ABSTRACT Machine learning (ML) algorithms have been increasingly used to predict learning disability (LD) risk across various disciplines, but the effectiveness of different algorithms remains unclear. We summarize the literature on ML applications for the identification and classification of LDs using behavioral (e.g., phoneme manipulation and sound blending), behavioral biometric (handwriting, keyboarding, eye‐movements, facial expressions), and neuroimaging (e.g., EEG, fMRI) data. We use multi‐level meta‐analysis to understand how classification statistics performed across data sources, populations, study designs and purpose, algorithms (e.g., support vector machines), and model building approaches (e.g., feature selection, hyperparameter tuning), among other study‐ and model‐level characteristics. The meta‐analysis included 41 primary studies and 331 ML models. Across algorithms, pooled accuracy estimates ranged from 0.75 to 0.88. Factors such as sample composition (e.g., grade level), types of data (e.g., brain imaging vs. checklists), and inclusion of academic skills (e.g., spelling and executive functions) had varying effects depending on the algorithm, but even the most predictive models were subject to substantial residual heterogeneity. Overall, the findings provide a baseline snapshot of an emergent literature and the conclusions are expected to evolve as primary studies adopt more standardized and comprehensive reporting beyond model accuracy, use larger and more diverse samples (particularly for biological data sources), incorporate external validation, and more explicitly address equity and generalizability in model development and evaluation.
Ahmed et al. (Fri,) studied this question.