基于机器学习的重症患者脓毒症实时风险预测模型

Sepsis Real⁃Time Risk Prediction Model for Intensive Care Unit Patients Based on Machine Learning

脓毒症是人体对感染反应失调导致的器官功能障碍综合症,具有较高的发病率和死亡率。传统的评分系统存在特异性低的问题。本研究基于LightGBM机器学习框架,提出了一种对脓毒症进行早期预测和风险评估的模型,以便对具有脓毒症潜在风险的患者进行及时干预。为了实现该模型,提出基于LASSO特征选择和滑动窗口路径重积分的时间序列特征构建方法,以及基于动态时间规整算法的时间序列聚类采样方法。选择MIMIC-Ⅲ数据库29239位病人和PhysioNet/CinC 2019挑战赛数据集40336位病人的临床信息来训练和验证模型。所提出的模型在MIMIC-Ⅲ和PhysioNet/CinC 2019独立测试集上的灵敏度、特异性、操作特征曲线下面积(AUC)分别为0.7377、0.7304、0.8147和0.8026、0.7891、0.8730。与目前最先进的EASP方法相比,AUC分别提高了3.62%和2.83%。本研究模型可以实时预测脓毒症发生的风险,揭示影响脓毒症发生的重要因素,为脓毒症风险人群的及时干预提供依据。

Sepsis is a syndrome of organ dysfunction caused by the body′s dysfunctional response to infection,with high morbidity and mortality.The traditional scoring system has low specificity.Based on the LightGBM machine learning framework,this study proposed a model for early prediction and risk assessment of sepsis to provide timely intervention for patients with potential risk of sepsis.In order to realize the model,a time series feature construction method based on LASSO feature selection and sliding window path reintegration and a time series clustering sampling method based on dynamic time regularization algorithm were proposed.We selected clinical information from 29239 patients in the MIMIC⁃III dataset and 40336 patients in the PhysioNet/CinC 2019 challenge dataset to train and validate the model.The sensitivity,specificity,and area under the receiver operation characteristic curve(AUC)of the proposed model on the MIMIC⁃III and PhysioNet/CinC 2019 independent test sets were 07377,07304,08147 and 08026,07891,08730,respectively.Compared with the state⁃of⁃the⁃art method EASP,the improvement of AUC was 362%and 283%respectively.In conclusion,the established model could predict the risk of sepsis in real time,reveal the important factors affecting the occurrence of sepsis,and provide a basis for timely intervention of people at risk of sepsis.