Explainable Machine Learning Framework for Predicting Cardiometabolic Risk Using Meal Timing and Eating Habits

Technical Summary: Explainable Machine Learning Framework for Predicting Cardiometabolic Risk Using Meal Timing and Eating Habits

Problem Statement
Traditional cardiometabolic risk assessment strategies predominantly rely on static anthropometric and clinical indicators (e.g., BMI, blood pressure) and dietary content, often neglecting the temporal dimension of eating behavior. While emerging chrono-nutrition research suggests that meal timing and regularity significantly influence metabolic health—specifically through circadian rhythm alignment—these factors are rarely integrated into predictive models. Furthermore, existing statistical techniques often fail to capture the complex, nonlinear relationships between behavioral patterns and metabolic outcomes. Additionally, the "black-box" nature of advanced machine learning (ML) models limits their clinical utility, as healthcare providers require interpretable insights to guide interventions.

Methodology
This cross-sectional study analyzed data from 300 adults, integrating demographic, anthropometric (BMI, Waist-Hip Ratio), clinical (blood pressure), and chrono-nutritional data (meal timing, frequency, regularity, and behavioral awareness).

• Data Preprocessing & Feature Engineering: Raw data was cleaned, with missing values imputed using mean or mode. Categorical variables were encoded, and ordinal variables were numerically sorted. Novel features were engineered to quantify behavioral variability, including a Circadian Misalignment Score, a Meal Timing Variability Index, and a Diet Quality Index. Continuous predictors underwent Z-score normalization and log-transformation for skewed distributions.
• Model Development: Three supervised learning algorithms were developed and compared: Logistic Regression (LR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost).
  • LR: Tested with L1, L2, and elastic-net regularization (optimal $C=1.0$).
  • RF: Configured with 500 trees, max depth of 20, and specific splitting criteria to balance bias/variance.
  • XGBoost: Fine-tuned via grid search with a learning rate of 0.05, max depth of 6, subsampling ratio of 0.8, and regularization parameters $\alpha=0.1, \lambda=1$.
• Validation & Interpretability: Models were evaluated using nested 10-fold stratified cross-validation. Performance metrics included AUC-ROC, accuracy, precision, recall, F1-score, and Brier score for calibration. To ensure interpretability, Shapley Additive Explanations (SHAP) were applied to the best-performing model to quantify feature contributions and identify actionable behavioral drivers.

Key Results

• Model Performance: The XGBoost model demonstrated superior performance, achieving an AUC-ROC of 0.98, 93% accuracy, 92% precision, 91% recall, and an F1-score of 0.92. It also exhibited the best calibration with a Brier score of 0.03. Random Forest followed with an AUC of 0.94 and 89% accuracy, while Logistic Regression achieved an AUC of 0.92 and 87% accuracy.
• Predictors of Risk:
  • Physiological: Waist-Hip Ratio (WHR) and BMI remained the most significant physiological predictors (ORs of 4.39 and 2.74, respectively).
  • Behavioral/Chrono-nutritional: SHAP analysis identified later dinner timing (OR = 2.5), irregular meal timing (OR = 1.6), and reduced adherence to recommended meal frequency (OR = 1.8) as top independent predictors of cardiometabolic risk.
  • Protective Factors: Increased meal frequency and heightened awareness of the significance of meal timing were associated with reduced risk.
• Interactions: The study found a compounding effect where individuals with higher WHR or BMI were more negatively impacted by irregular meal patterns and late-night eating, suggesting that circadian misalignment exacerbates metabolic vulnerability.

Key Contributions

1. Integration of Chrono-nutrition: The study successfully incorporates meal timing and regularity into a predictive ML framework, demonstrating that these temporal factors provide significant predictive value beyond traditional anthropometric measures.
2. Explainable AI Framework: By combining high-performance ensemble learning (XGBoost) with SHAP, the study bridges the gap between predictive accuracy and clinical interpretability. It translates complex model outputs into specific, modifiable behavioral targets (e.g., avoiding late dinners, maintaining regularity).
3. Identification of Awareness as a Factor: The research highlights "meal-timing awareness" as a distinct, protective behavioral factor, independent of body size or food quality, suggesting a cognitive component to metabolic regulation.

Significance and Claims
The paper claims that integrating meal timing and behavioral regularity into explainable machine learning models significantly improves the accuracy of cardiometabolic risk prediction. The authors assert that the XGBoost-SHAP framework not only offers high predictive performance but also elucidates the interactions between body composition and circadian eating rhythms.

Crucially, the study posits that meal-timing awareness is a low-cost, scalable, and modifiable target for public health interventions. The findings suggest that synchronizing food consumption with the body's natural physiological cycles—specifically by avoiding late dinners and maintaining irregular schedules—may serve as an effective strategy for the long-term prevention of cardiometabolic disease, even in individuals with normal BMI. The authors emphasize that these behavioral interventions can mitigate metabolic impacts in modern, erratic lifestyles without necessarily requiring severe dietary restrictions.

Note: As this is a preprint, the authors explicitly state that the research has not been certified by peer review and should not be used to guide clinical practice.