Blue Cross expert warns EMR prompts may inflate diagnoses such as sepsis and postpartum anemia

Dr. J.B. Sobel told lawmakers that some electronic medical record systems and embedded AI features can extract keywords from clinical encounters and suggest diagnoses that are not always supported by clinical evidence, creating what he called "discordant treatment" when expected therapies are not documented.

"There's a new phenomenon called ambient listening that's been introduced by some of these software companies," Sobel said, describing tools that can record encounters, flag keywords and recommend diagnoses; he warned that those features can be difficult to audit because audio files may be deleted. He cited two examples to illustrate the problem: an increase in acute post-hemorrhagic anemia coding after newborn deliveries that did not correspond with the frequency of blood transfusions, and a rise in sepsis diagnoses that occurred alongside shorter lengths of stay and, in many cases, no antibiotic treatment in the record.

Sobel also described a trend in outpatient evaluation-and-management (E&M) coding where previously bell-curved intensity levels have shifted upward, with level 4 and 5 visits becoming more common at some providers after new billing systems were introduced. He linked the change in part to EMR features that over-document review-of-systems data or otherwise make it easy to add higher-intensity codes.

Committee members pressed for clinical outcomes data; Sobel said Blue Cross's risk-adjusted analyses did not show a change in disease burden across Medicaid, Medicare and marketplace products but that he did not have data on mortality trends available in the hearing.