Dyslexia, which affects 10% to 20% of the global population, poses significant challenges to learning, underscoring the need for accessible diagnostic tools. This study explores the use of eye-tracking technology combined with machine learning as a cost-effective and non-invasive approach for early dyslexia detection. By analyzing key eye movement patterns—such as prolonged fixations and erratic saccades—we proposed an enhanced feature framework and achieved 88.58% accuracy using a Random Forest Classifier. Hierarchical clustering was also applied to uncover varying dyslexia severity levels. The results, validated across diverse populations and settings, highlight the method’s scalability and potential for identifying even borderline dyslexia traits, offering a promising advancement in clinical diagnostics.