Itai Pessach, MD, PhD, MHA
Director, The Edmond and Lily Safra Children's Hospital
Sheba Medical Center, Israel
Disclosure information not submitted.
Ofer Chen, MD
VP Clinical Development
Clew Medical, United States
Disclosure information not submitted.
Gershon Celniker, M.Sc
Head of Data and AI Research
CLEW Medical, United States
Disclosure information not submitted.
Ari Lipsky, MD
Chief of Emergency Medicine
HaEmek Medical Center, United States
Disclosure information not submitted.
Andrea Forgacs, M.Sc
Data Scientist
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.
Amanda Edwards, PA-C
Director, Advanced Practice Providers
WakeMed Health & Hospitals, United States
Disclosure information not submitted.
David Kirk, MD
Associate chief medical officer
WakeMed, United States
Disclosure information not submitted.
Title: Using artificial intelligence to predict which patients are not going to deteriorate
Introduction: Predicting patient stability may sometimes be as challenging as predicting patient deterioration. Relieving providers from the need to actively monitor stable patients may allow enhanced attention to more critical and unstable patients. The goal of this study was to prospectively evaluate a previously developed, FDA-approved, predictive model for identifying ICU patients at low-risk (LR) for respiratory failure or hemodynamic instability.
Methods: In this prospective, observational study, the model’s performance was evaluated on 514 adult ICU stays from the WakeMed tele-ICU system in the period between May-July 2020. Each stay was divided into 1-hour evaluation periods. The first 12 hours of each stay as well as stays of patients defined as DNI/DNR were excluded. Respiratory failure and hemodynamic instability events were prospectively tagged by a senior on-site clinician. The model's performance in accurately predicting 8 hours of stability was compared to the expert tagging of each period.
Results: A total of 21,391 one-hour periods were evaluated. Of which 19,375 (90.7%) were defined as stable periods and 1,989 (9.3%) periods were defined as unstable periods (periods in which either respiratory failure or hemodynamic instability events occurred). Our model had a sensitivity, specificity and PPV of 0.46, 0.97 and 0.99, respectively for predicting low-risk periods. Events in which patients deteriorated while our model predicted they were stable were extremely rare. Only 59 (0.27%) of the 21391 periods were mislabeled by our model as low risk. On the other hand more than 40% of the periods were labeled as stable by our model.
Conclusions: This predictive model was able to identify, with high performance, ICU patients in low risk for significant deterioration, who were unlikely to experience major clinical deterioration or require significant intervention within the 8 hours following the prediction. Use of this model in clinical settings may allow providers to focus on the more unstable patients, providing them with better care, and better outcomes.