Supporting personalized medicine for patients with multiple chronic conditions: Machine Learning to Augment Benefit Harm Analysis
Description: Background: Most clinical guidelines and clinical decisions in multimorbidity are based on evidence from single disease studies. As a result, patients are treated for each condition in isolation, leading to a “treatment cocktail”, with unclear benefit-harm balance. An informed multidisciplinary approach and the use of high-dimensional routine care data could be potential solutions to address the challenges of multimorbidity.Objectives: This study aims to (1) apply machine learning (ML) methods to estimate best possible treatment combinations for defined multimorbidity from observational data. (2) inform the benefit-harm balance assessment with the ML results to identify treatment combinations with the best possible outcome for defined multimorbidity.Methods: To address objectives (1) Causal variational autoencoders (VAE) and conditional generative adversarial networks (GAN) will be used to understand, generate complete and representative data conditional on patient characteristics, the treatments and outcomes of interest. The generated data will then be used to create counterfactual ('what-if') scenarios to estimate effects under different treatment combinations. The effects will be anchored to state-of-the-art conventional methods available for estimating treatment effects from observational studies, as well as to field knowledge and prior evidence. Once we have confidence in the procedure for one or two treatments at a time and the results are consistent with conventional methods, we will use ML methods only for more complex conditions, as conventional methods cannot capture factorial complexity for more than two treatments. (2) We will perform the benefit-harm balance to weigh the benefit against the harms of the treatment combinations identified in step 1. The analyses will be performed for patient subgroups since the treatment effects, baseline risks, and patient preferences vary from patient to patient. Moreover, a personalized and interactive visualization dashboard will also be developed to assist treatment decision making for individual patient.Significance: This will be used as a case study to demonstrate that reductionist single-disease studies cannot address multimorbidity and a paradigm shift toward an informed multidisciplinary approach that makes use of conventional methods, advanced calculations, observational or routine care data, and field knowledge to address the challenges of multimorbidity.
Amount: CHF 114,200
Funding type: fellowship
Funder scheme: Return CH Postdoc.Mobility
Start date: 2023-03-01
End date: 2024-02-29
Investigator: Yebyo, Henock Gebremedhin (lead)
Institutions: University of Zurich, University of Zurich – ZH
Subfield: Epidemiology
Field: Medicine
Domain: Health Sciences
Funded outputs count: 0
Funder award ID: 210789
OpenAlex ID: G8823848339
