Transforming Patient Monitoring with Machine Learning
In June 2024, David Kim, MD, PhD was awarded an NIH R01 grant for "Integrating Acute and Ambulatory Care with Post-Discharge Monitoring and Machine Learning."
This new project aims to monitor certain high-risk ED patients after hospital discharge. Study participants will wear a wireless armband device to continuously monitor their physiology after leaving the hospital. Kim and the team will combine in-hospital and remote monitoring data to better understand how patient physiology evolves after discharge, and to identify remote monitoring signals that may represent important changes in health status.
The R01 grant will extend Kim’s research to the outpatient setting. Over the past four years, Kim and his team have developed software that synthesizes data from electronic health records and physiologic monitors in real-time to provide more specific and accurate information about a patient’s physiology during an emergency department (ED) visit.
Trained on tens of thousands of ED visits, PhysioHub learns the relationships between many aspects of a patient’s trajectory and produces a detailed representation of individual physiology and risks. Simultaneously modeling the relationships between ECG signals, lab results, vital sign trends, and subsequent diagnoses may enable emergency physicians to make more informed and accurate predictions about a patient’s likely response to an intervention.
Modern AI language models like GPT-4 calibrate themselves against countless next-word predictions. However, the complexity of emergency medicine, with simultaneous processes occurring over different timescales, presents challenges for the development of effective AI models. Kim emphasizes the need to develop tools that address the high intensity, variety, and unpredictability of emergency medicine.
Kim envisions using the software to automate repetitive processes and to extract more meaningful information about patient physiology, enabling physicians to spend less time on data collation and more time answering difficult diagnostic and therapeutic questions. The goal is not just to enhance the efficiency of data interpretation, but to achieve more accurate and specific diagnoses, and better patient outcomes.
In Kim’s view, practical experience and domain knowledge are vital for creating tools that physicians find useful in the emergency setting. Kim credits his students and collaborators, specifically Stanford Computer Science graduate students Tom Jin and Julia Reisler, with leveraging modern machine learning capabilities towards clinically ambitious goals. He advises innovators to focus on achieving better outcomes in real world settings, rather than focusing solely on test characteristics like accuracy alone.
Updated Spring 2024