Characterizing neural signatures of dyslexia and co-occurring math learning difficulties (MLD) with machine learning

Poster Session A - Saturday, April 13, 2024, 2:30 – 4:30 pm EDT, Sheraton Hall ABC

Margo Kersey¹ (margo.kersey@ucsf.edu), Janhavi Pillai¹, Rian Bogley¹, Marni Shabash¹, Dolce Vita Martin-Moreno¹, Elizabeth Carpenter¹, Boon Lead Tee¹, Jessica de Leon¹, Zachary Miller¹, Christa Pereira Watson¹, Maria Luisa Mandelli¹, Bettina Pedemonte¹, Maria Luisa Gorno Tempini¹, Pedro Pinheiro-Chagas¹; ¹University of California, San Francisco

Approximately 10% of children face persistent learning disabilities in reading and math. The most prevalent is developmental dyslexia, among which 40% of children concurrently experience math learning difficulties (MLD). This study aims to investigate the multifactorial nature of dyslexia through the neural mechanisms of dyslexia and MLD. Using a machine learning approach, we analyzed MRI data from 272 children including cortical/subcortical and diffusion tract metrics. All children completed an extensive battery of neuropsychological testing and a novel math battery designed to diagnose MLD. Children were diagnosed as dyslexia-only (n=126), control (n=50), and dyslexia with MLD (n=96). A repeated, 5-fold cross-validated model comparison identified the most effective composite model: Recursive Feature Elimination with a Linear Support Vector Classification (LinearSVC) estimator, followed by LinearSVC for classification. We achieved 67.4% accuracy in differentiating children with dyslexia-only from controls. Permutation feature importance analysis highlighted significant neural markers, such as the superior longitudinal fasciculus IP and right superior temporal gyrus, consistent with established language processing research. Furthermore, we reached 57.6% accuracy in distinguishing dyslexia-only from dyslexia with MLD, emphasizing regions such as the left temporal pole, right fusiform gyrus, and left middle temporal gyrus. These regions, implicated in language, semantic memory, and numeral processing, correspond with common struggles in arithmetic fact retrieval and calculation procedures found in children with MLD. Our findings corroborate known neural markers of dyslexia and reveal additional brain areas implicated in MLD, underlining the need for intervention strategies tailored to the unique neural patterns associated with specific learning challenges.

Topic Area: LANGUAGE: Development & aging