DysDiTect: Dyslexia Identification Using CNN-Positional-LSTM-Attention Modeling with Chinese Dictation Task.

Area of Science:

• Neuroscience
• Computational Linguistics
• Educational Psychology

Background:

• Handwriting difficulty is a key indicator of Chinese developmental dyslexia (DD).
• Previous models overlooked the temporal aspects of character writing in dictation tasks.
• Chinese characters' complex structure poses unique challenges for dyslexia assessment.

Purpose of the Study:

• To develop and evaluate a machine learning model for identifying Chinese developmental dyslexia (DD) in children.
• To incorporate temporal-sequential handwriting features into dyslexia detection.
• To assess the model's performance using both handwriting data and grade information.

Main Methods:

• Combined Convolutional Neural Network (CNN) with Long Short-Term Memory (LSTM) and attention mechanisms.
• Utilized transfer learning and positional encoding for feature extraction.
• Trained and tested the model on 100,000 Chinese characters from 1064 children (Grades 2-6).

Main Results:

• The model achieved 83.2% accuracy using handwriting features alone.
• Incorporating grade information improved classification accuracy to 85.0%.
• High sensitivity (83.3%) and specificity (86.4%) were observed with grade information.

Conclusions:

• Machine learning models can effectively identify children at risk for Chinese developmental dyslexia (DD).
• Integrating temporal-sequential handwriting analysis enhances dyslexia detection accuracy.
• Early identification of DD is feasible using advanced computational methods.