Informatics Grand Rounds with Dr. Matthew Robinson

Why It Matters

Robust, AI‑augmented stewardship can dramatically lower antimicrobial resistance, improve patient outcomes, and protect hospitals from costly penalties.

Key Takeaways

• Antibiotic resistance causes 5 million deaths annually, surpassing HIV
• Stewardship requires right drug, dose, timing, duration for each patient
• Misdiagnosed UTIs and pneumonia drive unnecessary antibiotic use in hospitals
• Structured EHR data is limited; unstructured notes hinder automation
• LLMs improve but need engineered pipelines to ensure safe, guideline‑consistent recommendations

Summary

In the May 2026 Grand Rounds at Johns Hopkins, infectious‑disease specialist Dr. Matthew Robinson examined how data science can strengthen antibiotic stewardship. He framed stewardship as a mandatory, four‑moment process—confirm infection, obtain cultures, select empiric therapy, then reassess and stop or narrow treatment—highlighting its role in curbing the global crisis of antimicrobial resistance. Robinson presented stark data: antimicrobial resistance now kills roughly five million people worldwide, outpacing HIV and malaria, and misdiagnosis of common syndromes fuels overuse. At Hopkins, up to 80 % of presumed community‑acquired pneumonia cases and a quarter of urinary‑tract infection diagnoses were later deemed inappropriate, extending hospital stays and triggering penalties. Traditional decision‑support tools, such as pre‑authorization and prospective audit, address only a fraction of prescriptions and rely heavily on structured EHR fields. He traced the evolution of AI in this space, noting the 1973 rule‑based system for antibiotic selection and more recent, modest rule engines that cut broad‑spectrum use. Early large‑language models (LLMs) performed poorly—only 64 % of empiric recommendations were correct—but newer models like GPT‑5.5 now reject unsafe prompts. Nonetheless, real‑world scenarios involve fragmented, unstructured data (clinical notes, radiology reports) that LLMs must parse reliably. Robinson concluded that effective stewardship will require engineered pipelines that translate clinical guidelines into computable rules, integrate both structured and free‑text data, and embed LLMs within rigorous safety checks. Such systems promise reduced drug costs, shorter lengths of stay, and compliance with Joint Commission and CMS mandates, while safeguarding patients from inappropriate therapy.

Original Description

This session with Dr. Matthew Robinson will discuss Building and Deploying Contextual AI for Antibiotic Prescribing. This session covers challenges in translating clinical guidelines into machine instructions, trade-offs in deterministic vs LLM-driven decision support, limitations in data availability in production EHRs, real-world barriers to running LLMs to support clinical decision-making in production, and differences in reporting vs production EHR data. #healthai #johnshopkins #healthinformatics https://bids.jhmi.edu/about/events-also-past-grand-rounds/