Multivariate prediction of motor diagnosis in Huntington's disease: 12 years of PREDICT-HD

Abstract

Background: It is well known in Huntington's disease that cytosine-adenine-guanine expansion and age at study entry are predictive of the timing of motor diagnosis. The goal of this study was to assess whether additional motor, imaging, cognitive, functional, psychiatric, and demographic variables measured at study entry increased the ability to predict the risk of motor diagnosis over 12 years.

Methods: One thousand seventy-eight Huntington's disease gene-expanded carriers (64% female) from the Neurobiological Predictors of Huntington's Disease study were followed up for up to 12 y (mean = 5, standard deviation = 3.3) covering 2002 to 2014. No one had a motor diagnosis at study entry, but 225 (21%) carriers prospectively received a motor diagnosis. Analysis was performed with random survival forests, which is a machine learning method for right-censored data.

Results: Adding 34 variables along with cytosine-adenine-guanine and age substantially increased predictive accuracy relative to cytosine-adenine-guanine and age alone. Adding six of the common motor and cognitive variables (total motor score, diagnostic confidence level, Symbol Digit Modalities Test, three Stroop tests) resulted in lower predictive accuracy than the full set, but still had twice the 5-y predictive accuracy than when using cytosine-adenine-guanine and age alone. Additional analysis suggested interactions and nonlinear effects that were characterized in a post hoc Cox regression model.

Conclusions: Measurement of clinical variables can substantially increase the accuracy of predicting motor diagnosis over and above cytosine-adenine-guanine and age (and their interaction). Estimated probabilities can be used to characterize progression level and aid in future studies' sample selection.

Keywords: Huntington's disease; assessment of cognitive disorders/dementia; clinical trials methodology/study design; prognosis.

Figure 1

Scatterplot of variable merit (strength of prediction) based on the random survival forest analysis. Minimal depth is shown as a function of variable importance for the group of 34 predictors (left) and 12 predictors (right). Variables are numbered by minimal depth (1 = best), variables in the lower right (red) have the most predictive strength, and variables in the upper left (blue) have the least predictive strength (detailed description of the variables is provided in the Supplemental Data). Key: [1], total motor score from the Unified Huntington's Disease Rating Scale (UHDRS); [2], putamen volume; [3], diagnostic confidence level from the UHDRS; [4], speeded tapping; [5], paced tapping; [6], caudate volume; [7], cytosine‐adenine‐guanine expansion; [8], Symbol Digit Modalities Test; [9], Stroop interference test; [10], accumbens volume; [11], site; [12], University of Pennsylvania Smell Identification Test; [13], Trail Making Test Part A; [14], Stroop color test; [15], Stroop word test; [16], Trail Making Test Part B; [17], emotion recognition test; [18], age (years); [19], Frontal Systems Behavioral Scale—executive subscale; [20], thalamus volume; [21], Symptom Checklist 90—obsessive‐compulsive scale; [22], hippocampus volume; [23], Symptom Checklist 90—Global Severity Index; [24], Frontal Systems Behavioral Rating Scale—disinhibition subscale; [25], Beck Depression Inventory—II; [26], Symptom Checklist 90—anxiety subscale; [27], education (years); [28], Symptom Checklist 90—hostility subscale; [29], Frontal Systems Behavioral Scale—apathy subscale; [30], Symptom Checklist 90—depression subscale; [31], functional activity scale from the UHDRS; [32], total functional capacity from the UHDRS; [33], sex; [34], scanner field strength.

Figure 2

Five‐year log cumulative hazard of motor diagnosis by total motor score, CAG, and age. The slabs of the marginal plots show ranges of the conditioning variables that are used for the interior scatterplot panels. Overlap of the slabs indicates that some participants are used for multiple panels. Smooth curves in the scatterplots (red) are cubic splines. Abbreviation: CAG, cytosine‐adenine‐guanine expansion.