A novel prediction model of pancreatic fistula after pancreaticoduodenectomy using only preoperative markers

For the training analysis, patients who underwent PD at the Department of Gastroenterological, Breast and Endocrine Surgery, Yamaguchi University Graduate School of Medicine, were selected. A total of 188 patients underwent PD between January 2010 and December 2020. Patients were excluded from the analysis if they met any of the following criteria: hepatopancreatoduodenectomy (n = 2) and two-stage pancreatojejunostomy [14] (n = 6). Ultimately, 180 patients were enrolled in the training set.

For the validation analysis, 371 patients, who underwent PD at the Department of Gastroenterological Surgery, Osaka University Graduate School of Medicine, from January 2011 to December 2020, were selected. Inclusion and exclusion criteria for study participation were similar to those for the training set. Patients were excluded from the analysis if they met any of the following criteria: hepatopancreatoduodenectomy (n = 2), PD for remnant pancreatic lesions (n = 2), and PD concomitant with colectomy for colon cancer (n = 1). Finally, 366 patients were enrolled in the validation set.

The collected data were statistically analyzed and assessed at Yamaguchi University. Approval from the institutional review board was obtained from each institution (IRB number: H2021-163). The study was conducted in accordance with Good Clinical Practice guidelines and the Declaration of Helsinki.

Each patient was evaluated using routine preoperative computed tomography (CT) or magnetic resonance imaging images and clinical data, including preoperative characteristics, postoperative complications, and histopathological diagnosis, from a prospectively maintained computer database. The collected data included patient age, sex, height, weight, body mass index (BMI),　presence/absence of diabetes, serum albumin level, total peripheral blood lymphocyte count, histopathological diagnosis, main pancreatic duct (MPD) size, stump thickness, stump width, preoperative biliary drainage (yes/no), and preoperative therapy. The prognostic nutritional index (PNI) [15] and controlling nutritional status (CONUT) score [16], indicators of nutritional status, were assessed in accordance with the following equation, as described previously.

Preoperative radiological markers included visceral fat area (VFA) [17], parenchymal thickness, MPD size, MPD index [18], presence of sarcopenia, and preoperative diagnosis [19].

According to the 2016 ISGPS guidelines, POPF was defined as a drain output of any measurable volume of fluid with amylase levels > 3 times the upper institutional limit of normal for serum amylase for each specific institution and association with clinically relevant development [6]. POPF was classified into three grades (A-C). Grade A implied that there was no deviation from the normal postoperative procedure and no impact on postoperative hospital stay duration. Grade B POPF required a change in the management of the expected postoperative pathway, including persistent drainage for > 3 weeks, percutaneous or endoscopic drainage, and angiographic procedure for bleeding. Grade C POPF led to organ failure, secondary operation, or subsequent POPF-related mortality. Grades B and C were defined as CR-POPF.

We extracted predictive markers of CR-POPF from the previous literature, and the following 14 clinicopathological markers were used as CR-POPF predictors: age [20], sex [21], BMI [22], VFA [23], PNI [24], serum albumin level [22], parenchymal thickness [25], MPD size [26], MPD index [18], presence of diabetes mellitus [27], presence of preoperative biliary drainage [28], CONUT score [29], presence of sarcopenia [30], and preoperative diagnosis [21].

Among the candidate markers, a combination of two markers was selected under the conditions and examined. There were N training samples in total. Of note, marker selection was performed using only the training samples. In pattern recognition fields, markers cannot be selected based on their individual effectiveness [31]. Therefore, the combination of these markers should be carefully selected. As shown in Fig. 1, we used the leave-one-out method [32] to identify the optimal combination of markers. According to this method, one training sample was selected as a sub-test sample from N training samples, and the remaining N­1 training samples were assigned as sub-training samples. To explore marker combinations, we initially selected one combination of two markers. The discrete Bayes classifier (Additional File) was designed using N­1 sub-training samples, and the re-substitution estimate was obtained by classifying N­1 sub-training samples using the classifier. Next, feature criteria, such as sensitivity, specificity, and F1 measure, were calculated. One sub-test sample was not used for marker selection. This process was repeated until all two-marker combinations were evaluated. Among all two-marker combinations, a combination with either maximal sensitivity, subjected to a specificity ≥ 50%, or a maximal F1 measure was selected. In the leave-one-out method, the aforementioned process was repeated N times (i.e., until each training sample had been selected only once as a sub-test sample). Among the N resulting combinations, the most frequently selected combination in the leave-one-out loop was considered optimal. When the number of markers was 3, marker selection was repeated according to the same procedure. Additional details concerning the discrete Bayes classifier were provided in previous studies [13, 33, 34].

Table 1 showed the demographic and clinicopathological characteristics of 180 patients from Yamaguchi University and 366 patients from Osaka University who underwent PD. According to the ISGPF classification, CR-POPF occurred in 60 (33%) patients of the training set and 130 (36%) patients of the validation set, with no significant difference between two groups. In-hospital mortality was 1.7% in the training set and 0.6% in the validation set, with no significant difference between the two groups. There were no significant differences in the baseline and clinical data between the training and validation sets, except that the training samples had higher preoperative albumin levels (P < 0.001), higher PNI (P < 0.001), thicker pancreatic parenchyma (P = 0.0007), larger MPD size (P = 0.023), higher frequency of samples with CONUT score ≥ 3, higher frequency of samples with pancreatic ductal adenocarcinoma (PDAC) (P = 0.0004), and lower frequency of preoperative radiation (P < 0.0001). Regarding pancreatic-enteric anastomosis, all were pancreatojejunostomy with no pancreatogastrostomy in the training set, whereas in the validation set, pancreatojejunostomy and pancreatogastrostomy were performed in 335 (91.5%) and 31 (8.5%) samples, respectively. However, there was no difference in operative time and blood loss between the training and validation sets.

Initially, we performed analyses on all cases. Nevertheless, we could not develop highly accurate risk models (Additional Table). It has been reported that the pancreatic exocrine function is involved in pancreatic fistulas; generally, the exocrine function is often impaired in pancreatic cancer compared with non-pancreatic cancer [35]. Therefore, we differentiated between pancreatic and non-pancreatic cancers; moreover, we performed analyses separately within each group by categorizing all cases into pancreatic cancer and non-pancreatic cancer groups. Marker selection was performed using 84 PDAC training samples from a total of 180 available training samples from Yamaguchi University. According to the leave-one-out method with 84 PDAC training samples on the discrete Bayes classifier, the optimal combination of predictive markers was determined (Fig. 1), and classification performance for the combination was then evaluated. Table 2 shows the diagnostic potential of CR-POPF using various combinations of predictive markers in the discrete Bayes classifier. For PDAC, we could differentiate CR-POPF with 87% sensitivity, 81% specificity, 50% positive predictive value, 97% negative predictive value, and F1 measure of 0.63, using only the MPD index. In contrast, the results of the two-marker combination on the discrete Bayes classifier were as follows: 93% sensitivity, 73% specificity, 44% positive predictive value, 98% negative predictive value, and F1-measure of 0.60 for the combination of MPD index and BMI; 73% sensitivity, 94% specificity, 73% positive predictive value, 94% negative predictive value, and F-measure of 0.73 for the combination of MPD index and presence of sarcopenia; 87% sensitivity, 81% specificity, 50% positive predictive value, 97% negative predictive value, and F-measure of 0.63 for the combination of MPD index and PNI; and 87% sensitivity, 81% specificity, 50% positive predictive value, 97% negative predictive value, and F-measure of 0.63 for the combination of MPD index and presence of preoperative biliary drainage. Next, marker selection was conducted for 96 non-PDAC training samples. The design of the discrete Bayes classifier with the selected combination and the evaluation of its classification performance were performed in the same manner as for PDAC samples. MPD index as a single marker resulted in 78% sensitivity, 63% specificity, 65% positive predictive value, 72% negative predictive value, and F-measure of 0.71. In contrast, we could differentiate CR-POPF with 84% sensitivity, 57% specificity, 63% positive predictive value, 80% negative predictive value, and F-measure of 0.72 using the combination of MPD index and BMI. The results of the three-marker combination on the discrete Bayes classifier were as follows: 98% sensitivity, 39% specificity, 59% positive predictive value, 96% negative predictive value, and F-measure of 0.73 for the combination of BMI, VFA, and MPD size; 84% sensitivity, 57% specificity, 63% positive predictive value, 80% negative predictive value, and F-measure of 0.72 for the combination of MPD index, age, and BMI; and 84% sensitivity, 57% specificity, 63% positive predictive value, 80% negative predictive value, and F-measure of 0.72 for the combination of MPD index, presence of sarcopenia, and BMI.　Therefore, we adopted the MPD index as a single marker for PDAC and MPD index/BMI as a two-marker combination for non-PDAC. For all samples in the training set, this prediction model showed 85% sensitivity, 71% specificity, 70% positive predictive value, and 90% negative predictive value. Based on these results, we developed a prediction model as shown in Fig. 2.

We used a validation set of 366 patients from Osaka University, which was independent of the training set. The results are shown in Tables 3 and 4. For 130 PDAC test samples in the validation set, the prediction model using only the MPD index had 93% sensitivity, 87% specificity, 47% positive predictive value, and 99% negative predictive value. For 236 non-PDAC test samples in the validation set, the prediction model using the two-marker combination of MPD index and BMI showed 85% sensitivity, 53% specificity, 64% positive predictive value, and 78% negative predictive value. For all test samples in the validation set, this prediction model showed 86% sensitivity, 71% specificity, 62% positive predictive value, and 90% negative predictive value.