Machine Learning–Based Prediction of Reading Comprehension Deficits from Attention Control, Processing Speed, Phonological Awareness, and Academic Anxiety in Children with Dyslexia

Abstract

The present study aimed to predict reading comprehension deficits based on attention control, processing speed, phonological awareness, and academic anxiety using machine learning models in children with dyslexia. This study employed a correlational design with a machine learning predictive approach. The statistical population included children with dyslexia enrolled in elementary schools and learning disability centers in Tehran during the 2025–2026 academic year. A total of 240 children aged 8–12 years were selected using multistage cluster sampling and purposive screening procedures. Data collection instruments included the Reading Comprehension Subtest of the Gray Oral Reading Tests, the Attention Control Scale, the Processing Speed Index of the Wechsler Intelligence Scale for Children–Fifth Edition, the Comprehensive Test of Phonological Processing, and the Academic Anxiety Scale. Data were analyzed using SPSS-27 and Python machine learning libraries. Several algorithms including Random Forest, Support Vector Machine, Gradient Boosting, and Artificial Neural Network models were implemented. Model evaluation was conducted using accuracy, precision, recall, F1-score, root mean square error, and area under the receiver operating characteristic curve indices. The findings revealed significant relationships among all study variables. Reading comprehension deficits demonstrated significant negative correlations with attention control, processing speed, and phonological awareness, while academic anxiety showed a significant positive association with reading comprehension deficits. Among predictor variables, phonological awareness had the strongest relationship with reading comprehension deficits. The Gradient Boosting model demonstrated the highest predictive performance with 93% accuracy, 91% precision, 92% recall, an F1-score of 91%, and an area under the curve value of 0.96. The Random Forest model also showed high predictive performance with 91% accuracy and an area under the curve value of 0.94. Feature importance analysis indicated that phonological awareness was the strongest predictor, followed by attention control, academic anxiety, and processing speed. The findings support multifactorial neurocognitive models of dyslexia suggesting that reading comprehension deficits are influenced by integrated cognitive, executive, phonological, and emotional processes. Machine learning algorithms, particularly ensemble learning models, demonstrated high effectiveness in predicting reading comprehension deficits among children with dyslexia. These findings highlight the potential utility of machine learning approaches for early screening, individualized assessment, and targeted intervention planning in educational and clinical settings.