Get Full Government Meeting Transcripts, Videos, & Alerts Forever!
Get email alerts on the Health It Ai topic
No spam. Unsubscribe anytime.
Panelists: AI can reduce clinician burden but needs guardrails to protect patients
Summary
Researchers, vendors and safety officers told a P-TECH session that artificial intelligence can reduce cognitive load, help detect harms earlier, and improve follow‑up — but urged explainability, bias testing, lifecycle monitoring, and workflow integration to avoid new risks.
Experts speaking to a P-TECH advisory session convened by HHS described both the opportunities and hazards of applying artificial intelligence in health care.
Dr. Tejal Gandhi, chief safety and transformation officer at Press Ganey Associates, said AI could help in diagnosis, medication reconciliation, prediction of sepsis or falls, chart summarization, and patient engagement — all of which could reduce clinician cognitive load and upstream risks. “It is not going to really have benefit unless we can really tie it into clinical workflows and actionability,” Gandhi said, stressing that risk flags must be paired with clear guidance and workflow integration.
Melissa Swanfield, senior director for quality and regulatory compliance at Meditech, described vendor approaches to data capture, real‑time surveillance, AI‑generated summaries, and rural use cases where surveillance tools reduced catheter‑associated urinary tract infections and cut reporting time for infection control teams. She recommended “standardized plain‑language model cards,” disclosure when AI is used in care, patient consent processes, formal bias testing, and lifecycle risk assessment for models.
Why it matters: Panelists warned that AI can introduce new safety issues if governance, monitoring, and human‑in‑the‑loop assumptions are treated as a panacea. Gandhi noted limits of human vigilance: “Human beings are not good at vigilance and overseeing things that AI is going to produce,” a point used to argue for built‑in safeguards rather than relying solely on clinicians to catch model errors.
Panelists and committee members discussed regulatory alignment. Swanfield noted recent FDA guidance classifying some clinical decision support as medical devices and urged risk‑based oversight to match AI risk. Aneesh of the Arcadia Institute proposed technical standards and data‑sharing models (COIN/Conversational Interop, FIRE APIs, BCDA) to enable reasoning agents and safer measurement.
The committee’s discussion emphasized balancing promising pilots and measurement improvements with guardrails, clinician workload studies, and evaluation of unintended consequences before broad deployment.
