Review of LLM-generated synthetic Parkinson's patients for in silico trials

Heterogeneity in sporadic Parkinson's Disease (PD) is a critical problem that drives variable rates of progression and treatment response and complicates clinical trials. Access to large PD datasets that may help in clustering this heterogeneity is restricted by privacy and regulatory constraints. Simulated patients or digital twins may offer a solution. We developed a large language model (LLM)-framework to generate high-fidelity synthetic PD patients from the Parkinson's Progression Markers Initiative (PPMI) dataset based on the open-source Qwen3-8B-Base model. Using a relational, tree-structured representation and domain-specific fine-tuning, the model produces patient-level records with longitudinal clinical, imaging, and biomarker data. Fidelity was assessed through distributional similarity, correlation structure, and neurologist review. Utility was tested by training diagnostic classifiers, reproducing a published pharmacometric disease progression model applied to in silico trials, and by extracting a stringent dopamine-motor impairment relationship at early PD stages. Privacy was evaluated via identical match share, distance-to-closest-record, and membership inference attacks. These findings support the use of a dedicated LLM framework for patient simulation, contributing to the foundations of digital twins for PD in silico trials.