Early and accurate detection of Autism Spectrum Disorder (ASD) is essential for effective intervention and lifelong support. Recent advances in machine learning and eye-tracking technologies have enabled objective screening approaches based on gaze behavior, providing a non-invasive alternative to traditional diagnostic methods. This study introduces a hybrid ensemble learning framework that integrates bagging, boosting, and stacking strategies to leverage the complementary strengths of diverse classifiers, including decision trees, random forests, k-nearest neighbors, AdaBoost, and gradient boosting. The framework was evaluated on a benchmark eye-tracking dataset of 28 participants (14 ASD, 14 TD) using five-fold cross-validation. The proposed stacking model achieved a mean accuracy of 0.960 ± 0.080 and an F1-score of 0.952 ± 0.095 across the five folds, outperforming individual base learners and performing comparably to image-based deep learning models. By operating directly on raw scanpath data, the proposed model enhances interpretability and reduces preprocessing overhead by eliminating the need for image generation and transformation steps, making it suitable for practical and real-time screening environments. Future work will focus on validating the framework on larger and more diverse datasets and integrating it within an end-to-end ASD screening platform.
Alagha et al. (Tue,) studied this question.