Theoretical and Computational Neuroscience
Author: Fermin Travi | Email: fermintravi@gmail.com
Fermin Travi1°2°, Eduardo Castro3°, Hongyang Li3°, Anushree Mehta3°, Jenna Reinen3°, Pablo Meyer Rojas3°, Guillermo A. Cecchi3°, Pablo Polosecki3°
1° Facultad de Ciencias Exactas y Naturales, Departamento de Ciencias de la Computación, Universidad de Buenos Aires, Buenos Aires, Argentina
2° Laboratorio de Inteligencia Artificial Aplicada (LIAA), CONICET – Universidad de Buenos Aires, Instituto en Ciencias de la Computación (ICC), Buenos Aires, Argentina
3° IBM T. J. Watson Research Center, Yorktown Heights, New York, NY, United States
One of the main hypotheses across a variety of disorders, including dementias, bipolar disorder, and schizophrenia, is that their effects produce premature brain aging. Current approaches typically find correlations between features extracted from MRI images and subjects’ age, conflating phenotypes with progression. Different disorders are hard to distinguish when the degree of progression of each subject is not considered. Simply finding correlates of aging tends to group subjects by progression.
To tackle this problem, we leveraged invariant deep Generative Artificial Intelligence models with the goal of obtaining brain embeddings (i.e., abstract representations in the form of vectors) capable of capturing the phenotypes that underlie neurodegenerative disorders. By processing 44.178 brain MR images (23.115 females, 21.063 males; mean age 64, std. 7.7) from the UK Biobank dataset, we were able to obtain brain embeddings that contain little to no information about age (0.11 pearson’s correlation vs. 0.79 from the non-invariant model), albeit at a cost in sex prediction (down from 95% to 78%) and reconstruction error (up from 0.21 to 0.33). The ensuing evaluation will consist of predicting neurodegenerative disorders —were our hypothesis to be correct, there should not be differences in the performance of the classifier across age windows. This model holds the promise of paving the way for the development of early diagnostic tools.