Unsupervised Eye-Tracking Dyslexia Detection Achieving 91.52% Accuracy on the ETDD70 Dataset

Dyslexia is widely recognized as one of the most common specific learning disorders in children, manifesting primarily as reading and spelling difficulties despite adequate intelligence and educational opportunities. Leveraging eye-tracking technology and advanced machine-learning techniques has increasingly been shown to improve diagnostic accuracy and potentially guide more effective interventions. In this paper, we present a comprehensive methodology and empirical evaluation of an unsupervised pipeline that achieved 91.52% accuracy in distinguishing dyslexic from non-dyslexic readers within the publicly available ETDD70 (Eye-Tracking Dyslexia Dataset, 70 participants). Specifically, we employed no outlier filtering, balanced the dataset by undersampling, extracted up to 51 eye-movement metrics (including ratio features and interquartile statistics), performed 15-component Principal Component Analysis (PCA), embedded the participants into 2D space via Uniform Manifold Approximation and Projection (UMAP), and clustered with HDBSCAN. Our best pipeline removed only 11 participants as noise and successfully labeled 59 participants into two clusters that aligned with dyslexia diagnoses at over 91% unsupervised accuracy—a finding supported by a 0.74 Silhouette Score and a 0.68 Adjusted Rand Index. This study thus provides strong evidence that advanced unsupervised clustering, guided by robust feature engineering and dimensionality reduction, can reveal meaningful reading-group structures in children’s eye movements, circumventing the need for labeled training samples. We discuss the ETDD70 dataset, highlight the synergy of multiple reading tasks (Syllables, MeaningfulText, PseudoText), and offer perspectives on future expansions such as multimodal data integration and refined eye-movement metrics.