Assessing patterns of network degeneration in familial frontotemporal lobar degeneration using FDG‐PET and unsupervised machine learning

Abstract

AbstractBackgroundGenetic mutations causative of familial fronto‐temporal lobar degeneration (f‐FTLD) are highly predictive of a specific proteinopathy, but there exists substantial inter‐individual variability in their system‐level pathophysiology, which can be assessed with 18Fluorodeoxyglucose‐positron emission tomography (FDG‐PET). Determining data‐driven patterns of network degeneration suggestive of underlying specific FTLD‐related genetic mutations is paramount to guide complex clinical decision making supported by the quantification of FDG‐PET.MethodWe collected clinical and FDG‐PET data from 39 patients with f‐FTLD, including 11 carrying the C9orf72 hexanucleotide expansion (1 asymptomatic, 9 with a predominant behavioral phenotype, 1 with a predominant language phenotype), 12 carrying a GRN mutation (1 asymptomatic, 7 behavioral, 4 language), and 16 carrying a MAPT mutation (6 asymptomatic, 10 behavioral). We performed a spectral covariance decomposition analysis between FDG‐PET images to yield latent patterns reflective of a distribution of metabolism opposing poles of relative hyper‐ and hypo‐metabolism across the entire brain (“eigenbrains” or EBs). We then conducted linear discriminant analyses (LDAs) to perform EB‐based multiclass predictions of genetic mutation and clinical phenotype.ResultThe spectral decomposition analysis yielded five eigenbrains, which explained 58.79% of the covariance between FDG‐PET images (Fig. 1). Gradients indicative of relative hypometabolism in left fronto‐temporo‐parietal areas distinguished GRN mutation carriers from other genetic mutations (EB1, EB2, EB4) and were associated with clinical phenotypes predominantly involving language (EB1, EB2, EB3). Conversely, those indicative of relative hypometabolism in prefrontal areas (EB3) and temporopolar areas (EB2) with a right hemispheric predominance were mostly associated with clinical phenotypes predominantly involving behavior/personality, and EB2 distinguished MAPT mutation carriers from other genetic mutations. The LDAs yielded accuracies of 80% and 75% in predicting genetic status and predominant clinical phenotype, respectively (Fig. 2).ConclusionFive EBs explained a high proportion of covariance in patterns of network degeneration across FTLD‐related genetic mutations. These EBs contained information relevant to the system‐level physiology associated with FTLD‐related pathogenic mechanisms and that allowed the prediction of clinical phenotype with relatively high accuracy. This supports the utility of FDG‐PET to guide the development of tools supporting clinical decision making and ongoing efforts to develop network‐based biomarkers to track disease progression and risk of phenoconversion in f‐FTLD.