Itai Pessach, MD, PhD, MHA
Director, The Edmond and Lily Safra Children's Hospital
Sheba Medical Center, Israel
Disclosure information not submitted.
David Kirk, MD
Associate chief medical officer
WakeMed, United States
Disclosure information not submitted.
Ofer Chen, MD
VP Clinical Development
Clew Medical, United States
Disclosure information not submitted.
Ari Lipsky, MD
Chief of Emergency Medicine
HaEmek Medical Center, United States
Disclosure information not submitted.
Gershon Celniker, M.Sc
Head of Data and AI Research
CLEW Medical, United States
Disclosure information not submitted.
Craig Lilly, MD, FCCM
Director and Founder of the ICU Telemedicine Program
UMass Memorial Medical Center, United States
Disclosure information not submitted.
Eric Cucchi, MS, PA-C
Director of eICU Operations
UMass Memorial Health Care, United States
Disclosure information not submitted.
James Blum, MD, FCCM
Assistant Professor
Emory University Hospital, United States
Disclosure information not submitted.
Title: Development of AI based models for predicting life threatening events in critically ill patients
INTRODUCTION,
Delayed recognition of evolving instability is a driver of morbidity and mortality among critically ill adults. Clinical interventions triggered by artificial intelligence (AI) based early detection systems are expected to be associated with earlier and more effective management of time sensitive conditions. We developed two novel AI based algorithms that predict respiratory and hemodynamic deterioration and validated them among distinct cohorts of critically ill adults.
METHODS,
72,650 unique patient stays in one of 7 ICUs across the UMass Memorial Health Care system from 7-2006 to 9-2017 were used for model derivation (n=59,573) and for two separate prospectively designated internal validation cohorts (n=6,541 and 6,536 stays). An external validation cohort was selected from 3,172 stays in 7 adult intensive care units of 2 hospitals at the WakeMed Health System eICU, collected from 11-2019 to 7-2020. By utilizing a gradient boosting algorithm we created an automatic system that identified periods during which patients received interventions for respiratory failure (RF) or hemodynamic instability (HI).
RESULTS,
The study cohorts had similar demographic characteristics that were comparable to those reported for United States adult intensive care units that use telemedicine monitoring. 794 respiratory failure events as well as 2,112 hemodynamic instability events were detected during 16,249 stays of the 3 validation cohorts. Our HI prediction model predicted these events with an AUC of 0.96, 0.97 and 0.90 and a median lead time of more than 3.5 hours for the two UMass cohorts and the WakeMed cohort, respectively. The RF model had a median lead time of almost 4 hours for detecting respiratory failure with an AUC of 0.95, 0.96 and 0.95 for the two UMass cohorts and the WakeMed cohort, respectively.
CONCLUSIONS,
Artificial Intelligence based early warning systems were able to predict HI and RF events hours before they occurred with favorable performance characteristics in 3 separate cohorts of critically ill adults admitted to 14 different ICUs across 2 independent healthcare systems serving unique populations. The clinical use of this clinical vent prediction system will lead to early detection of HI and RF events and when timely interventions are made would lead to improved outcomes and reduced costs.