Title: Enhancing Medical Decision Support Systems for Sepsis Patients in the ICU: Real-Time Detection and Algorithmic Bias Mitigation

Date: Wednesday, November 13th, 2024

Time: 1:00 PM – 3:00 PM ET

Location: https://gatech.zoom.us/j/99264574785

Jeffrey Smith

Machine Learning PhD Candidate

H. Milton Stewart School of Industrial and Systems Engineering
Georgia Institute of Technology

Committee:

Dr. Yao Xie (Advisor), H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Pinar Keskinocak, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Gian-Gabriel Garcia, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Jing Li, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology

Dr. Andre Holder, Department of Biomedical Informatics, Emory University School of Medicine

Abstract:

The digital revolution in healthcare has significantly improved the precision and efficiency of healthcare delivery. Over the past decade, healthcare data from sources such as electronic health records (EHRs), imaging, patient notes, and genetic databases has grown exponentially. This surge, paired with advancements in machine learning (ML) and artificial intelligence (AI), offers unprecedented opportunities to significantly improve patient outcomes and alleviate the burdens on medical professionals.

However, integrating these technologies, especially in critical care settings like the Intensive Care Unit (ICU), presents substantial challenges. For instance, accurately detecting severe illnesses such as sepsis in a timely manner and addressing biases in medical-AI models are complex issues that require attention. To address these issues, this thesis proposes novel ML and statistical methods for developing explainable ML models. 

The first study introduces a novel approach using Jensen-Shannon divergence to monitor health decline through real-time EHR data analysis. This method assesses health deterioration for individual patients and subgroups, aiming to identify the specific onset of sepsis and its contributing factors.

The second study focuses on the ethical considerations of AI in healthcare, particularly addressing clinical and algorithmic biases. We present a bias detection framework integrating uncertainty quantification with Classification and Regression Trees (CART). This approach is especially important in healthcare settings where biases in decision support systems can lead to inaccurate diagnoses or treatments, further amplifying existing disparities. Experiments using synthetic data and real-world data from Grady Memorial and Emory University Hospitals demonstrate its effectiveness in detecting bias.