Machine learning Modeling of Time-dependent Patient Trajectories

The increasing availability of large-scale medical datasets has fueled the hope of an acceleration toward precision medicine, driven by data. Indeed, the systematic collection of Electronic Health Records (EHR) across several healthcare institutions has resulted in large numbers of patient records even for low-prevalence diseases and allowed uncovering more specific patterns in the evolution of the disease of individual patients. Ultimately, these uncovered patterns can be used to forecast the future medical trajectory of individual patients and support individual treatment recommendations to improve medical outcomes. Nevertheless, the data collected in daily clinical practice present some specific features that raise challenges for typical machine learning architectures. Modeling of clinical data, and patient clinical trajectories in particular, therefore requires dedicated machine learning methods. In this thesis, I focus on two main requirements that clinical machine learning models should fulfill: the ability to handle irregular sampling and the ability to perform causal inference. Clinical data collected over time is indeed usually irregularly sampled, as it is only collected at medical visits, which do not necessarily happen at fixed time intervals. Moreover, not all possible measurements are collected at each visit. While classical machine learning methods for time series do not naturally support irregular sampling, we propose new architectures based on neural ordinary differential equations that operate in continuous time and can therefore elegantly model irregularly-sampled incomplete time series. An important application of modeling clinical trajectories resides in the ability to recommend treatments to individual patients, based on their clinical history. However, as most available clinical data arise from daily clinical practice rather than strictly curated clinical trials, inferring the effect of a treatment on individual patients is not as straightforward as mere forecasting. Estimating the individual treatment effect is a causal question and thus requires a causal framework. We thus propose three new causal machine learning models for patient trajectories: a longitudinal individual treatment effect prediction model focused on uncertainty quantification, a model to estimate counterfactuals when patients can be divided into clinical subgroups, and a causal discovery algorithm based on convergent cross mapping. We motivate all our contributions from a theoretical perspective and demonstrate their improved performance over previous state-of-the-art approaches through a series of rigorous numerical experiments. Lastly, we perform an in-depth analysis of a specific clinical use case on multiple sclerosis. We develop a machine learning method to predict the risk of disability progression within two years.

Jaar van publicatie:2023

Toegankelijkheid:Open