SynthAgent: A Multi-Agent LLM Framework for Realistic Patient Simulation -- A Case Study in Obesity with Mental Health Comorbidities

Simulating high-fidelity patients offers a powerful avenue for studying complex diseases while addressing the challenges of fragmented, biased, and privacy-restricted real-world data. In this study, we introduce SynthAgent, a novel Multi-Agent System (MAS) framework designed to model obesity patients with comorbid mental disorders, including depression, anxiety, social phobia, and binge eating disorder. SynthAgent integrates clinical and medical evidence from claims data, population surveys, and patient-centered literature to construct personalized virtual patients enriched with personality traits that influence adherence, emotion regulation, and lifestyle behaviors. Through autonomous agent interactions, the system simulates disease progression, treatment response, and life management across diverse psychosocial contexts. Evaluation of more than 100 generated patients demonstrated that GPT-5 and Claude 4.5 Sonnet achieved the highest fidelity as the core engine in the proposed MAS framework, outperforming Gemini 2.5 Pro and DeepSeek-R1. SynthAgent thus provides a scalable and privacy-preserving framework for exploring patient journeys, behavioral dynamics, and decision-making processes in both medical and psychological domains.

💡 Research Summary

SynthAgent introduces a novel multi‑agent system (MAS) for generating high‑fidelity synthetic patients that simultaneously model obesity and common mental‑health comorbidities such as depression, anxiety, social phobia, and binge‑eating disorder. The authors argue that real‑world clinical data are fragmented, biased, and subject to privacy constraints, limiting comprehensive research on the intertwined metabolic and psychological dimensions of obesity. To overcome these limitations, SynthAgent integrates four complementary data sources: (1) twelve cycles of the CDC’s National Health and Nutrition Examination Survey (NHANES) covering demographic, anthropometric, lifestyle, and mental‑health variables; (2) a de‑identified claims database (PurpleLab) comprising 70 000 patients with at least one obesity‑related claim, providing longitudinal diagnosis, procedure, and medication trajectories; (3) population‑level epidemiological probabilities from CDC’s Behavioral Risk Factor Surveillance System, the World Obesity Federation, and the National Comorbidity Survey; and (4) PubMed case‑report literature that supplies narrative clinical evidence.

The MAS consists of five specialized agents that operate sequentially. The Summarizer Agent creates a probabilistic demographic profile and matches it to the most similar NHANES respondent and three claim‑based disease trajectories, producing a structured blueprint. The Generator Agent expands this blueprint into a complete patient record, including demographics, BMI class, comorbidity list, symptom chronology, laboratory values, treatment plans, and personality scores derived from HEXACO, Reinforcement Sensitivity Theory, and the Temperament and Character Inventory. The Augmenter Agent enriches the record by querying PubMed for ten case reports per disease keyword, filtering them for age and gender compatibility, and synthesizing evidence‑based additions such as the link between anxiety and sleep disturbance. The Evaluator Agent performs automated quality assurance, checking demographic plausibility, temporal consistency (e.g., treatments follow diagnoses), clinical coherence, psychological alignment, and lifestyle realism, flagging issues as major, moderate, or minor. Finally, the Refiner Agent resolves the flagged inconsistencies and outputs a validated synthetic patient profile.

Performance benchmarking of large language models (LLMs) used as the core engine shows that GPT‑5 and Claude 4.5 Sonnet achieve the highest fidelity, as measured by Train‑on‑Synthetic‑Test‑on‑Real (TSTR) metrics and expert‑rated clinical realism. Gemini 2.5 Pro and DeepSeek‑R1 lag behind, indicating that the most recent LLMs better handle the complex, multimodal prompts required for coordinated multi‑agent operation.

The authors generated over 100 synthetic patients, each exhibiting a coherent blend of demographic, metabolic, behavioral, and personality attributes. By embedding personality dimensions, SynthAgent can simulate how traits such as conscientiousness or emotionality influence treatment adherence, emotional regulation, and lifestyle choices—critical factors in obesity management. The framework thus enables systematic exploration of “what‑if” scenarios, rare phenotypes, and stress‑testing of clinical decision‑support tools while preserving patient privacy.

Limitations include the lack of direct integration with standardized electronic health record formats (e.g., FHIR), absence of formal differential privacy mechanisms, and limited external clinical validation beyond internal evaluator reports. Future work should focus on embedding regulatory‑compliant data schemas, applying privacy‑preserving noise addition, and conducting head‑to‑head comparisons with real cohorts to quantify utility. Extending the MAS to other multimorbidity clusters (e.g., diabetes with depression) could further demonstrate its generalizability. Overall, SynthAgent represents a significant step toward privacy‑preserving, psychologically informed synthetic patient generation, offering a scalable platform for hypothesis testing, model prototyping, and education in complex chronic disease research.