Unique combinations of shape morphometric features improves discriminability of ftld phenotypes

Abstract Background Frontotemporal lobar degeneration (FTLD) is associated with diverse clinical phenotypes underlain by multiple disease pathologies and genetic mutations. As such, traditional structural MRI analyses lack sensitivity and specificity for discriminating FTLD syndromes. Here, we use data‐driven methods to extract a concise set of MRI‐derived shape morphometric features and examine the discriminatory capability of their unique combinations in four FTLD clinical phenotypes. Method 190 patients with sporadic or familial FTLD spectrum disorders (i.e., behavioral variant (bvFTD, n = 107), non‐fluent variant primary progressive aphasia (nfvPPA, n = 27), semantic variant primary progressive aphasia (svPPA, n = 12) and progressive supranuclear palsy (PSP, n = 44)) and 27 controls without pathogenic mutations from the ALLFTD cohort were evaluated. MRI data were preprocessed with FreeSurfer software and four cortical measures were extracted and indexed to a normalized surface atlas: cortical thickness (CT), surface area (SA), surface curvature (SC) and jacobian white matter surface metric distortion (JW). For all phenotypes, each shape feature was contrasted with controls using linear models adjusted for age and gender. Principal component analysis (PCA) was then applied to each individual structural measure and the discriminatory power based on individual and combined measures was assessed using logistic regression and 10‐fold cross‐validation. Result Figure 1 displays FDR‐corrected T‐scores of differences from controls in each morphometric feature in bvFTD. Results reveal complementary patterns of CT, SA, SC and JW for each phenotype, with greatest differences observed in CT across groups. In Figure 2, we display the features derived from the PCA, weighted by eigenvector coefficients, in bvFTD. In Figures 3‐6, receiver operating characteristic curves for individual and combined measures alongside areas under the curve (AUCs) are shown for each phenotype. In bvFTD, nfvPPA and PSP, the superior model included CT and SC at AUCs of 88.3, 87, and 78.6, respectively. For svPPA, the superior model included CT, SC and JW at an AUC of 85.6. Conclusion Integrating additional MRI‐derived surface morphometric features improved classification performance in all FTLD phenotypes. The principal component analysis‐based approach indicated distinct brain regions contribute to discrimination for each shape feature, suggesting they may reflect unique aspects of neurodegeneration across groups.