Supervised Bayesian statistical learning to identify prognostic risk factor patterns from population data

Crooks, Colin J.

Abstract

Current methods for building risk models assume averaged uniform effects across populations. They use weighted sums of individual risk factors from regression models with only a few interactions, such as age. This does not allow risk factor effects to vary in different morbidity contexts. This study modified a supervised Bayesian statistical learning method of topic modelling, allowing individual factors to have different effects depending on a patient's other comorbidity. This study used topic modelling to assess more than 71,000 unique risk factors in a population cohort of 1.4 million adults within routine data. The model learnt prognostically important risk factor patterns that predicted 5 year survival, and the resulting model achieved excellent calibration and discrimination with a C statistic of 0.9 in a held out validation cohort. The model explained 92% of the observed variation in 5 year survival in the population. This paper validates using survival supervised Bayesian topic modelling within large routine electronic population health data to identify prognostically important risk factor patterns.

Citation

Crooks, C. J. (2020). Supervised Bayesian statistical learning to identify prognostic risk factor patterns from population data. In L. B. Pape-Haugaard, C. Lovis, I. Cort Madsen, P. Weber, P. Hostrup Nielsen, & P. Scott (Eds.), Digital personalized health and medicine: Proceedings of MIE 2020 (422-426). IOS Press. https://doi.org/10.3233/SHTI200195

• Acceptance Date: Feb 6, 2020
• Publication Date: Jun 16, 2020
• Deposit Date: Oct 4, 2020
• Publicly Available Date: Oct 5, 2020
• Publisher: IOS Press
• Pages: 422-426
• Series Title: Studies in Health Technology and Informatics
• Series Number: 270
• Book Title: Digital personalized health and medicine: Proceedings of MIE 2020
• ISBN: 9781643680828
• DOI: https://doi.org/10.3233/SHTI200195
• Public URL: https://nottingham-repository.worktribe.com/output/4747664
• Publisher URL: http://ebooks.iospress.nl/publication/54197