用于预测中低收入国家早发型新生儿脓毒症的机器学习模型:开发与验证研究
该研究基于印度德里新生儿感染研究(2011–2014)中疑似早发型脓毒症(EOS)的足月与早产儿数据,开发并外部验证了机器学习模型以预测培养证实的EOS。最佳模型(Boruta+随机森林,含28个围产-新生儿变量)在训练集敏感度90.3%、特异度40.6%,外部验证敏感度92.3%、特异度37.2%;模型使用可能将抗生素使用率从74.8%降至55.7%。
要点
- 最佳模型含28个围产-新生儿变量,训练集敏感度90.3%、特异度40.6%。
- 外部验证敏感度92.3%、特异度37.2%,阴性预测值95.7%。
- 模型使用可能使抗生素使用率从74.8%降至55.7%。
核心结论
- 机器学习模型在中低收入国家新生儿中预测培养证实的EOS表现出高敏感度和可接受的特异度,有望作为床旁决策支持工具。
研究设计与样本
- 对印度德里新生儿感染研究(2011–2014)进行二次分析,数据来自三家三级新生儿病房的疑似EOS新生儿,排除培养阴性脓毒症病例。
- 采用动态80:20训练-测试划分,对64个变量使用Boruta和Lasso两种特征选择方法,并应用五种机器学习技术。
- 模型以90%敏感度为目标选择最优模型,并整合至网页应用,在2015–2021年出生的外部队列中进行验证。
主要发现
- 共纳入2924名新生儿,其中548名(18.7%)为培养证实的脓毒症;平均胎龄35.3(±3.8)周,平均出生体重2112(±754)g。
- Boruta与随机森林分类器组合产生最佳模型,包含28个围产-新生儿变量,训练集敏感度90.3%、特异度40.6%。
- 外部验证队列n=147(26例培养证实脓毒症),模型敏感度92.3%、特异度37.2%、阳性预测值24.0%、阴性预测值95.7%。
- 在仅具有围产期EOS风险因素的无症状新生儿中,模型敏感度为100%。
- 模型使用可能使抗生素使用率从74.8%降至55.7%(风险差:–19.1%;95% CI –8.3至–29.7)。
关键数值与推荐
- 最佳模型选择目标为达到90%敏感度。
- 训练集敏感度90.3%、特异度40.6%;外部验证敏感度92.3%、特异度37.2%、阳性预测值24.0%、阴性预测值95.7%。
- 无症状但具有围产期EOS风险因素的新生儿中敏感度为100%。
- 抗生素使用率可能从74.8%降至55.7%(风险差–19.1%;95% CI –8.3至–29.7)。
临床怎么用
- 该模型可作为中低收入国家床旁决策支持工具,辅助及时识别培养阳性脓毒症。
- 高敏感度和高阴性预测值可能有助于减少不必要的抗生素使用,但特异度较低需结合临床评估。
局限性
- 研究为二次分析,数据来源于印度三家三级新生儿病房,外推至其他中低收入国家需谨慎。
- 外部验证队列样本量较小(n=147,26例培养证实脓毒症)。
- 模型特异度较低(训练集40.6%,外部验证37.2%),可能限制其单独用于排除脓毒症。
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)。
原文
Objective Early-onset sepsis (EOS), which occurs within the first 72 hours of life, can often be fatal for neonates. Machine learning (ML) models demonstrate promise for timely diagnosis. However, current ML models primarily rely on data from high-income countries, which reduces their applicability to low-income and middle-income countries (LMICs) that have a higher burden and different disease profiles. We developed an ML model for the timely prediction of culture-proven EOS in LMICs. Methods We conducted a secondary analysis of the Delhi Neonatal Infection Study (2011–2014) carried out in three level-3 neonatal units in India. Data from inborn neonates suspected of having EOS were extracted, and cases of culture-negative sepsis were excluded. By implementing a dynamic 80:20 (train:test) data split, two feature selection methods were employed—Boruta and Lasso—across 64 variables, and five ML techniques were applied. The aim was to achieve 90% sensitivity to identify the optimal model based on performance metrics. The developed model was integrated into a web application and validated in an external cohort of neonates born between 2015 and 2021. Results Of 2924 neonates, 548 (18.7%) had culture-proven sepsis. The mean gestation and birth weight were 35.3 (±3.8) weeks and 2112 (±754) g, respectively. The Boruta and random forest classifier yielded the best model, which included 28 perinatal–neonatal variables. The sensitivity and specificity of the model were 90.3% and 40.6%, respectively. In external validation (n=147; 26 culture-proven sepsis cases), the model’s sensitivity, specificity, positive predictive value and negative predictive value were 92.3%, 37.2%, 24.0% and 95.7%, respectively. The sensitivity was 100% in asymptomatic neonates with only perinatal risk factors for EOS. The use of the model could have reduced antibiotic usage from 74.8% to 55.7% (risk difference: –19.1%; 95% CI –8.3 to –29.7). Conclusions The ML model demonstrated high sensitivity and acceptable specificity in predicting EOS. This prediction model has the potential to assist in the timely and reliable identification of culture-positive sepsis and may serve as a bedside decision support tool in LMICs.
引用
- BMJ Paediatrics Open:A machine learning model for prediction of early-onset neonatal sepsis in low-income and middle-income countries: development and validation study
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。