Development And Application Of Classification And Prognostic Model Among Severe Trauma Patients

Background:Trauma is a global epidemic,accounting for 9.2%of the total deaths worldwide,costing more than US$500 billion a year,and placing a heavy economic burden on society.Timely and effective pre-hospital assessment and triage are of great significance for good prognosis of trauma patients.The American College of Surgeons Committee on Trauma（ACS-COT）guidelines state that the goal for trauma triage is to have an under-triage rate of less than 5%and an over-triage rate of less than 35%.Previous studies have sought to develop predictive models to assist Emergency Medical Service（EMS）personnel in early clinical assessment and medical triage.However,on the one hand,most of the existing trauma triage schemes are based on the summary of clinical experience,on the other hand,they are mostly based on the small sample data of a single center,or adopt simple generalized linear statistical models for fitting,or contain variables that are relatively difficult to measure in the clinical settings.Therefore,these models are not well applied in practical work,and under-triage and over-triage often occur.Given the complex association between individual differences in injury and their potential synergistic effect on mortality,Machine learning is widely used in medical modeling because of its ability to capture complex,non-linear relationships between features.Objective:In this study,this paper aims to use the National Trauma Data Bank（NTDB）to develop a prediction model,so as to improve the efficiency of injury classification and provide a theoretical basis for the rational allocation of human and medical resources.First,introduce the variable distribution in NTDB.Secondly,to develop a prognostic model of adverse clinical outcomes in severe trauma patients based on machine learning algorithm;Thirdly,8-hour and 24-hour survival prediction models for patients with severe trauma were developed based on machine learning survival algorithm（considering time factors）.Methods:1.Identify trauma patients based on ICD-9 or ICD-10 codes,and conduct corresponding statistical analysis on the overall distribution of trauma patients and the subgroup distribution of trauma patients（annual trend of trauma patients,trauma locations,trauma mechanisms,severity of injuries,survival outcomes and complications,and multiple injuries of different patients）.This study aimed to comprehensively depict the basic information of trauma patients in NTDB.2.In the study of adverse outcomes of severe trauma patients,demographic,trauma type,pre-hospital vital signs,hospital outcomes and other indicators of patients in NTDB from January 1,2017 to December 31,2018 were collected.Nine machine learning algorithms,namely Neural Network（NNET）,Naive Bayes（NB）,Logistic Regression（LR）,Gradient Boosting Machine（GBM）,Adaptive Boosting（Ada）,Random Forest（RF）,Bagging Tree（TB）,Categorical Boosting（Cat Boost）and Extreme Gradient Boosting（XGB）,were used to construct the predictive model in order to predict the clinical outcome of patients with severe trauma.3.In the study of the 8-and 24-hour survival of severe trauma patients,demographic,trauma type,pre-hospital vital signs,hospital outcomes and other indicators of patients in NTDB from January 1,2015 to December 31,2018 were collected,and survival tree（ST）,random survival Forest（RFS）,GBM and Cox proportional risk algorithm were used to construct prediction model.C index,differentiation and calibration curve were used to evaluate the model performance from multiple dimensions.Results:1.From 2015 to 2018,a total of 3,917,656 trauma patients were identified from NTDB,including 3,177,141 males（81.11%）and 739,901 females（18.89%）respectively.The male-female ratio was about 4.29:1.From 2015 to 2018,a p-trend test of trauma numbers<0.001.The trauma patients were mainly young adults aged 16-49 years（40.73%）,and white people（74.36%）.In addition,patients were mostly admitted to non-profit teaching hospitals with more than 600 beds.From the cause of injury,traffic injury（32.78%）ranked first.The head（27.57%）and lower limb（24.42%）were the most vulnerable locations to injury.In terms of pre-hospital indicators,the pre-hospital time of trauma patients was generally shorter,such as,emergency medical system（EMS）response time of 9.00（5.00,18.00）minutes.In terms of pre-hospital scores,90.91%of patients had ISS scores of 13-15,and 84.30%had GCS scores of 0-16.In terms of in-hospital outcomes,1.07%of patients died in the emergency room and 3.05%in the hospital.2.A total of 50,429 patients were included,of which 2808 had adverse outcomes.The risk factors for adverse outcomes were trauma type,EMS transport time,systolic blood pressure,oxygen saturation,respiratory rate,EMS response time,pulse rate,EMS field time,ISS score,GCS score,and age.The AUCs of NNET,NB,LR,GBM,Ada,RF,TB,Cat Boost and XGB were 0.893（0.8888,0.899）,0.863（0.856,0.870）,0.891（0.886,0.898）,0.542（0.530,0.553）,0.865（0.858,0.872）,0.888（0.882,0.895）,0.837（0.829,0.846）,0.776（0.766,0.787）,0.863（0.856,0.870）,among which,NNET model has the best differentiation.In terms of calibration,NNET also shows good performance.Compared with traditional scores,the NNET model showed better predictive performance（0.893 vs 0.747 vs 0.813）.In order to convenient for clinical use,we set up an online based on the"Shiny"packet visualization platform（https://pcstudy.shinyapps.io/Probability＿of＿adverse＿outcome/）.3.From 2015 to 2018,a total of 191,240 patients with severe trauma were included.GCS score,ISS score,injury type,age,Sa O₂,respiratory rate,systolic blood pressure,EMS transport time,EMS field time,pulse rate and EMS response time were identified as the main predictors.For predicting 8-hour survival,The C indices of Cox,ST,RFS and GBM were 0.866（0.859,0.873）,0.851（0.842,0.860）,0.886（0.878,0.894）and 0.872（0.864,0.880）,respectively.The 24-hour survival model performed similarly,while the RFS model also had a lower error rate.In terms of calibration,the RFS model performs well.In order to convenient for clinical use,we set up an online calculator based on the"Shiny"package visualization platform（https://pcstudy.shinyapps.io/survival＿prediction＿model/）.Conclusions:From 2015 to 2018,the number of trauma patients was on the rise,and most of them were white young adults.Nine machine learning algorithms all performed well in predicting adverse outcomes of patients with severe trauma,among which NNET model was the best.Trauma type,EMS transport time,systolic blood pressure,oxygen saturation,respiratory rate,EMS response time,pulse rate,EMS field time,ISS score,GCS score and age could be used as predictors.It is worth noting that these prehospital indicators are easy to measure clinically.For survival prediction considering time factors,RFS has the best performance and has strong clinical application value.Although the model of this study is based on the validation of the large sample size of the multi-center,further clinical verification is needed.