My research is focused on the meaningful use of Electronic Health Records data with an interest in both deriving inference from EHR data and developing risk prediction models and clinical decision support tools with EHRs. From an inferential standpoint, I am interested in understanding the potential and limitations of EHRs for clinical research and adapting methods for the analytic challenges that arise. From a risk prediction standpoint, I am interested in best practices for developing, implementing and evaluating clinical decision support tools. I also have a growing interest in best practices for implementing tools into clinical environments to enhance usability and acceptability. Overall, my research sits at the intersection of biostatistics, biomedical informatics, machine learning, epidemiology and implementation science.

I enjoy collaborating with both clinicians and methodologists and involving students into these projects.

Informed Presence

A particular focus is identifying biases that may arise in the use of EHRs. One such bias is what we term informed presence bias. This arises from the observation that people only interact with the health system when they are sick. While this is a missing data problem – we are missing healthy observation – the potential for bias lies in what we observe as opposed to what we miss. One of my foci is identifying situations in which this bias can arise, characterizing the problem it can engender and ultimately developing solutions for addressing it.

Dynamic and Longitudinal Risk Prediction

One of the key ways that EHR data have been used is to develop risk prediction models. EHRs allow us to observe patients repeatedly  over time. We can exploit these repeated measurements to better characterize a patient’s risk profile and/or develop dynamic (time updated) risk models. I have been working on understanding what are the best way to incorporate these repeated measures into risk models. I am also interested in developing effective and robust approaches for dynamic prediction. In one collaboration, we have been working with clinicians at Duke Hospital to evaluate and improve a risk tool for time updated detection of patient deterioration.

Health system analytics: Clinical Decision Support Tools & Quality  Analytics

I have strong interest in using routinely collected health data to inform and improve the way clinical care is administered. I serve as the Chief Data Scientist for Quality, providing analytic support for quality improvement and health equity projects. I also work closely with members of the Duke University Health System to develop, implement  and evaluate clinical decision support (CDS) tools. I co-lead the evaluation committee for DUHS’s CDS governance committee, ABCDS.

Algorithmic  Bias, ML Ethics and Implementation Science

CDS tools have the power to transform how healthcare is delivered. They also have the potential to exacerbate existing inequalities in how healthcare is delivered. We have been working to understand how inherent biases within EHR data to lead to biased and inequitable algorithms. I also have begun to explore how the process by which we implement CDS tools can impact usability and acceptability. This work uses both statistical and informatics principles as well as qualitative methods of focus groups with tool users, developers and patients.

Pharmaco-Data Science

EHR data present a unique opportunity to understand and assess pharmaceutical interventions in a real world environment. We have conducted comparative effectiveness studies of pharmaceutical interventions, assessed the degree of inequity in the usage of medications, developed approaches to capture adverse events, mined EHR databases to identify medications related to positive outcomes, and examined the impacts of polypharmacy in children.

Development of Resources to Access and Analyze EHR Data

EHR data represent a unique and valuable data source to understand a person’s clinical status. Unfortunately EHR data are stored in complex and hard to access format. Our group is working to make Duke EHR generally and reproducibly available to clinical researchers. We have stood up the Duke Clinical Research DataMart (CRDM) which allows users to set up regular data pulls for both research purposes or tracking of clinical populations.