Perioperative risk of pancreatic head resection—nomogram-based prediction of severe postoperative complications as a decisional aid for clinical practice

Patient data from 956 consecutive pancreatic head resections for several indications were collected in a prospectively-maintained database and evaluated retrospectively. Operations were performed between August 2001 and October 2018. The study was reviewed and approved by the Ethics Committee of the University of Freiburg (EK 85/20).

According to the study protocol, the following clinical and laboratory data were extracted from the prospectively-maintained database: demographic data (age, gender), clinical data (indication for surgery, ASA score [17], body mass index (BMI), performed operation including reconstruction, intraoperative blood loss, intraoperative blood transfusions, operation time, texture of pancreatic tissue, postoperative complications, diagnostics of complications, and treatment of complications and all available laboratory data of routine laboratory monitoring. Complications were classified according to the Clavien-Dindo general classification of complications [18] and the comprehensive complication index (CCI) [19]. Pancreatic surgery-specific complications such as postoperative pancreatic fistula (POPF) [20], delayed gastric emptying (DGE) [21], and post-pancreatectomy hemorrhage (PPH) [22] were classified according to International Study Group of Pancreatic Surgery (ISGPS) definitions.

All available results of the routine perioperative laboratory monitoring were collected from the preoperative laboratory evaluation to postoperative day (POD) 14. Laboratory samples were categorized according to the postoperative time, beginning from the preoperative sample (POD 1) to the postoperative samples on POD 1 to 3 (POD 1–3), POD 4 to 6 (POD 4–6), POD 7 to 9 (POD 7–9), and POD 10 to 14 (POD 10–14). For each time interval, the maximum observed value of a certain parameter was chosen as the value for the corresponding time interval. The following parameters of routine laboratory monitoring from blood-samples were collected in the study database: white blood cell (WBC) count, C-reactive protein (CRP), procalcitonin (PCT), serum amylase (amylase), serum lipase (lipase), aspartate aminotransferase/glutamate oxaloacetate transaminase (AST/GOT), alanine aminotransferase/glutamate pyruvate transaminase (ALT/GPT), gamma-glutamyltransferase (GGT), alkaline phosphatase (AP), serum total bilirubin (bilirubin), creatinine and blood urea nitrogen (BUN), the preoperative carbohydrate antigen 19-9 (CA19-9) value, the preoperative International Normalized Ratio (INR) value, and the preoperative partial thromboplastin time (PTT). Drainage samples were evaluated for drain amylase on the first three postoperative days. Acute kidney injury on POD 1–3 was diagnosed when creatinine increased 1.5 times over the preoperative value according to RIFLE-criteria [23].

Patients were randomly assigned to two groups of about 80% and 20% of the entire cohort. A total of 772 patients (80.8%) were assigned to the test cohort and 184 (19.2%) were assigned to the validation cohort. Results are expressed as median ± interquartile range (IQR) or as number (percent). Categorical variables were compared using Pearson’s chi-square test. Mann-Whitney U-test or Kruskal-Wallis test with Mann-Whitney U-post hoc test and Bonferroni correction were used for descriptive analysis of non-parametric variables.

As the aim of this study was the development of a diagnostic screening tool to select patients at risk for postoperative complications after pancreatic head resection, patients of the test cohort were grouped according to Clavien-Dindo classification Grades 0–II and III–V. Univariate analysis was performed using Pearson’s chi-square test or Mann-Whitney U-test as applicable. Variables with a p-value < 0.1 were included in multivariate analysis. In addition to this variable selection, only one variable of possibly confounding variables (e.g., only one variable for each organ system) was selected for multivariate analysis and—in respect of the study aim—only preoperatively-known parameters were included in the multivariate analysis for the preoperative nomogram and only parameters with a known outcome on POD 3 were included in the variable selection for the multivariate analysis of the postoperative nomogram. For multivariate analysis and nomogram development, continuous variables were dichotomized by the optimal cutoff-value according to the results of the receiver operating characteristics (ROC) analysis with Youden index calculation [24]. Cutoff-values below laboratory threshold values were set to the laboratory threshold value (GPT POD 1 and AP POD 1). Multivariate analysis was then performed by binary logistic regression with backward stepwise variable selection. Nomograms were constructed based on the developed multivariate models [25]. Internal validation in the test cohort was performed by bootstrap resampling with 100 repetitions. External validation was performed with the validation cohort and the area under the curve (AUC) of the ROC analysis. The fistula risk score (FRS) was calculated as previously reported by Callery et al. [26]. Statistical analysis was performed using SPSS version 24.0.0.1 (IBM Corp., Armonk, NY, USA) and R version 4.0.0 (R Foundation for Statistical Computing, Vienna, Austria) with RStudio (RStudio Inc., Boston, MA, USA) and additional packages rms, ggplot, and pROC. Differences were considered statistically significant at p < 0.05.

The clinicolaboratory characteristics of the test cohort are summarized in Tables 1 and 2 and of the validation cohort in “Supplementary information” Table S1. During the perioperative course, several parameters of routine laboratory monitoring of the patients showed statistically significant differences between patients with and without severe postoperative complications. WBC showed an early postoperative peak on POD 1–3—as expected—with decreasing values in both groups of patients on POD 4–6 followed by a second increase in both groups on POD 7–9. While the patients without severe complications also showed an increase in WBC on POD 7–9, patients with severe complications showed persisting leukocytosis on POD 10–14 and significantly elevated WBC during the entire postoperative course (Fig. 1A). CRP showed higher elevation on POD 1–3 in patients with severe complications and persisting CRP-values > 100 mg/l after POD 10 in contrast to adequately decreasing CRP-values in patients without severe complications (Fig. 1B). PCT showed elevated values in patients with severe complications as well, but was excluded from further statistical evaluation due to the low number of performed tests and a high degree of patient selection. Serum amylase showed a peak on POD 1–3 in patients with severe complications, indicating postoperative pancreatitis as a sign for developing POPF while serum amylase-values below clinical threshold value on POD 1–3 were associated with an uneventful postoperative course. Interestingly, although serum amylase normalized below clinical cutoff-value in both groups, high-normal amylase values (≥ 7 U/l) were seen in the group with postoperative complications while significantly lower-normal values were observed in those patients without complications (Fig. 1C). Serum lipase values showed similar variations as the serum-amylase during the postoperative period. A significant difference between the patients with and without severe postoperative complications could be observed for the renal retention parameters creatinine and BUN for the entire perioperative time, indicating that chronic and acute renal failure are associated with perioperative risk of complications. The GOT showed an early peak on POD 1–3 in patients with complications, indicating its usefulness in diagnostics of postoperative liver perfusion complications. Bilirubin showed no differences in the early postoperative course but patients with complications showed elevated bilirubin after POD 7. An interesting observation was made concerning alkaline phosphatase. In our study cohort, higher preoperative values are associated with lower risk for postoperative complications. A statistically significant association between the texture of the pancreatic tissue and the preoperative value of alkaline phosphatase could be detected (p < 0.001). Correspondingly, a significant difference between the POPF and no POPF-groups was observed as well (p < 0.001, Kruskal-Wallis ANOVA with Mann-Whitney U-post hoc test, Supplementary information Table S2).

Univariate analysis revealed several parameters associated with perioperative complications Clavien-Dindo Grade III or higher (Tables 1 and 2). In the preoperative setting, the patient’s age, the indication for surgery, CA19-9 value, ASA score, BMI, WBC POD 1, GPT POD 1, AP POD 1, creatinine POD 1, and BUN POD 1 were significantly associated with postoperative complications. The indications for surgery were associated with different risks for postoperative complications. Pancreatic surgery for IPMN, bile duct cancer, duodenal cancer, and “Other indication” (including MCNs, SCNs, schwannoma, GIST, and sarcoma) was associated with a higher risk for complications, while surgical resection of PDAC, chronic pancreatitis, pancreatic neuroendocrine tumors (pNET), and pancreatic metastases was associated with a reduced perioperative risk. Therefore, the indications were classified as indications with a high risk (IPMN, bile duct cancer, duodenal cancer, and “Other indications”) and indications with a low risk (PDAC, CP, pNET, and metastases) for postoperative complications. Multivariate analysis identified indication (risk-stratified), ASA score, BMI, WBC POD 1, and preoperative alkaline phosphatase as independent risk factors for postoperative complications (Table 3).

The associations of perioperative parameters with a clinically known outcome on the third postoperative day and postoperative complications were evaluated. Added to the preoperatively known parameters, operation time, the need for intraoperative blood transfusion, texture of pancreatic tissue, maximum drain amylase on POD 1–3, WBC POD 1–3, CRP POD 1–3, PCT POD 1–3, (serum) amylase POD 1–3, (serum) lipase POD 1–3, GOT POD 1–3, GPT POD 1–3, creatinine POD 1–3, BUN POD 1–3, and acute kidney injury on POD 1–3 are associated with postoperative complications. Multivariate analysis revealed that the indication (risk-stratified), ASA score, BMI, the need for intraoperative blood transfusion, prolonged operation time, WBC POD 1–3, (serum) amylase POD 1–3, and the preoperative alkaline phosphatase were independent prognostic factors of postoperative complications Clavien-Dindo Grade III or higher (Table 4). The texture of the pancreatic tissue is correlated with AP POD 1 and serum amylase POD 1–3, and therefore, these quantifiable variables without interobserver variability were used for multivariate analysis.

Based on the results of the multivariate analyses, two nomograms were developed (Fig. 2A and B, a step-by-step user instruction is available in the “Supplementary information” section Fig. S1). Calibration plots with bootstrap resampling were generated to correlate predicted and empirically observed probabilities of postoperative complications for internal validation. The calibration plots demonstrated high reliability in predicting postoperative complications for the preoperative and postoperative nomogram (Fig. S2A and B).

External validation was performed by ROC analysis in the validation cohort. The AUC of the preoperative nomogram was 0.673 (95%, CI: 0.626–0.721; p < 0.001) in the test cohort and 0.676 (95%, CI: 0.586–0.766; p = 0.001) in the validation cohort. The AUC of the postoperative (POD 1–3) nomogram was 0.734 (95%, CI: 0.691–0.778; p < 0.001) and 0.677 (95%, CI: 0.591–0.762; p = 0.001) in the validation cohort. WBC showed an AUC of 0.552 on POD 1 (95%, CI: 0.502–0.603; p = 0.041), 0.569 on POD 1–3 (95%, CI: 0.518–0.619, p = 0.008), 0.670 on POD 4–6 (95%, CI: 0.623–0.716; p < 0.001), and 0.770 on POD 7–9 (95%, CI: 0.728–0.812; p < 0.001). The predictive value of the postoperative nomogram was therefore comparably accurate on POD 3 as the diagnostic value of WBC on POD 7–9. In the validation cohort, the AUC of the WBC was 0.562 on POD 1–3 (95%, CI 0.465–0.659, p = 0.212) and 0.706 (95%, CI: 0.618–0.793; p = 0.045) on POD 7–9 (Fig. 3A and B).

Although not designed for this specific task, prediction of Clavien-Dindo Grades IV and V also was possible with the nomograms (Supplementary information Table S1 and Supplementary information Fig. S3). We compared our nomogram with the FRS—which is calculated postoperatively—in a cohort of 57 patients from our validation cohort, for which all data needed for the calculation of the FRS were available. AUC for prediction of complications Clavien-Dindo Grade III or higher was 0.672 (p = 0.030) for the FRS, 0.665 (p = 0.033) for the preoperative nomogram, and 0.674 (p = 0.020) for the postoperative nomogram. Therefore, both postoperative scores were almost equally efficient in the prediction of postoperative complications and the preoperative nomogram only slightly less accurate. Combination of the FRS and our nomograms by simple addition of normalized score values showed that the FRS and the preoperative nomogram improved the predictive accuracy to AUC 0.696 (p = 0.012) and the FRS and the postoperative nomogram to AUC 0.682 (p = 0.017).

When comparing the risk for certain postoperative complications, different patterns of complications due to the different preoperative risk factors can be observed. While patients with high-risk indications or a low preoperative alkaline phosphatase have a higher risk for POPF and subsequently a higher risk for postoperative sepsis, intensive care-associated complications such as pneumonia, thromboembolism, and acute renal failure are more likely to occur in patients with a higher ASA score. The first diagnostic procedure to diagnose postoperative complications was performed on POD 8 (4–11.5) (computed tomography: n = 212 and endoscopy: n = 54) and the first intervention for treatment of postoperative complications was also performed on POD 8 (4–12). The first postoperative interventions are summarized in Table S6. Elevated WBC (≥ 12.5 × 10³/μl) and CRP (≥ 85 mg/l) levels in the later postoperative course POD 7–9 implicate a higher rate of operative revision procedures and less interventional procedures, indicating a higher severity of the clinical condition of patients with these laboratory findings (Supplementary information Tables S5, S6, and S7).