International multicentre validation of the left pancreatectomy pancreatic fistula prediction models and development and validation of the combined DISPAIR-FRS prediction model.

Background: Every fifth patient undergoing left pancreatectomy develops a postoperative pancreatic fistula (POPF). Accurate POPF risk prediction could help. Two independent preoperative prediction models have been developed and externally validated: DISPAIR and D-FRS. The aim of this study was to validate, compare, and possibly update the models.

Methods: Patients from nine high-volume pancreatic surgery centres (8 in Europe and 1 in North America) were included in this retrospective cohort study. Inclusion criteria were age over 18 years and open or minimally invasive left pancreatectomy since 2010. Model performance was assessed with discrimination (receiver operating characteristic (ROC) curves) and calibration (calibration plots). The updated model was developed with logistic regression and internally-externally validated.

Results: Of 2284 patients included, 497 (21.8%) developed POPF. Both DISPAIR (area under the ROC curve (AUC) 0.62) and D-FRS (AUC 0.62) performed suboptimally, both in the pooled validation cohort combining every centre's data and centre-wise. An updated model, named DISPAIR-FRS, was constructed by combining the most stable predictors from the existing models and incorporating other readily available patient demographics, such as age, sex, transection site, pancreatic thickness at the transection site, and main pancreatic duct diameter at the transection site. Internal-external validation demonstrated an AUC of 0.72, a calibration slope of 0.93, and an intercept of -0.02 for the updated model.

Conclusion: The combined updated model of DISPAIR and D-FRS named DISPAIR-FRS demonstrated better performance and can be accessed at www.tinyurl.com/the-dispair-frs.