Rare neurodegenerative diseases, including the hereditary cerebellar ataxias, pose significant challenges for disease monitoring. Small patient cohorts, heterogeneous progression patterns, and slow rates of progression make it difficult to track disease change using conventional biomarkers. Although clinical rating scales remain the standard for assessing severity, they are subjective, prone to measurement noise, and often lack sensitivity to subtle longitudinal decline. Recent work demonstrates that machine‑learning (ML) methods can integrate information from multiple sources, such as multimodal Magnetic Resonance Imaging (MRI), clinical assessments, and demographic factors, to derive weighted composite biomarkers that are more sensitive to disease progression than any individual measurement. These early studies highlight the promise of ML‑derived composites as complementary or surrogate markers of disease severity and progression. However, current approaches are still preliminary: most rely on small datasets, handcrafted imaging features, and linear regression models that cannot capture the complex, multisystem patterns characteristic of neurodegeneration. This underscores a clear need for systematic validation and for more advanced ML methods capable of learning robust, biologically meaningful biomarkers directly from high‑dimensional data.

Project Description

Magnetic resonance imaging (MRI) offers a non-invasive approach to assessing brain health. This project aims to develop and evaluate DL models that learn composite biomarkers of disease progression directly from raw multimodal MRI data, without relying on derived imaging features. Using longitudinal scans from patients (baseline and 1- or 2- years follow‑up), the project will build modality‑specific neural network encoders for structural MRI (captures brain's physical anatomy, its shape, size, and tissue structure), diffusion MRI (tracks water molecule movement to map neural pathways), resting‑state fMRI (captures spontaneous brain activity and functional connectivity at rest), and susceptibility MRI (detects iron deposits and blood vessels). These encoders will learn representations of brain structure and function, as well as their longitudinal changes over time. A multimodal fusion framework will integrate these learned representations, alongside clinical and demographic information, to produce a single composite progression score for each individual. The project will compare the sensitivity of this composite biomarker against traditional clinical measures and individual imaging modalities, and will explore model interpretability techniques to identify neurobiological contributors to the learned score. The overall goal is to assess whether DL-derived composite biomarkers can provide a more sensitive and reliable measure of disease progression for rare ataxias.

Impact

This project will advance the development of objective, data‑driven composite biomarkers for monitoring progression in rare neurodegenerative disease. A sensitive composite biomarker derived directly from raw MRI has the potential to reduce sample sizes required for clinical studies, improve early detection of disease worsening, and enhance personalised tracking of patient trajectories. Beyond its direct relevance to hereditary ataxias, the methodology developed here could be adapted to other neurodegenerative conditions where multisystem involvement and subtle progression challenge existing measurement tools.

Skills Developed 

The student will develop skills in multimodal neuroimaging, deep learning model development, longitudinal analysis, validation, and interpretability. By project completion, they will have hands-on experience with real clinical MRI data and advanced ML methods for neurodegenerative disease research.