Healthcare AI Safety Gap · Don't break patients
    Peer-reviewed / official21 Jun 2021·US

    Widely deployed sepsis AI missed two-thirds of cases

    Epic Sepsis Model: real-world AUC 0.63 vs the vendor's claimed 0.76–0.83.

    missed 2/3
    of sepsis cases, in external validation

    What the research found

    External validation of the Epic Sepsis Model — deployed at hundreds of US hospitals — found it missed roughly two-thirds of sepsis cases, with a real-world AUC of 0.63 versus the vendor's claimed 0.76–0.83, while generating alerts on 18% of patients.

    Why it matters for providers

    An independent check on a marketed clinical-AI tool found performance far below the vendor's claims — the 'cleared/marketed ≠ proven' gap, in one landmark case.

    Original source
    JAMA Internal Medicine
    Read the full report ↗
    Peer-reviewed or official source — the most reliable tier.