| ObjectiveAnemia is a common clinical manifestation of sepsis,which is closely related to poor prognosis of septic patients.Red blood cell(RBC)transfusion is an important treatment for anemia.However,due to the complexity and heterogeneity of sepsis,different patients have different responses to RBC transfusion therapy.Our study will use unsupervised clustering algorithm of machine learning to find the potential clinical subtypes in septic patients,and evaluate the effect of RBC transfusion on clinical outcome,specifically for potential clinical subtypes in septic patients.The aim of this study is to make more precise RBC transfusion strategy in septic patients.MethodsSeptic patients in Medical Information Mart for Intensive Care Ⅲ(MIMIC Ⅲ)were selected as the derivation cohort,and septic patients who admitted to ICU of the first and second affiliated hospital of Dalian Medical University from January 2018 to October 2021 were selected as the validation cohort.Patients with sepsis in derivation cohort and validation cohort who met our study criteria were screened,and clinical variables that might affect the therapeutic prognosis of RBC transfusion were extracted,including basic clinical information,clinical scores,comorbidity,laboratory fingdings,site of infection,site of operation and special treatment.First,feature engineering was performed on the dataset of derivation cohort,including missing value imputation,variables discretization and feature selection.Then,cluster analysis was performed using PAM algorithm on the derivation cohort to exploring the potential clinical subtypes.The differences of each subtype were compared and their clinical characteristics were summarized.Kaplan-Meier survival analysis and Cox proportional hazards model were used to evaluate the effect of RBC transfusion on clinical prognosis in overall and each subtype.Using the same variables which were selected for clustering analysis in derivation cohort,clustering analysis was carried out on validation cohort to validating the clinical subtypes,and then we evaluated the effect of RBC transfusion on clinical prognosis in overall and each subtype.The validation cohort was used to verify the stability and repeatability of the clustering results.Results1 Clinical data and clustering analysis results of derivation cohort.1.1 A total of 6821 septic patients meeting our study criteria were collected in the derivation cohort.The median age was 69.15(56.60,80.11)years old,and the median of Sequential Organ Failure Assessment(SOFA)was 5.00(4.00,8.00).The14-day,28-day,and 90-day mortality were 14.88%,21.52%,and 31.10% respectively.The median length of ICU stay was 4.00(2.00,8.00)days,and the total length of hospital stay was 11.00(7.00,20.00)days.Among them,3874 patients(56.80%)received RBC transfusion,3738 patients(54.80%)received mechanical ventilation,and 412 patients(6.04%)received renal replacement therapy.The scores of SOFA,logistic organ dysfunction score(LODS)and simplified acute physiological score Ⅱ(SAPS Ⅱ)in transfusion group were higher than those in non-transfusion group.The median levels of hemoglobin and hematocrit were 83.00g/L and 24.50% in transfusion group,which were lower than those in non-transfusion group with statistically significant difference(P<0.001).The 14-day mortality was lower in transfusion group(13.45% vs.16.76%,P<0.001),but there were no significant differences in 28-day and 90-day mortality compared with non-transfusion group.1.2 Boruta algorithm was used to select the appropriate variables for clustering analysis.According to silhouette coefficient,PAM algorithm was used to group the patients in derivation cohort into three subtypes.Cluster A,cluster B and cluster C contained 1835 cases(26.90%),3043 cases(44.61%)and 1943 cases(28.49%)respectively.Cluster A were characterized by advanced age and heart problems.The median age of cluster A patients was >70 years old.72.86% of them were complicated by heart disease,and 63.87% had received cardiac operation.Cluster B were characterized by mild disease and relatively higher hemoglobin level.The medians of SOFA,LODS and SAPS Ⅱ in cluster B were the lowest among the three subtypes,and the median level of hemoglobin was >90g/L.Cluster C were characterized by severe status,high proportion of bloodstream infection,coagulopathy,hyperlactemia,and the history of abdomen operation.The medians of SOFA,LODS and SAPS Ⅱ in cluster C were the highest among the three subtypes.The medians of activated partial thrombin time(APTT),prothrombin time(PT)and international standardized ratio(INR)were higher than those of the other two types,and the median level of lactic acid was >3mmol/L.The proportion of patients with bloodstream infection in cluster C was 72.31%,and 61.71% of patients had received abdomen operation.The 14-day,28-day and 90-day mortality of cluster C were the highest among the three subtypes.1.3 There were 1332 cases(72.59%),1168 cases(38.38%)and 1374 cases(70.72%)receiving RBC transfusion in cluster A,cluster B and cluster C respectively.Using Kaplan-Meier survival analysis and Cox proportional hazards model,in overall,RBC transfusion could reduce the risk of death at 14 and 28 days,with hazard ratio(HR)[95% confidence interval(CI)] of 0.67(0.58,0.78)and 0.82(0.73,0.93),but there was no statistically significant reduction in the risk of death at 90 days with HR(95%CI)of 0.92(0.83,1.02).In cluster C,RBC transfusion significantly reduced the risk of death at 14,28 and 90 days,with HR(95%CI)of 0.50(0.41,0.61),0.61(0.51,0.72)and 0.67(0.58,0.78),while it did not reduce the risk of death at 14,28,and 90 days in cluster A and B.2 Clinical data and clustering analysis results of validation cohort2.1 A total of 874 sepsis patients meeting our study criteria were collected in the validation cohort,including 342 patients from the first affiliated hospital of Dalian Medical University and 532 patients from the second affiliated hospital of Dalian Medical University.The median age was 69.73(56.64,80.17)years,and the median of SOFA was 6.00(4.00,8.00).The 14-day,28-day and 90-day mortality were17.73%,24.60% and 34.32% respectively.The median length of ICU stay was 4.00(2.00,9.00)days,and the total length of hospital stay was 12.00(7.00,19.00)days.Among them,520 patients(59.50%)received RBC transfusion,497 patients(56.86%)received mechanical ventilation,and 47 patients(5.38%)received renal replacement therapy.2.2 Using the same variables which were selected for clustering analysis in derivation cohort,the patients in the validation cohort were grouped into three subtypes using PAM algorithm,including 330 cases(37.75%)in cluster v A,176cases(20.14%)in cluster v B and 368 cases(42.11%)in cluster v C.The clinical features of each subtype were similar with derivation cohort.Cluster v A were characterized by mild disease and relatively higher hemoglobin level.Cluster v B were characterized by severe status,high proportion of bloodstream infection,coagulopathy,hyperlactemia,and the history of abdomen operation.Cluster v C were characterized by advanced age and heart problems.2.3 There were 151 patients(45.76%),135(76.70%)and 234(63.59%)receiving RBC transfusion in cluster v A,cluster v B and cluster v C respectively.Using Kaplan-Meier survival analysis and Cox proportional hazards model,in overall,RBC transfusion could reduce the risk of death at 14 days with HR(95%CI)of 0.66(0.45,0.98),but there was no statistically significant reduction in the risk of death at28 days and 90 days,which HR(95%CI)was 0.78(0.56,1.09)and 0.86(0.64,1.14)respectively.In cluster v B,RBC transfusion significantly reduced the risk of death at14,28 and 90 days with HR(95%CI)of 0.36(0.18,0.73),0.44(0.24,0.81),and 0.47(0.27,0.80),while it did not reduce the risk of death at 14,28,and 90 days in cluster v A and v C.Conclusion1 In general,RBC transfusion only improved the short-term clinical prognosis of septic patients.2 There were three potential clinical subtypes in septic patients: cluster I was characterized by advanced age and heart problems;cluster Ⅱ was characterized by mild disease and relatively higher hemoglobin level;cluster Ⅲ was characterized by severe status,high proportion of bloodstream infection,coagulopathy,hyperlactemia,and the history of abdomen operation.3 The patients in cluster Ⅲ can benefit from RBC transfusion to improve their short-term and long-term outcomes.4 Besides hemoglobin,risk factors such as severe status,high proportion of bloodstream infection,coagulopathy,hyperlactemia,and the history of abdomen operation should be considered when developing the strategy for RBC transfusion in septic patients.5 According to our study,the k-medoids algorithm has good stability and repeatability in the study of sepsis,which may help in understanding heterogeneity of treatment effects and facilitating future randomized controlled trials. |
PDF Full Text Request