Visceral obesity anthropometric indicators as predictors of acute pancreatitis severity

Abstract

Background: Acute pancreatitis (AP) severity assessment upon admission is crucial for prognosis, yet existing clinical scoring systems have limitations like delayed results, complexity, or low sensitivity. Obesity correlates with AP severity, but traditional body mass index (BMI) fails to accurately reflect visceral fat distribution. Although anthropometric indicators for visceral obesity offer alternatives, their predictive value for AP severity across all etiologies is poorly studied.

Methods: This retrospective cohort study analyzed 629 AP patients admitted to a tertiary hospital (2016-2023). Patients were classified as mild AP (MAP, n = 531) or moderately severe/severe AP (MSAP/SAP, n = 98) based on organ failure (modified Marshall score ≥ 2). Eleven anthropometric indicators and six clinical scoring systems were evaluated. Patients were randomly divided into training group (n = 441) and validation group (n = 188). LASSO regression identified key predictors from 37 clinical variables. Six machine learning (ML) models were built and evaluated using receiver operating characteristic (ROC) analysis, area under the ROC curve (AUC), calibration curves, and decision curve analysis (DCA).

Results: Nine anthropometric indicators [waist circumference, body roundness index, BMI, conicity index, lipid accumulation products (LAP), waist triglyceride index (WTI), cardiometabolic index (CMI), visceral adiposity index (VAI), chinese visceral adiposity index] and all clinical scoring systems (Ranson score, Glasgow score, SIRS, BISAP, APACHE II, JSS) significantly differed between MAP and MSAP/SAP groups (p < 0.05). VAI demonstrated the highest predictive AUC among anthropometric indicators (0.737 vs. SIRS 0.750, JSS 0.815), but superior to Ranson score, Glasgow score, BISAP, and APACHE II. LAP, WTI, and CMI also showed strong AUCs (0.729, 0.722, 0.736 respectively). LASSO selected 15 variables. Among ML models, XGBoost model performed best on the validation group (AUC = 0.878), and relatively good calibration curve and DCA results.

Conclusion: VAI, CMI, LAP, and WTI are independent predictors of AP severity, with VAI showing the highest individual predictive capability among them. The XGBoost model, incorporating VAI and routinely available clinical variables, achieved excellent performance (AUC = 0.878) for early severity assessment, offering a potentially rapid and cost-effective clinical tool. This supports the utility of visceral obesity anthropometric indicators and ML models for improving early risk stratification in AP.

Keywords: acute pancreatitis; anthropometric indicator; machine learning algorithm; organ failure; severity; visceral adiposity index.

Figure 1

Flowchart of patient inclusion and exclusion.

Figure 2

Correlation heatmap of anthropometric indicators with visceral obesity. WC, Waist circumference; BRI, Body roundness index; BMI, Body mass index; ABSI, A body shape index; TG, triglyceride; WWI, Weight-adjusted waist index; LAP, Lipid accumulation product; WTI, Waist triglyceride index; CMI, Cardiometabolic index; VAI, Visceral adiposity index; CVAI, Chinese visceral adiposity index.

Figure 3

Correlation analysis of anthropometric indicators with visceral obesity. LAP, Lipid accumulation product; WTI, Waist triglyceride index; CMI, Cardiometabolic index; VAI, Visceral adiposity index; CVAI, Chinese visceral adiposity index; BRI, Body roundness index; BMI, Body mass index; ABSI, A body shape index; WWI, Weight-adjusted waist index.

Figure 4

ROC curve analysis of anthropometric indicators (A) and clinical scoring systems (B). ROC, receiver operating characteristic; AUC, area under the curve; WC, Waist circumference; BRI, Body roundness index; BMI, Body mass index; CI, Conicity index; LAP, Lipid accumulation product; WTI, Waist triglyceride index; CMI, Cardiometabolic index; VAI, Visceral adiposity index; CVAI, Chinese visceral adiposity index; SIRS, Systemic inflammatory response syndrome; BISAP, Bedside index of severity in acute pancreatitis; APACHE II, Acute physiology and chronic health evaluation II; JSS, Japanese severity score.

Figure 5

Selection of clinical variables via the LASSO regression method. (A) The coefficient values for 37 variables are shown in relation to log(λ); (B) The left vertical dashed line demarcates the position where the minimum cross-validation error emerges, while the right vertical dashed line specifies the minimum error plus one standard deviation. LASSO, Least Absolute Shrinkage and Selection Operator.

Figure 6

Performance comparison of ML models. ROC curves on the training group (A) and validation group (B). Calibration plots on the training group (C) and validation group (D). Decision curves on the training group (E) and validation group (F). ROC, receiver operating characteristic.

Figure 7

Variable Importance Plots in the XGB Model. (A) The Gain Metric; (B) The Cover Metri; (C) The Frequency Metric; (D) Shapley Additive exPlanations visualization.