Decoding Alzheimer's: Interpretable Visual and Logical Attention in Picture Description Tasks

Recent studies have started to incorporate imagery information from picture-description tasks in clinical interviews to automate Alzheimer's disease detection in the elderly. However, the high-level logical flow of visual-attention cognition mechanisms has not yet been investigated for enhanced interpretability. In this study, we systematically analyze the elements of picture-description tasks and propose a set of top-to-bottom human-interpretable features to describe the cognitive behaviors of patients, focusing on visual attention patterns, description quality, and repetition characteristics. These features achieve 85% accuracy in AD detection without specialized equipment, offering valuable insights for clinical practices and non-expert caregivers. Our results demonstrate that these high-level descriptive features, particularly those related to visual attention and the logical flow of speech, serve as effective biomarkers for AD detection.