Research On Prognostic Factors Analysis And Early Warning Model Establishment Of Sepsis Based On Big Data

Objective:In this study,we used the ideas and techniques of big data to find the prognostic factors associated with sepsis,and established a model of early death warning for sepsis patients based on the random forest method,and discussed the predictive power of the model.Methods:In this retrospective study,we selected sepsis patients with sepsis meeting sepsis-3 who were admitted to the ICU within 24 hours from the MIMICIII database,tracked their progression and prognosis,recorded common clinical laboratory indicators after admission,and divided them into the training group and the verification group in a ratio of about 0.7:0.3.SPSS19.0 was used for single and multiple factor analysis of the data,and specific parameters were selected to create an early warning model of sepsis by logistic regression analysis.Boruta algorithm was used to select RF importance variables and build RF warning model.By comparing the two models and evaluating their sensitivity,specificity,ROC and AUC respectively,the validity of the RF model was evaluated with the validation group.Results :(1)No statistical difference is found in the training group when compared with the verification group.(2)By comparing the survival group and the death group,univariate and multivariate analyses are performed to obtain 14 variables including "bmi","hr＿max","t＿max","gcs＿min","be＿min","ph＿min","hb＿min","lym＿min","p＿min"," wbc＿max","scr＿max","na＿min","bil＿max" and "tni＿max",which are selected as specific parameters for logistic regression modeling.(3)Importance for variable selection,we adapt Boruta algorithm,determine the RF variable for "age","na +＿min","bun＿max","ph＿min","ag＿max","hco3-＿min","ldh＿max","cl-＿min" a total of eight important variables,by comparison with the conventional algorithm,it is concluded that the front eight variables are the most significant,adding variable model accuracy improvement is not obvious,so the choice of eight important variables simplify instead of 41 random forest early warning model is established.(4)The random forest warning model AUC=0.82(95%CI 0.81-0.83).The sensitivity and the specificity is tested,which are respectively 74.4% and 79%.Binary logistic regression model AUC=0.64(95%CI 0.63-0.66),and the sensitivity was 47.1% and the specificity was 75.62%.Conventional SOFA score AUC=0.59(95%CI 0.57-0.60),with a sensitivity of 47.1% and specificity of 68.68%;Comparing the two warning models with SOFA score,respectively,the conclusion is that the RF modeling effect is the most optimal,followed by logistic regression,and SOFA score prediction effect is the worst.(5)The prediction model is obtained by internal verification: AUC=0.80(95%CI 0.153-0.225),sensitivity 71.84%,specificity 75.56%.Conclusion:(1)The method based on the combination of big data mining and random forest can effectively predict the 28-day prognosis of sepsis patients.(2)After internal verification,the prediction efficiency of the sepsis warning model created based on the random forest method is better than the logistic regression model,and both models are better than the SOFA score.(3)In this study,the use of big data-driven machine learning methods is superior to existing clinical decision rules and traditional analysis techniques,and is used to predict the hospital mortality of patients with sepsis.