Author:
Mohammad Motaghianfar
Abstract
Heart disease is a major global health issue, and lifestyle changes like quitting smoking are often recommended to prevent it. However, these one-size-fits-all recommendations may not work equally well for everyone. Using a cutting-edge approach called causal machine learning (CML), we studied how quitting smoking affects heart disease risk differently across individuals. We analyzed data from 929 people, looking at factors like age, weight, and health habits. Our findings show that while quitting smoking doesn’t significantly lower heart disease risk for everyone, about 25% of people benefit greatly, while another 25% may not see much benefit or could even face risks. Older individuals with lower body weight tended to benefit more. This study shows how CML can help create tailored prevention plans to make heart disease interventions more effective.
Keywords: causal machine learning, personalized medicine, heart disease prevention, smoking cessation, precision health
1. Introduction
Heart disease is the leading cause of death worldwide, and lifestyle changes are a key way to prevent it. Traditionally, doctors recommend the same advice—like quitting smoking—to everyone based on average results from large studies. But people respond differently to these changes, and blanket recommendations might not be the best approach for everyone.
Recent advances in causal machine learning (CML) let us dig deeper into how treatments affect individuals, not just groups. Unlike older methods, CML can analyze real-world data to estimate how specific people respond to interventions like quitting smoking. While this technology is promising, it hasn’t been widely used for heart disease prevention yet. Our study uses CML to explore how quitting smoking impacts heart disease risk differently across people, aiming to pave the way for personalized prevention plans.
2. Methods
2.1 Data and Participants
We studied data from 929 people who underwent medical exams. The data included details like age, sex, height, weight, blood pressure, cholesterol, glucose levels, and lifestyle habits such as smoking, drinking, and exercise. We also tracked whether participants had heart disease.
2.2 Data Preparation
To ensure accuracy, we cleaned the data by removing errors, like cases where blood pressure readings didn’t make sense or body mass index (BMI) values were extreme (below 15 or above 50). We also converted measurements, like turning age from days to years, and created a heart disease risk score based on cholesterol and glucose levels.
2.3 Causal Machine Learning Approach
We used a method called the X-Learner, paired with a tool called XGBoost, to estimate how quitting smoking affects heart disease risk for each person. The X-Learner is great for handling messy real-world data and uneven groups (like more smokers than non-smokers). In our analysis, “treatment” was whether someone smoked (smoker = 1, non-smoker = 0), the outcome was heart disease status, and we considered factors like age, sex, BMI, blood pressure, and heart disease risk score.
2.4 Testing and Validation
To ensure our results were trustworthy, we ran several checks:
-
E-value analysis to test if unmeasured factors could skew our findings.
-
Placebo tests where we randomly shuffled who was a smoker to see if fake data gave fake results.
-
Subgroup checks to see if results held up across different age groups.
We also measured the overall effect of quitting smoking and looked at how individual effects varied.
2.5 Ethical Notes
We used anonymous data, and the study was approved by an ethics board to ensure participant privacy.
3. Results
3.1 Overall Effect
On average, quitting smoking had a small, non-significant effect on lowering heart disease risk (-0.033, with a range of -0.072 to 0.006). This suggests that, for the average person, quitting smoking doesn’t dramatically change heart disease risk (see Figure 1).
3.2 Individual Differences
Despite the small average effect, we found big differences in how quitting smoking affected individuals. Effects ranged from a strong benefit (-1.52) to potential harm (+1.32), with a standard deviation of 0.60. About 25% of people saw significant benefits (effect > 0.1), while 25% might face risks or no benefit (effect < -0.1). This shows clear groups of “responders” and “non-responders” (see Figure 2).
3.3 Who Benefits Most?
People who benefited most from quitting smoking were typically older (average age 51.4 years) and had lower BMI (26.3) compared to those who didn’t benefit as much (average age 48.3 years, BMI 27.8). Age and BMI were the biggest factors driving these differences (see Figure 3).
3.4 Validation Results
Our tests showed our findings were moderately robust. The E-value (1.22) suggested some sensitivity to unmeasured factors, but placebo tests (which showed no effect with fake data) and age-group checks supported our results. Older adults (50+ years) showed a stronger benefit (-0.095) compared to younger adults (see Table 1).
|
Age Group |
Number of People |
Average Effect |
Confidence Interval |
|---|---|---|---|
|
<50 years |
465 | 0.049 | (-0.015, 0.113) |
|
≥50 years |
464 | -0.095 | (-0.158, -0.032) |
Table 1. Effect of Quitting Smoking by Age Group
4. Discussion
Our study reveals that not everyone benefits equally from quitting smoking to prevent heart disease. While the average effect was small, about one in four people saw major benefits, while another one in four might not benefit much or could even face risks. This challenges the idea that the same advice works for everyone and shows the power of personalized approaches.
Using the X-Learner method, we could estimate individual responses in a way that’s practical for real-world data, where controlled experiments aren’t always possible. This approach is a step forward from traditional studies that focus on group averages.
From a doctor’s perspective, our findings suggest focusing smoking cessation efforts on those likely to benefit most, like older adults with lower BMI. We created a decision map (see Figure 4) to help doctors use patient details, like age and weight, to tailor advice.
4.1 Limitations and Next Steps
Our study has some limitations. Since we used real-world data, not a controlled experiment, we can’t be 100% sure about cause and effect, even with our rigorous tests. The moderate E-value also suggests some risk of unmeasured factors affecting results. In the future, we’d like to test these findings in controlled trials and explore other lifestyle changes, like diet or exercise, for heart disease prevention.
5. Conclusion
This study shows that quitting smoking affects heart disease risk differently for different people. By using causal machine learning, we can identify who’s likely to benefit most, paving the way for personalized prevention plans. This approach could make heart disease prevention more effective by focusing resources on those who need them most. Future research should test these ideas in diverse groups and with other lifestyle changes.
References
[1] Chernozhukov, V., et al. (2018). Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal.
[2] Kunzel, S. R., et al. (2019). Meta-learners for estimating heterogeneous treatment effects using machine learning. Proceedings of the National Academy of Sciences.
[3] VanderWeele, T. J., & Ding, P. (2017). Sensitivity Analysis in Observational Research: Introducing the E-Value. Annals of Internal Medicine.
[4] Prosperi, M., et al. (2020). Causal inference and counterfactual prediction in machine learning for healthcare. Nature Medicine.