Making personalized predictions of poor outcome post resection of pancreatic ductal adenocarcinoma (PDAC) : a prognostic bayesian network with pre- and post-operative application

Tools

Bradley, Alison and Van der Meer, Robert and McKay, Colin and Jamieson, Nigel (2019) Making personalized predictions of poor outcome post resection of pancreatic ductal adenocarcinoma (PDAC) : a prognostic bayesian network with pre- and post-operative application. Pancreatology, 19 (S1). S122. P6-13. ISSN 1424-3903 (https://doi.org/10.1016/j.pan.2019.05.326)

[thumbnail of Bradley-etal-Pancreatology-2019-Making-personalized-predictions-of-poor-outcome-post-resection]
Preview
 Text. Filename: Bradley_etal_Pancreatology_2019_Making_personalized_predictions_of_poor_outcome_post_resection.pdf
Accepted Author Manuscript
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 logo
Download (1MB)| Preview

Abstract

Background and Objectives: The high-risk field of pancreatic cancer surgery, where surgical benefits are often nullified by early disease reoccurrence, mandates better patient selection for surgical intervention. Existing predictive models are limited in value and scope, relying heavily on post-operative information. The objective of this study was to combine PubMed and patient level data to create and validate a Bayesian Network that can make accurate personalized predictions of poor prognosis (12 months or less) post resection of PDAC preoperatively and perform prognostic updating postoperatively. Materials and Methods: A weighted Bayesian network, based on PubMed post-resection survival analysis studies (n=31,214), was created using AgenaRisk software. Input variables included: inflammatory markers, tumour factors, tumour markers, patient factors and, if applicable, response to neoadjuvant treatment for pre-operative predictions. Prognostic updating was performed by inclusion of post-operative input variables including: pathology and adjuvant therapy. The model was validated against the database of a prospectively maintained tertiary referral centre (n=387). Results: For pre-operative predictions an Area Under the Curve (AUC) of 0.70 (P value: 0.001; 95% CI 0.589-0.801) was achieved accepting up to two missing data points in the pre-operative validation dataset. For prognostic updating an AUC 0.79 (P value: 0.000; 95% CI:0.710-0.870) was achieved when validated against a dataset that had up to 6 missing pre-operative data points but full post-operative data. This dropped to 0.72 (P value: 0.000; 95% CI:0.660-0.788) when the validation dataset had up to 6 missing pre-operative, and up to 3 missing post-operative data points. Conclusion: The Bayesian network presented here demonstrates a predictive performance rivalling existing models. Benefits over existing models include: pre-operative application, application to neoadjuvant and upfront surgery management pathways, and greater generalizability. As patient databases mature globally and our understanding of disease at genomic level deepens so too will the accuracy of predictions of this model with associated benefits at clinical level by supporting better shared decision making. The future application of this work will be to include emerging genomic data and combine this with clinical and pathological data to make personalized predictions of outcome, hence effectively creating a vehicle to deliver precision medicine.

ORCID iDs

Bradley, Alison, Van der Meer, Robert ORCID logoORCID: https://orcid.org/0000-0002-9442-1628, McKay, Colin and Jamieson, Nigel;
CORE (COnnecting REpositories)