A Pipeline for Generating Longitudinal Synthetic Clinical Notes Using Large Language Models

Overview

This paper introduces a scalable, modular pipeline designed to generate synthetic clinical documentation. By combining structured patient generation (using Synthea UK), semi-structured patient journey simulation, and unstructured clinical note generation (using LLMs), the authors create longitudinal records that mimic real-world hospital stays. The primary goal is to provide researchers with high-quality synthetic data for developing and testing clinical AI tools—such as summarization, coding, and decision support systems—without the privacy risks and regulatory hurdles associated with real patient data.

The Generation Pipeline

The pipeline operates in five distinct stages to ensure both medical plausibility and longitudinal coherence:

1. Patient Generation: Uses Synthea UK to create demographic foundations, supplemented by LLM-generated details like next-of-kin information.
2. Admission Generation: Defines the reason for hospital entry, specialty, and ward details, which can be customized based on specific research needs.
3. Journey Simulation: Generates a sequence of clinical events (e.g., ward rounds, surgeries, registrar reviews) using iterative LLM calls to ensure the patient's clinical status evolves logically over time.
4. Clinical Note Generation: Transforms the event descriptions into realistic clinical notes. The pipeline uses versatile prompts that adapt to the specific event type and clinical context.
5. Augmentation: Introduces clinical personas—such as 'Concise,' 'Narrative,' or 'Bullet Points'—to simulate the natural variation in documentation styles found among different healthcare staff.

Validation and Utility

The authors emphasize the importance of internal consistency, ensuring that clinical notes across a single patient's journey reflect the same evolving clinical picture. To maintain quality, the pipeline incorporates LLM-based validators that check for realism and faithfulness to the simulated journey. The resulting dataset is released in tiers of validation (Bronze, Silver, and Gold), allowing users to select data based on their specific requirements for realism versus scale. The authors provide the code on GitHub and the dataset on Hugging Face to facilitate further research and iterative improvement of synthetic clinical data generation.