Elder-Sim: A Psychometrically Validated Platform for Personality-Stable Elderly Digital Twins

A research team from multiple institutions has published a paper on arXiv detailing ELDER-SIM, a novel platform designed to create personality-stable digital twins of elderly individuals. The system addresses a critical problem in AI-assisted geriatric care: personality drift, where conversational agents exhibit inconsistent traits over repeated interactions, undermining their reliability for simulating long-term care scenarios. ELDER-SIM combines several technical components, including Big Five (OCEAN) personality trait specifications, a Cognitive Conceptualization Diagram (CCD) based on Beck's Cognitive Behavioral Therapy framework, and a MySQL-based long-term memory module. The platform was built using n8n workflow orchestration with local LLM inference through Ollama/vLLM.

The researchers conducted ablation studies across four configurations to measure effectiveness. The baseline model achieved 83.3% role discrimination accuracy. Adding long-term memory improved this to 88.9%, while incorporating the CCD structure boosted it to 94.4%. The most effective configuration involved fine-tuning a model using LoRA on 19,717 instruction pairs from the CHARLS dataset, achieving 97.2% accuracy. Psychometric validation showed the CCD produced the largest consistency gain (mean Cronbach's α improved from 0.702 to 0.892), while the fine-tuned model achieved the highest overall consistency (α 0.940; ICC 0.958). This represents a significant advancement in creating reliable, longitudinal simulations for elderly mental health care, providing a reproducible method for in silico evaluation before real-world clinical deployment.