Article Text

Download PDFPDF
E13 Predicting Huntington’s disease state using structural MRI: it’s more than just the striatum
  1. Maitrei Kohli1,
  2. Dorian Pustina2,
  3. John H Warner2,
  4. Daniel C Alexander1,
  5. Rachael I Scahill3,
  6. Sarah J Tabrizi3,
  7. Peter A Wijeratne1
  1. 1Centre of Medical Image Computing, Dept. of Computer Science, University College London, UK
  2. 2CHDI Management/CHDI Foundation, Princeton, NJ, USA
  3. 3Huntington’s Disease Centre, Institute of Neurology, University College London, UK


Background Volumetric measures derived from structural MRI (sMRI) enable identification of Huntington’s disease (HD)-related brain alterations. Stacked machine learning (ML) models allow data-driven patient-specific predictions of disease state. Quantifying feature importance aids interpretation of model outcomes by identifying which regions carry most discriminative information.

Aims We aimed to (a) classify HD state using stacked ML model and features obtained from sMRI, and (b) identify which features are most discriminative for each HD state. We performed fine-grained (PreHD-A; PreHD-B; HD1; HD2), and binary (PreHD-A vs PreHD-B; PreHD-B vs HD1; and HD1 vs HD2) classification.

Methods We trained and evaluated our stacked ML model using baseline cross-sectional sMRI data from 184 HD gene-positive participants from the TRACK-HD dataset. Performance was assessed using repeated stratified 5-fold cross-validation. We performed recursive feature elimination with replacement-based feature importance analysis to identify most relevant features out of 15 sMRI measures for each disease state.

Results A component of the striatum is always amongst the most important features. However, maximum accuracy required at least 2 other features (e.g., Occipital lobe and Lateral ventricles); and may peak with up to 7 features depending on the task. Highest predictive-classification accuracy was achieved between PreHD-A vs PreHD-B (76.7%±8.0) and PreHD-B vs HD1 (75%±9.0); whereas HD1-HD2 accuracy was lower (68%±10.0) and the fine-grained classification was the hardest (55%±6.0).

Conclusions We demonstrate the utility of our stacked ML model and sMRI measures for fine-grained classification and suggest more accurate predictive-classification of HD disease states are achieved by including structures outside the striatum.

  • Classification
  • Striatum
  • sMRI
  • Stacked machine learning
  • feature importance

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.