Colour-Coded Reaction-Time Images Predict Multiple Sclerosis Disability Progression Better Than Summary Scores

Each of the 746 people in this study ended up with a picture of their own. It was not a scan of the brain. It was a small image built from how quickly they pressed a screen, test after test, over months and years. The pixels were red, green and blue, and the pattern those colours made turned out to say something about whether their multiple sclerosis would worsen.

The work, published in PLOS Digital Health on 25 September 2026, comes from a team led by Chao Zhu and Xing Zhang, with senior contributors including Anneke van der Walt, Bruce Taylor, Mastura Monif and Tomas Kalincik. The study asks a practical question. People with relapsing-remitting MS often show changes in thinking speed, attention and memory, sometimes years before physical disability is clear. Digital tests can capture those changes repeatedly, even at home, but they generate a lot of data, and that data is usually boiled down into summary scores that can discard useful information.

The participants were recruited from four tertiary MS clinics in Australia between March 2016 and December 2019. They used MSReactor, a web-based battery of three reaction-time tasks: simple reaction time for psychomotor function, choice reaction time for attention, and a one-back task for working memory. Participants responded to on-screen stimuli as fast as they could, and the system logged speed in milliseconds along with accuracy. Each task continued until at least 30 correct responses were recorded, and a full session typically took about five minutes.

Testing happened at clinic visits roughly every six months, on iPads supplied by each site. Participants could also test at home on their own tablet or computer, as often as they liked, with monthly testing suggested. People with upper-limb, visual or cognitive deficits that prevented touchscreen use were excluded, as were those who completed fewer than three tests or were enrolled for less than 180 days.

The outcome the researchers tried to predict was confirmed disability progression on the EDSS, the standard clinical disability scale. The size of the required increase depended on the starting score, and it had to be confirmed at two later assessments at least six months apart. To avoid mistaking a temporary dip for a baseline, they excluded EDSS scores recorded within 30 days of a relapse.

The core innovation is how the test data were handled. The team organised each person's results by month, keeping only the last test if there were several in one month and carrying forward the last observation to fill gaps. They then encoded each person's history as an image. Red, green and blue stood for the three tasks. The horizontal axis held 30 responses per task, and each row of pixels represented one month. A convolutional neural network extracted features from these images, and a transformer-based survival model, which the authors call MS-TranSurv, estimated each person's risk of progression over time.

Keeping the detail from each test paid off. The approach outperformed methods that rely on summary measures of performance. The authors conclude that repeated, low-burden cognitive testing can help identify people at higher risk of worsening disability, and that this could support earlier clinical review and better selection of participants for clinical trials. They also suggest the image-based method might be adapted to other kinds of digital health data.

The disclosures are extensive. Several authors report ties to MS drug makers including Biogen, Novartis, Roche and Merck, and the original MSReactor cohort was part-funded by Biogen and Novartis Australia through unrestricted funding. The authors state that the funders had no role in this study's design, analysis or publication. Patient-level data cannot be shared publicly, though qualified researchers can request access through MSBase. The next step is validation beyond this cohort.