Deep generative hidden Markov models for synthetic patient data generation: a novel approach for medical AI research

Abstract

Background High-quality healthcare data is often hard to access due to privacy rules and limited availability, making it difficult to develop and test clinical AI models. Objective We propose a Deep Generative Hidden Markov Model (DG-HMM) that combines deep neural networks with probabilistic sequential modeling to generate synthetic patient data that is clinically realistic while protecting privacy. Methods DG-HMM uses a deep encoder-decoder structure with a flexible HMM layer to model temporal patterns and mixed clinical features. We trained and tested it on three real datasets: MIMIC-III ICU records, long-term diabetes data, and mental health progression records. We measured statistical similarity, temporal consistency, clinical rule compliance, and privacy protection. Results DG-HMM demonstrated superior performance compared to baseline methods, preserving about 94.2 of correlations, following clinical rules in 96.3 of cases, and resisting membership inference attacks in 89.4 of attempts. Predictive models trained on synthetic data were only 2-5 less accurate than those trained on real data. Conclusion DG-HMM offers a robust framework to create synthetic healthcare data for research and collaboration where privacy limits real data sharing. It can help support medical AI development in constrained settings.