Designing and Evaluating a Conversational Agent for Early Diagnosis of Alzheimer's Disease and Related Dementias

TL;DR

The paper tackles the problem of enabling timely ADRD diagnosis by collecting rich patient narratives via a voice-based conversational agent powered by large language models. It develops a socio-technical interview system tailored for older adults and multi-party sessions, and evaluates it in a within-subject study (n=30) against clinician interviews. Results show promising agreement in symptom elicitation (83.5% sensitivity, 100% specificity) and high user acceptance, with sequential prompting improving depth but risking topic coverage. The work demonstrates feasibility and provides design considerations for integrating agent-led history taking into clinical workflows, while acknowledging limitations and outlining directions for scaling, separation of voices, and validation of clinical utility.

Abstract

Early diagnosis of Alzheimer's disease and related dementias (ADRD) is critical for timely intervention, yet most diagnoses are delayed until advanced stages. While comprehensive patient narratives are essential for accurate diagnosis, prior work has largely focused on screening studies that classify cognitive status from interactions rather than supporting the diagnostic process. We designed voice-interactive conversational agents, leveraging large language models (LLMs), to elicit narratives relevant to ADRD from patients and informants. We evaluated the agent with 30 adults with suspected ADRD through conversation analysis, user surveys, and analysis of symptom elicitation compared to blinded specialist interviews. Symptoms detected by the agent showed promising agreement with those identified by specialists. Users appreciated the agent's patience and systematic questioning, which supported engagement and expression of complex, hard-to-describe experiences. While these findings suggest potential for conversational agents as structured diagnostic support tools, further validation with larger samples and assessment of clinical utility is needed before deployment.

Figures (7)

• Figure 1: Web interface of the agent as seen by patients and informants during an interview. The left panel shows the animated agent avatar, while the right panel displays the real-time transcript, audio waveform, and playback controls.
• Figure 2: High-level interview instructions provided to the conversational agent, outlining the agent’s role, tone, and interaction constraints for the interview with older adults and their informants.
• Figure 3: Example of a topic-specific script (Memory Difficulties) containing a main question and conditional follow-ups, designed to elicit diagnostically relevant details.
• Figure 4: Patient satisfaction ratings (n=19) across six questionnaire items, each measured on a 5-point Likert scale. (5 = very positive, 1 = very negative)
• Figure 5: Box plots comparing per-symptom ambiguity rates (top row) and not-discussed rates (bottom row) between chatbot and clinician interviews (n=23). Left column (A, D) shows overall distributions across all architecture-assigned participants. Middle column (B, E) shows rates stratified by chatbot architecture type (Single Prompt vs Multi Prompt). Right column (C, F) shows the distribution of rate differences (Agent - Clinician) by architecture type. Box plots display median, quartiles, and outliers for per-symptom rates. Positive differences indicate higher rates for the agent compared to clinician interviews.