Artificial intelligence for the early detection of cognitive impairment in multiple sclerosis

A study combined diffusion MRI and machine learning to show that the condition of tiny brain fibers —the U-fibers— makes it possible to estimate working memory performance in people with multiple sclerosis, opening the door to an imaging biomarker for early monitoring.

Cognitive impairment is one of the earliest —and at the same time most difficult to detect— manifestations of multiple sclerosis (MS). It is often so subtle that it goes unnoticed in routine clinical evaluation, partly because the brain itself manages to compensate for it for a time, and because fatigue, common in the disease, makes it harder to recognize.

Research carried out by a multidisciplinary team from iHEALTH, Pontificia Universidad Católica de Chile, Universidad Andrés Bello, UTEM, and Hospital Dr. Sótero del Río proposes a way to make it visible before evident symptoms appear.

The focus: the fibers that connect what is nearby

Much of the research on MS has concentrated on the large white matter tracts that connect distant brain regions. This work, by contrast, focuses on a far less explored territory: the U-fibers, or short-range association fibers, superficial arc-shaped structures that link neighboring gyri of the cortex.

These fibers are key to the brain's local communication and are among the last to myelinate, which makes them especially vulnerable to the inflammatory and neurodegenerative processes of MS. Their damage could represent a very early stage of "disconnection" that precedes more extensive alterations and that conventional MRI protocols fail to capture.

For this reason, the proof-of-concept study, published in the journal Multiple Sclerosis and Related Disorders, analyzed 35 healthy controls and 58 patients with MS (with and without cognitive impairment), matched by age and educational level. Using diffusion tensor imaging, the team measured fractional anisotropy (FA) —an indicator of white matter integrity— in 100 regions rich in U-fibers.

To relate these measurements to cognitive function, the researchers used the PASAT, a validated test that assesses working memory and information processing speed. Fifteen machine learning models were then trained and compared, evaluated with cross-validation to ensure robust results.

The analysis identified the left superior temporal and right inferior parietal regions as the most informative —two areas that form part of the brain's attention and working memory networks. In simple terms: when these fibers preserve their integrity, cognitive performance tends to be better; when they show signs of microstructural damage, estimated performance declines. Using explainable artificial intelligence techniques (SHAP), the team was able to show that this pattern was consistent across the whole group, bringing transparency to the model's decisions.

The finding supports the idea that cognitive impairment in MS is not the result of a single lesion, but of the coordinated degradation of multiple brain networks. Above all, it suggests that the characteristics of these superficial fibers could become a non-invasive imaging biomarker for detecting and monitoring cognitive involvement at early stages, complementing neuropsychological testing.

Because this is an MRI protocol already acquired routinely in the clinical care of MS patients, the approach has concrete translational potential, especially with a view toward more personalized medicine.

The authors emphasize that this is a cross-sectional study with a limited, single-center sample, so the results should be interpreted as hypothesis-generating. The next step will be to validate them in longitudinal studies and independent cohorts, with higher-resolution protocols, to confirm whether these markers can identify the people at greatest risk of future cognitive impairment.

Montalba C., Franco P., Caulier-Cisterna R., Cruz J.P., Cárcamo C., Andía M.E., Ciampi E. Microstructural changes in juxtacortical white matter regions and their relationship with PASAT score in multiple sclerosis: A proof of concept using machine learning. Multiple Sclerosis and Related Disorders, vol. 112 (2026), 107333. DOI: 10.1016/j.msard.2026.107333