Explainable Machine Learning for Cardiometabolic Multimorbidity: Prediction, Phenotyping, and Identification of High-Risk Individuals: A Cross- Sectional Study

Technical Summary: Explainable Machine Learning for Cardiometabolic Multimorbidity

Problem Statement
Cardiometabolic multimorbidity (CMM), defined as the coexistence of at least two conditions among diabetes mellitus, hypertension, and heart disease, represents a growing public health challenge with significantly higher risks of mortality and disability than single conditions. While machine learning (ML) has shown promise in predicting individual cardiometabolic diseases, existing tools often lack the ability to predict their co-occurrence with high precision or provide the interpretability required for clinical decision-making. Furthermore, current approaches frequently overlook the characterization of specific multimorbidity phenotypes and the potential "latent" high-risk status of individuals who may not yet meet formal diagnostic criteria but exhibit intermediate clinical profiles.

Methodology
The study utilized a large-scale, integrated dataset of 278,701 individuals from the "Multi-Class Chronic Disease Data Warehouse" (Mendeley Data). The data was stratified into training (51%), validation (17%), test (17%), and independent holdout (15%) subsets, maintaining a consistent CMM prevalence of 1.5% across all partitions.

• Model Development: A CatBoost gradient boosting classifier was employed to predict CMM. The model was trained on 25 predictors, including demographics, lifestyle factors, and laboratory biomarkers. To address the severe class imbalance (1.5% positive cases), a class weight of 45 was assigned to the multimorbidity class. The model utilized 500 iterations, a learning rate of 0.03, and Logloss as the loss function, with Precision selected as the primary evaluation metric.
• Interpretability: To ensure clinical transparency, the study integrated SHAP (SHapley Additive Explanations) using TreeExplainer. This provided both global feature importance rankings and local explanations for individual predictions.
• Phenotyping and Error Analysis: The study categorized confirmed CMM cases into four mutually exclusive phenotypes (DM+HTN, DM+HD, HTN+HD, and DM+HTN+HD) to analyze metabolic heterogeneity. Crucially, the authors conducted a systematic error analysis on "false-positive" predictions from the holdout set to investigate whether these individuals represented a latent high-risk population with intermediate clinical profiles.

Key Results

• Predictive Performance: The CatBoost model demonstrated excellent discriminative ability with an Area Under the Curve (AUC) of 0.985 on the independent holdout dataset. By optimizing the decision threshold to 0.90, the model achieved a high precision of 0.727, meaning nearly 73% of positive predictions were correct. However, this optimization resulted in a lower recall of 0.192, identifying 120 out of 626 true CMM cases.
• Feature Importance: SHAP analysis identified Age as the most influential predictor, followed by blood pressure, CRP levels, LDL, glucose, gender, and HDL. Glycemic control markers (HbA1c, glucose), lipid components, and inflammatory markers were confirmed as primary drivers of the model's predictions.
• Phenotypic Heterogeneity: Analysis revealed substantial variation across CMM subtypes. The triple-comorbidity group (DM+HTN+HD) exhibited the most adverse metabolic profiles, including the highest BMI and triglycerides. Conversely, the HTN+HD group showed the highest median age but lower glycemic markers compared to diabetes-inclusive groups.
• Latent High-Risk Identification: Individuals misclassified as false positives (predicted as CMM but not meeting criteria) displayed clinical profiles intermediate between true negatives and true positives. Notably, they had higher median ages (65.0 vs. 49.0 in true negatives), elevated HbA1c (6.8% vs. 5.7%), and higher BMI (31.7 vs. 27.3), suggesting they constitute a "latent" high-risk population that may benefit from early intervention.

Key Contributions

1. High-Precision Prediction Framework: The study presents a robust, explainable ML framework capable of predicting CMM with an AUC of 0.985 and high precision (0.727), addressing the need for reliable screening tools in primary care.
2. Integration of Explainability: By leveraging SHAP, the study moves beyond "black box" predictions to provide transparent, quantifiable insights into how specific risk factors (e.g., age, CRP, LDL) contribute to individual risk scores, facilitating clinical trust and adoption.
3. Phenotype Characterization: The research delineates distinct metabolic and clinical profiles across different CMM combinations, highlighting the necessity for phenotype-specific risk stratification rather than a one-size-fits-all approach.
4. Identification of Latent Risk: A novel contribution is the characterization of false-positive cases not merely as errors, but as a potential transitional subgroup with intermediate risk profiles, offering a new target for preventive strategies.

Significance and Claims
The authors claim that this framework offers a "clinically actionable tool" for early detection and targeted prevention, particularly in primary care and population health settings. The study emphasizes that by combining high predictive performance with phenotype characterization and the identification of latent high-risk individuals, the model addresses critical gaps in current risk stratification. The authors assert that the integration of explainable AI (XAI) allows for the translation of complex, nonlinear interactions between demographic, lifestyle, and clinical factors into transparent insights suitable for clinical decision-making. The study concludes that while the model achieves high precision, the identification of the "latent" high-risk group among false positives suggests a valuable opportunity for early intervention before formal multimorbidity criteria are met.