Abstract
AbstractDeveloping effective treatments for Huntington’s disease (HD) requires reliable markers of disease progression. Striatal atrophy has been the hallmark of HD progression, but volumetric anomalies are also found in other brain regions. Little is known about the potential increase in predictive biomarking accuracy when volumetric scores from multiple brain regions are combined to predict the HD status of individual participants. We used cross-sectional structural MRI data from 184 HD gene-positive participants to a) test a novel ensemble machine learning model in classifying participants in one of four HD progression states (PreHD A; PreHD B; HD1; HD2), and (b) identify the brain regions that carry HD biomarking signal from 15 regions. We used 5-fold cross validation and backward feature elimination to find the optimal predictors and investigated the stability of the findings through repeated analyses. The ensemble predictive model systematically matched or outperformed the accuracy of nine standard machine learning models, reaching 55.3%±6.1 balanced accuracy in 4-group classification. The accuracy was higher for binary classifications (PreHD vs HD: 83.3%±6.3; PreHD A vs PreHD B: 76.7%±8.0; PreHD B vs HD1: 75.9%±8.5; HD1 vs HD2: 70.9%±9.4). Striatal structures (caudate and putamen) were systematically found to be top predictors. However, the accuracy increased substantially when we included other regions in the model (e.g., occipital cortex, lateral ventricles, cingulate, temporal lobe). Optimal models frequently included 2-7 brain regions from different areas. Overall, the accuracy of classifications remained stable across repetitions but the list of selected brain regions could vary, likely due to collinearities in volumetric scores. This is the first study to demonstrate the improvement of classification accuracy when predicting HD progression with a stacked ensemble model. Our findings indicate that HD progression is marked not only by striatal atrophy but also by volumetric changes outside the striatum, without which biomarking models cannot achieve optimal results. The robust methods applied here exposed instability in the selection of brain regions despite the sizeable sample size (n=184); such instabilities could lead to different conclusions in different studies when single analyses are applied on smaller sample sizes. From a translational perspective, our study informs on the selection of candidate endpoints or target regions for therapeutic intervention in future clinical trials.
Publisher
Cold Spring Harbor Laboratory
Cited by
1 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献