UNC NRP January 2024 Education Session: BERRY: A 6-month randomized controlled trial of the effect of blueberries on mild cognitive decline

Please join the UNC Network for Research Professionals as Carol Cheatham, PhD, returns to share her research on nutrition and the aging brain.

Register

The Cecil G. Sheps Center for Health Services Research is hosting Data Science Week, a focused series of discussions tailored for anyone studying access, disparities, quality, or the cost of health care, and is interested in emerging trends in data science.

The week’s agenda includes in-depth exploration of current themes in healthcare analytics, such as the application of causal machine learning techniques, the utilization of compliant synthetic data in research settings, the intricacies of Medicare claims data analysis, and the development of algorithmic approaches in clinical decision-making.

Sheps welcomes contributions from colleagues in the field, offering an environment conducive to scholarly exchange and professional development. This event is an opportunity for researchers, students, and practicing data scientists to engage with complex topics, refine analytical methodologies, and discuss the implications of advanced data science in the realm of health services research.

Presentations will begin at 12:00pm each day, followed up by a discussion period. Attendees may join in-person at the Sheps Center in room 2002 or via Zoom.

Wednesday, January 17, 12:00 – 1:30pm
Research Journey: Health Services Research Studies in Women’s Health, Disabilities, Surgical Outcomes, and More – Leveraging Large Observational Data
Speaker: Neil Kamdar, MA

Large administrative claims, electronic medical records, and other large clinical and surgical registries possess a breadth of information about our patient population. While these data sources may not offer the depth that may be afforded through studies involving primary data collection, they represent an important and fundamental strength of large sample sizes and the ability to study vulnerable populations and/or rare diseases. I will go through studies involving several data sources and populations broken into thematic units: vulnerable populations (e.g. disabled and those with diseases of aging), women’s health, and surgical outcomes and quality. This talk is especially geared for those who wish to better understand the breadth of published studies that can be tackled using these data, the key findings, and illustrate some other new areas I have explored with these registries over the last several years. I will finally touch upon some of the new avenues for exploration using these data and the potential synergies between Sheps analytic data assets and resources with our community of researchers.

Thursday, January 18, 12:00 – 1:30pm
Long COVID Phenotypes Leveraging a Large Primary Care Electronic Medical Record Registry, the American Family Cohort
Speaker: Neil Kamdar, MA

Long COVID or Post-Acute Sequelae following initial SARS-COV-2 (COVID-19) has continued to be a challenge for diagnosing in the primary care setting. Various published studies have incorporated different approaches and methods which have defined some potential opportunities for classification of COVID-19 afflicted patients into relevant sub-groups based on their symptom presentation during the follow-up period. Using a large primary care registry, the American Family Cohort, I will walk through the key existing challenges for classification in the literature, and the incorporation of a latent class analysis approach to examine extant differences in racial composition, sex, and social deprivation index in group membership. Diagnostic phenotypes of potential Long COVID patients and their comparisons to likely influenza-like (ILI) controls will be explored and examined. This talk was presented as a Distinguished Paper at the National American Primary Care Research Group (NAPCRG) in November 2023.

Friday, January 19, 12:00 – 2:00pm
Methods Session: Detecting and Mitigating Bias in Machine Learning
Speaker: Ashley Avis, MS

Statistical models utilize historic and current data to identify patterns and predict them forward. Historic and present-day biases can be perpetuated without careful evaluation and bias mitigation. We will discuss how to prepare and plan a project utilizing responsible data science methods and other techniques from peer reviewed journals and gray literature to support bias evaluation and mitigation. In this presentation, we will explore how to identify patterns of bias, how predictive models incorporate patterns of bias, and methods for resolving bias. We will explore definitions of algorithmic fairness, the conflicts between different fairness definitions, and the importance of diverse perspectives on final fairness evaluation. We will discuss methods for identifying unequal or biased model decisions. Lastly, we will discuss methods for correcting biased models and how to practice data science responsibly.

An Application and Example of Glassbox Explainable Boosting Machines (EBMs) in a Surgical Cohort Registry
Speaker: Neil Kamdar, MA

Predictive models have become important for postoperative occurrences after surgery, such as readmissions, 30-day reoperation, and other complications. Frequently, the application of predictive models has an opaqueness which remains extraordinarily difficult to achieve stakeholder buy-in, which is usually hospital administrators, clinicians, and other policymakers. In this tutorial-based, mostly graph-based presentation with code snippets, a statement of the key issues surrounding prediction, a description of the sample data source, the research question in-progress, and model outputs including graphs and interaction term interpretations will be discussed using Explainable Boosting Machines (EBMs). Advantages of this modeling approach for these contexts will be discussed, with special attention to some use cases within the Sheps context focusing on surgical and other clinical outcomes.

Register