📄 paper: academic.oup.com/jamiaopen/ar...

study made possible by the
@UWEmerMed
&
@uwisye
team
big credit to first author Dann Hekman!

#AIinMedicine #NLP #EHR #Falls #HealthcareAI #JAMIAOpen

when we build algorithms on top of these phenotypes, we risk baking in these issues

your cohort definition is your ground truth. if you get that wrong, everything downstream is shaky

its not just about falls

it's a broader reminder about phenotyping: diagnosis codes are easy to extract, but they're often shaped by documentation shortcuts, billing practices, and workflow quirks

the patients missed by ICD codes?

they were sicker, higher comorbidity scores, higher 30-day mortality, and more likely to have a medical illness (like sepsis or AKI) as the cause of their fall

finally, our findings aren't limited to CDI or infection prevention, it’s foundational work on how we design, implement, and evaluate AI tools that operate inside the hospital.

it’s a blueprint for making clinical AI actually safe & effective

#AIinMedicine #HealthIT #CDI #AI

this was a big, collaborative effort involving physicians, pharmacists, engineers, administrators, and analysts

healthcare AI work takes a village

we didn’t just evaluate performance metrics, we studied implementation:
✔️ usability
✔️ fidelity
✔️ provider perceptions
✔️ engagement across roles
this work builds on our prior studies on both prospective model validation & clinical implementation

while alerts to physicians (via BPAs) weren’t always acted upon, pharmacists did engage.
→ medication reviews led to meaningful de-escalation of high-risk antibiotics and PPIs 💊
→ another win for pharmacist-driven stewardship interventions 🔥

pharmacists really are awesome!

our system was used to flag patients at high risk for CDI, triggering a bundle of interventions:
🧼 enhanced hand hygiene
📋 best-practice advisories (BPAs) for providers
💊 pharmacist-led de-escalation of antibiotics & acid suppression
🍦 yogurt recommendations (yes, really)

we developed a machine learning model to predict CDI risk in hospitalized patients, implemented it into instance of
@hey.epic.com (EHR), and used the outputs to implement a multi-component infection prevention bundle — think of it like AI-powered stewardship

“The Dishwasher Loading Task (DLT) is one uniquely suited for the skills and intellect of human beings. For this reason Singh et al. have suggested replacing the Turing test with the DLT as the ultimate arbiter of AGI.”

wah wah waaAaAAaawwWw