Electronic Health Records (EHRs) provide rich longitudinal patient data with potential for predictive modeling, yet remain underutilized due to privacy restrictions on data sharing, irregular sampling with informative missingness, and heterogeneous structure requiring extensive feature engineering. This thesis addresses these challenges through four methodological contributions organized along two complementary axes: privacy-preserving synthetic EHR generation and supervised learning on irregularly sampled clinical time series.
For synthetic EHR generation, we first introduce TimEHR, which reformulates irregularly sampled time series as two-channel images encoding observed values and missingness patterns. A two-stage conditional GAN explicitly models non-random observation processes by decoupling missingness generation from value generation, addressing a key source of bias in clinical data. We then present SynEHRgy, which extends time-series-only generation to synthesize complete real-world EHRs comprising static covariates, clinical event codes, and irregular time series across multiple patient visits. By unifying heterogeneous EHR components through a tokenization scheme and formulating synthesis as causal language modeling, SynEHRgy leverages decoder-only transformers to generate realistic longitudinal records without modality-specific generative architectures. Both methods improve fidelity, utility, and privacy over state-of-the-art baselines on real-world EHR datasets.
For supervised learning, we pursue two strategies. TEE4EHR augments deep learning models with a Transformer Event Encoder that treats measurement timing as a marked temporal point process, capturing informative observation patterns to improve outcome prediction and patient representation learning. In contrast, FeatEHR-LLM leverages large language models to generate interpretable, executable feature construction rules from task descriptions and variable-level metadata, enabling clinically grounded feature engineering without patient-level data access or manual design. Across multiple tasks, FeatEHR-LLM-generated features improve downstream performance by up to 6 AUROC points relative to standard baselines.
These methods enable privacy-compliant data sharing, robust prediction under irregular sampling, and integration of clinical knowledge into automated machine learning pipelines, with applications demonstrated on large-scale ICU and hospital EHR datasets.
