Enhanced Snow Geese Algorithm and Transformer Models for Dyslexia Detection

Dyslexia detection systems face significant challenges due to the need for domain-specific discriminative features from high-dimensional data. Traditional dyslexia detection methods often fail to capture complex non-linear relationships and temporal dependencies inherent in dyslexia-related datasets. To address these limitations, we develop a dual approach that integrates deep learning-based feature extraction with an enhanced snow geese algorithm (SGA) for feature selection. In the feature extraction phase, the transformer with a multi-layer perceptron (MLP) architecture is applied for tabular data feature extraction and a vision transformer for image-based feature learning. These transformers have a high ability to extract relevant representation from dyslexia. Additionally, we introduce an enhanced feature selection technique using a modified SGA optimized with a multi-operator neighborhood-based opposition-based learning approach. Our method leverages the hierarchical feature mapping capabilities of SwinV2 transformers and the self-attention mechanisms of MLP-enhanced transformers to improve classification performance. We evaluate our model on the Eye-Tracking Dyslexia Dataset (ETDD70), which contains eye-tracking recordings from 70 participants performing reading tasks. Experimental results demonstrate that our approach outperforms other state-of-the-art feature selection approaches across multiple evaluation metrics. For example, the proposed model has average accuracy, F1-score, and sensitivity of 92.45%, 91.96%, and 92.05%, respectively. In addition, they offer a clinically viable solution for dyslexia detection.