Introduction
Coronary Heart Disease (CHD) remains a leading cause of global mortality. Its physiological roots often develop decades before clinical symptoms manifest. Identifying high-risk individuals during early adulthood (ages 18–30) offers a critical window for proactive, personalized medical intervention. However, analyzing longitudinal health data presents significant challenges for traditional statistical methods, which often struggle with the complex, non-linear interactions inherent in temporal measurements like blood pressure or cholesterol levels.
Methodology
To address these challenges, this research utilizes Statistical Relational Learning (SRL), an AI subfield that integrates logic for handling complex data structures with probability for managing uncertainty. The study evaluates two primary algorithms:
- Relational Probability Trees (RPT): An extension of classical decision trees capable of representing dependencies across different objects and time points.
- Relational Functional Gradient Boosting (RFGB): An advanced ensemble method that learns a sequence of relational regression trees to minimize prediction error.
A key innovation is the use of Temporal Modeling, treating features as "fluents" (e.g., weight(Person, Value, Time)). This allows the models to capture how factors at different time points interact. Additionally, Expert-Informed Discretization was applied, using medical domain knowledge to categorize continuous variables into clinically meaningful bins.
Dataset: The CARDIA Study
The models were trained and validated using data from the Coronary Artery Risk Development in Young Adults (CARDIA) study. This longitudinal dataset follows 5,115 participants from 1985–86 (ages 18–30) through 2005–06 (ages 38–50). Features include clinical measurements (BMI, cholesterol, blood pressure), lifestyle factors (exercise, smoking), and socio-economic indicators. The target variable is the Coronary Artery Calcification (CAC) level at Year 20, a subclinical marker of arterial plaque buildup.
Results and Analysis
The Predictive Power of Early Adulthood
A significant finding is the "Youth Signal": risk factors measured between ages 18 and 30 are nearly as predictive of middle-age heart health as the full 20-year history. This confirms that early adulthood is a vital period for cardiovascular risk assessment.
Model Performance Comparison
The SRL approaches significantly outperformed traditional machine learning models such as SVMs and Logistic Regression, which struggled with class imbalance and temporal complexity. RFGB achieved the highest performance with an AUC-ROC of 0.819, followed by RPT at 0.778. In contrast, standard models often failed to accurately identify the high-risk minority.
The models also revealed complex interactions, such as gender-specific differences in how smoking history interacts with cholesterol levels to influence future CAC risk. Preliminary analysis further suggested that socio-economic factors, like home ownership, are strong predictors of physiological outcomes.
Discussion and Future Work
While RFGB provides superior accuracy, its ensemble nature makes it less interpretable than a single decision tree. Future research will focus on consolidating these models into more interpretable formats for clinical use. Additionally, as the CARDIA cohort ages, the authors plan to extend these models to predict direct clinical events, such as acute myocardial infarction and heart failure, and incorporate a broader range of dietary and psychosocial variables.
Summary updated