Researchers have developed a language-based digital twin that uses large language models to mimic the conversational behavior of older adults and help identify signs of early cognitive decline.
The work, published as a preprint on arXiv, focuses on Mild Cognitive Impairment, a condition that can precede dementia and is often difficult to detect early with conventional clinical tools. The team argues that speech and language offer a practical, non-invasive window into cognitive health because changes in fluency, word choice, pauses and tempo can reflect underlying decline.
The approach builds a personalized virtual model of each participant’s speech patterns rather than relying only on standard prediction. The researchers describe the system as a digital twin that can reproduce an individual’s linguistic style, timing and conversational habits over time. To do that, they fine-tuned a version of GPT-4.1-mini using conversational transcripts, along with participant metadata such as age, gender, interview date and topic.
A notable part of the method is the use of stylometric markers. The transcripts were augmented with tokens that encode pause length and speaking tempo, giving the model more information about how a person speaks, not just what they say. The authors say those features can be useful because they are linked to cognitive status and may capture changes that are not obvious in a single assessment.
To judge whether the digital twin was accurate, the researchers introduced a second model based on a conditional variational autoencoder, or cVAE. That evaluator was designed to measure how closely generated responses matched real ones and to estimate cognitive scores at the same time. In the paper, the authors say this setup lets them assess both conversational fidelity and cognitive relevance.
The team evaluated the framework using the I-CONECT dataset, which includes older adults ages 75 and above who were either cognitively normal or diagnosed with Mild Cognitive Impairment. The dataset comes from a randomized clinical trial centered on conversational engagement.
According to the paper, the digital twin preserved identity-specific speech characteristics and produced reconstruction and cognitive prediction errors comparable to those seen in real data. It also performed better than baseline responses generated by GPT.
The researchers say these results suggest the system could support scalable, continuous monitoring of cognitive health without requiring invasive testing or frequent in-person assessments. They frame the work as a step beyond previous digital twin systems that have mostly focused on physiological signals or broad population-level modeling.
The authors position the model as a tool for personalized cognitive assistance and longitudinal monitoring rather than a standalone diagnostic device. They say language-based digital twins could eventually help clinicians detect subtle changes over time, especially in older adults who may not receive regular formal cognitive testing.
The paper also highlights the growing role of LLMs in healthcare applications where unstructured data, such as conversation, can carry clinical meaning. By combining linguistic content, temporal speech features and participant metadata, the proposed framework aims to capture both style and cognitive state in a single model.
The implementation details were made available by the researchers on GitHub. As with other preprint-stage studies, the work has not yet been described as peer reviewed, but it adds to a broader wave of research exploring how generative AI can be adapted for personalized health monitoring.