Personalized prognostic bayesian network for pancreatic cancer : delivering personalized pancreatic cancer management throughout the patient journey

Bradley, Alison and Van der Meer, Robert and McKay, Colin and Jamieson, Nigel (2019) Personalized prognostic bayesian network for pancreatic cancer : delivering personalized pancreatic cancer management throughout the patient journey. Pancreatology, 19 (S1). S31-S32. P1-40. ISSN 1424-3903

[img] Text (Bradley-etal-Pancreatology-2019-Personalized-prognostic-bayesian-network-for-pancreatic-cancer)
Bradley_etal_Pancreatology_2019_Personalized_prognostic_bayesian_network_for_pancreatic_cancer.pdf
Accepted Author Manuscript
Restricted to Repository staff only until 5 June 2020.
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 logo

Download (1MB) | Request a copy from the Strathclyde author

    Abstract

    Background and Objectives: The aim of this study is to create the first Personalized Prognostic Bayesian Network for Pancreatic Cancer (PPBN-PC) to provide personalized predictions of 3-year or more survival time post resection. PPBN-PC’s ability to handle the dynamic nature of the care processes, with predictions evolving as more information becomes available, was assessed at the pre and post-operative stage of the patient journey. Materials and Methods: Parent nodes were identified from PubMed survival analysis studies (n=48691) and included: tumour factors, patient factors, tumour markers, inflammatory markers, neoadjuvant therapy, pathology and adjuvant therapy. Variables underwent a two-stage weighting process to summarise both the weight of the evidence against conflicting findings and a normalized weighting process placing each variable’s weighting within the entirety of the existing body of evidence. Priors for the model were calculated using the normalized weight for each variable as the weighted mean of the TNormal distribution for the corresponding parent node. Results: The PPBN-PC was validated against a dataset of 365 patients who presented to a tertiary referral centre with potentially resectable pancreatic cancer. Model performance measured by Area Under the Curve (AUC) ranged from 0.94 (P-value 0.002; 95% CI 0.859-1.000) for 0 missing data points to AUC 0.74 (P-value 0.000; 95% CI 0.660-0.809) accepting more than 4 missing data points in the validation dataset, for accuracy of pre-operative predictions. PPBN-PC performance for prognostic updating based on post-operatively available information ranged from AUC 0.97 (P-value 0.000; 95% CI 0.908-1.000) for 0 missing data points in pre and post-operative validation dataset to AUC 0.75 (P-value 0.000; 95% CI 0.655-0.838) accepting more than 4 missing data points in the pre and up to and including 2 missing data points in the post-operative validation dataset. The latter was the only point at which AUC fell below 0.80. Validated against every other combination of missing pre and post-operative data points PPBN-PC maintained an AUC greater than 0.8 (range 0.97-0.80) with P-value consistently below 0.001. Conclusion: This marks an important step towards achieving the delivery of precision medicine, as the next step will be to incorporated genomic data into the model hence combining genetic, pathology and clinical data, creating a vehicle to deliver personalized precision medicine.