Examining Racial Disparities in Predictive Modeling Among Survivors of Critical Illness

University Of Pittsburgh At Pittsburgh, Pittsburgh PA

Investigators

Abstract

MODIFIED PROJECT SUMMARY/ABSTRACT SECTION: Survivors of critical illness often face significant challenges, such as neurocognitive disorders, physical disabilities, and respiratory limitations. Demographic subgroups may experience heightened challenges. Risk prediction tools designed to identify high-risk patients for interventions such as post-acute care clinics may lack generalizability among demographic subgroups. Intending to enhance prediction tools for survivors of critical illness, our team brings expertise in health services research focusing on predictive modeling and causal inference. In previous work, we have highlighted demographic impact in patients with cardiovascular disease and COVID-19 across different healthcare systems, including the Veterans Health Administration (VHA). Furthermore, we have begun exploring outcomes among survivors of critical illness including mortality and re-admission rates and developed an innovative post-ICU care model showing early indications of reducing hospital readmissions, increasing hospital-free days, and reducing mortality across patient populations. As an F32 grant recipient, I will integrate and build on the expertise of my mentorship to identify and characterize demographic differences within three datasets of critical care illness survivors as defined by mortality, 90-day re-admissions, and hospital-free days (HFDs) at 90 days. In parallel, I will compare the generalizability of two statistical models used to stratify patients by one-year mortality: (1) the Care Assessment Needs (CAN) score, a mortality risk model widely used to guide interventions among Veterans, and (2) the PREDICT score, a simplified one-year mortality risk model used at first patient contact to guide interventions such as palliative care consultation. Algorithmic generalizability is an emerging concept geared toward improving model applicability across heterogenous cohorts within statistical models and algorithms. To improve generalizability within our models, our team will employ novel approaches to achieve broader applicability, including double prioritization. In addition to identifying differences within a population of increasing care complexity, this proposal investigates statistical models as a source of worsened outcomes, aiming to improve outcomes through generalizable approaches. In doing so, we propose establishing a care delivery framework for critical illness survivors. With close mentorship from an experienced team in predictive modeling, healthcare delivery, and innovative research methodologies at the University of Pittsburgh, this training plan forms a foundation for a Career Development Award and a future career as a physician-scientist.