🧠 neurology

Using MRI whole brain atrophy and clinically reported outcomes in combination to assess interim treatment response in multi-arm multi-stage trials in progressive multiple sclerosis

This paper proposes and evaluates a multivariate mixed model that combines MRI whole brain atrophy with three clinical outcomes to enhance the resilience and statistical power of interim treatment response assessments in multi-arm multi-stage trials for progressive multiple sclerosis.

Original authors: Burnell, M., Nicholas, J., Burton, R., Chataway, J., Apap Mangion, S., Carpenter, J.

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Burnell, M., Nicholas, J., Burton, R., Chataway, J., Apap Mangion, S., Carpenter, J.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: The "Taste Test" Problem

Imagine you are a chef developing a new soup. You have three different experimental recipes, and you want to know which one is the best before you spend a fortune cooking it for a huge banquet.

In a standard cooking contest, you might wait until the very end to taste the final soup (the "primary outcome"). But in medical trials for serious diseases like Progressive Multiple Sclerosis (PMS), waiting until the end takes years and costs millions.

So, scientists use a "Multi-Arm Multi-Stage" (MAMS) trial. Think of this as a cooking competition with two rounds:

Round 1 (Interim): A quick taste test to see if the soup has any promise. If a recipe tastes terrible here, you stop making it immediately to save money and time.
Round 2 (Final): If a recipe passes Round 1, you cook it longer and do a full, detailed tasting to see if it's actually the best.

The Problem: Relying on Just One Ingredient

In the OCTOPUS trial (the study this paper discusses), the scientists needed a way to judge the "Round 1" taste test.

Originally, they decided to judge the soup based on only one ingredient: the rate at which the brain shrinks (called "whole brain atrophy"), measured by an MRI scan.

The Logic: They thought, "If the drug works, the brain shouldn't shrink as fast."
The Risk: The authors realized this was like judging a complex soup only by how much salt is in it. Sometimes, a great soup might have the perfect amount of salt, but other ingredients (like texture or flavor) might be off. Conversely, a bad soup might accidentally have the right amount of salt.

In real-world trials, they found cases where a drug successfully stopped patients from getting disabled (the "flavor"), but the brain shrinkage (the "salt") didn't change much. If they had relied only on the MRI scan, they might have thrown away a good drug too early.

The Solution: The "All-Hands" Approach

To fix this, the authors proposed a new strategy: Don't just look at the MRI; look at the MRI plus three other clinical tests.

They combined:

The MRI: Brain shrinkage.
The EDSS: A doctor's score of disability.
The Walk Test: How fast a patient can walk 25 feet.
The Peg Test: How fast a patient can put pegs in a board.

The Analogy: Instead of judging the soup by just the salt, you now taste for salt, sweetness, spice, and texture all at once.

The Technical Magic: The "Universal Translator"

Here is the tricky part: You can't just add these things together easily.

You can't add "millimeters of brain shrinkage" to "seconds it takes to walk." It's like trying to add "miles" to "apples."

The authors invented a statistical "Universal Translator" (a multivariate mixed model).

Standardization: They converted all the different measurements into a common language (like converting everything to "points of improvement").
Weighting: They looked at how these tests usually move together. If the walk test gets worse, does the peg test usually get worse too? They used this relationship to blend the data.
The Result: They created a single "Super-Score" that represents the overall health of the patient, combining the MRI and the three physical tests into one number.

What Did They Find?

They ran the numbers using data from a previous trial (MS-STAT2) to see if this new "Super-Score" would have worked better than just the MRI.

The Power Boost: By adding the three physical tests to the MRI, they increased the chance of spotting a working drug from 83% to 90%.
The Safety Net: More importantly, this method makes the trial more resilient. If the drug works great on walking but the brain shrinkage is weirdly slow, the "Super-Score" still catches the success. If the drug fails on walking but helps the brain, it still catches that. It spreads the risk so you don't accidentally reject a good drug just because one specific test didn't show a result.

The Verdict

The authors conclude that while this method doesn't magically turn a bad drug into a good one, it makes the "Round 1" check much smarter.

Old Way: "Did the brain shrink less?" (If no, stop the trial).
New Way: "Did the brain shrink less, OR did the patient walk better, OR did the peg test improve?" (If any of these show promise, keep the trial going).

This approach ensures that in the high-stakes game of finding a cure for Progressive MS, scientists are less likely to throw away a winning recipe just because they were only looking at the salt.

Technical Summary: Combining MRI Whole Brain Atrophy and Clinically Reported Outcomes in MAMS Trials for Progressive Multiple Sclerosis

Problem Statement
Multi-arm multi-stage (MAMS) trials require interim analysis outcomes that are sensitive enough to detect early drug activity, allowing for the continuation of effective treatments while dropping ineffective arms. However, relying on a single interim outcome measure poses significant risks. In the context of the Optimal Clinical Trials Platform for Progressive Multiple Sclerosis (OCTOPUS) trial, the original design proposed using the rate of whole brain atrophy (WBA) measured via MRI as the sole interim outcome.

The authors identify a critical vulnerability in this approach: historical data from Phase 3 trials in progressive multiple sclerosis (PMS), such as ORATORIO and HERCULES, demonstrate that treatments can show significant clinical benefits (reduced disability progression) without showing concordant reductions in WBA. Consequently, relying solely on WBA creates a risk of prematurely rejecting potentially effective treatments if the interim WBA signal is weaker than anticipated or absent, despite clinical efficacy. The challenge is to integrate multiple outcome types—specifically continuous clinically reported outcomes (CROMs) and MRI-based atrophy rates—into a unified interim decision-making framework that increases statistical power and resilience to uncertainty in effect sizes.

Methodology
The paper proposes a statistical framework for jointly analyzing multiple continuous outcomes with differing scales and measurement types (repeated clinical measures vs. directly measured changes in brain volume) within a MAMS design.

Multivariate Mixed Effects Model:
The authors propose a joint multivariate mixed effects linear regression model.
- Clinical Outcomes (CROMs): For repeated measures like the Expanded Disability Status Score (EDSS), Timed 25-Foot Walk (T25FW), and 9-Hole Peg Test (9HPT), a standard model with random intercepts and slopes is used.
- Imaging Outcome (WBA): For whole brain atrophy, which is measured as a direct difference between specific MRI scans (e.g., baseline to month 12), the model utilizes a formulation for "directly measured change." This accounts for the specific correlation structure where changes sharing a scan time point are correlated, distinct from standard repeated measures.
- Joint Structure: These sub-models are linked by appending the random effects of the imaging outcome to the vector of random effects for the clinical outcomes. This creates an unstructured covariance matrix ( $G$ ) that captures the correlations between the random slopes of the disparate outcomes.
Standardization and Linear Combination:
To combine outcomes with different units and variances, the authors propose scaling each outcome by its respective random slope standard deviation. This standardization allows for the linear combination of treatment effects ( $\hat{\beta}_{trt}$ ) into a single test statistic.
- The treatment effect is defined as a weighted sum of the individual treatment effects ( $\beta_{2q}$ ) for each outcome.
- While inverse-variance weighting (IVW) is possible, the authors find that an equal-weighting approach is effective after standardization, as the outcomes are placed on a commensurate scale of rate of change.
Power Calculation:
The paper details a method for calculating power for this multivariate design. It extends methods described by Frost et al. to account for:
- The marginal variance-covariance matrix derived from the joint model.
- Staggered recruitment and dropout patterns, modeled using a pattern-mixture approach where participants are grouped into cohorts based on their follow-up status at the interim analysis.
- The calculation of the variance of the combined treatment effect to determine the required sample size or power for a given significance level ( $\alpha$ ).

Key Results
The methodology was applied to data from the MS-STAT2 trial (a Phase 3 trial of simvastatin in secondary PMS) to estimate variance components and simulate power for the OCTOPUS trial.

Power Improvement: Using variance data from MS-STAT2 and assuming plausible treatment effects (15% reduction in atrophy rate and 30% reduction in CROMS rates), the joint model increased the power of the interim analysis from 83% (atrophy alone) to 90% (combined outcomes) at a one-sided alpha of 0.45. This represents a 7% absolute increase in power.
Robustness to Variability: When the analysis was repeated using observed effect sizes from the EXPAND and ORATORIO trials, the power gains were maintained or slightly improved (reaching up to 93.9% in the ORATORIO scenario). This demonstrates that the joint approach mitigates the risk of underpowered decisions when individual outcome effects vary significantly from assumptions.
Retrospective Application: When the proposed model was applied retrospectively to the MS-STAT2 data, it correctly identified a lack of treatment effect across all four outcomes (combined test statistic $p=0.634$ ), supporting the decision to drop the simvastatin arm. No individual component showed a signal either.
Correlation Structure: The analysis confirmed that CROMs are more highly correlated with each other than with WBA. While high correlation reduces the marginal benefit of adding an outcome, the inclusion of WBA still provided a measurable power boost.

Significance and Claims
The authors argue that while the power gain in their specific example was moderate (approx. 7%), the strategy offers significant methodological value:

Resilience to Uncertainty: The primary benefit is not necessarily a massive boost in power, but rather the mitigation of risk associated with uncertain effect sizes. By spreading the "liability" across related but distinct outcomes, the interim decision becomes less susceptible to unexpected negative results in any single metric.
Generalizability: The approach is presented as a generalizable solution for MAMS trials, particularly in neurodegenerative diseases where single surrogate endpoints may not perfectly align with clinical outcomes.
Practical Implementation: The authors note that the complex model can be fitted using standard statistical packages (e.g., MLwiN) and provide a user-written Stata command (mvmixedpower) to facilitate power calculations for similar designs.
Interim vs. Final Analysis: The authors emphasize that this combined metric is best suited for interim "stop/go" decisions where the goal is signal detection rather than precise clinical interpretation of effect size. They cite the EJS ACT-PD trial in Parkinson's disease as an example of a similar strategy being adopted for early-stage testing.

In conclusion, the paper demonstrates that a multivariate mixed model can successfully integrate MRI and clinical data to create a more robust interim outcome measure for MAMS trials, reducing the probability of erroneously discarding effective treatments in progressive multiple sclerosis.