Statistical and machine learning models for predicting disease progression

Abstract

Predicting the progress of rare diseases is a crucial task that has raised a lot of attention lately. It can help clinicians develop tailored treatment plans, have a better understanding of the functionality of the disease, and design better clinical trials. Amyotrophic Lateral Sclerosis (ALS) and Duchenne Muscular Dystrophy (DMD) are both rare neuromuscular diseases, having different trends in the course of the disease. The capacity to predict the course of the condition in both instances is crucial for individualized treatment planning and for the creation of current, pertinent research. Predictive models for rare disease development have been established recently, however, machine learning models have not been extensively applied for this purpose. Real-world patient data from DMD and ALS patients are used in this thesis with the purpose of comparing the predictive ability of a statistical survival analysis model and a machine learning-based survival analysis model, namely Cox Proportional Hazards (CoxPH) and Random Survival Forests (RSF), on two target events; death for ALS and the decline below a certain point in the NSAA test. We compare the models in terms of their ability to predict the likelihood that a target event will occur and estimate the time of the occurrence of those events, introducing a threshold strategy. Potential balance issues are also investigated with these datasets. Although RSF outperformed CoxPH in each case, it is evident that more complex and specialized solutions are needed in order for strategies like these to be implemented, even though certain elements of the results look hopeful, in order to create a trustworthy pipeline to handle medical cases like these.