Development and validation of a clinical prediction model for fatty pancreas disease: from CT-based indicators to an easy-to-use prediction application

Abstract

Background

Fatty pancreas disease (FPD) is closely associated with the pathogenesis of pancreatic cancer (PC) and diabetes mellitus (DM). However, current imaging modalities for assessing intra-pancreatic fat deposition (IPFD) have limitations and are not widely used in clinical practice. Developing a practical prediction tool for FPD would facilitate the evaluation of pancreatic health.

Methods

A total of 852 subjects were included in the model construction cohort, and 202 PC patients were included to investigate the association between fat deposition and PC. Quantitative assessment of IPFD and diagnosis of FPD were based on the pancreas-to-spleen attenuation ratio (P/S ratio) from CT scans. Multivariate logistic regression was used to identify risk factors associated with FPD, and a web-based application was deployed based on these factors. The performance of the model was validated in an internal cohort with respect to discrimination, calibration and clinical benefit.

Results

The incidence of PC and DM significantly increased as the P/S ratio decreased, particularly when P/S ratio fell below 0.8. Multivariate analysis identified seven independent risk factors: age > 65 years, abnormal waist circumference, abnormal gamma-glutamyl transferase, fasting plasma glucose > 6.1mmol/L, NLR > 1.97, fatty liver index > 24.7, and mFIB-4 > 3.05. A web application was deployed based on these risk factors. The model demonstrated good discrimination, with an AUC of 0.750 (95% CI: 0.707–0.793) in the development cohort and 0.723 (95% CI: 0.652–0.795) in the validation cohort, along with satisfactory calibration and clinical net benefit. Furthermore, regression analysis revealed significant correlations between model-predicted risk values and IPFD severity, indicating that the FPD model effectively captures the severity of fat deposition in the pancreas.

Conclusions

This multidimensional predictive model enables comprehensive evaluation of FPD risk using routinely available clinical parameters, which can serve as an effective preliminary screening tool to guide subsequent examinations and interventions.

Introduction

In recent decades, the global prevalence of obesity has surged, driven by shifts in dietary patterns, sedentary lifestyles, and altered work environments [1]. Excessive lipids accumulate in ectopic organs like the liver and pancreas, giving rise to a cascade of health issues [2, 3]. The detrimental effects of fatty liver (FL) are widely acknowledged by the general public. However, fatty pancreas disease (FPD), a disorder of lipid metabolism within the pancreas, remains relatively poorly recognized [4].

FPD refers to excess intra-pancreatic fat deposition (IPFD) that is scattered throughout the pancreas and above the upper limit of normal [5], which primarily results from metabolic disorders [6] and is closely associated with the development of diabetes mellitus (DM) and pancreatic cancer (PC) [7–9]. The most recent epidemiological meta-analysis indicates that the global prevalence of excessive IPFD is 28.7% (95% CI: 20.8%–36.7%) [10], moreover, it can reach as high as 61.4% in obese populations [11] and 73.1% in the cohorts of individuals with T2DM [10]. FPD is pathologically defined by abnormal lipid accumulation within pancreatic acinar cells and islets of Langerhans [12, 13], contributing to β-cell dysfunction. Evidence from animal and translational studies indicates that the inflammatory microenvironment triggered by IPFD plays a critical role in driving pancreatic carcinogenesis [14, 15]. PC leads to poor prognosis due to its aggressive nature and the lack of effective therapeutic interventions. It is projected to soon become the second leading cause of cancer-related mortality in Western countries [16].

Given these challenges, early recognition of modifiable risks is imperative. However, FPD remains underrecognized in clinical practice, which is not routinely documented in imaging reports or discussed with patients. The clinical utility of CT and MRI in quantitatively evaluating IPFD [17] is constrained by radiation risks and high costs, which is inappropriate for population-level screening. Abdominal ultrasound is a more commonly used imaging method for assessing IPFD, which is typically achieved by comparing the echogenicity of the pancreas and renal cortex. Since the pancreas and kidneys are often difficult to display on the same screen, the liver is used as a reference for comparing the two, a practice that increases difficulty and introduces more biases [18]. While ultrasound can qualitatively indicate IPFD through features like increased echogenicity, it cannot accurately quantify fat deposition as precisely as CT or MRI, which utilize CT values or fat quantification methods. In the present study, the CT-based pancreas-to-spleen attenuation ratio was employed to evaluate IPFD in the entire pancreas. This metric utilizes splenic attenuation as an internal reference for calibration, and it has become a widely recognized CT parameter for IPFD assessment in the existing literature [5, 19, 20].

Since FPD results from obesity and systemic dysregulation of lipid metabolism [4], we hypothesize that routinely available clinical parameters, including anthropometric measures and metabolic biomarkers, may serve as predictive indicators of IPFD. Furthermore, developing a web-based prediction application for FPD assessment could enable cost-effective, non-invasive stratification of IPFD severity.

Therefore, this study aims to achieve two primary objectives: first, to identify multidimensional risk factors for a comprehensive evaluation of FPD; second, to develop and validate a practical prediction model for FPD, with the ultimate goal of informing early intervention strategies and personalized risk management.

Materials and methods

Study design and patients

The clinical data of study subjects who underwent abdominal CT scans at Fujian Medical University Union Hospital between July 2022 and May 2023 were reviewed, to build a development cohort for the prediction model. Another cohort consisting of subjects who visited our hospital between July 2023 and May 2024 was included for model validation. The following exclusion criteria were applied: (1) pancreatic tumor indicated by abdominal ultrasound, MR or CT; (2) history of pancreatic surgery, acute/chronic pancreatitis, or alcoholic FPD. Medical histories and diagnoses were retrieved from the electronic medical record system. Alcoholic FPD was diagnosed based on ultrasound findings of increased echogenicity of the pancreatic body relative to the kidney [21], combined with a history of alcohol intake greater than 20 g/day. Additionally, patients with pancreatic cancer (PC) were included into the PC group, whose preoperative abdominal CT images were used to evaluate the correlation between the extent of IPFD and PC. This study was approved by the Ethics Committee of the Fujian Medical University Union Hospital. All patients signed an informed consent. The detailed research design for this study is depicted in Fig. 1.

Fig. 1.

Flow chart of the development and validation of a FPD assessment model. PC, pancreatic cancer; FPD, fatty pancreas disease; DM, diabetes mellitus; P/S ratio, ROC curves, receiver operator characteristic curves

Data collection

Clinical data were obtained from electronic medical records, including sex, age, waist circumference (WC), body mass index (BMI), history of DM, hypertension, and results of laboratory test. The blood sample for the laboratory test were collected when the subjects were fasting in the morning, including alanine aminotransferase (ALT), aspartate aminotransferase (AST), gamma-glutamyl transferase (GGT), triglycerides (TG), total cholesterol (TC), high-density lipoprotein (HDL), low-density lipoprotein (LDL), fasting plasma glucose (FPG), platelet count (PLT), neutrophil count, and lymphocyte count. The neutrophil-to-lymphocyte ratio (NLR) is a simple and widely used inflammatory biomarker calculated by dividing the absolute neutrophil count by the absolute lymphocyte count. The extent of fatty liver was evaluated using fatty liver index (FLI) [22], modified fibrosis-4 index (mFIB-4) [23], and the Forns index (FORNS) [24]. These indices use easily accessible clinical and laboratory measures to assess the severity of fatty liver disease. The formulas for these indices were as previously reported [22–24]. Continuous variables were dichotomized based on predefined criteria: laboratory biomarkers were categorized as abnormal or normal using established reference ranges; for other continuous variables, including WC, NLR, FLI, mFIB-4, and FORNS, the optimal cutoff values were determined by maximizing the Youden index (sensitivity + specificity − 1) via receiver operating characteristic (ROC) curve analyses. Sex-specific cutoff values for WC were set to define normal and abnormal WC, which is 88.4 cm for females and 93.4 cm for males. Overweight was defined as a BMI ≥ 25 kg/m².

CT protocols and assessment of FPD

FPD was defined using unenhanced abdominal CT scans, with data managed through the Picture Archiving and Communication System (PACS). The ratio of the mean pancreatic CT number to the mean splenic CT number (P/S ratio) was used to diagnose FPD [25]. The mean CT numbers of the pancreas and the spleen were calculated from CT numbers of three regions of interest (ROIs), each measuring 100 mm², located outside overlapping areas of the pancreas and spleen, avoiding vessels, cystic lesions, and neoplastic lesions [20, 25]. For pancreatic CT number measurements, the ROIs are located in the head, body, and tail of the pancreas separately; for spleen, ROIs are selected from the lower, middle and upper parts of the spleen. Image analysis was conducted by two abdominal radiologists (H.D. and T.L., with 7 and 5 years of experience, respectively). H.D. performed initial measurements while blinded to clinical information and biochemical parameters. For inter-observer reliability assessment, a random sample of 100 cases was independently re-measured by T.L. Agreement was determined using the intra-class correlation coefficient (ICC). An example image is shown in Fig. 2. As the degree of IPFD increases, the P/S ratio decreases. However, there are no consensus criteria for the CT diagnosis of FPD; both a P/S ratio of < 0.7 [19] and < 0.8 [20] had been used in previous studies as thresholds indicating the presence of fatty pancreas. The pancreas contains a certain amount of fat under normal physiological conditions and the varying degrees of IPFD evidently have different impacts on pancreatic health. Different P/S ratios imply different clinical significance and the necessity for potential interventions. It is essential to determine a clinically relevant P/S ratio threshold (0.7 or 0.8) for diagnosing FPD based on large-sample data. This will help address the question of at what threshold fat deposition becomes significantly associated with PC and DM. In this study, subjects were categorized into 6 subgroups based on their P/S ratios: G1: P/S < 0.6, G2: 0.6 ≤ P/S < 0.7, G3: 0.7 ≤ P/S < 0.8, G4: 0.8 ≤ P/S < 0.9, and G5: 0.9 ≤ P/S < 1.0, G6: P/S ≥ 1.0. The prevalence of PC and DM in these subgroups was then compared. ROC curves were utilized to identify the optimal P/S ratio thresholds for predicting these two diseases. The final diagnostic P/S ratio for FPD was determined by synthesizing the findings from both subgroup analysis and ROC curve analysis.

Fig. 2.

Example of CT number measurements in regions of interest (ROI). HU, Hounsfield Units

Model construction and visualization

Multivariate logistic regression analyses were used to identify independent predictors of FPD and develop the prediction model. The optimal prediction model was implemented into a web-based application using the Streamlit Python-based framework. This publicly accessible app will facilitate individual assessment of the probability of FPD, whereby users input the relevant features and the app automatically generates the risk assessment results.

Statistical Analysis

Continuous variables were expressed as mean ± standard deviation (SD) or median (interquartile range) based on distribution, and categorical variables were presented as frequencies and percentages. Mann–Whitney U or Student’s t-test were used for comparing continuous variables depending on the variable distribution. Categorical variables were compared using the Chi-square test or Fisher’s exact test. The diagnostic performance of variables was assessed using ROC curves and the area under the ROC curve (AUC). Multicollinearity among the predictor variables was assessed using the Variance Inflation Factor (VIF) and tolerance. The ROC analysis was applied to assess the discrimination ability of prediction models. Model calibration was assessed with the Hosmer–Lemeshow goodness-of-fit test and calibration plots. The decision curve analysis (DCA) was conducted to estimate the clinical utility of the model. Subjects with missing data were excluded from the analysis via complete case analysis. Statistical analysis was performed using SPSS software (Version 26.0), Python (Version 3.11.0) and R (version 4.2.1). The calibration curves were plotted using the rms package, and decision curve analysis was performed with the rmda package in R. Linear regression and correlation coefficients were used to assess the association between CT-quantified IPFD and predicted risk. Statistical significance was defined as a two-tailed p-value less than 0.05.

Results

Baseline characteristics of the study population

A total of 1054 subjects were included in the study. Among them, 629 subjects identified from July 2022 to May 2023 and 223 subjects identified from July 2023 to May 2024 were categorized into the development group and validation group, respectively; 202 patients diagnosed with PC were enrolled into the PC group. The baseline characteristics of the three groups were shown in Table 1.

Table 1.

Clinical characteristics of the study population

Parameters	PC Group (N = 202)	Development Group # (N = 629)	Validation Group # (N = 223)	p 1	p 2
Age, year	62 (55, 69)	58 (47, 67)	58 (48, 67)	< 0.001	0.923
Sex, male	118 (58.4%)	362 (57.6%)	125 (56.1%)	0.829	0.698
BMI, kg/m2	21.9 (20.2, 24.1)	23.1 (21.0, 25.7)	23.5 (21.7, 25.6)	0.013	0.405
WC, cm	89.6 (84.0, 95.0)	90.0 (83.0,96.0)	90.0 (84.0,95.7)	0.590	0.879
Hypertension	62 (30.7%)	179 (28.5%)	63 (28.3%)	0.542	0.953
DM	50 (24.8%)	87 (13.8%)	35 (15.7%)	< 0.001	0.495
ALT, IU/L	24.0 (13.8, 62.5)	18.0 (12.0, 30.0)	19.0 (14.0, 30.0)	< 0.001	0.283
AST, IU/L	23.0 (16.0, 46.2)	20.0 (15.0, 27.0)	21.0 (16.0, 26.0)	< 0.001	0.495
GGT, U/L	40.5 (19.0, 227.8)	25.0 (16.0, 48.5)	25.0 (17.0, 49.0)	< 0.001	0.577
TG, mmol/L	1.34 (1.00, 1.78)	1.28 (0.95, 1.87)	1.31 (0.92, 1.81)	0.599	0.847
TC, mmol/L	4.66 (3.80, 5.50)	4.75 (4.13, 5.41)	4.83 (4.09, 5.61)	0.570	0.355
HDL, mmol/L	1.10 (0.88, 1.34)	1.15 (0.96, 1.39)	1.21 (0.99, 1.47)	0.028	0.150
LDL, mmol/L	2.90 (2.14, 3.72)	3.11 (2.51, 3.64)	3.06 (2.48, 3.74)	0.111	0.871
FPG, mmol/L	6.2 (5.4, 7.8)	5.1 (4.6, 5.8)	5.2 (4.7, 5.8)	< 0.001	0.210
PLT, *109/L	217 (167, 274)	238 (187, 282)	223 (187, 268)	< 0.001	0.056
Neut, *109/L	4.30 (3.08, 5.45)	3.68 (2.79, 4.78)	3.47 (2.71, 4.67)	< 0.001	0.193
Lym, *109/L	1.52(1.18, 1.93)	1.77 (1.43, 2.25)	1.72 (1.37, 2.14)	< 0.001	0.250

^# Subjects without pancreatic tumors. p₁: PC group vs. development group; p₂: development group vs. validation group

BMI Body mass index, WC Waist circumference, DM Diabetes mellitus, ALT Alanine aminotransferase, AST Aspartate aminotransferase, GGT γ-glutamyl transferase, TG Triglyceride, TC Total cholesterol, LDL Low-density lipoprotein cholesterol, HDL High-density lipoprotein cholesterol, FPG Fasting plasma glucose, PLT Platelet count, Neut Neutrophil count, Lym Lymphocyte count

Compared to the subjects in development group, PC patients exhibited a significantly older age, lower BMI and higher prevalence of DM, while WC did not differ significantly between the groups. Laboratory analysis revealed that the PC group had significantly elevated FPG, neutrophil count, and liver enzymes (ALT, AST, GGT) compared to the development group. Conversely, PLT and lymphocyte count were significantly lower in the PC group. No significant differences were observed in serum lipid contents (TG, TC, LDL), except for lower HDL in the PC group.

The baseline characteristics between the development group and the validation group showed no significant differences.

The relationships of the P/S ratio with PC and DM

The interobserver intraclass correlation coefficients for the P/S ratio was 0.846 (95% CI: 0.771–0.896), indicating good consistency. Patients were categorized into subgroups according to different P/S ratios to evaluate the correlation between the extent of IPFD and presence of pancreatic disease (PC, DM): G1: P/S < 0.6, G2: 0.6 ≤ P/S < 0.7, G3: 0.7 ≤ P/S < 0.8, G4: 0.8 ≤ P/S < 0.9, and G5: 0.9 ≤ P/S < 1.0, G6: P/S ≥ 1.0 (Fig. 3a and c). The decreasing P/S ratio was associated with a trend of increased risk for both PC and DM. In the subgroup analyses for PC (Fig. 3a), no significant increase of PC prevalence was observed when the P/S ratio decreased from > 1.0 to 0.8 (G6: 15.3%, G5: 14.8%, G4: 19.9%, p > 0.05). However, when the P/S ratio fell below 0.8, the prevalence of PC surged significantly (G4 vs. G3: 19.9% vs. 38.8%, p < 0.001; P/S ratio ≥ 0.8 vs. <0.8: 17.4% vs. 39.3%, p < 0.001). Subsequent decreases in P/S ratio below 0.8 were associated with moderate further increases in PC prevalence (G3: 38.8%, G2: 36.2%, G1: 47.1%, p > 0.05). For DM, a similar threshold effect was observed. The increase in DM prevalence was not statistically significant as the P/S ratio decreased from > 1.0 to 0.8 (G6: 10.6%, G5: 11.8%, G4: 12.8%, p > 0.05). However, the prevalence of DM increased significantly when the P/S ratio dropped below 0.8 compared to those with a P/S ratio ≥ 0.8 (12.1% vs. 26.0%, p < 0.001). Moderate increments were observed among subgroups with P/S ratio < 0.8. The ROC curve analyses identified 0.788 as the optimal cutoff value for P/S ratio to predict PC (sensitivity: 0.793, specificity: 0.485, Fig. 3b) and DM (sensitivity: 0.764, specificity: 0.467, Fig. 3d). Therefore, a P/S ratio less than 0.8 was identified as a landmark related to the development of pancreatic disease and was determined as the cutoff value for diagnosing FPD, which is consistent with the criteria outlined in Otsuka et al.‘s study [20].

Fig. 3.

Determination of the optimal cutoff value of P/S ratio for defining FPD using cutoff values related to PC and DM. a, b Decreased P/S ratio was associated with an increased prevalence of PC, and there was a threshold effect when the P/S ratio fell below 0.8. The ROC curve determined 0.788 as the optimal cutoff value significantly associated with PC (specificity: 0.485, sensitivity: 0.793). c, d A similar association between P/S ratio and disease prevalence was observed in the DM group, with the optimal cutoff value determined to be 0.788 by the ROC curve (specificity: 0.467, sensitivity: 0.764). PC, pancreatic cancer; DM, diabetes mellitus; ROC, receiver operating characteristic

Multidimensional indices for predicting the degree of IPFD

We then focus on identifying risk factors related to FPD from multidimensional clinical indices, which may offer a comprehensive assessment of IPFD in the pancreas and facilitate early detection and interventions. Patients in the development group (N = 629) were categorized into FPD (N = 159) and non-FPD groups (N = 470) to further analyze risk factors associated with FPD (Table 2). Patients with FPD was characterized by older age (p < 0.001), BMI (p = 0.013), and greater WC (p < 0.001). Additionally, these patients had significantly higher prevalence of DM and hypertension than those without FPD. Among indices related to metabolism, higher levels of TG (median: 1.36 vs. 1.24 mmol/L, p = 0.001) and FPG (median 5.4 vs. 5.0 mmol/L, p < 0.001) were related to FPD. In terms of liver enzymes, higher levels of ALT (median 20 vs. 17 IU/L, p = 0.038), AST (median 21 vs. 19 IU/L, p = 0.021), and GGT (median 30 vs. 23 U/L, p < 0.001) suggested an increased risk of FPD. Additionally, FPD was associated with changes in blood routine examination, including elevated NLR (median 1.93 vs. 2.20, p = 0.033) and decreased PLT (median 243 vs. 225 *10⁹/L, p = 0.015).

Table 2.

Comparison of clinical characteristics between non-FPD and FPD groups

Parameters	Non-FPD group (N = 470)	FPD group (N = 159)	p
Age, > 65 years	129 (27.4%)	80 (50.3%)	< 0.001
Sex, male	274 (58.3%)	88 (55.3%)	0.515
BMI, kg/m2	23.0 (20.8, 25.4)	23.8 (21.5, 26.3)	0.013
WC, cm	88.7 (82.0,95.0)	93.0 (86.5, 99.0)	< 0.001
Hypertension	118 (25.1%)	61 (38.4%)	0.001
Diabetes mellitus	49 (10.4%)	38(23.9%)	< 0.001
ALT, IU/L	17 (12, 27)	20 (12, 37)	0.038
AST, IU/L	19 (15, 26)	21 (16, 31)	0.021
GGT, U/L	23 (15, 46)	30 (19, 55)	< 0.001
TG, mmol/L	1.24 (0.89, 1.80)	1.36 (1.07, 2.05)	0.001
TC, mmol/L	4.74 (4.11, 5.39)	4.76 (4.14, 5.53)	0.390
HDL, mmol/L	1.16 (0.96, 1.40)	1.13 (0.95, 1.35)	0.183
LDL, mmol/L	3.08 (2.51, 3.63)	3.19 (2.53, 3.80)	0.489
FPG, mmol/L	5.0 (4.6, 5.5)	5.4 (4.8, 6.6)	< 0.001
PLT, *109/L	243 (191, 287)	225(177, 266)	0.015
Neut, *109/L	3.69 (2.72, 4.69)	3.62 (2.92, 5.26)	0.202
Lym, *109/L	1.78(1.45, 2.26)	1.73(1.38, 2.11)	0.107
NLR	1.93 (1.48, 2.64)	2.20 (1.55, 3.02)	0.033
FLI	0.58 (0.18, 1.53)	1.05(0.50, 2.49)	< 0.001
mFIB-4	2.52 (1.56, 4.08)	3.18(1.79, 4.64)	0.006
FORNS	7.11 (5.85, 8.24)	7.84 (6.92, 9.01)	< 0.001
P/S ratio	0.90 (0.85, 0.96)	0.73 (0.68, 0.77)	< 0.001

BMI Body mass index, WC Waist circumference, DM Diabetes mellitus, ALT Alanine aminotransferase, AST Aspartate aminotransferase, GGT γ-glutamyl transferase, TG Triglyceride, TC Total cholesterol, LDL Low-density lipoprotein cholesterol, HDL High-density lipoprotein cholesterol, FPG Fasting plasma glucose, PLT Platelet count, Neut Neutrophil count, Lym Lymphocyte count, NLR the neutrophil-to-lymphocyte ratio, FLI Fatty liver index, mFIB-4 Modified fibrosis-4 index, FORNS the Forns index, P/S ratio the ratio of the mean pancreatic CT number to the mean splenic CT number

FLI, FORNS, and mFIB-4 have been evaluated as predictors of fatty liver. These indices may also reflect the extent of IPFD, therefore, the associations between these indices and FDP were evaluated in the present study. As is shown in Table 2, FPD was significantly associated with elevated FLI (median: 0.58 vs. 1.05, p < 0.001), FORNS (median: 7.11 vs. 7.84, p < 0.001), and mFIB-4 (median: 2.52 vs. 3.18, p = 0.006), indicating the potential value of these indices for predicting FPD. Prior to multivariate analysis, collinearity diagnostics were performed to evaluate potential multicollinearity among FLI, FORNS, and mFIB-4. The results indicated no significant multicollinearity, with VIF values of 1.115, 1.244, and 1.229, and tolerance values of 0.896, 0.804, and 0.814 for FLI, FORNS, and mFIB-4, respectively. All values remained well within acceptable thresholds.

Based on the ROC curve analyses, the effectiveness of various biochemical indices, inflammatory markers, and anthropometry parameters in predicting FPD was evaluated, indicated by AUC values in Fig. 4.

Fig. 4.

The ROC curves and AUC values of inflammatory markers (a), body measurements (b), lipid markers, liver enzymes (c), and FL indices (d) in predicting FPD. FL, fatty liver. BMI, body mass index; WC, waist circumference; DM, diabetes mellitus; ALT, alanine aminotransferase; AST, aspartate aminotransferase; GGT, γ-glutamyl transferase; TG, triglyceride; TC, total cholesterol; LDL, low-density lipoprotein cholesterol; HDL, high-density lipoprotein cholesterol; FPG, fasting plasma glucose; PLT, platelet count; N, neutrophil count; L, lymphocyte count; NLR, the neutrophil-to-lymphocyte ratio; FLI, Fatty liver index; mFIB-4, modified fibrosis-4 index; FORNS, the Forns index; ROC, receiver operating characteristic; AUC, area under the ROC curve

Multivariate analysis for risk factors associated with FPD

The significant risk factors (P < 0.05) identified by univariate analysis were further subject to multivariate analysis to identify independent risk factors for FPD, including age, BMI, WC, DM, hypertension, GGT, FPG, PLT, NLR, FLI, FORNS, and mFIB-4. As is shown in Table 3. Multivariate logistic analysis revealed that age > 65 years (OR 2.112, 95% CI: 1.350–3.350, p = 0.001), abnormal WC (OR 2.057, 95% CI: 1.308–3.233, p = 0.006), GGT (OR 1.639, 95% CI: 1.026–2.618, p = 0.039), FPG > 6.1mmol/L (OR 2.468, 95% CI: 1.567–3.888, p < 0.001), NLR > 1.97(OR 1.581, 95% CI: 1.056–2.368, p = 0.026), FLI > 24.7 (OR 1.802, 95% CI: 1.045–3.108, p = 0.034), and mFIB-4 > 3.05 (OR 2.077, 95% CI: 1.313–3.284, p = 0.002) were independently associated with the presence of FPD. These variables were subsequently incorporated into the FPD model for individualized risk prediction.

Table 3.

Univariate and multivariate analyses for factors related to FPD

Parameters	Univariate analysis	Multivariate analysis
	p	OR	95% CI	p
Age, > 65 years	< 0.001	2.112	1.350, 3.305	0.001
Sex, male	0.515
Overweight	0.018	-	-	0.966
Abnormal WC	< 0.001	2.057	1.308, 3.233	0.006
DM	< 0.001	-	-	0.220
Hypertension	0.001	-	-	0.980
Preoperative labs
ALT, > 40 IU/L	0.053
AST, > 46 IU/L	0.098
GGT, > 50 IU/L for males; > 32 IU/L for males;	0.007	1.639	1.026, 2.618	0.039
TG, > 1.86 mmol/L	0.124
FPG, > 6.1 mmol/L	< 0.001	2.468	1.567, 3.888	< 0.001
TC, > 6.10 mmol/L	0.217
LDL, > 3.50 mmol/L	0.970
HDL, < 0.9 mmol/L vs. 0.9 < HDL < 1.9 mmol/L vs. > 1.9 mmol/L	0.110
PLT, > 300*10^9/L	0.020	-	-	0.081
NLR, > 1.97	0.007	1.581	1.056, 2.368	0.026
FLI, > 24.7	< 0.001	1.802	1.045, 3.108	0.034
FORNS, > 5.17	< 0.001	-	-	0.158
mFIB-4, > 3.05	< 0.001	2.077	1.313, 3.284	0.002

BMI Body mass index, WC Waist circumference, DM Diabetes mellitus, ALT Alanine aminotransferase, AST Aspartate aminotransferase, GGT γ-glutamyl transferase, TG Triglyceride, TC Total cholesterol, LDL Low-density lipoprotein cholesterol, HDL High-density lipoprotein cholesterol, FPG Fasting plasma glucose, PLT Platelet count, NLR the neutrophil-to-lymphocyte ratio, FLI Fatty liver index, mFIB-4 Modified fibrosis-4 index, FORNS the Forns index

Construction and performance of the prediction model

Based on the multivariable logistic regression model, a web application (Fig. 5) was deployed (https://blank-app-4sc5kpvketekbv3smwbnpl.streamlit.app/). Each predictor in the model is assigned with weighted points according to its regression coefficient, and the cumulative score provides a direct estimate of an individual’s risk of FPD. The discriminative ability of the FPD model was evaluated by AUC, which reached 0.750 (95% CI: 0.707–0.793) in the development group and 0.723 (95% CI: 0.652–0.795) in the validation group (Fig. 6a and b), indicating a good discriminatory performance that outperformed those indices alone. Calibration curves demonstrated excellent agreement between predicted and observed probabilities. As is shown in Fig. 6c and d, the x-axis represents the predicted risk, while the y-axis indicates the observed outcome. The bias-corrected lines in both cohorts are close to the ideal 45-degree diagonal line, demonstrating good agreement between the predicted probabilities and actual observations. The calibration of the model was also confirmed by the Hosmer-Lemeshow test with a p value of 0.831. DCA in the development group (Fig. 6e) revealed that the model provided a higher net benefit across a wide range of threshold probabilities compared to the “treat all” or “treat none” strategies. DCA in the validation group (Fig. 6f) also indicated a good net benefit of the FPD model, supporting its clinical applicability for risk stratification and decision-making.

Fig. 5.

A web-based application incorprating multidimensional indicators for predicting FPD. a User interface of the web application; b Results of risk evaluation; c A risk factor radar chart showing the contributions of each factor to the final risk. FPG, fasting plasma glucose; GGT, γ-glutamyl transferase; WC, waist circumference; FLI, Fatty liver index; mFIB-4, modified fibrosis-4 index; NLR, neutrophil-to-lymphocyte ratio

Fig. 6.

ROC Curves, calibration curves, and decision curve analyses for the FPD prediction model. a, c, e Development gourp; b, d, f Validation group

Association between CT-quantified IPFD and predicted risk

The clinical utility of the model in evaluating IPFD was further validated by examining the correlation between model-predicted risk scores and CT-quantified IPFD, represented by P/S ratio. Linear regression analyses (Fig. 7) revealed a significant positive correlation between these variables in both the development group (R = -0.424, p < 0.0001) and the validation group (R = -0.419, p < 0.0001). Specifically, elevated predicted risk scores were associated with higher P/S ratios, indicating that the FPD model effectively captures the severity of fat deposition in the pancreas.

Fig. 7.

Correlation between model-predicted risk and CT-quantified IPFD (P/S Ratio). a Development gourp; b Validation group

Discussion

In this study, we developed and validated a clinical prediction model for FPD based on multidimensional predictors, incorporating demographic (age), anthropometric (WC), biochemistry (FPG, GGT), inflammatory (NLR) indicators and fatty liver indices (FLI, mFIB-4). This model shows a dose-response relationship between the predicted risk and the severity of IPFD, highlighting its precision in translating multidimensional indicators into quantifiable risk assessments, as well as its value as a non-invasive tool for clinically evaluating FPD.

FPD is characterized by abnormal fat accumulation in pancreatic tissue, with adipocytes infiltrating the intralobular and interlobular regions and lipid droplets accumulating in acinar cells [8]. The pathogenesis of IPFD primarily involves two distinct pathways: (1) fatty replacement, an irreversible process characterized by progressive adipocyte infiltration replacing damaged acinar cells; (2) metabolic-associated fatty deposition – a potentially reversible condition frequently observed in the contexts of obesity and metabolic syndrome, clinically termed FPD [26]. The second pathologic change constitutes the principal focus of the present study. Figure 8 illustrates the possible mechanisms behind FPD development and explains why the identified risk factors could effectively predict FPD.

Fig. 8.

Fat deposition in the pancreas: mechanisms, health implications, and multidimensional predictive factors. Diet, physical activity, aging, and genetic factors collectively contribute to the development of metabolic syndrome and obesity. In these circumstances, the deposition of fat in the liver and pancreas is often observed. A mutually reinforcing relationship exists between the two pathological conditions. Fetuin-A, a protein secreted by the liver, can act on adipocytes within the pancreas, leading to an increased release of pro-inflammatory chemokines such as IL-8, IL-6, MCP-1, and TNF-α, which in turn enhances immune cell infiltration. The resulting inflammatory microenvironment in the pancreas impairs β-cell function and promotes carcinogenesis. Additionally, under conditions of metabolic dysfunction, glucotoxicity and lipotoxicity also play a role in damaging β-cell function. During the development of fatty pancreas, certain indicators associated with key pathological changes may serve as reliable biomarkers for assessing IPFD and overall pancreatic health. These indicators were incorporated into the prediction model of the present study. MCP-1, monocyte chemotactic protein‐1; TNF-α, tumor necrosis factor-α

The causal relationship between FPD and PC have been confirmed by the results of Mendelian randomization study [27]. Several possible mechanisms have been proposed for explaining the oncogenesis impact of FPD. The PANDORA hypothesis [9], first proposed by Petrov, systematically elucidates the core mechanism of IPFD as a common driver of pancreatic diseases; specifically, IPFD triggers the development of diabetes and pancreatic cancer through lipotoxicity accumulation and microenvironmental disruption. The most frequently studied way is that hypertrophic adipocytes in the fatty pancreas secrete various adipokines and cytokines, resulting in fibrotic response formation, exacerbation of the inflammatory state, and promotion of cell proliferation, all of which can contribute to cancer development or progression [28, 29]. Additionally, patients with FPD frequently exhibit concurrent obesity and diabetes, both of which are established risk factors for pancreatic cancer [30]. Pancreatic cancer cells also exhibit close crosstalk with surrounding adipocytes, where cancer cells reprogram adipocytes into cancer-associated fibroblasts [31] and promote adipocyte lipolysis via exosomal signaling [32]. Given these tumor-microenvironment interactions that may influence the evaluation of steatosis, pancreatic tumor patients were not included in the development group for model development.

Under physiological conditions, a normal pancreas may have a fat content of 5.0 to 10.4% [8], but the threshold for pathological fat deposition that defines FPD remains controversial, necessitating further investigation to distinguish between normal physiological fat distribution and disease-associated pathological changes. The P/S ratio derived from CT scans can quantitatively assess the severity of IPFD, which had been reported to be related to PC [19], but there is still no consensus on the threshold for diagnosing fatty pancreas using this index, limiting its application. Both a P/S ratio < 0.7 [19] and < 0.8 [20] had been used in previous studies as thresholds indicating the presence of fatty pancreas, whereas, without validation. Thus, the present study conducted, to our knowledge, the largest sample size investigation to evaluate the association between P/S ratio and the occurrence of pancreatic disease (PC, DM). The results indicated that a P/S ratio less than 0.8 was significantly associated with an increased risk of PC and DM. By integrating our findings with the reported thresholds in the literature, we confirmed 0.8 as an effective disease-related threshold for diagnosing FPD.

FPD should be recognized as a critical health concern, yet it remains underappreciated. A PubMed search reveals that literature on fatty liver is nearly 20 times more abundant than that on fatty pancreas. Clinically, a significant proportion of fatty liver patients exhibit concurrent fatty pancreas, both serving as results of metabolic syndrome [8]. There is a close pathophysiological interplay between FPD and fatty liver disease, centered on metabolic syndrome, pancreatic β-cell dysfunction, and visceral fat deposition [8]. For example, a hepatokine secreted by fatty liver called Fetuin-A could activate adipocytes and macrophages within pancreatic islets, creating a pro-inflammatory microenvironment that exacerbates β-cell dysfunction [33], which further promotes insulin resistance and ectopic fat deposition in other tissues, including the liver [8]. This cyclic interaction between the liver and pancreas is known as the twin cycle hypothesis [34]. Given these mechanistic links, we hypothesized that indices predictive of fatty liver could also aid in identifying FPD. We found significant correlations between FLI, mFIB-4 and fatty pancreas, second only to FPG in predictive value; these are novel observations that have not been previously reported.

Researches have demonstrated that adipocytes could secrete pro-inflammatory adipokines/cytokines, such as leptin, interleukin 1β, tumor necrosis factor and monocyte chemotactic protein‐1 (MCP‐1) [35, 36], triggering a local and systemic inflammatory response. NLR, calculated as a simple ratio between the neutrophil and lymphocyte counts measured in peripheral blood, plays a role as a flag of immune system homeostasis. NLR had been demonstrated to reflected the chronic inflammation caused metabolic syndrome, with a positive correlation with the severity of metabolic syndrome [37]. Moreover, NLR was significantly associated with a western dietary pattern and parameters of obesity, including high body fat, high waist and hip circumference [38]. These evidences are consistent with our finding that patients with FPD were associated with an elevated NLR level, which has not been reported in the previous literature. To enhance the performance in predicting FPD, the model of this study integrates these variables described above together with age, waist circumference, FPG and GGT, which has been reported to associated with FPD [21, 39]. Notably, WC is a better indicator of the severity of IPFD than BMI, the results of Wang et al. also reveal that central obesity may represents a more important obesity index than BMI > 25 (Odds ratio: 1.908 vs. 2.163) [21].

A prediction score for fatty pancreas was proposed by Khoury et al. [40], which includes three risk factors: fatty liver, hyperlipidemia, and a BMI > 30 kg/m². While fatty liver and fatty pancreas are associated, using ultrasound as a model variable limits its utility due to simultaneous organ assessment. In this context, fatty liver indices may serve as more practical predictors. For individuals with FPD, the reversal of pancreatic fat deposition is achievable through dietary modification, physical exercise, and pharmacological approaches. The core of these interventions is weight reduction. Restricting dietary energy intake can significantly decrease pancreatic fat content and improve β-cell function [41]. Exercise is also effective in reducing pancreatic fat content and improving insulin sensitivity, as evidenced by randomized controlled trials [42, 43]. Regarding pharmacological interventions, although glucagon-like peptide-1 receptor agonists (GLP-1 RAs) have shown potential in reducing pancreatic fat [8], further evidence is required to confirm their efficacy.

There are several limitations to this study. First, while the model was developed using a large dataset and underwent internal validation with a temporally independent cohort, its single-center design remains a limitation to its broader generalizability. The patient characteristics and clinical practices at a single institution may not fully represent those in other settings. Therefore, external validation in independent cohorts is essential. We are currently planning multi-center studies to evaluate the model’s performance across geographically diverse regions and various ethnic populations. Secondly, despite evidence that the P/S ratio serves as a reliable indicator of the histological pancreatic fat fraction [25], this study lacked direct histological confirmation. The acquisition of pancreatic tissue was precluded by its invasiveness and ethical concerns. Finally, the clinical utility of the model for longitudinal assessment of IPFD changes in response to fat-lowering interventions requires further investigation through multicenter prospective studies.

Conclusions

This study developed and validated a web-based model for predicting FPD, which enables comprehensive and individualized risk assessment using multidimensional indicators that are widely accessible in clinical settings. The value of this tool lies in its ability to serve as a preliminary screening instrument in populations and to guide subsequent examinations and interventions. Future external prospective validation is warranted to confirm its clinical utility.

Abbreviations

FPD

Fatty pancreas disease

FL

Fatty liver

FP

Fatty pancreas

PC

Pancreatic cancer

DM

Diabetes mellitus

BMI

Body mass index

WC

Waist circumference

DM

Diabetes mellitus

ALT

Alanine aminotransferase

AST

Aspartate aminotransferase

GGT

γ-glutamyl transferase

TG

Triglyceride

TC

Total cholesterol

LDL

Low-density lipoprotein cholesterol

HDL

High-density lipoprotein cholesterol

FPG

Fasting plasma glucose

PLT

Platelet count

Neut

Neutrophil count

Lym

Lymphocyte count

NLR

Neutrophil-to-lymphocyte ratio

FORNS

the Forns index

FLI

Fatty liver index

mFIB-4

Modified fibrosis-4 index

P/S ratio

mean pancreatic CT number to the mean splenic CT number

ROC

Receiver operating characteristic

AUC

Area under the ROC curve

DCA

Decision curve analysis

ROI

Regions of interest

PACS

the Picture Archiving and Communication System

HU

Hounsfield unit

MCP-1

Monocyte chemotactic protein‐1

GLP-1 Ras

Glucagon-like peptide-1 receptor agonists

TNF-α

Tumor necrosis factor-α

Authors’ contributions

YZ and CC: Conceptualization, Investigation, Methodology, Formal analysis, Software, Writing-original draft, Writing-review & editing. FC: Investigation, Data curation. WX: Methodology, Formal analysis. HD: Methodology, Software, Data curation. HL: Formal analysis, Writing-review & editing. TT: Methodology, Writing-review & editing. HH: Project administration, Supervision, Writing-review & editing. FL: Funding acquisition, Conceptualization, Supervision, Writing-review & editing. All authors have reviewed and approved the final version of the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (Grant number: 82073139), the Joint Funds for the Innovation of Science and Technology, Fujian Province (Grant number: 2021Y9058) and the Medical Minimally Invasive Center Program of Fujian Province and National Key Clinical Specialty Discipline Construction Program, China.

Data availability

The datasets used and analyzed during the current study available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the World Medical Association Declaration of Helsinki, and with the approval of the Ethics Committee of Fujian Medical University Union Hospital. Informed consent was obtained from all individual participants included in the study.

Consent for publication

Not Applicable.

Competing interests

The authors declare no competing interests.