It was great to see Dr. Danny Lee and Bala Kulandaivel at their poster. They did the hard math to create a predictive model that detects pediatric diabetic ketoacidosis, and displayed it on the multi-provider schedule and on in patient trackboard units.
Congratulations to their achievement! This is hard work, difficult math, really hard tuning to make the alerts relevant to busy physicians/providers in an acute care setting.
Reflecting on our own journey with the Epic Sepsis model and other deterioration scores, we found:
- Signal to noise ratio tends to be low (in our case about 9% initially, later improved to 30%, still low) even with the best mathematicians and models
- Acute care clinicians are too busy to look at the scores, and don’t trust them “when my patient looks good right now. I have to go see the sicker patients down the hall, I will not act on your alert now”
- Consider giving these alerts to a separate virtual team that can surveil ALL patients in the hospital and know what to do for these early warning alerts.
- Be verrrry careful in the socio-political swoop in when you call a rapid response team to take over patient care from the primary team when you think a “true positive” is detected. This is where we stumbled a number of times.
- It takes publicizing some “good saves” that might have gone against the primary team’s instincts of “that patient looked fine, and then they crumped for no reason” especially if the virtual team spotted the possibility hours before, but were rebuffed.
These are some details of our difficult journey to the acceptance of predictive alerts like this.
To Dr. Lee and team, great work on a new predictive model! and best wishes on avoiding some of the potholes on this really important journey toward a successful intervention! We will be eagerly watching!
