In Silico Platform Shows Promise for Creating Personality-Stable Elderly Digital Twin Agents

BackgroundLarge language models (LLMs) enable patient-facing conversational agents, creating a plausible pathway toward patient digital twins that can capture older adults lived experiences, beliefs, and behavioral responses across time. A central barrier to clinical-grade digital twins is personality drift--inconsistent trait expression across repeated, longitudinal interactions--which can undermine the reliability of generated trajectories and intervention-response simulation in geriatric care contexts. ObjectiveTo develop ELDER-SIM as a multi-role elderly-care conversational platform oriented toward building personality-stable elderly digital twin agents, and to propose a psychometric validation framework for quantifying and improving personality consistency in LLM-based agents across repeated interactions. MethodsELDER-SIM was implemented using n8n workflow orchestration with local LLM inference (Ollama/vLLM), integrating (1) Big Five (OCEAN) trait specifications, (2) a three-layer Cognitive Conceptualization Diagram (CCD) grounded in Becks cognitive behavioral therapy framework, and (3) a MySQL-based long-term memory module for persistent states. We conducted a systematic ablation study across four conditions--Baseline (prompt-only), +Memory, +CCD, and +LoRA (domain-specific fine-tuning using 19,717 instruction pairs derived from CHARLS)--and evaluated personality consistency using Cronbachs (internal consistency), intraclass correlation coefficient (ICC) (test-retest reliability), and role discrimination accuracy . ResultsPersonality measurement reliability was acceptable to excellent across conditions (Cronbachs : 0.70-0.94), with consistently high test-retest stability (ICC: 0.85- 0.96). Role discrimination improved stepwise from 83.3% (Baseline) to 88.9% (+Memory), 94.4% (+CCD), and 97.2% (+LoRA). CCD produced the largest gain in internal consistency (mean 0.702[->]0.892), while LoRA achieved the highest overall internal consistency ( 0.940) and ICC (0.958). ConclusionsELDER-SIM provides a psychometrically validated approach for constructing personality-consistent elderly digital twin agents. By demonstrating that structured cognitive modeling and domain adaptation materially reduce personality drift, the framework supports more reliable longitudinal digital twins for elderly mental health and psychosocial care, enabling reproducible in silico evaluation of interaction trajectories and intervention strategies before clinical deployment.