基于大数据和机器学习的重症患者脓毒症的预警及治疗决策支持研究

Early identification and timely intervention is the key to improve the prognosis of septic patients and reduce medical expense. However, an early warning system for accurately and real-time predicting the occurrence of sepsis in ICU patients still lacks at present. On the other hand, more evidence from the real world setting is needed in clinical practice to support the decision-making of individualized treatment for sepsis. With the large-scale popularization of electronic medical records, a huge amount of data worth mining is continuously generated in daily healthcare practice. Due to the huge, complex and fast changing characteristics of big data, traditional statistical tools cannot effectively capture, manage and process the data. Machine learning algorithm is needed for reasoning and analysis. Our research group intends to construct an early warning model for sepsis with clinical value through the combination of big data (database of critically ill patients) and machine learning technology (random forest, recurrent neural network, etc.), and carry out model validation, and finally contributing to real-time and accurate early warning for the occurrence of sepsis in ICU wards, and providing valuable time windows for early clinical intervention. On the other hand, we will provide more perfected individualized decision support for the treatment of sepsis, especially the use of antibiotics, improve efficacy, and reduce the risk of death and the burden of medical treatment model for sepsis with clinical value through the combination of big data (database of critically ill patients) and machine learning technology (XGBoost, recurrent neural network, etc.), and carry out model validation, and finally contributing to real-time and accurate early warning for the occurrence of sepsis in ICU wards, and providing valuable time windows for early clinical intervention. On the other hand, we will provide more perfected individualized decision support for the treatment of sepsis, especially the use of antibiotics, improve efficacy, and reduce the risk of death and the burden of medical treatment.

早期识别并及时干预是改善脓毒症患者预后、降低医疗费用的关键两点。然而，目前在ICU患者中，仍然缺少一种可以准确、实时地预测脓毒症发生的预警系统；另一方面，临床实践亟需更多来自真实世界的循证证据，支撑脓毒症的个体化治疗决策。随着电子医疗病历的大规模普及，日常的临床工作产生了值得挖掘的海量数据。由于大数据的海量、复杂多样、变化快的特性，传统普通的统计工具已无法进行有效捕捉、管理和处理，需要采用机器学习算法引擎进行推理和分析。课题组拟通过大数据（重症患者的临床数据库）和机器学习技术（如XGBoost、递归神经网络等）的有机结合，开发构建具有临床应用价值的脓毒症预警模型，并进行模型验证，最终使其能够在ICU病房中实时、准确地预警患者脓毒症的发生，为临床早期干预赢得宝贵的时间窗，另一方面，为脓毒症的治疗，特别是抗菌药物的使用，提供更完善的个体化决策支持，提升治疗效果，降低患者死亡风险，减少医疗负担。

脓毒症是重症医学领域的关键问题，早期预警识别并及时干预能够显著改善脓毒症患者的预后、减轻医疗负担，具有较好的科学价值和社会意义。因此，项目组通过获取重症领域公开的大型数据库、有机结合机器学习算法，开发构建具有临床应用价值的脓毒症及其并发症的预警模型，并进行脓毒症患者的临床表型分类研究。主要研究进展如下：（1）获取了国外重症患者的大型公开数据库：eICU及MIMIC-III、IV数据库，并建立了 “八二”昆山爆炸的重症烧伤患者数据库；（2）针对不同重症患者群体，构建了多个不同机器学习算法的脓毒症预警模型。相较于其他算法，XGBoost这种基于梯度增强算法优化的机器学习算法，具有最为突出的预测性能和快速的训练速度，在训练集及验证集中表现优异；（3）构建了脓毒症患者发生急性肾损伤的预测模型并进行了预后分析。预测模型仅纳入5个预测变量，但在训练集和验证集中AUC值均超过0.8，预测性能良好且稳定。此外，项目组采用XGBoost算法，构建了患者28天死亡的预测模型。在验证集中，预测模型AUC值达到0.88，准确率、精准率、召回率、F1值分别为85.78%、73.26%、58.13%和64.82%；（4）运用潜类别分析方法完成了重症烧伤脓毒症患者的临床表型分类研究，研究将患者分为4个主要类别，为临床实施更有针对性的预防或治疗提供了依据。综上，项目组针对目前ICU中脓毒症及其并发症预警滞后以及治疗决策支持不足的困境，成功构建了基于机器学习算法的预测模型，并根据临床表型进行了重症烧伤脓毒症患者的分类研究。研究结果具有良好的理论意义和临床应用前景，有望为患者提供更准确快速的预警及个体化决策支持，提升治疗效果，改善患者预后，减少医疗负担。