MACHINE LEARNING FOR ALS SURVIVAL ANALYSIS

Amyotrophic lateral sclerosis (ALS) is a devastating neurodegenerative disease that slowly and progressively leads to death due to the loss of neuron functionality. Unfortunately, there is still no definitive cure for ALS, and the causes of the disease remain largely unknown. However, some treatments have been developed, most of them still in experimental stages, with the aim of alleviating symptoms so that patients can live as normal life as possible and, most importantly, delay death. To support this goal, the 'The Answer ALS Data Portal' has gathered hundreds of trillions of data points concerning thousands of patients, including those affected by ALS, patients with non-ALS neurodegenerative disease, asymptomatic ALS patients, and control patients. These data have been collected to provide researchers with resources to advance the study of the disease’s origin, as well as treatments for symptoms and potentially the disease itself. The objective of this thesis was to analyze the survival of patients with ALS and identify factors that contribute to a longer or shorter lifespan. Additionally, it aimed to develop a predictive model to find connections within the patient data, in order to predict survival for new patients based on their lifestyle and chosen therapies. For this purpose, survival was first examined using three methods that utilized time-fixed information available at the baseline: Kaplan-Meier (KM), Cox Proportional Hazard, and Random Forest for Survival, Regression and Classification (rf-src), and their performances were assessed using two indices: the C-index and the Brier Score. A recent development is the emergence of prognostic models capable of handling time- dependent covariates, which have drawn significant interest from researchers in recent years. These models enable survival prediction over time while accounting for changes in the data during disease preogression. To analyze time-dependent covariates, Cox Proportional Hazards models and DynForest models were implemented and analyzed.