Establish counterfactual and causal inference capabilities of GluFormer for dietary inputs

Establish whether GluFormer, a transformer-based foundation model trained on continuous glucose monitoring data and extended with dietary tokens, can perform valid counterfactual predictions and causal inference for dietary inputs, enabling rigorous estimation of intervention effects on glucose outcomes.

Background

The authors propose that integrating CGM and dietary data in GluFormer could enable individualized simulations of treatment responses and improve clinical trial design. However, they explicitly note that causal or counterfactual inference has not yet been demonstrated for the model, which is necessary to draw reliable conclusions about intervention effects.

Demonstrating counterfactual and causal inference capabilities would validate the use of GluFormer for pre-trial simulations and responder identification, and would move the model from predictive associations toward robust inference under interventions.