Methodology

This research is a 6-month feasibility study of the virtual trial emulation approach for an initial proof-of-concept.

The virtual trial emulations are are based on causality learning to capture causal relations between multiple data variables. The causal relations and structures learned from the real-world observational data are inherently incorporated into the generative AI model for synthetic data generation. This renders synthetic cohorts, in which all of the variables are statistically distributed according to their underlying causal relations.

The trial emulations compare treatment effects between a treatment and a control group, both of which consist of virtual synthetic populations. The synthetic cohorts are generated to meet the specific inclusion criteria that are defined by clinical experts to answer target clinical questions. This allows the clinical experts to gain quantitative insights into the treatment effect, predict potential outcomes for hypothetical interventions and make comparisons between different treatments.

We have mainly conducted two types of experiments on the virtual trial emulations in the context of T2DM treatment with different drugs. The first type of experiments is focused on the replication of an existing trial - LEAD-5 (Liraglutide Effect and Action in Diabetes), and the second type of experiments attempts to emulate counterfactual scenarios where different drugs are applied to the same patients to support clinical decision making. The effect sizes are estimated with both average treatment effects and difference-in-differences between pairwise drugs.

More details of the virtual trial emulation approach and its outcomes can be found in Deliverable: Trial Emulation Report.