Data to Dose: Efficient Synthetic Data Generation with Expert Guidance for Personalized Dosing

H. Satyam Verma, Holly Wiberg, Shixiang Zhu, Sridhar Tayur

Keywords: Generative Modeling, Bayesian Optimization, Personalized Dosing, Expert-in-the-loop, Preference learning

TL;DR: We introduce GenEx, a framework that combines preference-based Bayesian optimization with expert-gated generative models to efficiently learn personalized dosing under data scarcity.

Abstract: Effective personalized dosing is constrained by limited data, noisy outcomes, and heterogeneous patient profiles. We introduce GenEx, a hybrid framework that couples preference-based Bayesian optimization with a generative model fine-tuned via expert feedback. Expert pairwise rankings update the Gaussian Process (GP) surrogate and, via an expert–generator agreement test, gate uncertainty-guided synthetic data injection at the GP’s highest-variance doses. We provide guarantees of sublinear expected regret under decaying generator bias and validate them in numerical studies. We find that GenEx converges faster and improves decision quality over preference-only and synthetic-only baselines, enabling safer and more effective individualized treatment.

Submission Number: 197