Large language models for generating longitudinal synthetic data in low-risk pregnancy care

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Reproductive Medicine

Background:

• Access to comprehensive obstetric data is limited due to privacy and logistical constraints.
• Synthetic data generation offers a potential solution to overcome these limitations.
• Longitudinal data is crucial for understanding the continuum of obstetric care.

Purpose of the Study:

• To describe the generation and validation of a synthetic obstetric dataset.
• To provide a resource for research in maternal health, simulation studies, and machine learning development.
• To address data access and privacy concerns in obstetric research.

Main Methods:

• Utilized a large language model with zero-shot prompting and iterative prompt engineering.
• Defined a longitudinal schema, essential variables, and domain-specific generation rules.
• Incorporated expert feedback from General and Family Medicine and Data Science panels for refinement.
• Performed rule-based and plausibility-oriented validation for internal coherence and clinical assumptions.

Main Results:

• Generated a synthetic obstetric dataset for 5000 low-risk pregnant women, covering preconception to post-birth.
• Dataset includes trimester-specific clinical measurements (e.g., gestational weight, hemoglobin) and perinatal variables (e.g., consultations, delivery outcomes).
• Validated dataset ensures statistical coherence and clinical plausibility, mitigating privacy issues.

Conclusions:

• The synthetic obstetric dataset effectively overcomes data access and privacy barriers.
• This resource is suitable for simulation studies, methodological development, benchmarking, and educational machine learning.
• The dataset is not intended for direct clinical decision-making or model deployment in clinical settings.