Control arms are a mainstay of clinical trials. Accurately predicting Multiple Sclerosis (MS) disease progression for individual patients in the control arm is valuable for therapeutic development, and can be accomplished with a machine-learning model. The abundance of high-quality subject data, collected repeatedly from control arms in past clinical trials, provides an opportunity to build a model that can predict outcomes for control subjects in future trials.
To develop a machine-learning model that can accurately predict Multiple Sclerosis (MS) disease progression for individual subjects enrolled in the placebo arm of MS clinical trials.
We used a machine-learning model called the Conditional Restricted Boltzmann Machine (CRBM) that is designed to predict how clinical variables related to MS change over time. To develop this model, we extracted a machine-learning dataset from data provided by the Multiple Sclerosis Outcome Assessments Consortium comprising 2465 placebo arm subjects across 8 MS clinical trials. Our dataset contains 21 variables including demographic information, functional assessments, and disability measurements every 3 months for up to 48 months. We then trained the model and assessed its ability to generate statistically accurate predictions, using a set of test subjects not used during training.
Given baseline data for placebo subjects, the CRBM model can generate data for the future clinical state of individual subjects that are statistically indistinguishable from actual subject data along a number of key measures, suggesting the model is capable of generating digital subjects equivalent to actual subjects. Notably, the model accurately captures Expanded Disability Status Scale (EDSS) trajectories and relapse rates across MS subtypes.
We have created a subject-level model of MS disease progression that can predict the future state of placebo subjects in clinical trials. Our model has useful applications in clinical trials, including supplementing control arms with digital subjects generated from the model.