A look at the use of AI in some US hospitals, including adopting predictive models to help identify and prioritize ER and ICU patients who are most at risk
Laura Landro / Wall Street Journal : Tweets: @dukehospitalist , @wsj , and @aboutkp Tweets: @dukehospitalist : Congrats to our own Dr. Cara O'Brien, DUH hospitalist, and great colleagues featured in WSJ for their work on #SepsisWatch! AI is saving lives as emergency rooms and hospital ICUs turn to the technology to identify patients most at risk https://www.wsj.com/... via @WSJ @wsj : Hospitals are making a bet that AI can help identify and treat patients at highest risk in their ERs, inpatient wards and ICUs https://www.wsj.com/... @aboutkp : Vincent Liu, MD, shares with @WSJ how health systems are using AI to improve patient outcomes. Our Advance Alert Monitor predicts when a hospitalized patient's condition may worsen, allowing specially trained staff to intervene sooner. https://www.wsj.com/...
Context & Ripple Effects
This story sits mid-arc in health systems' slow adoption of clinical AI. Two years earlier, STAT reported that hospitals were rolling out AI-powered decision support tools that were often novel and unproven, with most patients unaware they were being used. The WSJ piece names the two flagship implementations that gave the approach credibility: Duke University Hospital's SepsisWatch, built around hospitalist Dr. Cara O'Brien's team, and Kaiser Permanente's Advance Alert Monitor under Dr. Vincent Liu.
First-order effects
- At Duke and Kaiser Permanente, triage decisions in the ER and ICU are now co-piloted by predictive models that flag sepsis risk and patient deterioration before clinicians would catch them manually, shifting which patients get seen first.
- Clinicians at these systems take on new workflow duties — validating or acting on model alerts — making the hospitalist and intensivist teams the de facto quality-control layer for the algorithms.
Second-order effects
- Peer hospital systems watching Duke and Kaiser's results face pressure to adopt similar deterioration-prediction models or explain why their triage lacks them, feeding a vendor market for clinical AI.
- The follow-on reporting matters here: by 2023 the same paper documented sometimes flawed AI-based diagnosis tools with clinicians feeling pressure from administrations to defer to the algorithm, showing the second-order cost of deploying these models faster than oversight matures.
Third-order effects
- If the pattern holds, in-house pilots give way to procurement: by early 2026 a survey found 27% of health systems paying for commercial AI licenses, turning hospital AI from a research project into a budgeted line item with vendor lock-in dynamics.
- That commercialization forces a governance reckoning — validation standards, clinician-deference norms, and disclosure to patients — because the 2020-era gap between what tools hospitals run and what patients know about has only widened as deployments scale.
The trend: Hospital AI is moving from bespoke, unproven pilots toward system-wide predictive triage and paid commercial licenses, with clinical governance and patient transparency lagging behind deployment.