PhD defence of Joshua Durso-Finley – Uncertainty Aware Causal Models for Prediction of Individual Treatment Effects for Image-based Precision Medicine

Abstract

This thesis presents, in sum, the first image-based longitudinal deep learning precision medicine model. The model uses modern deep learning methods and classical treatment effect estimation theory to estimate individual treatment effects (ITE) using high-dimensional patient MRI sequences as input. Before, treatment effect analysis focused on differences across entire populations of treated patients or subgroups of patients with simple distinguishing factors, e.g., sex. With unique patient data, such as MRI sequences, treatment effects can be estimated for the individual, which offers a fine-grained estimate of patient response. Currently, the models built using high-dimensional imaging data have only focused on predicting disease outcomes for only a single treatment. These models help doctors manage patient risk, but predicting a single outcome only is insufficient for treatment effect estimation. Treatment effect analysis takes a step beyond single outcome prediction by predicting outcomes for all potential therapies.

To achieve the goal of personalized treatment predictions, we developed the first imaging-based precision medicine model. The model is trained on data spanning different stages of multiple sclerosis (MS) and has been designed to predict several clinically relevant outcomes. Our first advancement estimates the number of future brain lesions for the Relapsing Remitting form of MS (RRMS) on five treatments and placebo. We demonstrate accurate prediction when the model makes a prediction for the patient's assigned therapy (factual outcome) and on the identification of groups that have heterogeneous responses to treatment. We deem such patients responders. In some cases, we found subgroups of responders whose outcomes were statistically significantly improved, even for drugs that did not have significant effects at the group level.

Our next advancement better aids clinical decision-making by quantifying and validating the uncertainty in predictions of treatment effect. Uncertainty quantification is critical in high-risk applications, such as treatment recommendation, as time spent on non-working treatments can have long-lasting irreparable effects. We show that treatment recommendations made with additional uncertainty information improves decision-making and leads to better outcomes when treating a population of patients.

The final contribution models the disease trajectory over time. Modeling the trajectory of patients becomes far more important when considering longitudinal diseases, where the intermediate outcomes can be just as important as the outcomes at the end. As MS is a chronic disease, predicting the entire trajectory better estimates the patient's experience throughout the disease. We validate these predictions on new lesion development and an outcome that better reflects the patient's quality of life, the Expanded Disability Status Scale (EDSS). The EDSS is more difficult to predict as it is not an image-based measure of MS; however, it is the primary outcome of interest in the Second Progressive (SPMS) and Primary Progressive (PPMS) subtypes of MS. Finally, it should be noted that while this work focuses on multiple sclerosis outcomes, the framework and design considerations can be applied broadly for any personalized medicine tools built using complex data.