Background
Pancreaticoduodenectomy (PD) is one of the most difficult surgical procedures. In recent years, the post-PD mortality rate at high-volume centers has decreased to < 2% owing to improvements in surgical techniques and perioperative management [
1,
2]. However, post-PD morbidity rates remain high (16–50%) [
3‐
5]. Among the most important post-PD complications, postoperative pancreatic fistula (POPF) can cause intra-abdominal hemorrhage and abscesses, leading to surgery-related deaths. Post-PD pancreatic fistula occurred in 15–45% of patients, according to the 2016 edition of the International Study Group of Pancreatic Surgery (ISGPS), and was associated with a mortality rate of up to 9% [
6].
According to the ISGPS diagnostic criteria, POPF is classified as biochemical leak (without adverse clinical consequences) or clinically relevant POPF (CR-POPF), which is severe and requires various treatments. In recent years, proactive, risk-based management of pancreatic anastomosis has been advocated before serious complications occur with CR-POPF [
7,
8]. Therefore, it is important to preoperatively predict CR-POPF risk, which is difficult with a single factor because POPF is associated with a number of confounders, including disease-related factors (such as pancreatic texture) and patient-related factors (such as obesity) [
9,
10].
Therefore, in recent years, prediction models using multiple markers have been developed to predict POPF. However, most of them consist of intraoperative or postoperative markers, and few have preoperative predictive utility [
11,
12]. This study aimed to predict CR-POPF with a high degree of accuracy by focusing solely on preoperative clinicoradiological data. We attempted to prevent missing CR-POPF. We extracted predictive markers of CR-POPF from previous literature and differentiated CR-POPF using a unique classifier [
13].
Discussion
This study established a novel prediction model of CR-POPF after PD by using a discrete Bayes classifier with only preoperative parameters. As a result, the prediction model was constructed using the preoperative diagnosis of disease, MPD index (ratio of pancreatic duct diameter/pancreatic parenchymal diameter) [
18], and BMI to stratify the training samples into high- and low-risk groups. This prediction model was as accurate in the validation set as in the training set. The ISGPS recommends the use of an external pancreatic duct stent and somatostatin analogs for high-risk samples after PD [
8]. Recent reports also suggested that pancreatojejunostomy was better than pancreatogastrostomy for high-risk samples [
7]. However, there are several advantages to predicting the risk of CR-POPF not only in terms of the choice of surgical technique and perioperative management but also in other aspects. For example, based on the risk of CR-POPF, preoperative informed consent can be provided more cautiously. Specifically, in frail and vulnerable patients with benign or borderline tumors, or with uncertain pathological behavior, the preoperative risk prediction may be helpful when deciding whether to pursue conservative treatment or surgical intervention. To implement such a policy, it is important to identify high-risk samples preoperatively. Several useful single risk markers for pancreatic fistula were previously reported. However, POPF is confounded by multiple markers, and its accurate prediction based on single risk markers is difficult. Therefore, it is important to create a more accurate predictive model by combining multiple risk markers with documented usefulness. Several risk models for POPF using multiple markers have been reported. Many previous reports used intraoperative or postoperative parameters, such as intraoperative blood loss or drain findings, which make it difficult to preoperatively schedule the details of the surgical procedure or prepare the medications to be used [
36]. Nevertheless, in recent years, there have been several reports of predictive models using only preoperative parameters [
37‐
39]. These models combined independent risk markers extracted from univariate and multivariate analyses. However, the combination of multiple markers extracted in this manner is not always optimal. We differentiated CR-POPF by using an original discrete Bayes classifier [
13], which is unique and can handle both non-numerical and numerical data on the basis of Bayes’ decision theory using posterior probability. The leave-one-out method was adopted in the model design to explore the optimal combination of markers and evaluate the distinguishability. Using this estimation, we can eliminate ambiguity in the determination of CR-POPF and obtain an objective probability distribution.
Our risk model first classified PDAC and non-PDAC during the development of our risk model. As shown in the Additional Table, the classification performance for the training set with stratified PDAC vs. non-PDAC samples was superior to that of the training set with two groups combined. This finding might result from a difference in the mechanism by which pancreatic fistulas develop between PDAC and non-PDAC samples.
As previously reported, many PDAC samples have increased physical strength due to fibrosis caused by pancreatitis associated with pancreatic duct obstruction, resulting in “hard pancreas” [
40]. Additionally, in PDAC samples, the distal MPD is dilated and easily sutured during surgery. The passage of pancreatic juice through the anastomosis might be smooth. In PDAC samples, the decline of exocrine function may lead to a decreased frequency of CR-POPF [
41]. In contrast, most non-PDAC samples are “soft pancreas” with little MPD dilatation, little fibrosis of parenchyma, and preserved exocrine function. Therefore, PDAC or non-PDAC may be major predictors of a soft or hard pancreas, respectively. In PDAC samples, the classification performance of a single marker (MPD index) was optimal without requirement for combination with other markers. As mentioned above, most pancreatic samples are hard pancreas. However, not all PDAC samples are hard pancreas, and lesions in the uncinate process of the pancreas or groove region may not be hard pancreas because there is no MPD obstruction. As a result, the MPD index will be low in such samples because the pancreatic parenchyma does not undergo atrophy without MPD dilatation. Therefore, a lower MPD index may be a predictor of soft pancreas, which is susceptible to CR-POPF, in PDAC samples, and is clinically reasonable.
In contrast, for non-PDAC samples, the two-marker combination was superior to a single marker, but a three-marker combination was not necessary. As described above, most non-PDAC samples were soft pancreas with a thin MPD and non-atrophied parenchyma, resulting in a lower MPD index. These samples may be at a high risk for CR-POPF because of the fragile pancreas and preserved pancreatic exocrine function. However, even in non-PDAC samples, when the MPD is obstructed by an invading lesion, the MPD index is higher because of MPD dilatation and fibrotic atrophy of the pancreatic parenchyma, which results in a hard pancreas. Such samples are less likely to develop pancreatic fistulas because they have the same mechanism as PDAC samples. However, non-PDAC samples, such as some samples of intraductal papillary mucinous neoplasm, may sometimes have a dilated MPD and high MPD index but are soft pancreas without obstructive pancreatitis and fibrosis [
42]. In such samples, a high MPD index can be a false negative for CR-POPF risk. Therefore, BMI may be a complementary risk factor in such samples. BMI was reportedly associated with soft-textured fatty pancreas and CR-POPF. The physical fragility of the fatty pancreas may be the cause of CR-POPF [
10].
Although significant differences could not be examined because this was a one-time result, the performance of our predictive model for the test samples in the validation set was similar to that of the training sample. While the performance of the predictive model generally degrades with test samples as compared to training samples, no degradation was observed because our prediction model was highly accurate. In particular, the prediction accuracy was maintained despite some background differences between the training and validation sets. Notably, the training set included only pancreatojejunostomy, while the validation set included both pancreatojejunostomy and pancreatogastrostomy. However, the performance of our prediction model for the validation set was not degraded.
POPF risk prediction can be influenced by a variety of factors, including the surgeon’s level of training and differences in management at the facility. In this study, the training and validation samples were obtained at different facilities, and the validation sample had a larger number of cases than the training sample; nonetheless, the validation sample had the same prediction accuracy as the training sample. The validation sample usually has a lower prediction accuracy than the training sample, and therefore this prediction system may be reproducible. For example, in this study, the cutoff for BMI was set at 25 kg/m
2, which corresponds to the standard threshold for obesity as defined by the Japanese Society for the Study of Obesity [
43]. However, this threshold may be different in other countries. Since our study sample was limited to Japanese participants, more research is required to verify our findings in individuals from other countries.
This study had several limitations. First, the data were collected retrospectively. There were variations in the observation points of the data (e.g., timing of blood biochemistry and CT scans). Additionally, owing to insufficient pathological samples of the remaining pancreatic transection, we were unable to verify the pathology of our hypotheses. Second, the frequency of CR-POPF at these two facilities is higher than that reported in recent years [
6]. This is because these two facilities used to be cautious in dealing with POPF and actively used octreotide and replaced drain based on strict criteria. This may have led to overtreatment for some cases. With the current improvement in treatment practices, the use of such procedures is gradually decreasing. Hence, the in-hospital mortality rate was lower in our study than that previously reported [
44]. Recent studies have reported the efficacy of drain replacement and octreotide administration in patients with CR-POPF before serious complications develop [
45,
46]. Therefore, it would be necessary to analyze the countermeasure for POPF; however, this was a retrospective study, and thus we could not perform the analysis due to missing data. Finally, because the results were obtained at only two facilities, there was some concern that the MPD index measurement results might vary if the results were verified in a larger number of facilities. Therefore, our predictive model should be validated in future prospective studies using a larger number of centers. Our next goal is to conduct a larger prospective study and identify high-risk patients for whom CR-POPF should be managed preoperatively and intraoperatively.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.