胸部X线AI机会性筛查骨质疏松:准确性与骨折死亡关联
该研究对一款基于胸部X线的商用AI工具进行回顾性外部验证,以双能X线吸收法为参考标准。在体检队列和慢性阻塞性肺疾病队列中分别纳入8618例和4941例患者,比较AI识别骨质疏松的诊断效能及其与后续骨折、全因死亡的关联。结果显示AI在体检队列AUC为0.94,在慢阻肺队列为0.81,其预测概率与骨折和死亡风险显著相关,且该关联并非由已诊断的骨质疏松完全中介。
要点
- 胸片AI机会性筛查骨质疏松在体检人群AUC达0.94,具有较好识别效能。
- AI预测概率升高与后续骨折和全因死亡风险增加显著相关。
- AI可能识别尚未被临床诊断为骨质疏松者的骨折风险。
核心结论
- AI可从单张胸部正位X线片中识别骨质疏松,且AI结果与后续骨折和全因死亡相关。
- AI评分对骨折或死亡的间接效应经DXA定义或临床诊断的骨质疏松中介不显著,提示AI可能识别尚未被临床诊断为骨质疏松人群的骨折风险。
研究设计与样本
- 该研究为回顾性外部验证研究,纳入一个含同日胸部X线与DXA配对数据的体检队列和一个有可用胸部X线的慢性阻塞性肺疾病队列。
- 评价的AI工具为商用化工具,采用InceptionV3骨干网络,使用55,600对胸部X线-DXA配对数据训练,用于从单张胸部正位X线片识别骨质疏松。
- 诊断效能以DXA为参考标准,采用受试者工作特征曲线下面积进行评估。
- 采用多变量Cox比例风险回归分析AI结果与后续骨折或全因死亡的关联,校正性别、年龄、低体重指数及慢阻肺分期。
关键影像表现/征象
- AI基于单张胸部正位X线片输出骨质疏松预测概率,无需额外骨密度检查即可进行机会性筛查。
- 在体检队列中AI识别骨质疏松的AUC为0.94,95%CI为0.93-0.95。
- 在慢阻肺队列中AI识别骨质疏松的AUC为0.81,95%CI为0.76-0.87。
诊断标准与关键数值
- 体检队列中AI识别骨质疏松的AUC为0.94(95%CI:0.93-0.95)。
- 慢阻肺队列中AI识别骨质疏松的AUC为0.81(95%CI:0.76-0.87)。
- 较高的AI预测骨质疏松概率与两个队列中后续骨折和全因死亡相关,P值均小于0.05。
- 中介分析显示AI评分经DXA定义或临床诊断骨质疏松介导的间接效应不显著,P值均大于0.05。
临床应用与阅片要点
- 对常规胸部X线检查者可利用AI机会性筛查骨质疏松,无需改变现有摄片流程。
- AI预测概率较高时应提醒临床关注骨折和死亡风险,即使患者尚未被临床诊断为骨质疏松。
- 在慢阻肺等慢性病人群中AI诊断效能低于体检人群,解读结果时需考虑合并症影响。
局限性
- 研究为回顾性设计,慢阻肺队列仅拥有可用胸部X线,未要求同日DXA配对。
- 研究纳入的体检队列与慢阻肺队列在年龄和性别构成上存在明显差异。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVES: To evaluate an AI tool for opportunistic osteoporosis screening using chest radiographs (CXRs), focusing on its diagnostic accuracy against dual-energy x-ray absorptiometry (DXA) and associations with long-term risks of fracture and mortality.
MATERIALS AND METHODS: This retrospective, external validation study included a health checkup cohort with same-day CXR-DXA pairs and a chronic obstructive pulmonary disease (COPD) cohort with available CXRs. We evaluated a commercialized AI tool (InceptionV3 backbone, trained with 55,600 CXR-DXA pairs) designed to identify osteoporosis from a single frontal CXR. Diagnostic performance was assessed against DXA using the area under the receiver operating characteristic curve (AUC). Associations between AI results and subsequent fractures or all-cause mortality were investigated using multivariable Cox proportional-hazard regression, adjusted for sex, age, low body mass index, and COPD stage. Mediation analyses were conducted to examine whether these associations were mediated by a diagnosis of osteoporosis.
RESULTS: In total, 8618 (male-to-female ratio, 2882:5736; mean age, 58 y; median follow-up, 2868 d) and 4941 (male-to-female ratio, 4110:831; mean age, 69 y; median follow-up, 2656 d) patients were included in the health checkup and COPD cohorts, respectively. The AI achieved AUCs of 0.94 (95% CI: 0.93-0.95) and 0.81 (95% CI: 0.76-0.87) for osteoporosis identification in the health checkup and COPD cohorts, respectively. Higher AI-predicted osteoporosis probability was associated with subsequent fracture and mortality in both cohorts ( Ps <0.05). Mediation analyses showed that indirect effects of AI scores on fracture or mortality mediated through DXA-defined or clinically diagnosed osteoporosis were not significant ( P s>0.05), suggesting that AI may identify fracture risk in individuals not yet clinically diagnosed with osteoporosis.
CONCLUSIONS: AI could identify osteoporosis from chest radiographs, and AI results were associated with subsequent fracture and mortality.
引用
- Lee, Kim, Kim, et al. Chest Radiograph-based Artificial Intelligence for Osteoporosis: Accuracy and Associations With Fracture and Mortality. Invest Radiol. 2026. PMID: 42452786. doi:10.1097/RLI.0000000000001303
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。