Machine learning-driven risk screening for Chinese developmental dyslexia in children: Theoretical framework, feature selection, and simulation-based methodological illustration

Developmental dyslexia affects Chinese children’s reading development, and school-based risk screening may help identify children who require further assessment and support. However, interpretable screening-oriented models remain limited by small samples, insufficient standardization, limited longitudinal evidence, and weak cross-school validation. This study proposes a simulation-based methodological framework for organizing and evaluating multidimensional cognitive indicators of Chinese developmental dyslexia. The framework integrates phonological processing, orthographic processing, attentional control, and working memory indicators within a feature-selection and explainable machine-learning pipeline combining statistical screening, LASSO-based embedded selection, ensemble-model importance analysis, and SHAP-based interpretation. Simulated experiments informed by prior empirical distributions were used to illustrate the pipeline’s internal behavior, methodological coherence, and interpretability under predefined data-generating assumptions. The simulation reproduced expected risk-feature patterns, including the prominence of rapid automatized naming, phonological awareness, orthographic processing, and cognitive-control indicators. These patterns, however, reflect recovery of predefined simulation signals rather than independent empirical identification of dyslexia predictors. Similarly, the reported AUC values, SHAP rankings, and threshold estimates are simulation-internal results, not evidence of real-world screening accuracy, clinical utility, or school-based readiness. The study therefore provides an illustrative methodological framework, rather than validating a practical screening tool.