Uncovering Latent Cognitive, Metacognitive, and Neural Bases of Problem-Solving Deficits in Children with Developmental Dyscalculia

Developmental dyscalculia (DD) affects 7-10% of children and is characterized by persistent difficulties in arithmetic problem-solving, which depends on the dynamic interplay between cognitive and metacognitive control processes governing strategy execution and selection. Conventional behavioral and neuroimaging approaches do not simultaneously quantify these latent processes and link them to distributed neural mechanisms, leaving the neurocognitive and metacognitive underpinnings of DD insufficiently characterized. We integrated Bayesian computational modeling with whole-brain decoding to uncover cognitive, metacognitive, and neural bases of problem-solving deficits in DD using a sample of 68 children aged 8-10 years (38 female). Children with DD showed deficits across all problem-solving strategies (counting, retrieval, and decomposition), indicating reduced strategy execution efficiency. They also exhibited significant metacognitive impairments: longer baseline switching times between strategies, reduced sensitivity to problem difficulty, and lower adaptivity in selecting optimal strategies. Whole-brain Elastic Net decoding accurately distinguished DD from TA peers based on distributed activation patterns across frontoparietal cortex, hippocampus, and ventral-visual regions, whereas univariate approaches did not. Linear multivariate brain-behavioral analyses and nonlinear transformer-based predictions converged in showing that these distributed signatures predicted counting efficiency and switching sensitivity. Notably, brain-behavior relations reversed in direction between groups, positive in TA children but negative in DD, most strikingly in the hippocampus. These findings demonstrate that arithmetic difficulties in DD arise from deficits in both strategy execution efficiency and metacognitive control, reflecting an altered mapping between distributed neural activity, and highlight the power of integrating computational modeling with brain-wide decoding to understand learning disabilities.