GPT-4-Assisted Versus Human Detection of Intraoperative Adverse Events From Spine Surgery Operative Reports
BACKGROUND AND OBJECTIVES: Manual chart review remains the primary method of clinical data collection in spine surgery research but is time intensive and prone to human error. Large language models may enable automated extraction of clinically meaningful events from unstructured operative text. We evaluated the performance of generative pre-trained transformer 4 (GPT-4) in detecting intraoperative adverse events from spine surgery operative notes compared with trained human reviewers. METHODS: A retrospective cohort of 1454 laminectomy cases performed at a single academic institution between 2016 and 2022 was analyzed. Deidentified operative notes were independently reviewed by 4 trained human reviewers and GPT-4 using standardized prompts to identify incidental durotomy and nondurotomy intraoperative adverse events. Discordant classifications were adjudicated by a senior reviewer to establish ground truth. Performance metrics including sensitivity, specificity, predictive values, and F1 score were calculated. Analyses were stratified by operative note length, and review time was estimated. RESULTS: Incidental durotomy occurred in 88 cases (6.1%). GPT-4 achieved 100.0% sensitivity compared with 89.8% for human reviewers, with lower positive predictive value (88.0% vs 100.0%) but similar specificity (99.1% vs 100.0%). For nondurotomy adverse events (n = 21), GPT-4 demonstrated higher sensitivity than humans (90.5% vs 57.1%) with equivalent specificity (100.0%). GPT-4 maintained stable performance across note length quartiles, whereas human reviewer sensitivity decreased for longer notes. Human reviewers required approximately 35 hours to review all notes, compared with approximately 2 hours for GPT-4. CONCLUSION: GPT-4 demonstrated higher observed sensitivity than trained human reviewers for detecting intraoperative adverse events from spine surgery operative reports while reducing review time by more than 90%. Large language model-assisted chart review may serve as a scalable human-in-the-loop adjunct to traditional chart review, improving the completeness and efficiency of neurosurgical outcomes research.
Authors
- Eric Karl Oermann (ORCID: https://orcid.org/0000-0002-1876-5963)
- Masami Nakatsuka (ORCID: https://orcid.org/0000-0002-0967-7099)
- Daniel D. Wiggan (ORCID: https://orcid.org/0000-0002-8294-8676)
- Caroline Cubells
- Madison R. Murata
- Anton Alyakin
- Ebunoluwa Morakinyo
Institutions
- Washington University in St. Louis (US)
- NYU Langone Health (US)
- New York University (US)
- Stanford University (US)
Publication Details
- Journal
- Neurosurgery Open
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1227/neuprac.0000000000000333
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- article
- Field-Weighted Citation Impact
- 0.00