#ACTRIMS2022 – Machine Learning Helps Predict Treatment Response in PPMS

Machine learning — using computer algorithms — can be used to identify people with primary progressive multiple sclerosis (PPMS) who are more likely to respond to treatment, a new study shows.

The ability to predict treatment response could allow clinical trials to be designed more efficiently, researchers said.

Jean-Pierre Falet, MD, a graduate student at McGill University, in Canada, presented the findings at the Americas Committee for Treatment and Research in Multiple Sclerosis (ACTRIMS) Forum 2022 on the conference’s opening day. His talk was titled, “Deep learning prediction of response to disease modifying therapy in primary progressive multiple sclerosis.”

PPMS is characterized by the gradual accumulation of disability over time. While more than a dozen therapies are approved to treat relapsing forms of MS, there is currently only one treatment — Roche’s Ocrevus (ocrelizumab) — approved for PPMS in the U.S.

Recommended Reading

February 9, 2022 by BioNews Staff

Expert Voices: Current state of MS treatments and cure research

“Part of the difficulty in identifying efficacious treatments [for PPMS] probably lies in the fact that disability progresses very slowly over time,” Falet said.

Given that disability progresses so slowly, a trial that is either very large or very long may be needed to identify a meaningful effect — and such trials often are not logistically feasible.

An alternative strategy is to specifically select patients who are most likely to respond to a given treatment in a study, which increases the trials’ ability to detect a meaningful effect. This technique is called predictive enrichment.

“The objective here is to use deep learning, which is a type of machine learning … to see if we can predict the response to treatment using readily available MRI metrics as well as clinical and demographic information,” Falet said.

In basic terms, the researchers’ machine learning algorithm involves taking a large collection of data — age, sex, height, weight, disability, and a range of symptom and MRI-based scores — for patients given a specific treatment or placebo, and feeding that data into a computer. Then, based on a set of mathematical rules, the computer “learns” from the data, detecting patterns that can then be used to make sense of other datasets.

To illustrate the utility of this approach, the researchers used data from two clinical trials that enrolled participants with PPMS: OLYMPUS (NCT00087529), which compared rituximab against a placebo, and ORATORIO (NCT01194570), in which Ocrevus was tested against a placebo.

Ocrevus and rituximab are both anti-CD20 monoclonal antibodies, a class of MS treatment that works by killing immune cells called B-cells. Rituximab is not approved for any MS indication, but is often used off-label in patients with MS.

“Given the similar mechanism of action, we were able to pool those two drugs, as well as the two placebo arms” for the analysis, Falet said. In total, this generated a dataset of 1,080 patients.

Using about 70% of that dataset, the researchers then “trained” the machine learning algorithm — feeding the data to a computer, and letting the computer generate rules. The goal was that the computer learned how a person’s clinical and demographic data could predict disability progression after treatment, which was assessed by changes in scores on the expanded disability status scale (EDSS) over time.

The rest of the data then was used to test the rules generated by the computer. In a number of analyses, the researchers showed that this method could distinguish between patients more or less likely to respond to treatment.

In fact, an analysis that looked at 24-week confirmed disability progression in the total population suggested that treatment with anti-CD20 therapy reduced the risk of progression by about 21%, compared with a placebo.

However, when that analysis included only the 25% of patients predicted to have the highest responses to treatment, the reduction in progression risk was much greater — about 60%. By contrast, there was no difference in progression risk between treatment and placebo groups in the patients predicted to have the lowest response to treatment.

In general, the patients who were anti-CD20 antibody responders were more likely to be younger and male. They also tended to have a shorter disease duration, with a greater initial disability, and more T2 lesions — reflecting damaged areas of the brain — on MRI scans.

In further analyses, the researchers showed that this general concept can be applied to other classes of MS medication.

The researchers calculated that a hypothetical year-long clinical trial that only included the top 50% of predicted responders would need just under 500 patients to be able to identify a statistically meaningful effect on disability progression. By comparison, a trial that included all patients, regardless of predicted response, would need more than 3,000 participants to detect a significant effect.

“We’ve shown how we can increase the efficiency of clinical trials using predictive enrichment,” Falet concluded, adding that this type of analysis also may be useful in clinical settings to aid in treatment decisions for individual patients.

But he notes certain limitations to the approach, including challenges in the interpretation of the algorithm. It’s unclear how the algorithm uses the data to make a prediction, he said. Additionally, there are implications for follow-up trials after a first clinical trial with enrichment shows a beneficial effect.

Editor’s note: The Multiple Sclerosis News Today team is providing in-depth coverage of the ACTRIMS Forum 2022 Feb. 24–26. Go here to see the latest stories from the conference.