Novel Statistical Methodology Development and Applications in ALS Research

Abstract

Being able to predict, with accuracy, the disease progression of patients with a given disease is extremely useful from the perspectives of clinicians, patients, and clinical trial investigators. We introduce a novel method of reducing the expected prediction error when using linear models, given approximate monotonicity of the response; we refer to this method as utilizing an “anchor.” We justify this method mathematically, and then show how to improve predictions arising from standard ordinary least squares (OLS) models when modelling disease progression in a population of patients with amyotrophic lateral sclerosis (ALS). We go on to show that using an anchor can be used in conjunction with more complex modelling schemes to further improve the predictions of ALS patients; an anchor improves both Bayesian hierarchical linear models and Bayesian mixture models. Furthermore, we explore potential covariates that may be included in the models to improve predictions, but find that only time of disease onset results in improved model performance. We also explore how well these models work in a clinical setting, rather than in a clinical trial. We first demonstrate the feasibility of automatically extracting patients’ data, pertaining to survival and disease progression, from the electronic medical record, as well as showing that our disease progression model is feasible for clinical patients. We then compare survival rates between the two populations and determine that, even after adjusting for several important covariates, there is a large difference between survival in the clinic setting and survival in ALS clinical trials. We assert that the two patient groups’ differences in disease progression and survival highlight the needs to understand better disease variability in the clinical setting and to refine the inclusion criteria in ALS trials. We determine an anchor can be used to improve predictive models in ALS disease progression, for both simple independent OLS regressions and for far more complicated Bayesian hierarchical linear models. We conclude that using a Bayesian hierarchical linear model with an anchor is useful in both a clinical trial population of ALS patients as well as a dissimilar population seen in the Midwestern academic medical center ALS clinic.