Theory-driven classification of reading difficulties from fMRI data using Bayesian latent-mixture models

Decades of research have led to several competing theories regarding the neural contributors to impaired reading. But how can we know which theory (or theories) identifies the types of markers that indeed differentiate between individuals with reading disabilities (RD) and their typically developing (TD) peers? To answer this question, we propose a new analytical tool for theory evaluation and comparison, grounded in the Bayesian latent-mixture modeling framework. We start by constructing a series of latent-mixture classification models, each reflecting one existing theoretical claim regarding the neurofunctional markers of RD (highlighting network-level differences in either mean activation, inter-subject heterogeneity, inter-region variability, or connectivity). Then, we run each model on fMRI data alone (i.e., while models are blind to participants' behavioral status), which enables us to interpret the fit between a model's classification of participants and their behavioral (known) RD/TD status as an estimate of its explanatory power. Results from n=127 adolescents and young adults (RD: n=59; TD: n=68) show that models based on network-level differences in mean activation and heterogeneity failed to differentiate between TD and RD individuals. In contrast, classifications based on variability and connectivity were significantly associated with participants' behavioral status. These findings suggest that differences in inter-region variability and connectivity may be better network-level markers of RD than mean activation or heterogeneity (at least in some populations and tasks). More broadly, the results demonstrate the promise of latent-mixture modeling as a theory-driven tool for evaluating different theoretical claims regarding neural contributors to language disorders and other cognitive traits.