TL;DR
This study explores the use of machine learning (ML) algorithms to predict high coronary artery calcium scores (CACS) in a healthy population, using only routine health examination data. The findings suggest that advanced ML models like XGBoost significantly outperform traditional binary logistic regression, offering a more accurate and potentially more accessible tool for early cardiovascular disease risk assessment.
What Problem Does This Paper Address?
Coronary artery calcium score (CACS) is a crucial indicator of future cardiovascular disease (CVD) risk. While CT-based deep learning methods have shown promise in predicting CACS from imaging, there's a gap in understanding the efficacy of machine learning algorithms using readily available clinical and laboratory data for predicting high CACS in a healthy, general population. Current risk stratification tools often rely on more complex assessments or are developed for symptomatic patients. This paper aims to bridge this gap by investigating whether ML models, beyond standard binary logistic regression (BLR), can effectively predict high CACS using only data from regular health screenings.
Methods
The researchers conducted a retrospective observational study using data from participants who underwent regular health screenings, including coronary CT angiography. They defined "high CACS" as an Agatston score of 100 or more.
The core of their methodology involved comparing three classification algorithms:
- Binary Logistic Regression (BLR): A standard statistical method for binary classification.
- CatBoost: An open-source gradient boosting library designed to handle categorical features effectively.
- XGBoost (Extreme Gradient Boosting): A highly optimized and widely used gradient boosting library known for its speed and performance.
The dataset was split into training (70%) and testing (30%) sets, with stratification to ensure proportional representation of high CACS cases in both. All ML models were trained using 5-fold cross-validation and hyperparameter optimization via grid search to prevent overfitting and ensure robust performance. The primary metric for evaluating the models' predictive power was the Area Under the Receiver Operating Characteristic (AUROC) curve.
Datasets
The study utilized a dataset of 2133 participants who underwent regular health screenings and coronary CT angiography at three tertiary academic centers between January 2014 and December 2019. The input features for the ML models included anthropometric measures (age, sex, blood pressure, body mass index) and laboratory parameters (complete blood count, liver and renal function tests, lipid profile, atherosclerosis markers, and diabetes screening tests). Notably, clinical risk factors like hypertension, diabetes, dyslipidemia, and smoking were not included as direct input features, as their effects were assumed to be captured by the laboratory parameters.
Key Insights and Ideas
- Beyond Imaging: The study highlights that valuable predictive insights for CACS can be extracted from routine clinical and laboratory data, potentially reducing reliance on more resource-intensive imaging techniques for initial risk stratification.
- ML Superiority: Advanced ML algorithms like XGBoost demonstrate a significant advantage over traditional BLR in predicting high CACS, suggesting that complex, non-linear relationships within the data are better captured by these methods.
- Feature Importance: Age, systolic blood pressure, and LDL cholesterol emerged as consistently important predictors across all ML models, aligning with established CVD risk factors. However, the study also revealed that ML models can identify other potentially overlooked features.
- The "Healthy" Population Nuance: The finding that high CACS individuals had lower total and LDL cholesterol than low CACS individuals is counterintuitive and warrants further investigation, possibly related to statin use or other primary prevention strategies in this healthy cohort.
Key Results
- Predictors: In multivariable BLR, age (OR 1.12), male sex (OR 2.93), systolic blood pressure (OR 1.02), and LDL cholesterol (OR 1.00) were significant predictors of high CACS.
- Performance:
- XGBoost achieved the highest AUROC of 0.823.
- CatBoost followed with an AUROC of 0.750.
- Binary Logistic Regression (BLR) had an AUROC of 0.585.
- The performance difference between XGBoost and BLR was statistically significant (p < 0.001).
Limitations and Future Work
The authors acknowledge several limitations:
- Drug Usage Data: The study did not include information on drug usage (e.g., statins) or regular exercise. This is a significant limitation, especially given the counterintuitive finding regarding LDL cholesterol, and suggests that future work should incorporate this data.
- Generalizability: The study was conducted on a Korean population, and its findings may not be directly generalizable to other ethnic groups.
- Selection Bias: Participants were recruited from tertiary hospitals, potentially introducing selection bias, even though the focus was on healthy individuals.
Future work could focus on:
- Incorporating drug history and lifestyle factors into ML models.
- Validating these ML models in diverse ethnic populations.
- Exploring the interpretability of ML models to uncover novel biological pathways or risk indicators.
- Investigating the potential for these ML models to be integrated into routine health screening workflows.
Summary updated