Optimizing Screening for Intrastriatal Interventions in Huntington's Disease Using Predictive Models.

Abstract

Intrastriatal delivery of potential therapeutics in Huntington's disease (HD) requires sufficient caudate and putamen volumes. Currently, volumetric magnetic resonance imaging is rarely done in clinical practice, and these data are not available in large research cohorts such as Enroll-HD. The objective of this study was to investigate whether predictive models can accurately classify HD patients who exceed caudate and putamen volume thresholds required for intrastriatal therapeutic interventions. We obtained and merged data for 1374 individuals across three HD cohorts: IMAGE-HD, PREDICT-HD, and TRACK-HD/TRACK-ON. We imputed missing data for clinical variables with >72% non-missing values and used the model-building algorithm BORUTA to identify the 10 most important variables. A random forest algorithm was applied to build a predictive model for putamen volume >2500 mm3 and caudate volume >2000 mm3 bilaterally. Using the same 10 predictors, we constructed a logistic regression model with predictors significant at P < 0.05. The random forest model with 1000 trees and minimal terminal node size of 5 resulted in 83% area under the curve (AUC). The logistic regression model retaining age, CAG repeat size, and symbol digit modalities test-correct had 85.1% AUC. A probability cutoff of 0.8 resulted in 5.4% false positive and 66.7% false negative rates. Using easily obtainable clinical data and machine learning-identified initial predictor variables, random forest, and logistic regression models can successfully identify people with sufficient striatal volumes for inclusion cutoffs. Adopting these models in prescreening could accelerate clinical trial enrollment in HD and other neurodegenerative disorders when volume cutoffs are necessary enrollment criteria. © 2024 The Authors. Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.