Learning nonparametric individualized treatment response curves

Abstract

Thanks to modern medical devices, clinicians are able to obtain accurate and frequent measurements of the patient’s physiological state. Precision medicine aims to individualize the treatment for each patient and design optimal treatment regimes, using the vast amount of data stored in EHRs. Learning individualized treatment responses accurately is an essential step to achieve the goals of precision medicine.

In the literature, the majority of treatment response methods use parametric functions to model the response curves. The functions are designed using domain knowledge about the clinical behavior of the treatment and make strong assumptions about the response curve’s shape. Our goal is to develop a nonparametric model for treatment response curves that achieves competitive performance against parametric models while allowing patient-specific customizations.

We analyze the differences between directly modeling the treatment responses with a Gaussian Process (GP) and modeling the treatment dynamics using a Latent Force Model (LFM). We evaluate our models on a challenging blood glucose prediction dataset.

Additionally, we use the treatment’s covariates to scale the response curve model. We run experiments comparing two GP regression models as well as several ways of sharing the treatment response and treatment covariate model between patients. Our code and data are public for reproducibility and as a building block for future work. We obtain State-OfThe-Art (SOTA) performance on our dataset and discover that modeling the treatment dynamics with a LFM does not significantly improve the predictive performance.

Our results support the case for nonparametric models in treatment response curve estimation, and lay a solid foundation for more sophisticated, GP-based methods. By providing better estimation of physiological states, we hope to empower clinicians and provide better, faster, and cheaper healthcare.