Predicting cardiovascular risk under intervention: Development and internal validation of the CHARIOT Model in 19 million adults

The CHARIOT model, developed and internally validated using electronic health records from over 19 million UK adults, uniquely predicts 10-year cardiovascular risk reductions under specific interventions like statins or lifestyle changes to enable actionable, patient-centred primary prevention.

The Gist

Imagine you are standing at a crossroads, looking at a map of your future health. For decades, the maps doctors have given us for heart disease have been very simple. They point to a spot and say, "You are in the danger zone. You need to change something." But they couldn't tell you what to change, or how much it would help. It was like being told, "Your car is going to break down soon," without being told if you should change the oil, get new tires, or just drive slower.

Enter CHARIOT. Think of CHARIOT not just as a map, but as a high-tech flight simulator for your heart.

Here is the simple breakdown of what this new tool does, how it works, and why it matters.

1. The Problem with Old Maps

Traditional heart risk calculators (like the ones currently used in the UK) are like rear-view mirrors. They look at your history—your age, your weight, your smoking habits—and tell you the odds of having a heart attack in the next 10 years based on what has already happened.

If a patient asks, "What if I quit smoking?" or "What if I start taking blood pressure pills?", the old tools can't answer. They just say, "You are high risk." They can't simulate the future.

2. The CHARIOT Solution: A "What-If" Machine

CHARIOT is different. It combines two powerful ideas: Prediction (guessing the future) and Causality (understanding cause and effect).

Think of CHARIOT as a video game with a "Save State" feature.

• The Baseline: It first calculates your current risk score (e.g., "You have an 18% chance of a heart event in 10 years").
• The Simulation: Then, it lets you press "Reset" and try different scenarios.
  • Scenario A: "What if I start a statin?" -> The simulator runs the numbers and says, "Your risk drops to 14%."
  • Scenario B: "What if I quit smoking?" -> "Your risk drops to 15%."
  • Scenario C: "What if I do all three?" -> "Your risk drops to 5%."

It doesn't just guess; it uses real data from millions of people to calculate exactly how much each specific action lowers your personal risk.

3. How It Was Built (The Secret Sauce)

The researchers didn't just guess these numbers. They built a massive engine using electronic health records from 19 million adults in the UK. That's a huge crowd!

• The "Time Travel" Trick: One of the hardest things in medicine is that people change their minds. Someone might start taking pills, then stop. Old models get confused by this. CHARIOT uses a special mathematical trick (called "causal inference") to separate what happened because of the treatment from what happened naturally. It's like a detective who can tell the difference between a suspect's actions and the actions of everyone else in the room.
• The "Memory" Feature: Most health apps forget you after you close the window. CHARIOT has a memory. If you visit your doctor today, and then come back in a year, it remembers your starting point. It can say, "Last year you were at 18%. You quit smoking and lost weight. Now you are at 11%. But your blood pressure is still high. Let's see what happens if we treat that."

4. A Real-Life Example

Imagine a 70-year-old woman named Sarah.

• The Old Way: The doctor says, "Sarah, your risk is 19%. You should probably take a statin." Sarah feels overwhelmed and doesn't know if the pill is worth it.
• The CHARIOT Way: The doctor opens the app and says, "Sarah, right now your risk is 19%. But look at this: If you quit smoking, your risk drops to 15%. If you start a statin, it drops to 14%. If you do both, it drops to 10%."
• The Result: Sarah sees the numbers. She realizes quitting smoking gives her almost the same benefit as the pill, but with fewer side effects. She feels empowered. She chooses to quit smoking. The tool becomes a conversation starter, not just a warning label.

5. Why This Matters

The biggest breakthrough here is motivation. Psychology tells us that people are more likely to change their behavior if they believe their specific actions will make a difference. This is called "response efficacy."

CHARIOT gives patients a clear, personalized "cause and effect" story. It turns a scary statistic ("You might get sick") into an empowering plan ("If I do X, I will be safer").

6. The Bottom Line

CHARIOT is a massive step forward. It moves us from passive prediction (telling you what might happen) to active decision support (telling you what will happen if you take action).

It's like moving from a weather report that just says "It might rain" to a smart app that says, "If you buy an umbrella, you stay dry. If you don't, you get wet. Here is exactly how much dryness you get for your money."

In short: CHARIOT is a personalized simulator that helps you and your doctor play out different health futures, so you can choose the one that leads to a longer, healthier life.

Technical Summary

1. Problem Statement

Conventional cardiovascular disease (CVD) risk prediction models (e.g., QRISK, Pooled Cohort Equations) are purely associational. They estimate a patient's risk based on current observations but cannot answer critical clinical questions regarding counterfactual interventions. Specifically, they fail to quantify:

• How much a specific action (e.g., starting statins, quitting smoking, losing weight) would reduce an individual's risk.
• The comparative benefit of different intervention strategies for a specific patient.

Current workarounds, such as applying average treatment effects to a baseline score, suffer from methodological flaws, including the inability to handle "treatment drop-in" (patients starting or stopping medication during follow-up) and the inability to model complex lifestyle changes. This limits their utility as decision-support tools for motivating patient behavior change.

2. Methodology

The authors developed CHARIOT (Causal Health Assessment and Risk Intervention Optimization Tool), a model designed to predict 10-year CVD risk under hypothetical interventions.

Data Source and Cohort

• Dataset: Retrospective cohort study using the Clinical Practice Research Datalink (CPRD) Aurum (UK primary care electronic health records).
• Sample Size: 19,410,403 adults (approx. 9.5 million males, 9.9 million females) from 1,478 English general practices.
• Inclusion: Adults aged 18+ without prior CVD diagnosis.
• Outcome: Incident CVD (composite of coronary heart disease, ischemic stroke, TIA, and related deaths).
• Follow-up: Up to March 1, 2020.

Model Architecture

CHARIOT consists of two distinct components:

A. Initial Risk-Estimation Component

• Model Type: Cox proportional hazards model with restricted cubic splines for non-linear predictors (age, SBP, BMI, non-HDL cholesterol).
• Handling Treatment Drop-in: To address the issue of patients changing treatment status during follow-up (which biases standard models), the authors used a method inspired by Xu et al. (2023). They created interval-censored data at treatment change points and introduced time-varying offset terms.
• Causal Integration: The coefficients for these time-varying offsets were fixed to causal average total effects derived from systematic reviews and meta-analyses (e.g., statins, antihypertensives). This allows the model to estimate risk under the strategy of "continuing current intervention" while accounting for the causal impact of treatment changes observed in the data.

B. Intervention Component (Prediction Under Intervention)

• Mechanism: This component does not fit new models but applies causal odds ratios to the baseline risk estimates.
• Causal Structure: Utilizes a Directed Acyclic Graph (DAG) to map interventions (statins, smoking cessation) to modifiable risk factors (SBP, BMI, non-HDL, smoking status).
• Flexibility: It decomposes effects into direct and indirect pathways. For example, it can estimate the risk reduction of stopping smoking by calculating the direct effect of smoking cessation plus the indirect effects mediated through changes in BMI and blood pressure.
• Memory Property: Designed for longitudinal use. At follow-up visits, the model:
  1. Remembers baseline risk factor values (Visit 0).
  2. Updates non-modifiable factors (e.g., age).
  3. Estimates risk based on achieved modifiable factors.
  4. Estimates hypothetical future risks based on proposed interventions.

Statistical Techniques

• Missing Data: Multiple Imputation by Chained Equations (MICE) with 10 chains.
• Validation: Internal validation using a 70/30 split. Evaluation included calibration (ICI, E50, E90), discrimination (Harrell's C-index), and subgroup analysis (sex, age, ethnicity, region).
• Temporal Validation: The intervention component was tested by simulating follow-up visits at 1, 3, and 5 years to assess performance over time.

3. Key Contributions

1. Causal Prediction Framework: CHARIOT is one of the first large-scale models to integrate causal inference directly into the risk estimation process, moving beyond association to predict outcomes under specific interventions.
2. Handling Treatment Drop-in: The novel use of time-varying offsets with fixed causal coefficients allows the model to remain valid for patients who are already on treatment or change treatment status, a common failure point in traditional models.
3. Arbitrary Intervention Combinations: Unlike previous tools (e.g., Million Hearts, Qintervention) that offer pre-defined intervention lists, CHARIOT allows users to input arbitrary changes to modifiable risk factors (e.g., "reduce SBP by 15mmHg and lose 10kg") to see the combined effect.
4. Longitudinal "Memory" Design: The model is explicitly architected for repeated clinical encounters, storing baseline values to accurately calculate risk reduction relative to a patient's specific starting point.
5. Scale and Representativeness: Developed on a cohort of nearly 20 million adults, making it one of the largest CVD prediction models to date.

4. Results

• Discrimination: The model demonstrated strong discrimination, with Harrell's C-indices of 0.874 for females and 0.859 for males. Performance remained robust across ethnic and regional subgroups, though it naturally decreased in younger age groups (<30) where absolute risk is low.
• Calibration: The model was well-calibrated across most subgroups. Minor under-prediction was observed in high-risk individuals in the North West of England, and slight miscalibration in the 17–30 age group (attributed to spline knot placement in a low-risk demographic).
• Intervention Component Performance:
  • Calibration remained strong for lifestyle modifications (BMI, non-HDL, smoking).
  • Slight over-prediction was observed for high-risk individuals regarding systolic blood pressure modifications at follow-up visits.
• Worked Example:
  • A 70-year-old female smoker with a baseline risk of 18.97%.
  • Statin therapy reduced risk to 14.32%.
  • Smoking cessation reduced risk to 15.77%.
  • Antihypertensive therapy (10mmHg reduction) reduced risk to 15.26%.
  • Combination therapy (statin + BP reduction + weight loss) reduced risk to 5.49%.

5. Significance and Clinical Implications

• Actionable Decision Support: CHARIOT transforms risk prediction from a static "flag" into a dynamic conversation tool. It addresses the psychological mechanism of response efficacy, helping patients visualize the tangible benefit of specific actions (e.g., "If I quit smoking, my risk drops by X%").
• Personalized Prevention: By quantifying the trade-offs between medication and lifestyle changes, it empowers shared decision-making between clinicians and patients.
• Scalability: The model is publicly available as an interactive Rshiny application and is designed for integration into NHS Health Checks and patient-facing digital health platforms.
• Future Directions: The authors identify the need to extend the model to competing risks (to estimate lifetime risk and multiple outcomes like diabetes) and to validate the model in non-English UK nations.

In summary, CHARIOT represents a substantive methodological shift in CVD prevention, bridging the gap between epidemiological risk prediction and causal decision support, enabling clinicians to prescribe not just a diagnosis, but a quantified path to risk reduction.