A machine learning analysis of structural brain MRI identified three MRI-based subtypes of Friedreich ataxia with differing patterns of neurodegeneration and clinical associations, according to a multicenter study of 565 participants.
Friedreich ataxia (FRDA) is a progressive, inherited neurodegenerative disease with substantial variability in clinical course. To better characterize this heterogeneity, researchers analyzed MRI scans and clinical data from 275 patients with FRDA and 290 healthy controls using the Subtype and Stage Inference (SuStaIn) algorithm, a machine learning approach that groups patients based on shared patterns of brain atrophy.
In Radiology, Ian H. Harding, PhD, of QIMR Berghofer Medical Research Institute in Brisbane, Australia, and Sirio Cocozza, MD, of the University Federico II in Naples, Italy, and colleagues reported a cohort with a mean age of 32 years, about half of whom were women. Structural MRI scans were used to measure 128 brain regions, with 21 regions selected for further analysis based on differences compared with controls. These included the brainstem, cerebellar peduncles, dentate nucleus, corticospinal tracts, and selected cerebral gray and white matter regions.
The optimal model identified three subtypes. The “classical” subtype, comprising 67% of patients, showed a pattern of atrophy beginning in the brainstem and cerebellar pathways and later extending to the cerebellum and cerebrum. The “early cerebral” subtype (26%) showed earlier involvement of brain pathways and motor cortex before cerebellar changes. The “early cerebellar” subtype (8%) showed early cerebellar atrophy preceding involvement of the brainstem and other regions.
All three subtypes demonstrated early changes in key brain pathways and the medulla but differed in how atrophy spread over time. The subtype patterns were consistent across analyses, supporting the robustness of the model.
Clinical and demographic characteristics—including age, sex, symptom duration, disease severity, and genetic repeat length—were similar across subtypes. These findings suggest that MRI-based subtypes may capture disease-related heterogeneity not reflected in standard clinical or genetic measures.
MRI-derived disease stage was associated with both symptom duration and disease severity, with higher stages linked to longer disease duration and greater severity. However, this relationship varied by subtype. Patients in the early cerebral subtype showed a weaker association between MRI stage and both duration and severity than those in the classical and early cerebellar subtypes, suggesting differences in disease progression patterns.
After adjustment for age and sex, MRI stage showed a stronger association with disease severity than a commonly used imaging marker, medulla atrophy. In combined models, MRI-based subtype and stage, together with age and sex, accounted for approximately one quarter of the variance in clinical scores.
The study had several limitations. Its cross-sectional design did not allow assessment of longitudinal progression or prediction of outcomes. Clinical scores were harmonized across different rating scales, which may have reduced granularity. The analysis relied on structural MRI alone and did not include other imaging modalities that may detect more subtle changes. In addition, cross-site data harmonization and feature selection on the full dataset may introduce bias, and the findings were not validated in an independent cohort.
Despite these limitations, the findings suggest that combining structural MRI with machine learning can identify meaningful disease subtypes in FRDA that are not apparent from clinical data alone.
“Subtyping can reveal latent biologic heterogeneity, uncovering clinically meaningful differences that would otherwise remain undetected,” the researchers wrote.
Study supported by multiple international funding sources, including national research agencies and Friedreich’s Ataxia Research Alliance grants. Author disclosures are available in the original publication.
Source: Radiology