CHARM: Practical AI for Healthcare Data

​How can AI move beyond experimentation and solve practical problems using real-world healthcare data?

This session will introduce CHARM (Collaboration to Harmonize Antimicrobial Registry Measures), an antimicrobial stewardship initiative that transforms healthcare data into actionable insights through interactive dashboards and analytics. Attendees will see how CHARM uses data from participating healthcare organizations to examine antibiotic prescribing patterns, identify opportunities for improvement, and support antimicrobial stewardship efforts.

The session will then showcase two AI projects being developed around the CHARM data ecosystem. The first uses locally hosted, open-source large language models (LLMs) to standardize prescription data while keeping sensitive healthcare data within a controlled environment. The project combines structured data processing, retrieval-augmented generation (RAG), fine-tuning, systematic evaluation, and human review to improve accuracy while reducing hallucinations and manual data-cleaning effort.

The second project, CHARM Outlier, uses machine learning to identify antibiotic prescriptions that differ from expected prescribing patterns. Flagged records are presented through a web application where researchers and healthcare professionals can review them and provide feedback. That feedback is then used to improve the model’s future predictions, creating a human-in-the-loop approach to identifying potentially inappropriate prescribing.

Together, these projects demonstrate how dashboards, locally deployed LLMs, machine learning, and human expertise can work together to turn complex healthcare data into useful information while addressing practical concerns around privacy, accuracy, and responsible AI implementation.