Flexible modelling of individualized treatment-response curves

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis

Date

2023-06-12

Department

Major/Subject

Machine learning, data science and artificial intelligence

Mcode

SCI3044

Degree programme

Master’s Programme in Computer, Communication and Information Sciences

Language

en

Pages

60

Series

Abstract

These days the field of digital medicine is growing at a fast pace. It happens due to the exponential increase of healthcare data, growth in available computational resources, and development of state-of-the-art algorithms for medical data analysis. Precision medicine is an innovative approach to making disease treatment and prevention individualized, using the vast amounts of data stored in electronic health records (EHRs). Learning individualized treatment-response functions plays an important role in precision medicine. Most approaches to work with treatment-response curves (TRCs) are parametric Bayesian models. However, non-parametric models are gaining popularity due to their flexibility in approximating complex functions, requiring fewer assumptions about the data. Existing work concentrates on incorporating these non-parametric methods to estimate flexible treatment responses. However, they are unable to distinguish what effect a particular treatment type has on the overall response, do not find the differences between the behavior of individual treatment functions and do not take into account treatment dosages. To address the limitations of the existing literature, we develop flexible models, built on top of currently available approaches, which are more sophisticated and can differentiate between various treatment types. Furthermore, we analyze the differences between parametric, GP-based, LFM treatment-response models, and check their behavior on a real-world blood glucose time series dataset. We show that the non-parametric approach fits treatment-response data better, providing a flexible way to model responses to various treatment types. Our code is reproducible and serves as a building block for future work on building even more sophisticated models. We believe that our work helps clinicians by providing better, faster, more accurate estimation of physiological states.

Description

Supervisor

Marttinen, Pekka

Thesis advisor

Hizli, Caglar

Keywords

treatment-response curves, gaussian process, bayesian models, latent forces, probabilistic machine learning, time series modelling

Other note

Citation