融入临床先验知识的ICU内脓毒症早期预警模型

An early warning model of sepsis in ICU with clinical prior knowledge

目的基于重症监护病房(ICU)内脓毒症数据集,融入临床先验知识,构建ICU内脓毒症早期预警模型,提高模型的准确性,并增加模型的可解释性。方法使用MIMIC-Ⅲ数据库作为内部验证集,选取入住ICU的患者,根据纳入和排除标准确定最终入选病例(n=13737)。然后在集成学习方法的基础上,将临床先验知识与其他生理特征相结合,建立一个ICU内患者发生脓毒症的早期预警集成学习模型,使用十折交叉验证方法,同时设计消融实验和基线对比,并基于平均信息增益和Shapley值对脓毒症早期诊断影响因素进行解释分析。利用内部验证集对模型进行检验,并采用dtchina数据库(n=670)对模型性能进行检验。结果本研究通过使用不同的分类器进行建模,结果显示,融入先验知识后模型的受试者工作特征曲线下面积(AUC)值能增加2%~3%,而特征融合模型取得的效果优于其他任意模型,在内部验证集中,其AUC为0.879(95%CI 0.859~0.901),在外部验证集中,其AUC为0.793(95%CI 0.713~0.863)。同时,根据可解释性分析得出临床先验知识的权重最高,WBC、收缩压、HCO-3、呼吸频率是影响脓毒症早期诊断的重要因素。结论通过将临床先验知识和生理特征进行结合,构建的特征融合集成模型性能较好,并通过实验设计和分析说明加入临床先验知识的必要性。

Objective To build an early warning model of sepsis in ICU based on sepsis data set and clinical prior knowledge for improving the accuracy of the model and increasing the interpretability of the model.Methods MIMIC-Ⅲdatabase was used as an internal validation set,and the final enrolled cases(n=13737)admitted to ICU were determined according to inclusion and exclusion criteria.On the basis of ensemble learning method,an ensemble learning model for the early warning of sepsis in ICU patients was established by combining clinical prior knowledge with other physiological characteristics.Ten fold cross-validation method was used,and ablation experiment and baseline comparison were designed.Based on the average information gain and Shapley value,the factors influencing the early diagnosis of sepsis were explained and analyzed.The internal validation set was used to test the model,and dtchina database(n=670)was used to test the model performance.Results In this study,different classifiers were used for modeling.The results showed that the AUC value of the model could be increased by 2%-3%after incorporating prior knowledge,and the effect of the feature fusion model was better than that of any other models.In the internal verification set,the AUC of the model was 0.879(95%CI 0.859-0.901).In external validation set,the AUC of the model was 0.793(95%CI 0.713-0.863).At the same time,the weight of clinical prior knowledge was the highest according to interpretability analysis,and white blood cell count,systolic pressure,HCO-3 and respiratory rate were important factors affecting the early diagnosis of sepsis.Conclusions By combining clinical prior knowledge and physiological characteristics,the feature fusion integrated model constructed in this paper has good performance,and the necessity of adding clinical prior knowledge is demonstrated through experimental design and analysis.