Geometric brain signatures for characterising hereditary ataxia and predicting functional brain organisation: a retrospective multicohort study

This is the authors' abstract. We don't add key points for this paper.

Background Hereditary cerebellar ataxias (HCAs) are rare, neurodegenerative disorders with heterogenous presentations, overlapping phenotypes, and variable progression, contributing to diagnostic delay and inadequate disease progression monitoring tools. Conventional neuroimaging approaches often do not capture distributed, multiscale brain alterations. Hence, we aimed to evaluate a geometry-driven deep learning framework based on eigenmodes, which are natural spatial patterns in brain structure and function constrained by cortical surface geometry, to better characterise disease-related brain organisation in HCAs. Methods In this retrospective multicentre cohort study, we analysed structural MRI (sMRI) and functional MRI (fMRI) data from three independent cohorts in Belgium, Australia, five additional international sites (TRACK-FA), and Brazil (Campinas). The cohorts included patients with genetically confirmed HCA (Friedreich ataxia or spinocerebellar ataxia type 1 or 2) and healthy participants without neurological or psychiatric disorders; those with contraindications to MRI or poor-quality imaging were excluded. We decomposed cortical thickness maps and fMRI activation maps using geometric eigenmodes to derive structural and functional signatures and built neural network models for distinguishing patients with Friedreich ataxia (FRDA) from healthy participants. Models trained on the Belgium cohort were fine-tuned on the independent TRACK-FA cohort to assess cross-cohort generalisability. Structural signatures were also used for HCA subtype classification and longitudinal change detection. We also developed models to predict functional brain patterns from sMRI alone and evaluated their utility. Primary outcome measures were area under the receiver operating characteristic curve (AUC) for disease and subtype classification, Pearson's correlation coefficient ( r ), coefficient of determination (R 2 ) for structure-to-function prediction, and longitudinal sensitivity quantified using Cohen's d. Secondary outcomes included accuracy, precision, recall, F1-score, activation-map reconstruction accuracy, and comparisons with conventional MRI volumetric and clinical measures. Findings Data were originally collected between Aug 19, 2009 and Oct 20, 2024, with longitudinal analyses performed based on 12-month follow-up in the TRACK-FA cohort. Participants were aged 6–78 years with different disease durations. Data of 272 patients with Friedreich ataxia, 31 patients with spinocerebellar ataxia type 1, 88 patients with spinocerebellar ataxia type 3, and 113 healthy control participants were included. Models distinguished healthy participants from patients with Friedreich ataxia (maximum area under the receiver operating characteristic curve [AUC] 0·93 [95% CI 0·85–0·98]) and differentiated ataxia subtypes (AUCs up to 0·81 [0·65–0·93]), with cross-cohort generalisability. Structure-to-function prediction showed coefficients of determination up to 0·62 and correlation of up to 0·86 across functional features. Predicted functional signatures improved classification and enabled partial reconstruction of individual task-activation maps. Geometric signatures were more sensitive to longitudinal disease progression than conventional volumetric MRI measures. Interpretation The exploratory geometry-driven deep learning approach used in this study proposes to describe brain changes across multiple scales by integrating structural and functional information and predicts brain function when fMRI data are unavailable. Thus, geometric brain signatures might support biomarker development for HCA subtype differentiation and disease progression tracking in larger prospective studies.