胸部CT加入机器学习预测模型在住院前排除COVID-19的附加价值:一项回顾性大学网络研究
本研究基于德国12个中心RACOON联盟的4437例回顾性数据,构建监督机器学习模型预测COVID-19,参考标准为RT-PCR。在临床与实验室特征基础上加入胸部CT特征后,敏感度由0.82升至0.89,特异度由0.84升至0.89,AUC由0.92升至0.95,假阴性比例由2.6%降至1.7%(均P<0.0001)。结论认为胸部CT特征可提升住院前排除COVID-19的模型效能。
要点
- 临床+实验室模型加入胸部CT特征后AUC由0.92升至0.95,敏感度0.89、特异度0.89。
- 加入CT后假阴性比例由2.6%降至1.7%,有助住院前排除COVID-19。
- 模型基于12中心4437例回顾性数据,COVID-19阳性率15.6%。
核心结论
- 在机器学习预测模型中,将胸部CT特征加入临床与实验室特征可显著提升住院前排除COVID-19的效能。
研究设计与样本
- 参考标准为逆转录聚合酶链反应(RT-PCR)检测,COVID-19阳性692例(15.6%)。
- 采用监督机器学习构建三组诊断模型:仅胸部CT特征(模型R)、仅临床检查与实验室特征(模型CL)、以及三者联合(模型RCL)。
诊断标准与关键数值
- 模型CL与模型RCL的敏感度分别为0.82与0.89(P<0.0001)。
- 模型CL与模型RCL的特异度分别为0.84与0.89(P<0.0001)。
- 模型CL与模型RCL的阴性预测值分别为0.96与0.97(P<0.0001)。
- 模型CL与模型RCL的AUC分别为0.92与0.95(P<0.0001)。
- 模型CL与模型RCL的假阴性分类比例分别为2.6%与1.7%(P<0.0001)。
临床应用与阅片要点
- 在住院前排除COVID-19的场景中,可将胸部CT特征纳入机器学习模型以降低假阴性风险。
- 模型效能基于多中心回顾性数据,应用于日常阅片时需结合本地人群与PCR参考标准验证。
局限性
- 本研究为回顾性设计,数据库来自联盟中心,可能存在选择偏倚。
- 原文未报告模型的外部验证结果。
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
PURPOSE: During the coronavirus disease 2019 (COVID-19) pandemic, hospitals still face the challenge of timely identification of infected individuals before inpatient admission. An artificial intelligence approach based on an established clinical network may improve prospective pandemic preparedness.
METHOD: Supervised machine learning was used to construct diagnostic models to predict COVID-19. A pooled database was retrospectively generated from 4437 participant data that were collected between January 2017 and October 2020 at 12 German centers that belong to the radiological cooperative network of the COVID-19 (RACOON) consortium. A total of 692 (15.6 %) participants were COVID-19 positive according to the reference of the reverse transcription-polymerase chain reaction test. The diagnostic models included chest CT features (model R), clinical examination and laboratory test features (model CL), or all three feature categories (model RCL). Performance outcomes included accuracy, sensitivity, specificity, negative and positive predictive value, and area under the receiver operating curve (AUC).
RESULTS: Performance of predictive models improved significantly by adding chest CT features to clinical evaluation and laboratory test features. Without (model CL) and with inclusion of chest CT (model RCL), sensitivity was 0.82 and 0.89 (p < 0.0001), specificity was 0.84 and 0.89 (p < 0.0001), negative predictive value was 0.96 and 0.97 (p < 0.0001), AUC was 0.92 and 0.95 (p < 0.0001), and proportion of false negative classifications was 2.6 % and 1.7 % (p < 0.0001), respectively.
CONCLUSIONS: Addition of chest CT features to machine learning-based predictive models improves the effectiveness in ruling out COVID-19 before inpatient admission to regular wards.
引用
- Krämer, Ingwersen, Teichgräber, et al. Added value of chest CT in a machine learning-based prediction model to rule out COVID-19 before inpatient admission: A retrospective university network study. Eur J Radiol. 2023. PMID: 37087928. doi:10.1016/j.ejrad.2023.110827
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。