C36-08 AI Rewriting the Record in ICU: GPT 3.5- Enhances Documentation Precision and Clinical Documentation Improvement (CDI) Language Without Altering Diagnosis-Related Group (DRG) Severity

Rationale Artificial intelligence is increasingly integrated into clinical workflows, yet its measurable impact on documentation completeness and billing accuracy remains unclear. In critical care, precise discharge summaries drive Clinical Documentation Improvement (CDI) and influence Diagnosis-Related Group (DRG) reimbursement. We benchmarked GPT-3.5(an AI model trained on ICU records)-generated ICU discharge summaries against clinician-authored notes to assess improvements in documentation quality and reimbursement parity. Methods This retrospective study analyzed 5,000 real MIMIC-IV ICU discharge summaries (ICD-9-coded) and 5,000 GPT-3.5-generated summaries derived from equivalent diagnoses and procedures expressed in ICD-10 terminology. A curated CDI Lexicon defined four domains: Major Complications/Comorbidities (MCC), Complications/Comorbidities (CC), linkage/causality phrases, and specificity descriptors. Each note received a CDI Composite Score = 3×MCC + 1 × (CC + Linkage + Specificity) per 1,000 words. A diagnosis-matched ICD-10 subset (n = 123) was used to address coding-era bias. Paired comparisons used the Wilcoxon signed-rank test and Cohen’s d. DRG severity was modeled using proxy weights (1.70 for “with MCC,” 1.30 for “with CC only,” 1.00 for “without CC/MCC”). Each summary’s detected MCC/CC terminology determined its tier, producing per-admission relative weights analogous to MS-DRG categories. All data were obtained from MIMIC-IV (2008-2022, PhysioNet, MIT Laboratory for Computational Physiology) under a credentialed data-use agreement. Results GPT-3.5 summaries contained roughly twice the CDI-relevant term density of clinician notes (19.5 vs 9.6 points per 1,000 words; p = 1.8 × 10-202). In the diagnosis-matched subset, enrichment reached 140% (21.8 vs 9.1; p = 3.9 × 10-10; d = 0.88). Mean DRG proxy weights were identical (1.50 vs 1.50). GPT-3.5 upgraded 19.5% of encounters by adding MCC/CC-linked severity terms, while clinicians exceeded GPT-3.5 in another 19.5%, suggesting complementary documentation strengths. Conclusions GPT-3.5 significantly improved CDI-relevant documentation density without altering DRG severity classification. AI-generated summaries enhanced clarity, causal linkage, and specificity while maintaining billing neutrality. Clinician notes still captured temporal and contextual nuances that AI missed, supporting a collaborative human-AI documentation model. Future work includes prospective trials embedding GPT-assisted documentation into ICU workflows to assess impacts on after-hours charting, note completeness, and team communication. These findings parallel ongoing pilot programs at the Mayo Clinic, which integrate generative AI into real-time documentation workflows, underscoring the timeliness and translational relevance of this work.