Evidently - Your AI Is Only as Good as Its Intelligence Layer: Unearthing Quality-Impacting Diagnoses at Three Health Systems
Presentation Type: Industry Partner Session
Click to View Presentation
Presentation Time: 08:00 AM - 09:15 AM
Abstract Keywords: Clinical Decision Support, Natural Language Processing, Patient Safety, Informatics Implementation, Workflow, Artificial Intelligence, Machine Learning, Deep Learning
Today's patient records have outgrown the capacity of both clinicians and EHR systems to fully process them. Roughly 30–40% of clinical data resides in unstructured documents — scanned records, outside facility notes, historical reports — that current tools either ignore or merely summarize. This session presents results from a prospective, multi-site study across an academic medical center, a non-profit health system, and a regional IDN evaluating Clinical Data Intelligence (CDI), an AI knowledge layer that reads the entire longitudinal patient record, including unstructured documents, and generates evidence-traced clinical assertions. Using a novel methodology — independent comparison of AI-drafted H&Ps against provider-authored H&Ps with blinded clinician adjudication — the study found 4–5 quality-impacting diagnoses missed per provider H&P across all three sites. These are diagnoses that change care plans, risk stratification, reportable condition status, or handoff content. The session will detail how CDI differs from existing approaches (EHR-native AI, ambient scribes, chart summarization, traditional CDI tools), present the documentation accuracy improvements observed (MCC accuracy +24–41%, CC accuracy +40–54%, mortality benchmarking accuracy +49–81%), share patient safety stories from the study, and discuss implications for clinical workflow integration via SMART on FHIR.
Speaker(s):
Karen Sedivy, MDEvidently