Clinical Epidemiology/Healthy Ageing Division, Garvan Institute of Medical Research, Sydney, NSW, 2010, Australia.
Prince of Wales Hospital, Randwick, Sydney, NSW, Australia.
Arch Osteoporos. 2021 Jan 6;16(1):6. doi: 10.1007/s11657-020-00859-5.
Text-search software can be used to identify people at risk of re-fracture. The software studied identified a threefold higher number of people with fractures compared with conventional case finding. Automated software could assist fracture liaison services to identify more people at risk than traditional case finding.
Fracture liaison services address the post-fracture treatment gap in osteoporosis (OP). Natural language processing (NLP) is able to identify previously unrecognized patients by screening large volumes of radiology reports. The aim of this study was to compare an NLP software tool, XRAIT (X-Ray Artificial Intelligence Tool), with a traditional fracture liaison service at its development site (Prince of Wales Hospital [POWH], Sydney) and externally validate it in an adjudicated cohort from the Dubbo Osteoporosis Epidemiology Study (DOES).
XRAIT searches radiology reports for fracture-related terms. At the development site (POWH), XRAIT and a blinded fracture liaison clinician (FLC) reviewed 5,089 reports and 224 presentations, respectively, of people 50 years or over during a simultaneous 3-month period. In the external cohort of DOES, XRAIT was used without modification to analyse digitally readable radiology reports (n = 327) to calculate its sensitivity and specificity.
XRAIT flagged 433 fractures after searching 5,089 reports (421 true fractures, positive predictive value of 97%). It identified more than a threefold higher number of fractures (421 fractures/339 individuals) compared with manual case finding (98 individuals). Unadjusted for the local reporting style in an external cohort (DOES), XRAIT had a sensitivity of 70% and specificity of 92%.
XRAIT identifies significantly more clinically significant fractures than manual case finding. High specificity in an untrained cohort suggests that it could be used at other sites. Automated methods of fracture identification may assist fracture liaison services so that limited resources can be spent on treatment rather than case finding.
文本搜索软件可用于识别有再次骨折风险的人群。研究中的软件与传统病例发现相比，识别出的骨折人数高出三倍。自动化软件可以帮助骨折联络服务比传统病例发现识别出更多有风险的人群。
骨折联络服务解决了骨质疏松症（OP）治疗后的骨折缺口问题。自然语言处理（NLP）能够通过筛选大量放射学报告来识别以前未被识别的患者。本研究的目的是将一种 NLP 软件工具（XRAIT）与在其开发地点（悉尼王子威尔士医院[POWH]）的传统骨折联络服务进行比较，并在 Dubbo 骨质疏松症流行病学研究（DOES）的裁决队列中对其进行外部验证。
XRAIT 搜索放射学报告中的骨折相关术语。在开发地点（POWH），XRAIT 和一位盲法骨折联络临床医生（FLC）在同时进行的 3 个月期间，分别审查了 5089 份报告和 224 份 50 岁及以上人群的就诊报告。在 DOES 的外部队列中，无需修改即可使用 XRAIT 分析数字可读放射学报告（n = 327），以计算其敏感性和特异性。
在搜索了 5089 份报告后，XRAIT 标记了 433 处骨折（421 处真骨折，阳性预测值为 97%）。与手动病例发现相比（98 人），它识别出的骨折数量高出三倍以上（421 处骨折/339 人）。在未经外部队列（DOES）当地报告风格调整的情况下，XRAIT 的敏感性为 70%，特异性为 92%。
XRAIT 比手动病例发现识别出的临床意义更重大的骨折数量明显更多。在未经训练的队列中具有高特异性表明，它可以在其他地点使用。骨折识别的自动化方法可以帮助骨折联络服务，以便将有限的资源用于治疗而不是病例发现。
