用AI提升主动脉钙化检测准确率，帮医院多发现12%漏诊并追回31万年收入
A University of Texas Medical Branch Case Study on Aortic Calcification Detection
- 用AI图像模型与语言模型分析胸片报告，自动识别钙化迹象
- 发现12.4%报告漏编码、2.1%病例被漏诊，共涉及579名患者
- 推动医院全院部署AI系统，兼顾临床精准与财务收益
本案例研究描述了德克萨斯大学医学部（UTMB）与Zauron Labs公司合作，利用人工智能工具提升胸部影像中主动脉钙化（ACs）的检测与编码效率。尽管ACs对心血管疾病具有重要预后价值，但常被低估。研究采用高精度图像模型（AUC=0.938）及基于Meta Llama 3微调的语言模型，回顾性分析3,988名患者的影像与报告数据（共5,000次检查）。结果发现，495名患者（12.4%）的报告存在未正确编码的钙化指征，另有84名患者（2.1%）在初筛中被遗漏。该发现使UTMB有望每年追回31.4万美元的漏报收入，并改善患者诊疗质量。研究最终促使医院决定全院推广Zauron Guardian Pro系统，实现更精准的AI辅助质控与编码。本研究经德克萨斯大学健康科学圣安东尼奥分校机构审查委员会批准（Study ID 00001887）。
原文摘要 · Abstract (English)
This case study details The University of Texas Medical Branch (UTMB)'s partnership with Zauron Labs, Inc. to enhance detection and coding of aortic calcifications (ACs) using chest radiographs. ACs are often underreported despite their significant prognostic value for cardiovascular disease, and UTMB partnered with Zauron to apply its advanced AI tools, including a high-performing image model (AUC = 0.938) and a fine-tuned language model based on Meta's Llama 3.2, to retrospectively analyze imaging and report data. The effort identified 495 patients out of 3,988 unique patients assessed (5,000 total exams) whose reports contained indications of aortic calcifications that were not properly coded for reimbursement (12.4% miscode rate) as well as an additional 84 patients who had aortic calcifications that were missed during initial review (2.1% misdiagnosis rate). Identification of these patients provided UTMB with the potential to impact clinical care for these patients and pursue $314k in missed annual revenue. These findings informed UTMB's decision to adopt Zauron's Guardian Pro software system-wide to ensure accurate, AI-enhanced peer review and coding, improving both patient care and financial solvency. This study is covered under University of Texas Health San Antonio's Institutional Review Board Study ID 00001887.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。