Dear editor,

We appreciate your suggestions in the article titled, “Development of a Nomogram to Predict Clinically Relevant Postoperative Pancreatic Fistula After Pancreaticoduodenectomy on the Basis of the Visceral Fat Area and Magnetic Resonance Imaging.”

Clinical data from 205 patients who underwent pancreaticoduodenectomy (PD) were collected and randomly divided into a training set and a testing set. Data were missing due to the retrospective design of the study. Data on the inflammatory burden index (IBI) were collected before surgery and on postoperative day (POD) 3. Data on drainage fluid amylase concentrations (DFAC) and serum amylase concentrations (SAC) were collected on PODs 1 and 3, and the ratio of drainage fluid to serum amylase concentrations (DFSAC) was calculated. The subgroups with clinically relevant postoperative pancreatic fistula (CR-POPF) were compared using the Mann-Whitney U test.

The results showed that preoperative IBI did not differ significantly between the patients with and those without CR-POPF (P = 0.713). In turn, significant between-group differences in IBI were observed on POD3 (P < 0.001), in DFSAC on POD1 (P = 0.004), in DFSAC on POD3 (P < 0.001), in DFAC on POD1 (P < 0.001), and in DFAC on POD3 (P < 0.001) (Table 1).

Table 1 Comparison of clinical variables between the patients with and those without clinically relevant postoperative pancreatic fistula (CR-POPF)
Full size table

Patient characteristics did not differ significantly at baseline between the training and testing sets (P > 0.05; Table 2). Clinical variables were analyzed by uni- and multivariate logistic regression. Both IBI on POD3 and DFAC on POD3 were added to the predictive model such that the final model contained five variables: visceral fat area (VFA), pancreas-to-spleen signal intensity ratio (PSSI), main pancreatic duct diameter (MPDD), IBI on POD3, and DFAC on POD3 (Tables 3, 4). Among the patients with POPF, IBI on POD1, DFSAC on POD1, DFSAC on POD3, and DFAC on POD1 differed significantly but were not included in the model.

Table 2 Comparison of clinical variables between the training and testing sets
Full size table
Table 3 Analysis of clinical variables by univariate logistic regression
Full size table
Table 4 Analysis of risk factors by multivariate logistic regression
Full size table

The equation for the final five-variable model was

$$ \begin{aligned} X & = \left( {1.6990 \times {\text{IBI}} - {\text{POD}}3} \right) + \left( {0.5845 \, \times {\text{DFAC}} - {\text{POD}}3} \right) \\ & \quad + \left( {0.0122 \times {\text{VFA}}} \right){-}\left( {7.2302 \times {\text{PSSI}}} \right){-}\left( {0.7638 \times {\text{MPDD}}} \right). \\ \end{aligned} $$

The nomogram was plotted using the rms package in R version 3.5.0 (Fig. 1A). The area under the curve (AUC) of the final model in the training set was 0.950, slightly higher than the AUC of the original three-variable model (0.903). The AUC of the final model in the testing set was 0.919, slightly higher than the AUC of the original model (0.903) (Fig. 1B, C).

Fig. 1
figure 1

A The optimized prediction model in which five variables were scored, with higher total scores indicating a higher risk for occurrence of clinically relevant postoperative pancreatic fistula. B, C Comparison between the original three-variable model and the final five-variable model in the testing and training sets. The area under the curve (AUC) of the original and final models was 0.903 and 0.950 in the training set and 0.903 and 0.919 in the testing set. D, E Decision curve analysis of the two models in both datasets.

Full size image

Previous studies showed that DFAC on POD1 and POD3 determined the time when the drainage tubes were removal, consistent with our results. Decision curve analysis showed that the final model was superior to the original model in the training set (Fig. 1D, E). Integrated discrimination improvement (IDI) of the final model relative to the original model was 0.146 (95 % confidence interval [CI] 0.049–0.243; P = 0.003) in the training set and 0.021 in the testing set (95 % CI 0.058–0.099; P = 0.608) (Table 5).

Table 5 Integrated discrimination improvement (IDI) of the final model compared with the three-variable model
Full size table

The risk of POPF can be predicted by DFAC and SAC.1,2 Inadequate blood supply to the anastomosis and local inflammation after PD are reflected by high SAC, impairing anastomotic healing and potentially leading to POFP. Consistent with our findings, high SAC and acute pancreatitis were shown to correlate with POPF.3 Additionally, DFSAC on POD3 can predict CR-POPF.4 Although DSFAC differed significantly between the patients with and those without CR-POPF, this parameter was not included in the model.

The IBI is used to assess the inflammatory status and survival in cancer patients. Higher IBI correlates with poorer outcomes and reduces patients’ quality of life and physical function.5,6 The neutrophil-to-lymphocyte ratio and C-reactive protein (CRP) are measures of systemic inflammation. Large surgical incisions and prolonged surgical time may affect local and systemic inflammation, further impairing the nutritional status and local healing and leading to delayed anastomotic growth and fistula development.7

The inclusion of DFAC on POD3 and IBI on POD3 increased the predictive ability of the final model. Thus, this model can guide the early removal of drainage tubes and postoperative recovery. However, this model could not predict CR-POPF preoperatively due to the inclusion of postoperative indicators.

A good model must make predictions as early and accurately as possible using simple data and easily accessible methods. The predictions made by the original model were based on preoperative variables. Moreover, although the predictive performance of the original model in the training set was slightly lower than that of the final model, the former identified patients at high risk of CR-POPF earlier for treatment and decision-making in the immediate preoperative period.

A larger multicenter study was not performed because of limited resources, reducing the generalizability of the findings. Nonetheless, we intend to conduct a large prospective clinical trial using the improved model to reduce the incidence of pancreatic fistula in our center.