Abstract / Summary
Early detection of Alzheimer's disease (AD) remains a critical challenge in clinical practice. Standardized cognitive assessment instruments provide structured evaluation across cognitive domains, but their automation using artificial intelligence is limited by the lack of structured, annotated datasets capturing real clinical interactions. Existing resources depend primarily on unstructured narrative or conversational speech and do not capture the task-level prompt-response interactions present in clinician-administered assessments. This paper introduces a structured, UDS-aligned cognitive assessment dataset comprising 309 annotated segments from 11 publicly available cognitive evaluation videos, spanning 12 domains including Orientation, Memory, Language, and Working Memory. Each segment maps examiner prompts and participant responses to standardized cognitive constructs defined in the NACC Uniform Data Set (UDS-4), preserving temporal boundaries, response modality, correctness labels, and ASR risk metadata. To validate dataset utility, a zero-shot automated scoring experiment using Llama-3 achieves 70.0% accuracy and macro F1 of 0.54 across correct, incorrect, and partial response classes, demonstrating that the annotations contain sufficient clinical signal for downstream AI modeling.