An Interpretable Machine Learning Model for Predicting in-Hospital Progression in Initially Mild Hypertriglyceridemia-Induced Acute Pancreatitis Using Clinical and Non-Contrast CT Features

Abstract

Introduction

Early identification of patients with initially mild hypertriglyceridemia-induced acute pancreatitis (HTG-AP) who are at risk of an unstable disease course remains challenging using conventional assessment alone. This study aimed to develop an interpretable machine learning model based on baseline clinical and non-contrast computed tomography (CT) features for early prediction of in-hospital progression.

Methods

We retrospectively enrolled 164 patients with initially mild HTG-AP between October 2020 and October 2024. Baseline clinical variables, CT-based body composition parameters, and pancreatic radiomics features from non-contrast CT were collected at admission. Patients were classified into progression (n = 88) and non-progression (n = 76) groups based on clinically relevant worsening supported by clinical or CT evidence during hospitalization. Five-fold cross-validation was used for model development and internal validation. Four XGBoost-based models were constructed: clinical only, clinical + body composition, clinical + radiomics, and clinical + body composition + radiomics. Model performance was assessed using receiver operating characteristic analysis, calibration analysis, decision curve analysis, and Shapley additive explanations (SHAP).

Results

The clinical + body composition + radiomics model achieved the best overall performance among the four models, with an AUC of 0.830 and an accuracy of 0.768. Calibration analysis showed relatively good agreement between predicted and observed risks, and decision curve analysis demonstrated a higher net benefit across most clinically relevant threshold probabilities. SHAP analysis identified triglycerides (TG) and the visceral fat area-to-abdominal cavity area ratio (VFA/ACA) as the dominant contributors to model prediction.

Conclusion

The proposed interpretable multimodal model may improve early risk stratification for in-hospital progression in initially mild HTG-AP and help identify patients who require closer monitoring, although further external validation is needed to confirm its generalizability and clinical applicability.

Introduction

With the increasing prevalence of hypertriglyceridemia and related metabolic abnormalities, hypertriglyceridemia-induced acute pancreatitis (HTG-AP) has become the second most common etiology of acute pancreatitis in China, after biliary pancreatitis.¹ Compared with other common etiologies, HTG-AP is associated with higher risks of systemic inflammatory response syndrome, acute peripancreatic fluid collection, and acute necrotic collection.² However, the severity of acute pancreatitis is often difficult to assess reliably during the early disease course.³ Some patients initially classified as having mild disease may still experience persistent or worsening symptoms or show evolving pancreatic and peripancreatic findings during hospitalization.^3,4 Although these changes do not necessarily indicate an escalation in the Atlanta severity classification, they may suggest that the disease course has not yet stabilized and may precede the subsequent development of objective local or systemic complications.^3–5 Therefore, early identification of patients at risk of an unstable disease course may help guide close monitoring and timely reassessment in initially mild HTG-AP.

Currently, early risk stratification in acute pancreatitis relies mainly on clinical scoring systems and conventional imaging assessment. Bedside Index for Severity in Acute Pancreatitis (BISAP), Acute Physiology and Chronic Health Evaluation II (APACHE II), and Ranson scores are commonly used to estimate the risk of severe disease, but they mainly reflect systemic inflammation, physiological derangement, or organ dysfunction, and some require variables obtained over time, which may limit their ability to identify subtle early risk differences in patients initially classified as having mild disease.^6,7 The modified Pancreatitis Activity Scoring System (mPASS) can quantify disease activity, but it requires repeated assessments over the disease course.⁸ CT-based scores, including the Computed Tomography Severity Index (CTSI) and Extrapancreatic Inflammation on CT (EPIC) score, provide useful morphologic information; however, their early application may be limited because pancreatic necrosis and extrapancreatic inflammatory changes may not yet be fully apparent in the early phase of the disease.^9,10 Therefore, existing clinical and imaging scores may be insufficient for identifying subtle early heterogeneity and subsequent progression risk among patients with initially mild HTG-AP.

For patients with HTG-AP who underwent early non-contrast CT for diagnostic or differential diagnostic purposes, the acquired images can provide not only conventional morphologic information but also additional quantitative information through further image analysis. Abdominal body composition analysis and pancreatic radiomics can further quantify visceral fat distribution, metabolic phenotype, and subtle pancreatic heterogeneity that may not be fully captured by routine clinical or morphologic assessment.^11,12 Therefore, combining these imaging-derived quantitative features with clinical variables may provide a more comprehensive basis for early risk stratification.

Artificial intelligence provides a methodological framework for integrating clinical variables with quantitative imaging features.^13–15 Explainable artificial intelligence methods, such as SHapley Additive exPlanations (SHAP), further enhance this framework by quantifying feature contributions and improving model transparency.¹⁶ Recent HTG-AP studies have applied this integrative approach in prognostic modeling.^13,14,17 A clinical–CT random forest model showed good discriminatory performance for the early prediction of organ failure, whereas a model combining CT attenuation parameters with hematological markers outperformed the Ranson score in identifying severe HTG-AP.^13,14 More recently, MRI radiomics combined with clinical variables showed better performance than either single-modality models or conventional clinical scores for predicting HTG-AP severity.¹⁸ Nevertheless, previous studies have mainly focused on final severity classification or organ failure, and the value of multimodal models for predicting in-hospital disease progression among patients initially classified as having mild HTG-AP has not been sufficiently investigated.

Based on these considerations, this study integrated early clinical variables, non-contrast CT body composition parameters, and pancreatic radiomics features to develop a machine-learning model for predicting in-hospital disease progression in patients with initially mild HTG-AP. This approach was intended to support risk assessment and early identification of patients with a potentially unstable disease course, thereby helping guide close monitoring and timely reassessment during hospitalization.

Materials and Methods

Study Design and Population

This single-center retrospective cohort study was conducted at The First Hospital of Changsha to develop and internally validate an early risk stratification model for in-hospital disease progression. Consecutive patients with HTG-AP admitted between October 2020 and October 2024 were retrospectively screened, and their in-hospital course from baseline assessment to discharge was reviewed to determine whether in-hospital progression occurred.

Eligibility Criteria

Inclusion criteria were: (1) age ≥18 years, complete clinical and laboratory records, available in-hospital outcome data, and non-contrast abdominal CT performed within 24 hours before or after admission; (2) diagnosis of HTG-AP; and (3) classification as mild acute pancreatitis at baseline according to the 2012 Revised Atlanta Classification, based on clinical information obtained within 24 hours after admission and baseline non-contrast CT findings.³

HTG-AP was diagnosed when patients met at least two standard criteria for acute pancreatitis, including characteristic abdominal pain, serum amylase or lipase levels at least three times the upper limit of normal, and imaging findings consistent with acute pancreatitis, together with an admission triglyceride level ≥11.3 mmol/L after exclusion of other major etiologies, including biliary, alcoholic, drug-induced, and traumatic causes.¹⁹

Exclusion criteria were: local complications, systemic complications, or organ failure already present at baseline; chronic pancreatitis; previous pancreatic surgery; malignancy; other severe comorbidities; or imaging quality insufficient for quantitative analysis.

Outcome Definition

In-hospital progression was defined as clinically relevant worsening from an initially mild baseline status during hospitalization, based on an integrated assessment of clinical course, laboratory or systemic inflammatory changes, and follow-up imaging findings. Patients were assigned to the progression group when clinical reassessment was prompted by persistent or worsening symptoms or inflammatory response and was supported by objective evidence of disease worsening. Objective evidence included increased pancreatic or peripancreatic inflammatory changes or fluid collections on follow-up CT, new or increased ascites or pleural effusion, new local or systemic complications, or new organ dysfunction. Patients with stable or improved clinical course and imaging findings during hospitalization were assigned to the non-progression group.

Patient Management

All patients were initially treated in the general ward by the gastroenterology or gastrointestinal surgery team and received similar initial management according to institutional treatment principles for HTG-AP, including fasting, fluid resuscitation, analgesia, nutritional support, triglyceride-lowering therapy, and complication monitoring. Subsequent care was adjusted according to clinical evolution, including reassessment, follow-up CT when clinically indicated, and ICU transfer if necessary.

Clinical and Laboratory Data

Demographic variables and clinical history included age, sex, smoking history, alcohol history, and diabetes history. Laboratory variables included serum amylase and lipase, complete blood count, blood lipids, glucose, liver and renal function tests, calcium (Ca), procalcitonin (PCT), C-reactive protein (CRP), and baseline arterial blood gas results. These variables, together with time to follow-up CT and length of hospital stay, were retrieved from electronic medical records and laboratory information systems. Baseline clinical and laboratory variables were defined as the first measurements obtained in the emergency department or upon admission. The above laboratory tests were repeated during hospitalization when clinically indicated. BISAP and admission-based mPASS were calculated from baseline data according to their standard definitions and used as conventional comparators.

Conventional CT Assessment, Body Composition Analysis, and Radiomics Feature Extraction

All patients underwent baseline non-contrast abdominal CT within 24 hours before or after admission using a GE 64- or 256-slice CT scanner, and follow-up CT examinations used the same scanning protocol. The main scanning parameters were as follows: tube voltage, 120 kV; automatic tube current modulation; pitch, 0.984–1.375; rotation time, 0.5–0.75 s; and slice thickness, 1 mm.

CT body composition parameters were measured at the L3/L4 vertebral level using 3D Slicer, including subcutaneous fat area (SFA), visceral fat area (VFA), total fat area (TFA), abdominal cavity area (ACA), and mean attenuation values of the liver, spleen, and psoas major muscles. Two derived indices, VFA/TFA and VFA/ACA, were calculated to characterize relative visceral fat distribution (Figure 1A–D). In addition, we collected the required extrapancreatic inflammatory findings and calculated EPIC as a reference imaging-based comparator for exploratory analysis.

Figure 1.

Quantification of abdominal fat distribution and pancreatic segmentation. (A–D) CT-based body composition measurements at the L3/L4 level. (A) Visceral fat area (VFA), shown in green. (B) Subcutaneous fat area (SFA), shown in yellow. (C) Total fat area (TFA), comprising the green-highlighted visceral fat and yellow-highlighted subcutaneous fat. (D) Abdominal cavity area (ACA), shown in blue. (E–G) Pancreatic regions of interest, shown in green on axial, sagittal, and coronal CT images, respectively, generated using TotalSegmentator and reviewed and manually corrected by a radiologist when necessary. (H) Three-dimensional reconstruction of the segmented pancreas, shown in green.

Pancreatic regions of interest were automatically generated using TotalSegmentator²⁰ and then reviewed and manually corrected when necessary by a radiologist with more than 10 years of abdominal imaging experience to ensure anatomical accuracy (Figure 1E–H). After a one-month interval, the same radiologist repeated the review and correction procedure for all patients. Radiomics features were extracted from both segmentation sets using PyRadiomics, including shape, first-order, and texture features from original and wavelet-filtered images. Intraobserver reproducibility was assessed using a two-way random-effects, absolute-agreement, single-measurement intraclass correlation coefficient [ICC (2,1)], with ICC > 0.75 indicating acceptable reproducibility. Finally, 827 of 851 features were retained for subsequent feature selection.

Feature Selection and Model Development

A five-fold cross-validation strategy was used for internal validation. Within each fold, preprocessing, feature selection, hyperparameter optimization, and model training were performed using only the training subset, and the held-out fold was used for validation. Clinical variables and CT-based body composition parameters were screened by univariate analysis. After ICC filtering, pancreatic radiomics features underwent a multistep selection procedure within each training fold, including minimum redundancy maximum relevance, least absolute shrinkage and selection operator regularization, and recursive feature elimination. Features selected in at least three of five folds were retained as stable features to reduce fold-specific selection instability.

Four XGBoost-based models were constructed: clinical only, clinical + body composition, clinical + radiomics, and clinical + body composition + radiomics. Within each training fold, missing values were imputed and continuous predictors were standardized using training-fold statistics, and the fitted transformations were applied to the corresponding validation fold. XGBoost hyperparameters were tuned by grid search with three-fold stratified cross-validation, using AUC as the optimization metric. The search space included n_estimators = 50, 100, or 200; max_depth = 2, 3, or 4; learning_rate = 0.01, 0.05, or 0.10; min_child_weight = 1 or 3; subsample = 0.8 or 1.0; and colsample_bytree = 0.8 or 1.0. Given the relatively balanced outcome distribution, no oversampling, undersampling, or class-weighting method was applied.

Model performance was evaluated using five-fold out-of-fold predictions. XGBoost-based models were assessed using ROC analysis, calibration analysis, integrated discrimination improvement (IDI) analysis, and decision curve analysis (DCA), and were compared with BISAP, admission-based mPASS, and EPIC using ROC analysis. SHAP analysis was performed for the final best-performing model refitted on the full cohort.

Statistical Analysis

Statistical analysis was performed using Python 3.11. As this retrospective study included all eligible patients during the study period, no formal a priori sample size calculation was performed. A post hoc power assessment was conducted to evaluate whether the final cohort size was sufficient for ROC-based discrimination analysis. Variables with more than 20% missing data were excluded. Continuous variables are presented as mean ± standard deviation (SD) or median with interquartile range (IQR), and categorical variables as counts and percentages. Between-group comparisons were performed using appropriate parametric or nonparametric tests. The 95% confidence intervals (CIs) were estimated by bootstrap resampling based on out-of-fold predictions. Pairwise comparisons of AUCs were performed using the DeLong test. A two-sided p < 0.05 was considered statistically significant.

Results

Study Population

A total of 679 consecutive patients with acute pancreatitis were screened, and 393 with other etiologies were excluded, leaving 286 patients with HTG-AP. Of these, 87 patients with moderately severe or severe acute pancreatitis at admission were excluded. Among the remaining 199 patients with initially mild HTG-AP, 35 were further excluded because of chronic pancreatitis, previous pancreatic surgery, or malignancy (n = 19), severe comorbidities (n = 11), or insufficient imaging quality (n = 5). Finally, 164 patients were included, comprising 88 in the progression group and 76 in the non-progression group (Figure 2).

Figure 2.

Study flowchart.

Among the 88 patients with in-hospital progression, all showed worsening pancreatic or peripancreatic inflammatory changes or fluid collections on follow-up CT. New or increased ascites or pleural effusion occurred in 53 patients, local complications in 37, and new systemic complications or organ dysfunction in 27; some patients had multiple progression components. Among patients with systemic complications or organ dysfunction, 15 required ICU admission and 7 were transferred to higher-level care.

Clinical Characteristics and CT-Derived Measurements

Among the 164 included patients, 122 were male and 42 were female. No significant between-group differences were observed in age, sex, smoking history, alcohol history, or diabetes history (all p > 0.05). Among laboratory variables, amylase, lipase, hemoglobin, calcium, direct bilirubin, indirect bilirubin, triglycerides, and total cholesterol differed significantly between the two groups, whereas the remaining laboratory parameters did not. Compared with the non-progression group, the progression group had a longer hospital stay and a shorter interval from admission to follow-up CT for outcome assessment. Among CT-based body composition parameters, significant between-group differences were observed for all variables except splenic attenuation and muscle attenuation. Detailed patient characteristics are presented in Table 1.

Table 1.

Baseline Characteristics of Patients with and without in-Hospital Progression

Variable	Overall	Non-Progression	Progression	P-value
General information
Age [years, mean ± SD]	40.49 ± 9.26	41.80 ± 10.06	39.35 ± 8.40	0.095
Sex [n, %]				0.051
Male	122 (74.4%)	51 (67.1%)	71 (80.7%)
Female	42 (25.6%)	25 (32.9%)	17 (19.3%)
Smoking history [n, %]				0.358
Yes	39 (23.8%)	21 (27.6%)	18 (20.5%)
No	125 (76.2%)	55 (72.4%)	70 (79.5%)
Alcohol consumption history [n, %]				0.797
Yes	16 (9.8%)	8 (10.5%)	8 (9.1%)
No	148 (90.2%)	68 (89.5%)	80 (90.9%)
Diabetes [n, %]				0.093
Yes	69 (42.1%)	27 (35.5%)	42 (47.7%)
No	95 (57.9%)	49 (64.5%)	46 (52.3%)
Length of hospital stay [days, median (IQR)]	7.00 (6.00–9.00)	6.00 (5.00–8.00)	8.00 (6.00–11.00)	<0.001 (***)
Time to follow-up CT [days, median (IQR)]	4.00 (3.00–5.00)	5.00 (3.00–5.00)	3.00 (2.50–4.00)	<0.001 (***)
Laboratory parameters
Amylase [U/L, median (IQR)]	144.00 (69.25–300.50)	112.00 (58.00–247.00)	166.00 (86.00–360.00)	0.034 (*)
Lipase [U/L, median (IQR)]	275.80 (123.40–848.90)	246.10 (96.50–766.75)	406.90 (188.48–970.23)	0.008 (**)
WBC [×109/L, mean ± SD]	13.89 ± 4.19	13.25 ± 4.28	14.43 ± 4.05	0.073
RBC [×1012/L, mean ± SD]	4.90 ± 0.52	4.87 ± 0.51	4.94 ± 0.53	0.386
HGB [g/L, median (IQR)]	159.00 (146.00–169.00)	153.50 (140.50–163.00)	164.00 (147.75–173.25)	0.002 (**)
PLT [×109/L, mean ± SD]	238.82 ± 61.06	243.49 ± 59.16	234.78 ± 62.71	0.362
HCT [%, median (IQR)]	43.35 (40.27–45.90)	43.30 (39.98–46.15)	43.45 (40.40–45.73)	0.992
Ca [mmol/L, median (IQR)]	2.19 (2.10–2.27)	2.20 (2.13–2.30)	2.17 (2.08–2.24)	0.035 (*)
TBIL [µmol/L, median (IQR)]	9.30 (7.10–13.45)	10.10 (6.95–15.30)	9.00 (7.40–13.05)	0.488
DBIL [µmol/L, median (IQR)]	4.85 (3.32–7.50)	4.25 (3.30–6.12)	5.35 (3.55–8.30)	0.024 (*)
IBIL [µmol/L, median (IQR)]	4.50 (2.60–8.30)	5.50 (3.40–10.50)	3.45 (1.55–7.35)	<0.001 (***)
ALT [U/L, median (IQR)]	17.95 (10.97–29.38)	18.80 (11.00–31.05)	17.50 (11.05–29.25)	0.436
AST [U/L, median (IQR)]	18.50 (13.62–23.70)	18.20 (13.55–23.05)	18.70 (14.40–25.00)	0.653
Urea [mmol/L, median (IQR)]	4.00 (3.17–4.90)	4.12 (3.12–5.05)	3.92 (3.26–4.87)	0.691
Cr [µmol/L, median (IQR)]	55.30 (45.75–67.70)	54.20 (45.50–67.70)	56.65 (46.58–67.65)	0.380
TG [mmol/L, median (IQR)]	25.09 (14.66–42.80)	15.36 (12.71–28.71)	34.19 (21.15–52.27)	<0.001 (***)
TC [mmol/L, median (IQR)]	8.90 (6.58–12.40)	7.85 (5.98–9.57)	10.66 (7.88–14.24)	<0.001 (***)
GLU [mmol/L, median (IQR)]	8.00 (6.20–11.51)	7.40 (5.80–11.43)	8.50 (6.47–11.60)	0.146
PCT [ng/mL, median (IQR)]	0.07 (0.04–0.13)	0.06 (0.03–0.10)	0.08 (0.04–0.16)	0.084
CRP [mg/L, median (IQR)]	22.85 (5.91–59.31)	13.65 (4.10–53.94)	28.38 (7.01–76.36)	0.135
CT-based body composition
VFA [cm2, mean ± SD]	396.41 ± 146.60	333.35 ± 132.42	450.88 ± 136.66	<0.001 (***)
SFA [cm2, median (IQR)]	268.98 (208.65–339.97)	257.94 (191.48–320.19)	278.39 (220.23–361.22)	0.045 (*)
TFA [cm2, median (IQR)]	669.85 (532.81–835.87)	561.32 (444.67–719.17)	754.67 (620.44–870.43)	<0.001 (***)
ACA [cm2, mean ± SD]	708.53 ± 178.42	665.32 ± 175.64	745.84 ± 173.23	0.004 (**)
VFA/TFA [mean ± SD]	0.58 ± 0.12	0.56 ± 0.12	0.60 ± 0.11	0.024 (*)
VFA/ACA [median (IQR)]	0.56 (0.50–0.62)	0.51 (0.44–0.55)	0.60 (0.56–0.65)	<0.001 (***)
Liver attenuation [HU, median (IQR)]	38.00 (28.00–48.00)	41.00 (28.75–52.00)	35.50 (26.00–44.25)	0.008 (**)
Splenic attenuation [HU, median (IQR)]	51.00 (46.75–53.00)	51.50 (47.00–54.00)	50.00 (46.00–53.00)	0.117
Muscle attenuation [HU, mean ± SD]	56.01 ± 4.81	56.32 ± 5.10	55.74 ± 4.55	0.449
BISAP [score, median (IQR)]	0 (0–1)	0 (0–1)	0 (0–1)	0.989
EPIC [score, median (IQR)]	2 (1–3)	2 (1–2)	2 (1–3)	0.238
mPASS [score, median (IQR)]	85 (65–105)	70 (55–95)	87.5 (70–105)	0.005 (**)

Notes: Statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001.

Abbreviations: WBC, white blood cell count; RBC, red blood cell count; HGB, hemoglobin; PLT, platelet count; HCT, hematocrit; Ca, calcium; TBIL, total bilirubin; DBIL, direct bilirubin; IBIL, indirect bilirubin; ALT, alanine aminotransferase; AST, aspartate aminotransferase; Cr, creatinine; TG, triglycerides; TC, total cholesterol; GLU, glucose; PCT, procalcitonin; CRP, C-reactive protein; VFA, visceral fat area; SFA, subcutaneous fat area; TFA, total fat area; ACA, abdominal cavity area; HU, Hounsfield units; EPIC, extrapancreatic inflammation on CT score; mPASS, modified Pancreatitis Activity Scoring System; BISAP, Bedside Index for Severity in Acute Pancreatitis.

Feature Selection

The stable selected features are summarized in Figure 3. The features in each category were ranked by selection frequency across the five folds. The most frequently selected clinical features were total cholesterol (TC), triglycerides (TG), lipase, indirect bilirubin (IBIL), and hemoglobin (HGB). The most frequently selected CT-based body composition features were ACA, TFA, VFA, VFA/ACA, and liver attenuation. The selected radiomics features mainly included wavelet- and shape-derived features. Based on the predefined stability criterion of selection in at least three of five folds, 11 stable features were retained for final model construction.

Figure 3.

Feature selection frequency across five-fold cross-validation. The clinical features, CT-based body composition features, and radiomics features were ranked according to their selection frequency across the five folds.

Evaluation of Model Performance

The clinical + body composition + radiomics model achieved the highest AUC of 0.830 (95% CI, 0.764–0.890), improving by 0.123 over the clinical-only model, 0.092 over the clinical + radiomics model, and 0.024 over the clinical + body composition model (Figure 4A). BISAP, EPIC, and admission-based mPASS showed lower AUCs of 0.448, 0.545, and 0.627, respectively. IDI analysis showed that the full model improved discrimination over the clinical-only model (IDI = 0.254; 95% CI, 0.158–0.343), the clinical + body composition model (IDI = 0.074; 95% CI, 0.009–0.142), and the clinical + radiomics model (IDI = 0.189; 95% CI, 0.102–0.273). Calibration analysis showed good agreement for the full model, with a Brier score of 0.172 (Figure 4B). Decision curve analysis showed higher net benefit across most threshold probabilities (Figure 5). The DeLong test showed that the clinical + body composition + radiomics model had a higher AUC than the clinical-only model (p = 0.003), clinical + radiomics model (p = 0.008), BISAP (p < 0.001), EPIC (p < 0.001), and admission-based mPASS (p < 0.001), whereas the difference compared with the clinical + body composition model was not statistically significant (p = 0.351). In the post hoc ROC power analysis, all XGBoost-based models showed AUCs above the level detectable with 80% power. Detailed model performance metrics are shown in Table 2.

Figure 4.

Receiver operating characteristic curves and calibration curves based on five-fold out-of-fold predictions. (A) ROC curves compare the clinical-only, clinical + body composition, clinical + radiomics, and clinical + body composition + radiomics models, together with conventional clinical scoring systems. (B) Calibration curves show the agreement between predicted and observed probabilities for the four XGBoost-based models.

Figure 5.

Decision curve analysis based on five-fold out-of-fold predictions. The decision curves compare the net benefit of the four XGBoost-based models across threshold probabilities. The treat-all and treat-none strategies were used as reference curves.

Table 2.

Model Performance Based on Five-Fold Out-of-Fold Predictions

Model	AUC (95% CI)	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)
Clinical only	0.707 (0.622–0.786)	0.689 (0.628–0.762)	0.761 (0.523–0.899)	0.605 (0.443–0.845)	0.691 (0.610–0.817)	0.687 (0.562–0.826)
Clinical + body composition	0.806 (0.739–0.870)	0.756 (0.695–0.829)	0.739 (0.581–0.854)	0.776 (0.679–0.914)	0.793 (0.711–0.908)	0.720 (0.616–0.833)
Clinical + radiomics	0.738 (0.653–0.815)	0.713 (0.652–0.793)	0.784 (0.524–0.890)	0.632 (0.527–0.893)	0.711 (0.646–0.872)	0.716 (0.576–0.837)
Clinical + body composition + radiomics	0.830 (0.764–0.890)	0.768 (0.720–0.841)	0.784 (0.610–0.953)	0.750 (0.573–0.934)	0.784 (0.689–0.924)	0.750 (0.634–0.929)
EPIC	0.545	0.512	0.295	0.763	0.591	0.483
mPASS	0.627	0.640	0.818	0.434	0.626	0.673

Notes: Values for XGBoost-based models are presented with 95% confidence intervals; EPIC and mPASS are shown as point estimates.

Abbreviations: AUC, area under the receiver operating characteristic curve; PPV, positive predictive value; NPV, negative predictive value; EPIC, extrapancreatic inflammation on CT score; mPASS, modified Pancreatitis Activity Scoring System. BISAP was not included in the performance table because no clinically meaningful cutoff was identified in this cohort; it was therefore presented only in the ROC analysis as a reference comparator.

Interpretability of the Optimal Model

SHAP analysis was performed for the final clinical + body composition + radiomics XGBoost model refitted on the full cohort using the stable selected features. TG showed the greatest contribution to model prediction, followed by VFA/ACA, indicating that lipid metabolic burden and visceral fat distribution were the most influential factors. In the beeswarm plot, higher TG and VFA/ACA values were mainly associated with positive SHAP values, suggesting increased predicted risk of in-hospital progression. The selected radiomics feature provided complementary information for model prediction (Figure 6).

Figure 6.

SHAP analysis of the final clinical + body composition + radiomics XGBoost model. (A) Feature importance bar plot based on mean absolute SHAP values. (B) SHAP beeswarm plot showing the magnitude and direction of each feature’s effect on the model output. Positive SHAP values indicate a higher predicted risk of in-hospital progression, whereas negative SHAP values indicate a lower predicted risk.

Discussion

In this single-center retrospective study, we found that baseline clinical variables combined with non-contrast CT-derived quantitative imaging features could predict in-hospital progression in patients initially classified as having mild HTG-AP. The clinical + body composition + radiomics model achieved the highest AUC, while the clinical + body composition model performed close to the full model. These findings suggest that CT-derived body composition may provide prognostic information beyond routine morphological assessment.

The endpoint in this study reflected a spectrum of in-hospital worsening rather than a single severe endpoint. In this cohort, 22.6% of patients progressed to moderately severe or severe acute pancreatitis, whereas 31.1% developed persistent or worsening clinical or imaging abnormalities that required reassessment without crossing conventional severity thresholds. This suggests that final severity classification alone may not fully capture an unstable disease course in patients initially classified as having mild HTG-AP. The longer hospital stay in the progression group was also consistent with greater clinical management needs. Therefore, risk stratification should identify not only patients likely to develop moderately severe or severe disease, but also those who may require closer monitoring and timely reassessment.

The progression of HTG-AP likely reflects the combined effects of lipid metabolic burden, pancreatic injury, systemic inflammation, and abdominal fat distribution. In the SHAP analysis, TG and VFA/ACA were among the most influential predictors, suggesting the relevance of circulating lipid burden and relative visceral fat accumulation in identifying patients at risk of progression. Previous studies have shown that higher early TG levels and visceral adiposity-related indices are associated with more severe HTG-AP.^21,22 Mechanistically, excess circulating TG and visceral adipose tissue may jointly increase lipotoxic and inflammatory burden. TG hydrolysis generates free fatty acids that aggravate pancreatic injury, microcirculatory disturbance, oxidative stress, and inflammation, while visceral adipose tissue may further release free fatty acids, adipokines, and pro-inflammatory mediators.^21,23 Unlike previous studies that mainly used absolute visceral fat area,^24,25 the present study included VFA/ACA as an index of relative visceral fat accumulation, which may help characterize fat distribution within the abdominal cavity and reduce the influence of body size differences on area-based measurements.

In our study, the integrated clinical + body composition + radiomics model achieved the highest numerical AUC among the four machine-learning models, suggesting that quantitative imaging features may provide complementary information for predicting in-hospital progression. This result extends previous HTG-AP studies that integrated imaging with clinical or hematological features to predict organ failure or severe disease by focusing on in-hospital progression among patients initially classified as having mild HTG-AP.^13,14,17,18 Importantly, although all patients in the present cohort were initially classified as having mild disease, the integrated model still showed acceptable discriminatory performance. This suggests that patients with initially mild HTG-AP may have different underlying risk profiles that are not fully apparent from conventional severity classification. Compared with individual clinical variables or routine imaging assessment, the integrated model may help improve the identification of subsequent disease instability by combining multiple weak but complementary baseline signals.

A key finding was that the clinical + body composition model performed close to the full multimodal model, whereas adding pancreatic radiomics to clinical variables produced a smaller numerical improvement. This suggests that body composition parameters captured much of the additional information available from baseline non-contrast CT. Unlike pancreatic radiomics, which relies on detectable local texture and morphological heterogeneity, CT-derived body composition quantifies broader patient-level tissue characteristics, including fat and muscle distribution, and may serve as an imaging marker of systemic health and metabolic phenotype.^26,27 Previous studies have also shown that metabolic comorbidities and nutritional status are associated with the severity and prognosis of acute pancreatitis.²⁸ By contrast, although pancreatic radiomics has shown predictive value in acute pancreatitis,²⁹ its incremental contribution may be limited in patients initially classified as mild because early non-contrast CT may show less pronounced pancreatic heterogeneity. These findings suggest that body composition may provide a more robust source of early prognostic information, while pancreatic radiomics offers complementary but more limited incremental value.

Previous HTG-AP-related studies have shown that conventional tools such as APACHE II, SIRS, and BISAP showed moderate discriminatory ability for severe HTG-AP, with BISAP showing relatively better overall performance.^17,30 In our study, however, BISAP, EPIC, and admission-based mPASS showed lower AUCs than the machine-learning models. In this initially mild cohort, limited baseline systemic abnormalities and extrapancreatic inflammatory changes may have narrowed score distributions and reduced discrimination of subsequent progression risk. BISAP may also be limited by its lack of HTG-AP-specific information, particularly lipid metabolic burden; previous studies have shown improved performance after adding metabolic, inflammatory, or CT markers.^17,31 EPIC may have performed less well because extrapancreatic inflammatory findings showed limited variation among initially mild patients.¹⁰ In addition, because mPASS was assessed only at admission in our study, its ability to reflect serial changes in disease activity could not be fully utilized, which may have limited its prediction of subsequent clinical evolution.⁸ These factors may explain why conventional scores were less effective in differentiating progression risk among patients initially classified as mild.

From a clinical perspective, the proposed model may provide an auxiliary framework for early risk stratification in initially mild HTG-AP. By integrating baseline clinical variables with non-contrast CT-derived body composition and radiomics features, the model may help identify patients who are more likely to require closer monitoring and timely reassessment during hospitalization. These findings suggest the potential value of baseline multimodal information for supporting risk awareness and individualized observation strategies, although prospective validation is still needed before clinical implementation. Previous studies have highlighted the potential of bioinformatics and multimodality imaging approaches for improving disease assessment.^32,33 Future studies may explore whether molecular, bioinformatic, or broader multimodality imaging strategies can further refine risk characterization in HTG-AP.

Limitations

This study has several limitations. First, this was a single-center retrospective study with a relatively limited sample size, which may limit the generalizability of the findings. Although feature selection and internal cross-validation were performed to reduce overfitting, larger multicenter prospective studies are needed to confirm the robustness and clinical applicability of the proposed model. Second, follow-up CT examinations were performed according to clinical need, and the timing of imaging reassessment was not fully standardized. Third, the model was developed using early non-contrast CT images, and its performance in patients undergoing contrast-enhanced CT or different imaging protocols requires further investigation.

Conclusion

In conclusion, this study suggests that integrating clinical variables, CT-based body composition parameters, and pancreatic radiomics may provide a potentially useful approach for early risk stratification in patients initially classified as having mild HTG-AP. The findings highlight the potential relevance of lipid metabolic burden, abdominal fat distribution, and quantitative CT imaging features in in-hospital disease progression. Further prospective multicenter validation is needed before clinical implementation.

Funding Statement

This study was supported by the Scientific Research Program of the Changsha Municipal Health Commission (Grant No. KJ-A2023006). The funder had no role in the study design, data collection, data analysis, interpretation of the results, preparation of the manuscript, or decision to submit the manuscript for publication.

Data Sharing Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request. Due to ethical and legal restrictions related to the protection of participant privacy and confidentiality, the raw data cannot be made publicly available.

Ethics Approval and Consent to Participate

The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of The First Hospital of Changsha, China (Approval No. 2026-4). Due to the retrospective design and the use of de-identified clinical data, the requirement for written informed consent was waived by the ethics committee.

Disclosure

The authors report no conflicts of interest in this work.