Deep Learning Methods to Detect Dyslexia in Russian native-speaking School Pupils From Their Eye Fixation Data

This study explores the potential for using eye fixation data to diagnose dyslexia by focusing on different types of data representations. The goal was to determine the most appropriate type of data representation for detecting dyslexia. We conducted a detailed analysis and comparison of several types of neural networks, including convolutional (CNN), long short-term memory (LSTM) networks and convolutional long short-term memory (ConvLSTM), and ensemble models such as random forest (RF) and gradient boosting classifier (GBC) were also used in the work. Bayesian optimization methods were used to efficiently select model parameters. The data, that was used in the study, was collected by Center for Language and Brain, HSE University, forms the largest dataset on eye movement. The study concluded that representing eye fixations as time window series and processing them with LSTM neural networks resulted in high efficiency, achieving a 96.8% ROC AUC in the task of dyslexia identification.