Saba Amiri

This thesis addresses fundamental challenges in the development and implementation of Digital Health Twin (DHT) frameworks for distributed medical data analytics. The increasing volume and complexity of healthcare data, combined with stringent privacy regulations and data governance policies, necessitate innovative approaches to data access and utilization. We propose and evaluate methodologies for generating high-quality synthetic data using advanced generative models as an alternative to direct access of sensitive patient information.

The Digital Health Twin concept offers a promising framework for personalized medicine, enabling virtual representations of individual health status that can be continuously updated with diverse data streams. However, realizing this potential is hampered by practical difficulties in accessing and processing real-world medical data across different administrative domains. While federated learning approaches partially address these challenges by bringing computation to the data rather than centralizing sensitive information, they still face significant hurdles related to permission acquisition, coordination of training iterations, and continuous data availability.

To overcome these limitations, we investigate the use of generative models to create synthetic data that statistically mirrors the properties of real-world patient data without containing actual information about individuals. This approach creates a privacy-preserving abstraction layer between sensitive data and the model training process. Unlike federated learning, which requires repeated access to real data throughout the iterative model development process, synthetic data generation requires only a single authorization phase, after which model development can proceed independently without further access to sensitive information. Additionally, synthetic data enables asynchronous model development across distributed nodes, eliminating the coordination challenges inherent in federated approaches where all participants must be simultaneously available. This fundamental shift in data access patterns offers significant practical advantages in healthcare settings where obtaining and maintaining data permissions is particularly complex and time-consuming.

Given the potential advantages of synthetic data for DHT implementation, our research focuses on enhancing the utility and privacy characteristics of generative models specifically for healthcare applications. Healthcare data presents unique challenges that standard generative approaches struggle to address adequately, potentially limiting the practical value of synthetic data in clinical contexts. Our work addresses four critical challenges to increase the suitability of synthetic data for DHT frameworks: (1) accurately modeling mixed-tail behavior in healthcare data, where variables within the same dataset exhibit fundamentally different distributional characteristics; (2) balancing fidelity and diversity in synthetic data generation to ensure both accurate representation of the original data and sufficient variation; (3) preserving privacy while maintaining data utility; and (4) handling data heterogeneity in geographically distributed environments.

For the mixed-tail modeling challenge, we develop tail-adaptive normalizing flows that can simultaneously accommodate both heavy-tailed and light-tailed distributions within a unified framework. This approach aims to overcome the optimization instability, flexibility limitations, and likelihood balancing issues common in standard generative models when applied to healthcare data with diverse distributional characteristics.

To address the fidelity-diversity trade-off, we propose a model-agnostic framework for unsupervised learning of semantic constraints in target datasets. This includes a validator component that guides generative models toward respecting both explicit and implicit boundaries in the real data, enhancing the quality of synthetic outputs while maintaining appropriate variation.

For privacy preservation, we introduce a noise-free mechanism for generating differentially private tabular synthetic data using normalizing flows. This method improves upon current noise-adding approaches, potentially offering stronger privacy guarantees while preserving important statistical properties of the original data.

Finally, to handle data heterogeneity in distributed environments, we first demonstrate the sensitivity of privacy-preserving federated learning setups to data distribution shifts and their adverse impacts on utility and fairness metrics. Building on these insights, we develop and evaluate a proof-of-concept implementation using real-world private healthcare data that demonstrates how synthetic data can effectively replace real data in distributed analytics pipelines. We empirically evaluate this approach, comparing performance metrics between synthetic and real data scenarios to quantify the practical viability of synthetic data solutions in addressing challenges of data heterogeneity while maintaining both utility and fairness.

The thesis demonstrates the practical application of these methodologies through a proof-of-concept DHT framework optimized for clinical pathways. Our findings suggest that appropriately designed generative models can produce synthetic data with sufficient quality to support various downstream tasks while addressing privacy concerns and regulatory requirements. The proposed approaches may help unlock the potential of distributed healthcare data for personalized medicine and preventive care while respecting the fundamental principles of data privacy and sovereignty.

This research contributes to the broader fields of distributed data analytics, privacy-preserving machine learning, and healthcare informatics by offering novel technical solutions to long-standing challenges in data access and utilization. The methodologies developed may be applicable beyond healthcare to other domains where sensitive data and distributed processing present similar challenges.