Workflow-Aware Clinical Prediction Using Electronic Health Record Clickstreams: Representation Strategies across Machine Learning Frameworks

Abstract

Electronic health records produce two distinct information sources from every clinical encounter: patient data, which documents the patient's evolving health state, and EHR use metadata, which captures how clinicians interact with the record while delivering care. The most extensively studied form of this metadata is the audit log, a time-stamped record of every chart access, order entry, and note signed during care. Most clinical prediction models are built on patient data alone, leaving the documented process of care largely unexploited. This dissertation investigates whether audit logs carry complementary workflow-aware signals that improve clinical prediction and develops representation strategies that render these signals usable across distinct machine learning frameworks. While audit logs contain useful signal, that signal is hard to use, because the event streams are high-dimensional, sparse, and temporally irregular. This dissertation addresses this challenge through three aims, each pairing one model class with a tailored representation strategy. First, we transform raw event streams into fixed-dimensional tabular features for gradient-boosted trees, showing that recency-aware count aggregation adds complementary value for next-day hospital discharge prediction. Second, we serialize event sequences as text for pretrained large language models, identifying first-occurrence serialization as the strongest representation and a hybrid design that augments the prompt with a calibrated tree-based probability. Third, we develop a transformer-based clinical foundation model pretrained jointly on patient data and audit-log events, showing that joint pretraining improves downstream discharge prediction, 30-day readmission and acute kidney injury onset. Together, these contributions establish EHR audit logs as a predictive data modality, deliver model-specific representation strategies for an underused data source, and build toward workflow-aware clinical AI that learns from both patient state and clinician activity.