Development And Evaluation Of Predictive Mortality Risk Model And Scoring System For Sepsis Patients

Part ? Development and Validation of Predictive Mortality Risk Model and Scoring System for Sepsis Patients in Intensive Care UnitObjective:Sepsis is a major cause of death among patients in Intensive Care Unit(ICU).In order to improve clinicians'capability to predict the mortality risk in sepsis patients,we used machine learning approach to develop a predictive mortality risk model and scoring system to predict 30-day mortality among sepsis patients in ICU.All predictors we used were readily available in clinical practice,like demographic characteristic,initial vital signs and laboratory findings.We termed our model as Sepsis Mortality Risk Score 1.0(SMRS 1.0).Methods:Data was extract from Medical Information Mart for Intensive Care ?(MIMIC ?)database during the period from 2008 to 2012.All patients were?18 years old and admitted to ICU with sepsis,and the diagnosis of sepsis was based on the new criteria of Sepsis 3.0.We extracted patients'demographic characteristic,vital signs and laboratory findings from the first 24 hours after ICU admission,and the 30-day mortality was primary outcome.Patients were randomly assigned in a 7:3 ratio into training dataset or validation dataset,the logistic regression and random forest model were developed based on the data in training dataset.We used the data in validation dataset to calculate sensitivity,specificity,Area Under the Curve(AUC)of new predictive model,and existing scoring systems commonly used in clinic.We compared the prognostic accuracy of new model with existing scoring systems.Finally,we built a scoring system to assess the mortality risk of sepsis patients according to the risk factors identified by multivariate logistic regression,and we termed this scoring system as Sepsis Mortality Risk Score 1.0(SMRS 1.0).Results:This study included a total of 4,601 sepsis patients from ICU,30-day mortality was19.5%(895/4,601).In training dataset,the logistic regression model(AUC:0.803)and random forest(AUC:0.811)had good discrimination,which was better than sepsis related score(AUC:SOFA score 0.697;q SOFA score 0.596;SIRS score 0.585)and ICU severity score(AUC:SAPS II 0.793;APS ? 0.757;OASIS 0.760;LODS 0.743).In validation,the random forest model had the best discrimination(AUC:0.809),the AUC of logistic regression model(0.777)was better than sepsis related score,but was similar with ICU severity score.The multivariate logistic regression indicated that age,admission type,metastatic cancer,respiratory rate,blood pressure,body temperature,oxygen saturation,lactate,white blood cell(WBC)count,blood urea nitrogen(BUN),creatinine,anion gap were independent mortality risk factors for sepsis patients.Risk score criteria:age?70 years old,2 points;emergency admission,3 points;metastatic cancer,4 points;respiratory rate>20 times/minute,2 points;body temperature<36.5?,2 points;systolic blood pressure<90mm Hg,4 points;oxygen saturation<94%,3 points;lactate>2.7mmol/L,2 points;WBC count>12×10~9/L,1 point;BUN/creatinine>22,2points;anion gap>16mmol/L,2 points.The total score was 27 points and sepsis patients were stratified into mortality risk groups of low(0-4 points,mortality rate2.2%),moderate(5-8 points,mortality rate 10.3%),high(9-13 points,mortality rate29.2%),and very high(>13 points,mortality rate 70.0%).Mortality rates of the corresponding risk groups in the internal validation were 3.8%,11.8%,31.2%,and64.2%,respectively.After internal validation,the SMRS 1.0 was proved to have a good discrimination(AUC:0.772)and calibration(H-L test,P=0.650).Conclusion:Using MIMIC ? database and machine learning approach,a predictive 30-day mortality risk model is developed,which is superior to the sepsis related score(SOFA,q SOFA or SIRS)or ICU severity score(SAPS,APS ?,OASIS or LODS).In addition,a scoring system of sepsis mortality termed as Sepsis Mortality Risk Score 1.0(SMRS1.0)to identify distinct mortality subgroups of sepsis is established,which has a good discrimination and calibration and a potential of clinical application value.Part ? Performance of the MEDS Score in Predicting Mortality among Emergency Department Patients with a Suspected Infection: A Meta-AnalysisObjective:We conducted a meta-analysis to examine the prognostic performance of the Mortality in Emergency Department Sepsis(MEDS)score in predicting mortality among emergency department(ED)patients with a suspected infection.Methods:Electronic databases Pub Med,Embase,Scopus,EBSCO and the Cochrane Library were searched for eligible articles from their respective inception through February2019.Sensitivity,specificity,likelihood ratios,and receiver operator characteristic area under the curve were calculated.And summary estimates were derived using a hierarchical summary receiver operating characteristic curve.Subgroup analyses were performed to explore the prognostic performance of MEDS in selected populations.Results:We identified 24 studies involving 21,246 participants.The pooled sensitivity of MEDS to predict mortality was 79%(95%CI: 72%-84%);specificity was 74%(95%CI:68%-80%);positive likelihood ratio 3.07(95%CI 2.47-3.82);negative likelihood ratio0.29(95%CI 0.22-0.37);and area under the curve 0.83(95%CI 0.80-0.86).The MEDS score provided good discrimination,moderate sensitivity and specificity across all subgroups.However,there was significant heterogeneity among included studies.Meta-regression analyses showed that the time when MEDS score was measured and the cutoff value used were important sources of heterogeneity.Conclusion:The MEDS score has moderate accuracy in predicting mortality among ED patients with a suspected infection.