were naturally good at spotting smoking-related cancer risks too.

3. The "False Alarm" Check
To make sure the models weren't just guessing, the researchers tested them on something unrelated: accidental injuries (like breaking a leg).

• The Result: The models failed to predict injuries (which is good!). This proved they weren't just saying "everyone is at risk of everything." They were specifically picking up on the biological signals that lead to disease.

Why Does This Matter?

1. No Need to Reinvent the Wheel
Specialized cancer prediction tools exist, but many doctors don't use them because they aren't built into their computer systems. However, the "Heart Maps" are already installed in almost every doctor's office in the UK and US.

• The Metaphor: Instead of building a new, expensive bridge to cross the river, we realized the existing bridge is strong enough to carry the new traffic too. We just need to put up a new sign saying, "This bridge also leads to Cancer Prevention."

2. A Double Win for Prevention
If a doctor sees a patient has a high risk of a heart attack, they might say, "Let's lower your blood pressure and stop smoking."

• The New Approach: With this repurposed model, the doctor can say, "You have a high risk of a heart attack and a high risk of cancer. Let's lower your blood pressure and stop smoking to protect you from both." It gives patients a stronger reason to make healthy changes.

3. Smarter Screening
Currently, cancer screening (like lung scans) is often based on age alone. This study suggests we could screen people based on their actual risk.

• The Analogy: Instead of checking every car in a parking lot for a flat tire, we could use a sensor to check only the cars that look like they might have a flat tire. This saves time and money, allowing doctors to focus on the people who need help the most.

The Bottom Line

This research shows that we don't always need a brand-new tool to solve a new problem. By tweaking the tools we already have, we can catch more diseases earlier, save resources, and help people stay healthier for longer. It's a clever, efficient way to use our existing medical knowledge to fight two of the world's biggest killers at the same time.

Technical Summary

1. Problem Statement

Cancer and cardiovascular disease (CVD) are the leading causes of death globally and share common modifiable risk factors (e.g., smoking, hypertension, obesity). However, clinical management strategies differ significantly:

• CVD: Widely utilizes risk prediction models (e.g., QRISK3, SCORE2) to stratify risk and initiate primary prevention (e.g., statins, antihypertensives).
• Cancer: Relies primarily on population-based screening for specific cancers (breast, cervical, colorectal) to detect early-stage disease, rather than predicting future risk in asymptomatic individuals. Existing cancer risk models (e.g., QCancer) are underutilized in clinical practice.

The Core Question: Can widely implemented, clinically integrated CVD risk models be "repurposed" to predict the 10-year risk of incident cancer with performance comparable to dedicated cancer models, thereby enabling risk-stratified cancer prevention and screening within existing primary care workflows?

2. Methodology

Data Sources

• Primary Cohort: UK Biobank (UKB), split into a 20% training set (for recalibration) and an 80% testing set.
• External Validation: Clinical Practice Research Datalink (CPRD) Aurum database (primary care records from England).
• Population: Adults aged 40–84 years at study start (Jan 1, 2011).
• Outcomes: 10-year incidence of 11 specific cancer types (including colorectal, lung, breast, prostate, liver, etc.), composite "any cancer," CVD, and combined CVD/cancer events.

Models Evaluated

• CVD Models (Repurposed): QRISK3, Pooled Cohort Equations (PCE), SCORE2, and SCORE2-OP.
• Benchmark Model: QCancer (a dedicated cancer risk prediction tool).
• Negative Control: Accidental injury (to ensure models were not predicting general morbidity).

Statistical Approach

1. Recalibration: CVD models were recalibrated using the UKB training set to adjust for the difference in baseline incidence between CVD and cancer. This involved rank-preserving transformations of the intercept and slope.
2. Performance Metrics:
   • Discrimination: Measured by the C-statistic (Area Under the Curve).
   • Calibration: Measured by calibration intercept (bias) and slope (reliability of risk spread).
3. Feature Importance: Permutation importance analysis was conducted on QRISK3 to identify which risk factors drove cancer prediction.
4. Sensitivity Analyses: Included landmark analysis (excluding cancers diagnosed within the first 3 years to rule out reverse causality/subclinical disease) and imputation of missing data in CPRD.

3. Key Contributions

• Novel Repurposing Strategy: Demonstrates that CVD models, already embedded in Electronic Health Records (EHR) and national health checks (e.g., NHS Health Check), can effectively predict cancer risk without requiring new data collection.
• Comprehensive Validation: Validated across two massive, distinct datasets (UKB and CPRD) with over 5 million combined participants, ensuring robustness across different population structures.
• Risk-Stratified Screening Framework: Proposes a pathway to integrate cancer risk assessment into existing CVD prevention infrastructure, potentially enabling targeted screening for cancers lacking current population-wide programs (e.g., renal, laryngeal, gastro-oesophageal).
• Open Access Implementation: The recalibrated models are made available via a public API and web application.

4. Key Results

Discrimination (C-Statistic)

• CVD Performance: All CVD models performed well for their original purpose (C-statistics 0.71–0.74).
• Cancer Performance:
  • For "Any Cancer", all CVD models achieved a C-statistic of 0.63, which was comparable to the QCancer model (0.65).
  • Specific Cancers: CVD models showed strong discrimination (C-statistic 0.70–0.81) for gastro-oesophageal, liver/biliary, laryngeal, renal tract, and lung cancers.
  • Lung Cancer Nuance: While QCancer outperformed CVD models for the general lung cancer population (0.80 vs. ~0.75), performance was nearly identical for never-smokers (0.64 vs. 0.63).
  • Combined Outcomes: Models effectively predicted the joint risk of developing both CVD and cancer (C-statistic ~0.76–0.78).
  • Negative Control: As expected, models performed poorly for accidental injury (C-statistic ~0.53–0.55), confirming specificity.

Calibration

• After recalibration, the CVD models showed near-perfect calibration in the UKB test set (Median intercept: 0.01; Median slope: 1.00).
• Performance was highly consistent in the external CPRD validation, with only marginal differences in discrimination and calibration compared to UKB.

Feature Importance

• Top Predictors: After age, smoking status and systolic blood pressure (SBP) were the most influential predictors for cancer risk in the QRISK3 model.
• Other Factors: Total cholesterol/HDL ratio, Townsend Deprivation Index, and family history of CVD also contributed significantly.
• Implication: This confirms that shared risk factors drive the predictive power of these repurposed models.

5. Significance and Clinical Implications

• Infrastructure Efficiency: By leveraging existing CVD risk assessment tools (like QRISK3 in the UK), healthcare systems can immediately deploy cancer risk stratification without the resource burden of developing and integrating new, standalone cancer models.
• Preventative Synergy: Identifying high-risk individuals for both CVD and cancer simultaneously could reinforce lifestyle interventions (smoking cessation, diet, exercise) and pharmacological treatments (statins, antihypertensives, potentially GLP-1 agonists for obesity-related cancers).
• Targeted Screening: The models can identify high-risk subgroups for cancers where population screening is not currently recommended due to low incidence or test specificity (e.g., renal or laryngeal cancer). This could justify targeted low-dose CT screening or multi-cancer early detection (MCED) tests for specific high-risk cohorts rather than age-based screening.
• Future Directions: The authors suggest using these models to refine screening thresholds (e.g., for lung cancer) based on individual risk rather than age alone, and to prioritize patients for emerging early detection technologies.

Conclusion: The study successfully validates that widely used CVD prediction models, when recalibrated, serve as effective tools for predicting 10-year cancer risk. This offers a pragmatic, scalable solution for integrating cancer prevention into routine primary care.