The use of deep learning methods may help find participants with multiple sclerosis (MS) who are most likely to be responsive to specific treatments in clinical trials, according to a study published in Nature Communications.

The efficiency of early proof-of-concept clinical trials could be increased through a deep-learning predictive enrichment strategy to select patients most likely to benefit from therapies. This method could increase the power of these early studies and reduce the number of patients needed to achieve statistical significance.

“This approach can consistently identify and rank treatment effect among patients exposed to anti-CD20-Abs, and could reduce by several fold the sample size required to detect an effect in a short 1or 2-year long trial,” the authors said.


Continue Reading

Read more about MS experimental therapies

In the study, a multiheaded multilayer perceptron model was trained using previous clinical trial data to predict the change in Expanded Disability Status Scale over time for treatment with anti-CD20 monoclonal antibodies (anti-CD20-Abs) and placebo. The difference between the 2 predictions gives an estimate of the conditional average treatment effect for each individual.

Through smaller and smaller subsets of the most responsive patients, the statistical differences between treatment and placebo grew larger. The inclusion of all participants yielded P =.1179, inclusion of only the top 50% of responders yielded P =.0218, and inclusion of the top 30% yielded P =.00817.

A corresponding decrease in statistical difference was observed with the 50% least responsive patients having P =.744 between treatment and placebo and the 70% least responsive having almost no difference with P =.925.

The model trained on the anti-CD20-Abs dataset also generalized to data on treatment with laquinimod, which has a different mechanism of action. The hazard ratios (HR) progressively improved, with the selection of top responders from a whole-group HR of .667 to HRs of .492 and .338 when including only the top 50% and 30% of responders, respectively.

As a simulation of a phase 2 study, using the model and only including patients in the 50th percentile of responders for a 2-year-long clinical trial yielded a 6-fold reduction in the number of patients needing to be randomized and a 3-fold reduction in the number of patients needing to be screened.

Reference

Falet JPR, Durso-Finley J, Nichyporuk B, et al. Estimating individual treatment effect on disability progression in multiple sclerosis using deep learning. Nat Commun. 2022;13(1):5645. doi:10.1038/s41467-022-33269-x