Abstract

Electronic Health Records (EHRs) have transformed clinical care by enabling structured documentation and streamlining workflows. Characterizing the patterns of EHR use has relied on methods that treat actions as isolated events, limiting the ability to study the sequential dependencies that shape behavior. Modeling clinician EHR action sequences through the lens of a structured language—an approach inspired by prior work in human-computer interaction on the sequential and grammatical nature of user behaviors—offers a novel framework to capture the temporal and structural dependencies within workflows. This dissertation introduces the action-as-language framework and presents a three-part investigation into representing, quantifying, and modeling the structure of clinician behavioral patterns using various structured language modeling techniques. In Aim 1, I develop structure-sensitive sequence representations of EHR interactions to examine how patterns of interaction structure reflect clinical workflows. I apply frequency-weighted statistics and embedding techniques that capture action co-occurrence, evaluating their ability to distinguish work patterns across care settings. Findings show that even simple frequency-weighted representations can differentiate workflows, and embedding-based representations reveal clusters of tasks that align with known clinical activities. These results support the feasibility of using natural language-inspired methods to model clinical behavior in EHR use. In Aim 2, I introduce a novel metric—action entropy—derived from behavioral patterns to quantify the variability and cognitive effort in clinician behavior. By training a custom transformer-based model on directional progression of actions, I estimate how predictable or routine each action is based on learned interaction patterns. Initial validation in known high cognitive-effort scenarios shows significant increases in action entropy between cases and matched controls, suggesting its value as a proxy for cognitive effort in complex workflows. Finally, in Aim 3, I develop and evaluate a general-purpose modeling pipeline for clinician action sequences using large language models (LLMs). This pipeline is designed to learn from EHR action sequences using both symbolic and semantic representations, and is evaluated through two experiments: (1) next-action prediction to evaluate model learning of interaction patterns, and (2) wrong-patient error prediction to test its utility for detecting safety-critical events. Results show that the LLM-based approach more accurately predicts subsequent actions. However, all evaluated approaches—including both traditional and LLM-based models—perform poorly in detecting wrong-patient errors. This suggests that even though language models better capture sequential structure and variability in clinician actions, wrong-patient errors may be influenced by factors beyond what is encoded in prior EHR interactions. Collectively, this dissertation presents the action-as-language framework as a novel approach for understanding clinician behavior through structured language modeling. By capturing structural patterns in EHR action sequences, quantifying behavioral variability, and modeling sequential dependencies, this work lays the groundwork for future efforts to study clinical workflows, cognitive demands, and clinician interaction with EHR systems.

Committee Chair

Thomas Kannampallil

Committee Members

Aristeidis Sotiras; Julia Adler-Milstein; Philip Payne; Sunny Lou; Thomas Kannampallil

Degree

Doctor of Philosophy (PhD)

Author's Department

Biology and Biomedical Sciences

Author's School

Graduate School of Arts and Sciences

Document Type

Dissertation

Date of Award

8-12-2025

Language

English (en)

Author's ORCID

https://orcid.org/0000-0001-9447-4888

Download

Available for download on Wednesday, August 11, 2027

Included in

Biology Commons

Share

COinS