Nigel Kim
Program: Biomedical Informatics and Data Science
Current advisor: Thomas G. Kannampallil, PhD
Undergraduate university: Washington University, 2019
Enrollment year: 2021
Research summary
Adoption of electronic health records (EHRs) in clinical work environments has reshaped work. With EHR audit logs, we can study clinicians’ work behaviors at scale.
Clinical work environments, especially intensive care settings, are characterized by multiple competing demands, requiring clinicians to switch their attention between overlapping tasks. Given the limitations of human cognition, attention switching has been shown to be associated with decreased productivity and performance, increased cognitive burden, and errors in non-clinical settings. However, the consequences of attention switching are less understood in clinical settings due to reliance on labor-intensive observational studies to measure attention switching. We developed a scalable metric for attention switching based on passively collected electronic health record (EHR) audit log data, and used it to assess the downstream effects of attention switching in ICU clinicians.
We then sought to understand more complex patterns of clinical work using the EHR audit log data, which may provide us better context on understanding clinical workflows, and the potential factors that exacerbate workload burden, errors, and clinician burnout. We incorporated action-as-language framework, theories in human cognition, and techniques such as deep learning, statistical modeling, and language modeling to study long sequences of EHR-based actions. We began with analyzing individual action occurrences as part of a greater sequence of EHR-based workflow. A case study has been conducted to apply such approach in identify action sequence phenotypes associated with different clinical settings. We then expanded it to generate NLP- and graph-based embeddings that encode relative relationships between actions based on occurrence locations. A case study has been conducted to explore the utility and effectiveness of such technique in identifying clinical task phenotypes. Finally, we developed methodological pipelines to study autoregressive longitudinal structure of actions from long temporal window. Based on theories of behavioral entropy, we developed a novel action entropy metric, which is a proxy measure of required cognitive effort at a given point of EHR interaction. A case study has been conducted to establish discriminant validity of the action entropy metric, using EHR audit log data of intensive care clinicians.
The focus of current and future research is: 1) to develop and improve the temporal modeling pipeline, 2) to apply the pipeline to identify temporal behavioral phenotypes associated with various cases related to clinican outcomes of interest.
Graduate publications
Kim S, Warner BC, Lew D, Lou SS, Kannampallil T. 2024 Measuring cognitive effort using tabular transformer-based language models of electronic health record-based audit log action sequences. J Am Med Inform Assoc, (): Online ahead of print. PMCID:
Kim S, Lou SS, Baratta LR, Kannampallil T.. 2023 Classifying clinical work settings using EHR audit logs: a machine learning approach. Am J Manag Care, 29(1)::e24-e30. PMCID:
Lou SS, Kim S, Harford D, Warner BC, Payne PRO, Abraham J, Kannampallil T. 2022 Effect of clinician attention switching on workload and wrong-patient errors. Br J Anaesth, 129(1)::e22-e24. PMCID: