PatientFlow: Learning to generate mixed-type longitudinal clinical data with flow matching.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Clinical Data Science

Background:

• Longitudinal clinical data is crucial for deep learning models in complex diseases.
• Generating realistic synthetic patient data presents challenges in data structure modeling and privacy protection.

Purpose of the Study:

• To introduce PatientFlow, a novel generative model for creating synthetic longitudinal clinical data.
• To evaluate PatientFlow's ability to model complex patient data and protect privacy.

Main Methods:

• PatientFlow combines Variational Autoencoders for data representation and Flow Matching for patient generation.
• The model was evaluated on a large longitudinal cohort of Amyotrophic Lateral Sclerosis patients (N = 1560).
• Qualitative and quantitative assessments, including validation by expert clinicians, were performed.

Main Results:

• PatientFlow successfully generated high-fidelity synthetic longitudinal clinical data.
• Prognostic models trained on synthetic data matched or exceeded performance of models trained on real data across five endpoints.
• Expert clinicians validated the realism of the generated patient data.

Conclusions:

• PatientFlow effectively models longitudinal clinical data, offering a privacy-preserving solution for data augmentation.
• The method shows significant potential for advancing deep learning applications in healthcare by enabling secure data sharing and expansion.