Early Death Risk Prediction Model For Multiple Organ Dysfunction In The Elderly Based On Integrated Machine Learning

Background:Multiple organ dysfunction syndrome in the elderly(MODSE)is a clinical condition referring to the critical state of organ dysfunction in the elderly on the basis of organ aging and various chronic diseases.This syndrome can be triggered by seemingly trivial precipitating factors,in which two or more organs can sequentially or simultaneously deteriorate into dysfunction or failure in a short period of time.The high mortality makes it one of the main causes of death in elderly patients with severe disease.Therefore,the combination of early diagnosis,prognostic prediction and prompt comprehensive treatment is an important way to reduce its mortality.Commonly used scores of organ dysfunction in clinical practice,such as MODS scores,Sequential Organ Failure Assessment(SOFA)scores,and Acute Physiology and Chronic Health Evaluation Ⅱ(APACHE Ⅱ),are predominantly more suitable for non-elderly adults.MODSE is different from MODS,consequently commonly used organ dysfunction score has limitations in evaluating the severity of the disease and predicting the outcome in MODSE given the complexity of severe diseases in elderly patients and management environment in the ICU.The inclusion of monitoring comprehensive,multi-dimensional indicators,the creation and development of Electronic Health Records(EMR)has made this kind of research possible.Objectives:1)For MODSE patients,machine learning algorithm was used to analyze the risk factors of early death;2)The integrated learning model XGBoost was used to construct a predictive model to evaluate the risk of early death of MODSE,so as to better assist clinical decision-making and treatment.Methods:1)Using the published electronic medical record-based database "Medical Information Mart for Intensive Care(MIMIC)-Ⅲ",14329 patients with MODSE were included,of which 2360(14.73%)died in the hospital.According to the prognosis,patients were divided into death group(n=2360)and survival group(n=13654).A total of 122 clinical characteristics,including demographic features,vital signs on the first day of ICU,clinical interventions,and systemic inflammatory reaction syndrome(SIRS)scores were collected.Differences between both groups were analyzed.XGBoost model algorithm was used for the investigation of the distribution and ranking of death-related features.2)Based on the MODSE inclusion criteria,three MODSE patient datasets were generated from the MIMIC-Ⅲ database,the eICU collaborative research database,and the PLA General Hospital critical patient database.Among included patients,80%were randomly selected as the training set,and the remaining 20%were the test set.The 122 clinical features were set as model parameters for integrated machine learning model:The support vector machine(SVM)model,K Nearest Neighbor(KNN)model,and random forest(RF)model.A predictive algorithm was established using the XGboost model in the training set.Receiver operating characteristic(ROC)curve was used to evaluate the predictive value of the model for the risk of death in patients with MODSE in the validation set,which was then compared to the performance of existing clinical models such as SOFA,APACHE-Ⅲ,MODS,and SAPS.Finally,the best performing model was cross-validated in three databases.Results:Compared with the survival group,patients in the death group had significantly higher mean Glasgow coma score(GCS),mean age,and maximum heart rate,as well as lower BMI and systolic blood pressure(both P<0.01).The ten factors with the best predictive values in the XGBoost model were:Respiratory rate,activated partial thromboplastin time(APTT),age,body temperature,body mass index(BMI),systolic blood pressure,platelets,blood glucose,shock index,and white blood cell count.In the prediction of death in MODSE patients,the XGboost model had a sensitivity of 0.824,a specificity of 0.725,and an accuracy of 0.854 with an AUC of 0.853.In the cross-validation comparison,the model constructed by MIMIC-III database is the best in the PLA General Hospital critical patient database with an AUC of 0.882.Conclusion:Compared with traditional scores,the XGBoost model has better predictive performance and is more universal.