Use of machine learning to predict cognitive performance based on brain metabolism in Neurofibromatosis type 1

• Published in: PloS one Vol. 13; no. 9; p. e0203520
• Main Authors: Schütze, Manuel, de Souza Costa, Danielle, de Paula, Jonas Jardim, Malloy-Diniz, Leandro Fernandes, Malamut, Carlos, Mamede, Marcelo, Miranda, Débora Marques de, Brammer, Michael, Romano-Silva, Marco Aurélio
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 07-09-2018; Public Library of Science (PLoS)
• Subjects: Academic failure; Adolescent; Adult; Analysis; Artificial intelligence; Biology and Life Sciences; Brain; Brain - diagnostic imaging; Brain - metabolism; Child; Cognition; Cognition Disorders - diagnosis; Cognition Disorders - metabolism; Cognitive ability; Computed tomography; Female; Fluorodeoxyglucose F18 - metabolism; Gaussian process; Genetic disorders; Health aspects; Humans; Learning algorithms; Machine Learning; Male; Medical imaging; Medicine and Health Sciences; Memory; Metabolism; Neurofibromatosis; Neurofibromatosis 1 - diagnostic imaging; Neurofibromatosis 1 - metabolism; Neurofibromatosis 1 - psychology; Neuroimaging; Neurological disorders; Neuropsychological Tests; Permutations; Physical Sciences; Positron emission; Positron emission tomography; Positron Emission Tomography Computed Tomography; Predictions; Principal Component Analysis; Principal components analysis; Recklinghausen's disease; Regression analysis; Research and Analysis Methods; Social Sciences; Studies; Tomography; Tumors; Young Adult; Brazil
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0203520

Neurofibromatosis Type 1 (NF1) can cause a wide range of cognitive deficits, but its underlying nature is still unknown. We investigated the correlation between cognitive performance and specific patterns of resting-state brain metabolism in a NF1 sample. Sixteen individuals diagnosed with NF1 underwent 18F-FDG PET/CT brain imaging followed by a neuropsychological assessment. Principal component analysis was performed on 17 measures of cognitive function and a machine learning approach based on Gaussian Process Regression was used to individually predict the components that represented most of the variance in the neuropsychological data. The accuracy of the method was estimated using leave-one-out cross-validation and its significance through permutation testing. We found that only the first component could be accurately predicted from resting state metabolism (r = 0.926, p<0.001). Multiple and heterogeneous measures contribute to the first component, mainly WISC/WAIS Procedure and Verbal IQ, verbal memory and fluency. Considering the accurate prediction of measures of neuropsychological performance based on brain metabolism in NF1 patients, this suggests an underlying metabolic pattern that relates to cognitive performance in this group.