A novel model based on clinical and computed tomography (CT) indices to predict the risk factors of postoperative major complications in patients undergoing pancreaticoduodenectomy

Study population

The study population were retrospectively enrolled between January 2019 and December 2023 from two centers, Jinling Hospital and The Affiliated People’s Hospital of Jiangsu University. The flowchart of this study is shown in Fig. S1.

All patients admitted would first receive an assessment of surgical resectability by the multidisciplinary team (MDT) through imaging to determine curative resectability of tumors. The inclusion criteria are all patients that underwent upfront PD during the study period at the designated institutions by expert pancreatic surgeons. Finally, a total of 360 patients were included in the study. Subsequently, their medical records, laboratory results, and imaging data of these patients were collected and systematically organized for further analysis. In terms of the Clavien-Dindo system, the cohort was divided into those with or those without major complications. Written informed consent was obtained from patients prior to the research. This study conformed to the ethical guidelines of the 1975 Declaration of Helsinki and was approved by the Ethics Committee of Jinling Hospital (2023DZKY-049-03) and registered at ClinicalTrials.gov (NCT06378853).

Surgical procedures and perioperative management

All patients underwent an MDT discussion before treatment. The open Kausch-Whipple procedure is performed according to the NCCN guidelines (Tempero et al., 2021). The surgical procedure generally includes: opening the Kocher incision, organ resection, anastomotic reconstruction in hand-sewn by continuous suturing techniques, and irrigating the peritoneal cavity.

Drugs inhibiting secretion of digestive fluid such as proton-pump inhibitors and somatostain were administered routinely. Laboratory results, including routine blood and biochemical examinations, were collected on postoperative day (POD) 1. All patients in the study underwent abdominal CT or trans-abdominal ultrasound within the first week after surgery. Additional abdominal imaging was conducted when there was a sign of possible intra-abdominal complications. Oral diet, walking, and early withdrawal of drainage tubes were advocated early postoperatively. Liquid diet was gradually resumed around POD 2–5 and soft diet after defecation. Drainage tubes were removed appropriately on day 3 and day 5 after surgery if there were no signs of leakage and drainage fluid amylase levels were within normal range.

Data collection

The clinical information of all patients included age, sex, and BMI, diabetes, hypertension and ASA. The preoperative data closest to the time of surgery were collected for analysis, including white blood cell (WBC) count, red blood cell (RBC) count, platelet (PLT) count, hemoglobin (Hb) level, albumin (ALB) level, lymphocyte count, monocyte count, neutrophil count, triglyceride (TG) level, fasting plasma glucose (FPG) level, and CA199 level. The prognosis nutrition index (PNI) was calculated using the formula: PNI = serum ALB level (g/L) + 5 × total number of peripheral blood lymphocytes (×10⁹/L) (Onodera, Goseki & Kosaki, 1984), and the systemic inflammation response index (SIRI) was calculated using the formula: SIRI = total number of peripheral blood neutrophils (×10⁹/L) × total number of peripheral blood monocytes (×10⁹/L)/total number of peripheral blood lymphocytes (×10⁹/L) (Qi et al., 2016). The TyG index was calculated using the formula: TyG = Ln [TG (mg/dl) × FPG (mg/dl)/2] (Simental-Mendia, Rodriguez-Moran & Guerrero-Romero, 2008). The TyG-BMI index was defined as the TyG index multiplied by BMI (Kg/m²). The postoperative data such as maximum tumor size, length of hospital stay, and postoperative complications were analyzed. Postoperative complications were mainly evaluated based on the diagnostic criteria established by the International Study Group of Pancreatic Surgery (ISGPS) (Bassi et al., 2017; Koch et al., 2011; Wente et al., 2007a, 2007b).

Collection of CT data

A single trained radiologist analyzed CT images preoperatively using the HY-QCT medical software version V2.5.0. The software is based on high-quality CT imaging data and simulates the human visual nervous system’s method of detecting target objects in images through convolutional neural networks. It trains the computer to become a perceptual system composed of neurons that are extremely sensitive to the features of the spinal region in CT images. By processing, analyzing, and calculating the spinal region in CT images, it achieves segmentation detection of spinal vertebrae, morphological analysis of vertebrae, and tissue analysis. Then, the tissue composition, such as L3 SAV, VAV, abdominal wall muscular spatium fat, paravertebral muscular spatium fat, abdominal wall muscle, and paraspinal muscle were segmented automatically and their volume were calculated (Fig. 1).

Definition of cut-off value

The BMI ≤ 18.5 kg/m² indicates that weight is below the healthy range, suggesting the possibility of malnutrition, chronic illness, or other health issues according to clinical management and treatment of obesity in China (Zeng et al., 2021). CA199 ≥ 37 U/ml indicates that it exceeds the normal healthy range, suggesting the possibility of pancreatic cancer or other gastrointestinal tumors based on comprehensive guidelines for the diagnosis and treatment of pancreatic cancer (Pancreatic Cancer Committee of Chinese Anticancer A, 2021). Due to the lack of specific cut-off values in previous studies for VAV/SAV, SIRI, PNI, and TyG-BMI, we determined the optimal cut-off value for the classification performance of these indicators using the area under the receiver operating characteristic (ROC) curves. The optimal cut-off value was determined by maximizing specificity and sensitivity using Youden’s Index. Therefore, SIRI ≥ 0.86 was defined as high SIRI, PNI ≤ 47.65 as low PNI, TyG-BMI ≥ 200.11 as high TyG-BMI, and VAV/SAV ≥ 1.59 as high VAV/SAV.

Feature selection and nomogram development

From all demographic features and preoperative clinical data, to prevent overfitting and handle multicollinearity, these variables which showed differences between the two groups (P < 0.05) and other potential risk factors were evaluated by the LASSO regression. Then it produced 11 significant parameters (age ≥ 65, diabetes, hypertension, CA199 ≥ 37, NWR, SIRI, PNI, TyG-BMI, VAV/SAV, jaundice and Hb that had nonzero coefficients. And we obtained the lambda min value when the MSE is minimized, which indicates that the prediction error of the model is at its lowest.).

Then those significant parameters were inputted into the final multivariate regression analysis. The multivariate regression analysis conducted as described demonstrated that CA199 ≥ 37, high SIRI, low PNI, high TyG-BMI, and high VAV/SAV were independent risk factors for major complications after PD. Next, the stepAIC selection was performed to obtain the best prediction model. Based on the results, a nomogram model was established to predict major complications after PD.

Statistical analysis

Continuous variables with a normal distribution are presented as the mean and standard deviation (SD) and were analyzed using the Student’s t-test. Continuous variables with a non-normal distribution are presented as the median with interquartile range (IQR), and the Mann-Whitney U test was used for analysis between the two groups. Chi-squared or Fisher test was used to compare the distribution of categorical variables between groups. Two-sided p-values less than 0.05 were considered significant. All statistical analyses were conducted using IBM SPSS Statistics for Windows, version 25.0 (IBM Corp., Armonk, NY, USA) and R version 4.0.5 (R Core Team, 2021).

The model building followed a predefined plan. Firstly, data cleaning and imputation were conducted for datasets. Variables with missing percentages of more than 10% were deleted, and other variables were filled with multiple imputation methods (Mice package in R). Secondly, least absolute shrinkage and selection operator analysis was used to filter valuable diagnostic variables. Next, multivariate logistic regression analyses were used to determine independent risk factors for major complications in patients undergoing PD. StepAIC selection was then performed to obtain the best prediction model. Nomograms were constructed based on the proportional conversion of each regression coefficient from multivariate logistic regression onto a 0- to 100-point scale. Then, the nomogram was internally validated by the training and test database.

The optimal cut-off value of the nomogram score was calculated using the maximum Youden index method. The area under the curve (AUC) was subsequently corrected by bootstrapping validation (1,000 bootstrap resamples) (Harrell, Lee & Mark, 1996; Steyerberg et al., 2001). The Cox-Snell R², Nagelkerke R² and Hosmer-Lemeshow R² (HL R²) were used to fit of the model. The calibration curve, and decision curve analysis (DCA) were used to evaluate this model.

Clinicopathological characteristics of patients and incidence of postoperative complications

Of 415 PDs performed at the participating institutions during the study period, 360 (86.7% in relation to 415) met the criteria for analysis. The clinical pathological characteristics of these patients were shown in Table 1. There were 173 males (48.1%) and 187 females (51.9%). The median age of the patients was 61.0 ± 10.6 years. Among baseline characteristics, significant differences were noted for age, diabetes BMI, lymphocyte, neutrophil, FPG, TG, and CA199 between the two groups. Tumor size and duration of hospital stay differed significantly between the two groups.

All postoperative complications were present in Tables S1 and S2. There were 48 cases (13.3%) of major complications in the total cohort. A total of 53 (14.7%) patients experienced pancreatic fistula, 16 (4.4%) patients had bile leakage and 6 (1.7%) patients had intestinal fistula. The rates of abdominal and pulmonary infection were 14.2% and 6.9% respectively. The PPH occurred in 28 (7.8%) patients. The DGE and organ dysfunction rates were 9.2% and 2.5%, respectively.

Risk-associated factors of major complications after PD

The results of LASSO are shown in Figs. S2 and S3. Then those significant parameters were inputted into the multivariate regression analysis. The results of multivariate regression analysis for risk factor selection are as follows (Table 2). To obtain the best model, we used the stepAIC method to select variables: one variable is added at a time, but at each step, the variables are re-evaluated. Variables that do not contribute to the model will be removed. The AIC value is calculated to assess the quality of each candidate model, and this process continues until the optimal model is obtained. Cox-Snell R², Nagelkerke R² and HL R² are commonly used metrics to assess the goodness of fit of the selected model. As demonstrated in Table 3, the goodness-of-fit statistics for this model outperform those of the alternative models, suggesting that it exhibits superior explanatory power, accuracy, and a better fit to the data. The nomogram model to predict major complications after PD was established according to these results shown in Fig. 2. The total score was calculated using VAV/SAV, PNI, SIRI, TyG-BMI, and CA199. Each of these variables was assigned a score based on a scaled axis. The individual scores were then summed to obtain the total score. By mapping this total score to the corresponding point scale, we were able to estimate the risk of major postoperative complications in patients after PD. Contribution of each risk factor to the prediction model were shown in Table 3. The AUC and C-index of the nomogram was 0.854 (95% CI [0.800–0.907]), the positive and negative predictive value were 0.700 and 0.883 respectively, and accuracy was 0.878, indicating that the model exhibited good discrimination and substantial accuracy in predicting major complications after PD (Fig. 3).

Calibration and validation of the nomogram

The calibration curve via 1,000 bootstrap resampling was used to assess the predictive accuracy of the nomogram. The calibration curve for the probability showed good concordance between the predicted and actual observations and the absolute error between the simulated curve and the actual curve is 0.016 (Fig. 4), which demonstrated the nomogram had great accurate predictive ability. The DCA converts complex mathematical models into simple and easy-to-understand graphics to intuitively judge the practicability and net benefits of different models. Therefore, in order to further investigate the clinical utility of the prediction model, the DCA (Fig. 5) was drawn with the threshold probability as the abscissa and the net benefit rate of patients as the ordinate. It can be clearly shown that patients whose postoperative major complications were predicted were predicted based on the nomogram have a higher net benefit than any patient predicted by a single indicator, and the nomogram had a better potential for clinical application when the threshold probability was 0.10 to 0.80, indicating that the prediction model has great clinical value.

To test the applicability of our prediction model, patients were divided into the training database (n = 251) and test database (n = 109) randomly with a ratio of 7:3. Internal validation of this model showed that the AUC was 0.864 and data from the test database was used for cross validation of the model, showing the AUC of 0.824 (Fig. S4). The calibration curves of the training and test database also demonstrated that the prediction model was in great agreement with actual observations (Figs. S5, S6). The DCA curves also revealed that the model provided great clinical benefit in the training and test database (Figs. S7, S8).