Get Full Government Meeting Transcripts, Videos, & Alerts Forever!
Get email alerts on the Patient Safety Measurement topic
No spam. Unsubscribe anytime.
Experts urge shift from retrospective safety measures to assurance cases and real‑time AI monitoring
Summary
Federal and academic experts told a PTEC panel that traditional retrospective measures miss most harm and are limited for accountability in AI‑enabled systems; they recommended combining measures with assurance cases and deploying high‑PPV real‑time surveillance to detect and prevent harm.
Experts at a public PTEC session on patient safety argued that retrospective quality measures alone are inadequate for accountability in increasingly automated, AI‑enabled care and urged combining traditional measures with structured “assurance cases” and real‑time monitoring.
Jeffrey Geppert, senior research leader at Battelle and the measure science team lead, told the committee that accountability measures must be feasible, linked to material outcomes and demonstrate consistent, causally attributable performance. "Assurance cases help determine whether the system is credibly safe before the harm occurs," Geppert said, framing assurance cases as an evidence‑based complement rather than a replacement for measures.
David Classen, a professor at the University of Utah who leads work with patient safety organizations (PSOs), described evidence that commonly used systems miss large amounts of harm. He summarized comparisons showing the IHI global trigger approach detected far more adverse events than voluntary reporting or conventional PSIs: "the way that many hospitals measure harm... was missing 90% of harm," he said, and described PSO implementations that automated dozens of electronic safety measures to produce prospective alerts and unit‑level surveillance.
Panelists said AI and automated triggers make it feasible to detect and intervene on safety risks during care, not only after the fact. Classen described trials where patient‑facing dashboards and real‑time alerts were associated with lower 30‑day readmissions and mortality. He urged building high‑predictive‑value algorithms and local teams to triage signals so clinicians are not overwhelmed.
Speakers cautioned that AI‑driven systems require governance and validation. As one panelist put it, agentic AI can amplify small input changes into large variations in outputs and thus requires control mechanisms before its outputs are used for accountability or payment.
The panel did not recommend an immediate CMS mandate for real‑time reporting; several speakers and CMS representatives said piloting, careful evaluation, and attention to legal and reputational risks are needed before broad deployment. Next steps cited by panelists included multi‑site pilots, PSO‑mediated trials, and development of assurance‑case frameworks to accompany measures and AI systems.
