Interpretable Machine Learning of Non-Traditional Lipid Indices for Diagnostic Classification of CHD in Patients with Comorbid MASLD and T2DM: A Multicenter Study

Why It Matters

Integrating non‑traditional lipid markers into an interpretable machine‑learning tool improves early CHD detection for a high‑risk MASLD‑T2DM population, potentially guiding preventive therapies and reducing cardiovascular events.

Key Takeaways

• CRI-II linked to 2.4‑fold higher CHD odds.
• All eight non‑traditional lipids significantly predict CHD risk.
• Stacking model achieved AUC ~0.75 in validation.
• SHAP identified CRI‑II, eGFR, age, LCI as top predictors.
• Non‑linear relationships found for RC, AIP, LCI.

Pulse Analysis

The convergence of metabolic dysfunction‑associated steatotic liver disease (MASLD) and type 2 diabetes mellitus (T2DM) creates a perfect storm for coronary heart disease (CHD). Traditional lipid panels—total cholesterol, LDL‑C, HDL‑C—often underestimate the residual cardiovascular risk in this cohort, prompting researchers to explore alternative lipid metrics that capture nuanced atherogenic pathways. Non‑traditional indices such as the Castelli risk index‑II, remnant cholesterol, and the atherogenic index of plasma reflect particle composition and metabolic stress more accurately, offering a richer risk‑stratification landscape for clinicians.

In a robust multicenter retrospective study, 1,823 MASLD‑T2DM patients were examined, with 630 propensity‑matched subjects used for association analysis and the full cohort employed to train six machine‑learning algorithms. All eight evaluated lipid indices showed significant ties to CHD, but CRI‑II emerged as the most potent predictor, doubling the odds of disease. The final stacking ensemble, blending logistic regression, random forest, and gradient boosting, incorporated ten variables—including eGFR and age—and delivered balanced discrimination (AUC ≈ 0.75) across internal and external validation sets. Transparency tools like SHAP and LIME confirmed that CRI‑II, estimated glomerular filtration rate, age, and the lipoprotein combined index drove model decisions, reinforcing clinical relevance.

The study’s blend of interpretable AI and novel lipid biomarkers signals a shift toward precision cardiometabolic care. By delivering a transparent, validated risk model, clinicians can identify high‑risk patients earlier, tailor lipid‑lowering strategies, and potentially curb the progression of coronary lesions. Future work should test the framework prospectively, integrate it into electronic health records, and explore therapeutic interventions targeting the highlighted lipid pathways, thereby translating statistical insight into tangible health outcomes.