Predicting clinical diagnosis in Huntington's disease: An imaging polymarker

Abstract

Objective: Huntington's disease (HD) gene carriers can be identified before clinical diagnosis; however, statistical models for predicting when overt motor symptoms will manifest are too imprecise to be useful at the level of the individual. Perfecting this prediction is integral to the search for disease modifying therapies. This study aimed to identify an imaging marker capable of reliably predicting real-life clinical diagnosis in HD.

Method: A multivariate machine learning approach was applied to resting-state and structural magnetic resonance imaging scans from 19 premanifest HD gene carriers (preHD, 8 of whom developed clinical disease in the 5 years postscanning) and 21 healthy controls. A classification model was developed using cross-group comparisons between preHD and controls, and within the preHD group in relation to "estimated" and "actual" proximity to disease onset. Imaging measures were modeled individually, and combined, and permutation modeling robustly tested classification accuracy.

Results: Classification performance for preHDs versus controls was greatest when all measures were combined. The resulting polymarker predicted converters with high accuracy, including those who were not expected to manifest in that time scale based on the currently adopted statistical models.

Interpretation: We propose that a holistic multivariate machine learning treatment of brain abnormalities in the premanifest phase can be used to accurately identify those patients within 5 years of developing motor features of HD, with implications for prognostication and preclinical trials. Ann Neurol 2018;83:532-543.

Figure 1

High‐level schematic of the analysis approach. In the Cambridge cohort, both resting‐state fMRI and structural images were available, only structural images were available for the TRACK‐HD cohort. Colors represent the independent samples used for different aspects of the analysis (blue = Cambridge cohort [19 preHD, 21 controls]; green = TRACK‐HD cohort [118 preHD, 121 controls]). fMRI = functional magnetic resonance imaging; SVM = support vector machine.

Figure 2

(A) Schema ball depicting cross‐group differences in resting state network coupling (connections thresholded at p < 0.02 uncorrected). Blue curves represent reduced network coupling and the red curves represent increased network coupling for the preHD group relative to controls. (B) Table showing correlations for the hypoconnected network measures and estimated years to diagnosis. (C) Scatterplot of the correlation between the mean values for composite hypoconnected network measures and estimated years to diagnosis.

Figure 3

(A) Subcortical gray matter volume of the preHD‐near (light gray), preHD‐far (gray) and the matched control (light gray) groups. Each bar is accompanied by an image with the associated structure highlighted in black. Error bars report the standard error of the mean (**p < 0.01; *p < 0.05). (B) Scatterplot showing the correlation between Putamen (black dots), Caudate (gray dots), and Pallidum (light‐gray dots) volume (collapsed across hemisphere) with Estimated years to disease onset.

Figure 4

(A) Comparisons of cortical thickness (collapsed across hemisphere) between the Controls (dark gray), preHD‐far (medium gray), and preHD‐near (light gray) groups. Error bars represent the standard error of the mean (***p = 0.001). (B) A scatterplot showing the correlation between mean cortical thickness & estimated years to diagnosis.

Figure 5

Relationship between actual time of diagnosis and estimated years to diagnosis (A), CAG‐Age product scaled (B), caudate volume (C), and SVM classification strength (D). Yellow = expected to phenoconvert within 2 years or less of the analysis date. Red = early diagnosis. Blue = yet to phenoconvert. SVM = support vector machine. Yellow = expected to phenoconvert within 5 years or less of analysis date.

Figure 6

Cambridge data classified with models trained on independent data from the TRACK‐HD consortium. (A) Permuted null distribution F1 scores (pink) relative to the true model (yellow bar, N.B. Bar height and width are arbitrary) for the controls versus preHD‐far (Ai), preHD‐all subjects (Aii), and preHD‐near (Aiii) models. (B) Confusion matrices for each model. A model trained to classify preHD versus controls in the TRACK‐HD data was used to measure distance to SVM hyperplane when the model was tested on the Cambridge preHD. (C) Yellow = expected to phenoconvert within 2 years or less of the analysis date. Red = early diagnosis. Blue = yet to be diagnosed. SVM = support vector machine. Yellow = expected to phenoconvert within 5 years or less of analysis date.