Research On Prediction And Prewarning Model Establishment Of Traumatic Haemorrhagic Shock Based On Trauma Big Data

Purpose:In this research,data of hemorrhagic traumatic shock patients were extracted from massive trauma data by data retrieval technology.The key indicators of traumatic hemorrhagic shock were extracted by intelligent screening algorithm.Machine learning and deep learning algorithms were applied to analyze the time series data and verify the models.Therefore,this research can predict and warn traumatic hemorrhagic shock as well as evaluate the effectiveness of the models objectively to improve the ability of medical staff to perceive this complication in advance,realize early intervention and early treatment,reduce death rate and disability rate of the wounded caused by post-traumatic blood loss,and improve the prognosis of the wounded.At the same time,it provides some references for other clinical decision support researches using big data technology.Methods:(1)Data retrieval technology was used to extract the data of patients with traumatic hemorrhagic shock from MIMIC III database.Rough set algorithm was used to analyze the huge data set of indicators including vital signs,blood routine,blood gas analysis,coagulation function,blood biochemistry and urine routine.Then,the cellular genetic algorithm(CGA)was used to conduct independent repeated experiments,and the recognition ability of the indicators was determined per the number of times they were retained in the screening,so as to form a new set of key indicators.(2)According to the different ability of the selected key indicators to identify injuries,that is,the number of times the selected indicators are retained is different,the key indicators were reorganized and grouped,and the key indicators of time series are taken as the analysis object.At the same time,the indicators that cannot meet the research needs are removed.Applying the logistic regression,support vector machine,naive bayes and AdaBoost in machine learning,the outcome variables of traumatic hemorrhagic shock were predicted.The performance of the classifier model was evaluated,and then the accuracy,recall rate,precision rate and F value calculated by various algorithms under different grouping index sets were obtained respectively,and the results were compared.(3)Two neural network models,MLP and GRU,were used respectively,and the key indicators of the ladder type of the wounded in traumatic hemorrhagic shock were used.There are three different combinations:vital signs only,vital signs+blood gas analysis and vital signs+blood gas analysis+blood routine.The time span was set as 1 hour and the time step was adjusted to optimize the performance of the model.The performance of the model was evaluated to obtain the accuracy rate,recall rate,precision rate and F value calculated by various algorithms under different grouping index sets and to compare the results.(4)The traumatic data in the PLA General Hospital Emergency Rescue Database were used to verify the effectiveness of the models objectively.The external validation was obtained to compare with the internal validation so as to analyze the generalization ablity of the models in other database.Results:(1)Cellular genetic algorithm was applied to conduct 10 independent repeated attribute reduction tests.The key indicators of reduction for 10 times were respiratory rate and leukocyte.The key indicator of reduction and retention for 9 times was aspartic acid aminotransferase.The key indicators of reduction and retention for 8 times were PC02,PH,body temperature and urine specific gravity.The key indicators of reduction and retention for 7 times were plasma fibrinogen determination,systolic blood pressure and diastolic blood pressure.The key indicators of reduction and retention for 6 times were INR,lactic acid,chloride,glucose and heart rate.A total of 10 sets of optimal key indicators combination rules are discovered.(2)The time series indicators were divided into 4 data sets including all indicators,13 key indicators,8 key indicators and 6 key indicators,and 4 machine learning algorithms were applied for predictive analysis respectively.When the full indicators data set is applied,the AdaBoost method performs better than the other three methods in three aspects of accuracy rate,recall rate and accuracy rate.When applying different number of key indicators data sets,F1.5 value of AdaBoost reached 90.1%when applied to data sets containing 13 key indicators.When applying the full data set,F1.5 value reaches 91.8%,which is the best among the four algorithms.(3)Two kinds of neural network models,GRU and MLP,were used for prediction on three sets of different prediction and warning indicators data setsrespectively.When the time span was 1h and the time step was 5,the performance of the model was the best.When applying GRU to predict traumatic hemorrhagic shock,F1.5 value predicted by vital signs alone can reach up to 85.1%.When F1.5 value is guaranteed to be no less than 80%,the occurrence of traumatic hemorrhagic shock can be predicted 2 hours in advance.When using vital signs and blood gas analysis indicators for prediction,the F1.5 value can reach 86.3%at most.When the F1.5 value is guaranteed to be no less than 80%,the occurrence of traumatic hemorrhagic shock can be predicted 3 hours in advance.When using vital signs,blood gas analysis and blood routine indicators for prediction,F1.5 value can reach up to 90.1%.When F1.5 value is guaranteed not less than 80%,the occurrence of traumatic hemorrhagic shock can be predicted 4 hours in advance.F1.5 value of GRU was higher than that of MLP when the same index was used and the lead time was the same,indicating that the former had better performance than the latter.(4)Through external validation,it can be found that AdaBoost had the best performance and robustness in both external and internal validation.The F1.5 reached up to 0.892 and was near with the result in internal validation.On the opposite,the performance of deep learning models were not good enough,and only the values of accuracy were near the result of internal validation.Conclusion:(1)Screening of key indicators is an important basis for the prediction and warning of traumatic hemorrhagic shock,improving the prediction efficiency of the model,as well as an important measure and means to reduce the number of demand indicators and simplify the model as much as possible.(2)Machine learning can better predict the occurrence of traumatic hemorrhagic shock.When the key index set is applied to forecast,there may be an optimal index combination,and the economic benefit and time benefit of index collection should be considered comprehensively.(3)Deep learning is used to predict and warn traumatic hemorrhagic shock,and a prediction time window is introduced to put forward the concept of stepwise prediction index.When indexes that can be monitored in real time could be used and good prediction efficiency can be achieved by using algorithms,real-time dynamic prediction and warning can be implemented for traumatic hemorrhagic shock in theory.(4)Evaluating the generalization ability and performance through external validation is an effective method.The generalization ability was enhanced when using screened core indicators illustrating that applying Cellular genetic rough set algorithm to screen core indicators played a crucial part in enhancing the models' performance.In conclusion,through a series of studies,a variety of model algorithms that can predict and warn hemorrhagic shock for patients with clinical trauma have been obtained.The next research focus is to conduct rigorous clinical verification,test the actual prediction effect strictly,and to constantly optimize and improve the model.