Unlearn presents at the 2023 ACTRIMS Forum

Designing faster and more efficient Multiple Sclerosis clinical trials by incorporating Machine Learning predictions

Background

Multiple Sclerosis (MS) is the most common auto-immune disease affecting the central nervous system. Its clinical presentation is extremely heterogeneous, with the immune system damaging the myelin sheath of neurons at various sites throughout the body. While the disease worsens steadily over time for some patients, most patients experience these auto-immune attacks sporadically, as part of relapses, followed by a full or partial recovery. This complex clinical manifestation of the disease makes it very challenging to accurately measure the effects of treatments in clinical trials. Cost effective means of increasing the efficiency of MS clinical research are desperately needed.

Objectives

Demonstrate how machine learning can be used to build comprehensive models of MS progression and how these models can predict outcomes in control populations to support the primary and secondary analyses of clinical trials.

Methods

We used the data from nearly 2400 patients in the control arms of past clinical trials to train a machine learning model to predict the progression of MS. These data include patients with the three primary subtypes of MS: relapse-remitting, secondary-progressive, and primary-progressive. Given a patient’s baseline characteristics, our trained model generates comprehensive longitudinal clinical predictions, including Expanded Disability Status Score (EDSS), MS Functional Composite (MSFC), and relapse events. These predictions describe likely progressions for the patient when given a placebo, which we refer to as their Digital Twins. Digital Twins can be generated for patients with any of the subtypes of MS, and predict both the long-term effects of the disease and the sporadic short-term impairment observed during relapse events. 

Results

We evaluated Digital Twins on a held-out portion of the dataset (not used for training the model) and found their longitudinal predictions to be consistent with the observed data, across all the most common continuous, binary, and time-to-event endpoints, including EDSS, relapse rates, time to relapse, and the various functional components of MSFC.

Conclusions

Digital Twins accurately model MS progression and can be applied in a variety of ways to design faster and more efficient clinical trials. Such applications include simulating the outcomes for the control group across endpoints to aid in strategic decisions when designing trials, such as understanding the effects of inclusion/exclusion criteria, selecting cohorts/subgroups with specific types of progression, or to support sensitivity analyses. Additionally, Digital Twin predictions can be incorporated into the analysis of trial endpoints to increase the statistical power of those analyses.