Mehdi Ghayoumi and Kambiz Ghazinour, SUNY, USA

This paper proposes a deep learning paradigm for the early detection of Alzheimer’s disease (AD) through analysis of eye movement patterns. By using a publicly available dataset containing ocular data from both early-stage AD patients and healthy controls, we construct a balanced dataset that effectively encapsulates the temporal intricacies of saccades and fixations. The core of our framework is a Bidirectional Long Short-Term Memory (Bi-LSTM) architecture, enriched with a dynamic attention module that adaptively emphasizes salient ocular biomarkers, particularly subtle variations in saccade amplitudes and fixation durations. In contrast to conventional machine learning techniques, our model excels in extracting latent features and capturing complex temporal dependencies by leveraging the bi-directionality of LSTM layers. The inclusion of the attention mechanism further enhances interpretability and robustness, selectively weighting critical eye movement segments with the highest predictive relevance for AD classification. Empirical evaluations demonstrate that this Bi-LSTM–Attention model achieves superior performance across multiple metrics, including accuracy, precision, recall, F1 score, and area under the Receiver Operating Characteristic (ROC) curve, surpassing traditional statistical and machine learning baselines. These findings underscore the viability of eye movement data as a rich, non-invasive source of information for the early detection of neurodegenerative disorders. Beyond its immediate clinical applications, this work lays the foundation for the broader adoption of eye-tracking technologies in cognitive assessments, potentially revolutionizing both the diagnostic process and management strategies for Alzheimer’s disease and other related conditions.

Alzheimer’s Disease, Eye Movement Analysis, Deep Learning, Bidirectional LSTM, Attention Mechanism, Neurodegenerative Disease, Biomarkers, Non-Invasive Screening