Research And Application Of Recurrent Neural Networks For Survival Prediction In Trauma Critical Care Patients

Background:Trauma patients are characterized by urgent condition and rapid return,which impose strict requirements on the medical team’s experience,professional knowledge and tacit understanding in clinical resuscitation.Therefore,it is of high significance to use computer and big data technology to make accurate and rapid prediction on the progress of trauma patients’ condition to improve the survival rate of patients.In recent years,with the deep learning artificial intelligence research,recurrent neural network(RNN)models have made great progress in the prediction of temporal data including speech semantic recognition and human emotion judgment,etc.In addition,RNN has also been involved in a variety of ICU conditions in recent years,which makes it possible to establish dynamic models for prognosis and regression prediction of trauma critical care patients using RNN technology.Objective:The RNN algorithm is used to train the simulated data based on the augmented clinical information of trauma critical care patients,to build a dynamic prediction model of trauma critical care patients based on the RNN algorithm,and to explore the changes of model prediction performance under different RNN algorithms,model building structures,time window lengths,predicted outcomes and variable combinations,and then to find a reasonable way to build the model.Using the clinical data of real trauma critical care patients and building RNN prediction model again under the conclusion of simulation study,we compare the performance advantages and disadvantages of this model with the models of widely used prediction algorithms,and explore the self-improvement method of RNN model to verify the feasibility of RNN algorithm in the practice of trauma critical care.Methods:I.Trauma patients were screened in the MIMIC-III database and the data were augmented using the SMOTE algorithm.Two RNN algorithms,the Long Short Term Memory Network(LSTM)algorithm and the Gated Recurrent Unit(GRU)algorithm,were used to predict patient outcomes within 24 hours and final outcomes at 4,6,and 8 hours,respectively,using three structures: static modeling,dynamic "tandem",and dynamic "parallel".Three time windows were used to predict the patient’s outcome and final ICU outcome within 24 hours of the circulation node.After the prediction,the sensitivity,F1 value and AUC value of each model were calculated to analyze the performance difference of each RNN model under various structures,time windows or outcomes,and then select the suitable RNN model building strategy.II.Three machine learning algorithms,namely Random Forest(RF),XGBoost and Adaboost,and clinical experts suggested to screen and re-model the analysis variables of dynamic prediction models for trauma patients,and compare the changes of model prediction effects before and after variable screening.III.The clinical data of trauma patients were queried in the MIMIC-IV database.Using the RNN model construction scheme and variable combinations obtained in the simulation study,the LSTM algorithm model,GRU algorithm model,Hidden Markov Model(HMM)model,RF model,and Logistic model were trained to predict the immediate and final outcomes of trauma patients,and then the performance differences between the RNN algorithm model and other algorithm models were compared.Two secondary training schemes,including combined training and fine-tuned training,are adopted to discuss the continued performance improvement effect of the RNN algorithm model and the feasibility of the model based on the application in the context of trauma critical care patients.Results:I.Under the modeling conditions of this study,the average sensitivity,F1 and AUC values of the dynamic "tandem" structure of the RNN model were higher than those of the static and dynamic "parallel" structure models,and the time trends of the metrics were stable.In this study,the RNN dynamic model with 8-hour time window has a higher average performance than the 6-hour and 4-hour prediction performance indicators,and the prediction performance is positively correlated with the length of the time window set in this study.The RNN model in this study predicted the immediate outcome better than the final outcome but no difference in model performance was found between using the LSTM algorithm and the GRU algorithm.II.In this study,using the combination of "age,heart rate,systolic blood pressure,hemoglobin content,total bilirubin,serum creatinine and partial pressure of oxygen",the mean values of all indicators in the RNN model improved,with an average increase of over3.5% in F1 values,over 2.5% in AUC values,and an increase of over 1.5% in sensitivity,by combining algorithm screening and expert recommendations.The sensitivity was improved by more than 1.5%.III.Based on the model building conditions and the trauma patient data context of this study,the RNN model has a higher mean value of each evaluation metric than the HMM model,RF model and the algorithmic model of the Logistic model.IV.In this study,both the secondary training schemes of combined training and finetuning training could achieve the effect of continued improvement of model performance,and the combined training performed better than the fine-tuning training in this study,but in the context of trauma patients in this study,the overall improvement of model performance was about 1%.ConclusionAmong the various RNN construction schemes discussed in this study,the dynamic "tandem" structure and 8-hour time window modeling were the most effective,and the recent outcome prediction was better than the final outcome;the LSTM algorithm and GRU algorithm did not show any performance difference in this study;the variable screening in this study could improve the model prediction effect.The dynamic prediction model based on RNN algorithm in this study showed better applicability in the context of trauma critical illness,and both RNN models predicted better than the other models discussed in the context of real trauma critical illness,and both combined training and fine-tuning training could continue to improve the prediction performance of the model.