A Machine Learning Approach to Detecting Developmental Dyslexia: Achieving 100% Accuracy with CatBoost on Action Representation Data

Developmental dyslexia (DD) is a complex reading disorder characterized by persistent deficits in learning to read and write, often accompanied by a range of cognitive and motor impairments. Recent research, including the study by van de Walle de Ghelcke et al. (1), has highlighted possible deficits in internal action representation among adolescents with DD, suggesting broader sensorimotor integration challenges. In this paper, we present a novel machine learning (ML) pipeline aimed at classifying adolescents into dyslexic and non-dyslexic (control) groups based on a dataset of action representation and reading measures. Our dataset consists of 36 participants (18 with DD and 18 typically reading peers), and we employ a combination of state-of-the-art supervised ML algorithms—including Logistic Regression, Random Forest, XGBoost, CatBoost, and Support Vector Machine (SVM)—to discriminate between these two groups. We detail our data preprocessing steps, including imputation of missing values and oversampling via SMOTE, to address class imbalance. A thorough hyperparameter tuning via GridSearchCV is used to optimize each model. The CatBoost model achieves 100% accuracy, precision, and recall, culminating in an F1 score of 1.0000 and an AUC of 1.0000 on our test set—an outcome that suggests this model may be particularly well-suited to tasks of this nature. We discuss our findings in the context of sensorimotor theories of DD, emphasize the importance of advanced ML techniques in neuropsychological classification tasks, and provide directions for future research. Our study underscores that internal action representation impairments, measured via mental and actual movement times under Fitts’ law constraints, can be leveraged for accurate detection of DD, with potential clinical implications for early screening and intervention.