Unmasking cardiovascular risk beyond BMI in type 2 diabetes and MASLD: an interpretable multiorgan score for lower-BMI patients

Technical Summary: Unmasking Cardiovascular Risk Beyond BMI in Type 2 Diabetes and MASLD

Problem Statement
Cardiovascular disease (CVD) remains the leading cause of global mortality, with its burden increasingly linked to the overlapping conditions of obesity, type 2 diabetes mellitus (T2DM), and metabolic dysfunction-associated steatotic liver disease (MASLD). While Body Mass Index (BMI) is a dominant marker for metabolic severity in clinical practice, it fails to capture ectopic fat distribution, visceral adiposity, skeletal muscle reserve, and organ-specific injury. This limitation is particularly critical in Asian populations and patients with T2DM-MASLD, where a "lower-BMI" phenotype (non-obese or lean) may still harbor significant hepatic fibrosis, vascular injury, and systemic vulnerability. Current risk stratification tools often rely on BMI categories, potentially underestimating CVD risk in lower-BMI patients who possess a complex cardiometabolic profile. There is a need for an interpretable, multiorgan risk assessment tool that moves beyond body-size classification to identify high-risk individuals within this specific phenotype.

Methodology
This study employed a retrospective, single-center, hospital-based cohort design using data from Gansu Provincial Hospital, China.

• Cohort: The analytic cohort consisted of 1,466 adults with T2DM, documented MASLD, and a BMI < 27.0 kg/m². Patients with baseline documented CVD were excluded to focus on risk stratification for future or undiagnosed events.
• Data Splitting: The cohort was split 7:3 into a training cohort (n=1,026) and a locked internal-validation cohort (n=440) using stratification by CVD status and BMI stratum.
• Feature Engineering: Candidate predictors were organized into three pools. Pool A focused on routinely available variables for a parsimonious model. Pool B and C included extended metabolic, renal, and interaction terms for comparison.
• Model Development:
  • Primary Model (A2): A ridge-logistic regression model was developed using Pool A predictors. It was converted into an additive clinical score (nomogram-style) to ensure interpretability.
  • Comparators: The study compared A2 against standard logistic regression (A1), extended linear models (B1, B2), and machine-learning models (C1: ExtraTrees; C2: XGBoost) using Pool C.
• Validation Protocol: A strict "locked" framework was used. All preprocessing, feature screening, hyperparameter tuning, and threshold derivation occurred exclusively in the training cohort. The validation cohort was used only for final performance evaluation without any model refitting.
• Outcome: The primary outcome was binary documented CVD status (coronary artery disease, heart failure, atrial fibrillation, stroke, aortic disease, or peripheral arterial disease) by an administrative study cut-off.
• Metrics: Performance was assessed via discrimination (AUC), precision-recall (average precision), prediction error (Brier score), calibration (intercept and slope), and decision-curve analysis (net benefit).

Key Contributions

1. Development of the A2 Score: The study developed and internally validated an interpretable multiorgan risk score (A2) comprising eight routinely available predictors: diabetes duration, carotid plaque number, mean carotid intima–media thickness (IMT), age, FIB-4 index, estimated glomerular filtration rate (eGFR), LDL cholesterol, and albumin.
2. Multiorgan Risk Framework: The model integrates signals from cumulative diabetes exposure, subclinical atherosclerosis, hepatic fibrosis, kidney function, lipid burden, and systemic vulnerability, providing a holistic view of CVD risk that BMI alone cannot offer.
3. Validation in Lower-BMI Phenotypes: The study specifically demonstrated that the A2 score retains predictive performance and separates risk gradients even in patients with BMI < 25 kg/m², a group often considered low-risk based on body mass alone.
4. Interpretability vs. Complexity: The study showed that a parsimonious, interpretable ridge-logistic model (A2) achieved comparable or superior overall performance (balancing discrimination, calibration, and Brier score) compared to more complex machine-learning algorithms, supporting the utility of transparent clinical tools.

Results

• Population Characteristics: Among 1,466 patients, 240 (16.4%) had documented CVD. Baseline analysis revealed that cardiovascular event burden did not follow a simple lower-BMI/lower-risk gradient; event rates were highest in the BMI 23–<25 kg/m² stratum (18.1%) and lowest in the 25–<27 kg/m² stratum (13.7%).
• Predictor Selection: The final A2 model retained eight variables. Diabetes duration showed the strongest association (OR 2.12 per SD in the standard logistic model), followed by carotid plaque number, mean carotid IMT, and age.
• Model Performance (Validation Cohort):
  • A2 (Ridge Logistic): Achieved an AUC of 0.792 (95% CI 0.741–0.844), average precision of 0.420, and a Brier score of 0.116. Calibration was reasonable (intercept −0.307, slope 0.849).
  • Comparators: Machine-learning models (C1, C2) did not demonstrate a clear incremental advantage over A2. C1 (ExtraTrees) had a slightly higher average precision (0.438) but lower AUC (0.780) and worse calibration. A2 was selected as the primary model for its balance of performance, parsimony, and convertibility to a clinical score.
• Risk Stratification: Using training-derived tertiles, the A2 score successfully separated CVD burden in the validation cohort:
  • Low-risk stratum: 2.9% event rate.
  • Intermediate-risk stratum: 12.0% event rate.
  • High-risk stratum: 33.3% event rate.
• Subgroup Analysis: In the BMI < 25 kg/m² subgroup (n=277 in validation), A2 maintained robust discrimination (AUC 0.812) and clear risk separation (event rates: 1.3%, 10.8%, and 36.8% for low, intermediate, and high risk, respectively).

Significance and Claims
The paper claims that cardiovascular risk in lower-BMI patients with T2DM and MASLD is frequently underestimated when relying solely on BMI categories. The study posits that a "lean" phenotype does not preclude significant hepatic fibrosis, vascular injury, or systemic vulnerability. The A2 score offers a practical, interpretable tool to refine risk assessment by integrating multiorgan signals (liver, kidney, vasculature, and metabolic history) that are routinely available in hospital settings.

The authors emphasize that the study provides an internally validated framework. They explicitly state that the findings require external and prospective validation before clinical implementation. The significance lies not in the introduction of a novel algorithm, but in the demonstration that a parsimonious, multiorgan approach can effectively stratify risk in a clinically under-recognized population, potentially supporting earlier cardiovascular evaluation beyond traditional obesity definitions. The study concludes that BMI should not be interpreted as a marker of low cardiovascular risk in T2DM-MASLD patients and advocates for a shift toward an integrated cardiovascular–kidney–metabolic–hepatic assessment framework.