Chatbot AI helps screen older adults for memory decline symptoms
Conversational AI for Perceived Cognitive Decline Screening in Older Adults
Information Retrieval
Summary
It can be hard and expensive to detect early memory problems in older adults. The authors tested a chatbot that asks open questions and listens to people’s answers to spot signs of cognitive decline. They compared different computer methods for analyzing these conversations and found some promising results. The chatbot approach could help screen many people more easily, but larger studies are needed to confirm its accuracy and clinical usefulness.
What this means in practice
- •For health screening teams: Use chatbot interviews with older adults to gather language data that helps identify those with potential cognitive decline signs.
- •For clinical informatics teams: Incorporate automated text analysis from chatbot conversations into workflows to support early cognitive decline detection in outpatient settings.
Authors
Nikita Gautam, Moana Sargent, Kaitlyn Sievers, Ashley Sylvara, Doina Caragea, Heather Bailey, Tianjun Sun
Abstract
Early detection of cognitive decline in older adults remains challenging due to the resource demands of standard neuropsychological assessments. Chatbot-based Intelligent Cognitive Assistants (ICAs) provide a scalable alternative by eliciting open-ended responses rich in linguistic, semantic, and autobiographical cues. This study evaluates whether ICA question-answer (Q/A) interviews with older adults can screen for participant-reported perceived cognitive decline and related cognitive, functional, and personality outcomes. We compare traditional machine learning, frozen transformer embeddings, fine-tuned transformers, and large language model (LLM) reasoning strategies across Global Q/A, Single Q/A confidence averaging, and Aggregated Q/A pipelines. Among text-based methods, Gemini 3.1 Flash Lite (1-example-per-class, Aggregated Q/A) achieved the highest held-out test performance (0.742 accuracy, 0.739 balanced accuracy, 0.728 macro-F1). Given the small test set (n = 31), these results are interpreted descriptively. Frozen clinical transformer embeddings with a random forest head led non-LLM text methods, whereas fine-tuned models showed limited generalization. Models utilizing structured questionnaires, cognitive tasks, and personality traits outperformed text-based models in classification. Continuous regression analyses revealed modest predictive signals for AD8, FAQ, IQCODE, neuroticism, and conscientiousness. Exploratory prompt-level analyses identified both direct symptom-related and indirect autobiographical questions with strong test metrics, though sample sizes varied across prompts. These findings indicate that ICA chatbot interviews capture meaningful signals of perceived cognitive decline, supporting scalable screening research. Larger, clinically diverse cohorts with validated clinical outcomes are needed for clinical translation.