Predicting behavioral variant frontotemporal dementia with pattern classification in multi-center structural MRI data

Abstract

Purpose: Frontotemporal lobar degeneration (FTLD) is a common cause of early onset dementia. Behavioral variant frontotemporal dementia (bvFTD), its most common subtype, is characterized by deep alterations in behavior and personality. In 2011, new diagnostic criteria were suggested that incorporate imaging criteria into diagnostic algorithms. The study aimed at validating the potential of imaging criteria to individually predict diagnosis with machine learning algorithms.

Materials & methods: Brain atrophy was measured with structural magnetic resonance imaging (MRI) at 3 Tesla in a multi-centric cohort of 52 bvFTD patients and 52 healthy control subjects from the German FTLD Consortium's Study. Beside group comparisons, diagnosis bvFTD vs. controls was individually predicted in each subject with support vector machine classification in MRI data across the whole brain or in frontotemporal, insular regions, and basal ganglia known to be mainly affected based on recent meta-analyses. Multi-center effects were controlled for with a new method, "leave one center out" conjunction analyses, i.e. repeatedly excluding subjects from each center from the analysis.

Results: Group comparisons revealed atrophy in, most consistently, the frontal lobe in bvFTD beside alterations in the insula, basal ganglia and temporal lobe. Most remarkably, support vector machine classification enabled predicting diagnosis in single patients with a high accuracy of up to 84.6%, where accuracy was highest in a region-of-interest approach focusing on frontotemporal, insular regions, and basal ganglia in comparison with the whole brain approach.

Conclusion: Our study demonstrates that MRI, a widespread imaging technology, can individually identify bvFTD with high accuracy in multi-center imaging data, paving the road to personalized diagnostic approaches in the future.

Keywords: Atrophy; Behavioral variant frontotemporal dementia; Diagnostic criteria; FEW, family wise error; FTLD, frontotemporal lobar degeneration; Frontotemporal lobar degeneration; GMD, gray matter density; MNI, Montreal Neurological Institute; MPRAGE, magnetization-prepared rapid gradient echo; MRI; MRI, magnetic resonance imaging; Pattern classification; SVM, support vector machine; VBM, voxel based morphometry; bvFTD, behavioral variant frontotemporal dementia.

Fig. 1

Group comparison between behavioral variant frontotemporal dementia (bvFTD) vs. healthy control cohort (HC) for gray matter density (GMD). A: 19 patients with bvFTD vs. 19 center-matched control subjects. B: 52 patients with bvFTD vs. 52 control subjects. Family wise error (FWE) correction. Coordinates in Montreal Neurological Institute (MNI) space. Left side of the brain is shown on the left.

Fig. 2

Conjunction analyses across group comparison between behavioral variant frontotemporal dementia (bvFTD) vs. control cohort, leaving for each analysis one center, i.e. respective patients and control subjects, out. Differences in gray matter density (GMD). The scale illustrates the number of overlapping centers. A: 19 patients with bvFTD vs. 19 center-matched control subjects. B: 52 patients with bvFTD vs. 52 control subjects. Family wise error (FWE) correction. Coordinates in Montreal Neurological Institute (MNI) space. Left side of the brain is shown on the left.

Fig. 3

Weights of voxels most relevant for support vector machine (SVM) classification between patients with behavioral variant frontotemporal dementia (bvFTD) and healthy controls (HC). The most relevant voxels for classification as bvFTD are shown in red-yellow, for HC in blue. SVM classification was performed on all voxels within the gray matter mask (tissue probability > 0.4). A: 19 patients with bvFTD vs. 19 center-matched control subjects. B: 52 patients with bvFTD vs. 52 control subjects. Coordinates in Montreal Neurological Institute (MNI) space. Left side of the brain is shown on the right.