Related Experiment Videos
Generative AI in Admission Notes and Diagnostic Completeness: A Pilot Study
Alfredo Camargo Rodrigues1, Jason MIsurac2, Lindsey A Knake2
1University of Iowa Health Care, Health Care Information Systems, Department of Anesthesia, IA, United States, Iowa City.
Background:
Admission history and physical (H&P) notes influence inpatient documentation, risk adjustment, and reimbursement. In high-acuity settings, time constraints and fragmented chart data contribute to under-documentation of comorbidities. Generative AI systems that draft admission notes from existing electronic health record data may improve documentation completeness, but real-world inpatient performance remains insufficiently characterized.
Objective:
This study aimed to evaluate whether AI-generated admission notes identify documentation-relevant diagnoses supported by the medical record but not explicitly captured in provider-authored admission notes.
Methods:
In this single-center retrospective pilot study, we reviewed 22 matched pairs of AI-generated and provider-authored admission H&P notes at a large academic medical center. We assessed principal diagnosis concordance and identified net-new secondary diagnoses. Secondary diagnoses identified by the AI but absent from provider-authored notes were adjudicated by a clinical documentation improvement (CDI) team using standard institutional criteria.
Results:
AI-generated notes aligned with the provider-authored principal diagnosis in 91% of cases (20/22). The CDI team adjudicated 104 AI-identified secondary diagnoses, of which 97% (101/104) were supported for documentation. Ninety-four diagnoses were net-new, quality-relevant conditions not documented in provider-authored notes (median: 4.5 per admission). Net-new diagnoses were observed across all provider types, with numerically higher counts among advanced practice providers and residents; however, the sample was too small to support inferential comparisons. Despite its modest sample size, this pilot demonstrated high principal diagnosis concordance, though the AI misclassified the principal diagnosis in two cases. AI-generated drafts identified additional CDI-supported diagnoses not captured by providers, though narrative quality and factual accuracy were not evaluated.
Conclusion:
These findings highlight the potential of generative AI to surface documentation-relevant information at admission and underscore the importance of human oversight in AI-assisted documentation workflows. Larger multicenter studies are needed to assess generalizability and safety.