A Prediction Research Of Acute Kidney Injury Risk Based On Time Series

Acute Kidney Injury(AKI)is characterized by rapid onset,high mortality and poor prognosis.Clinicians usually pay attention to the changes in serum creatinine when diagnosing AKI,but patients often have AKI at the same time of diagnosis,missed the best opportunity for treatment,so it is necessary to study the method to quickly predict AKI in advance in clinical.The advent of the era of big data and the development of machine learning and deep learning make ai-assisted medicine possible.In recent years,scholars at home and abroad have carried out many studies on the prediction of AKI,but most of the studies are unable to continuously predict AKI in patients.To solve the above problems,this paper proposes a sequential model for AKI prediction.By using medical sequential data,a regression prediction model for serum creatinine can be established to assist doctors in diagnosis and advance treatment.The main research contents of this paper are as follows:(1)In this paper,Hi RID,the original data set published abroad,was cleaned and preprocessed to construct the timing data in line with the model input.Under the guidance of professional doctors,characteristics were selected and included in the training data,and important characteristics were screened through Xgboost to train clinical prediction models for AKI in different time periods,which could improve the applicability of the model.(2)This paper proposes an AKI continuous prediction algorithm based on the combination of attention mechanism and LSTM model.For multi-dimensional time series,the weight assigned by LSTM is the same,and the importance of information is not distinguished.Therefore,if long time series or multi-dimensional time series are trained,LSTM neurons may lose more important information.To solve the above problems,this paper introduces the Attention structure into the LSTM network structure.By assigning weight to information,the model can remember more information beneficial to the model effect.The experimental results show that LSTM+Attention model has the best performance,achieving the highest accuracy of 93.97% in the 8h prediction task,showing better prediction effect compared with other common deep learning models(CNN,LSTM).(3)This paper also proposes a new AKI deep network structure TCN+LSTM+Attention.Cycle the orderly recursive neural network training data characteristics lead to AKI time-series forecasting model of long training time,poor efficiency problem,this paper designs a new network structure,the introduction of sequence convolution module,through the study of TCN module time-series data characteristics,due to the existence of the convolution operation to improve the computation efficiency and compression characteristics.LSTM+Attention was used to further learn features.Finally,a residual module was used to reduce information loss,so as to ensure that the prediction effect of the model would not be reduced on the basis of improving the model's operation efficiency.The final results showed that TCN+LSTM+Attention model was 93.57% accurate in predicting serum creatinine after8 h,which was 0.4% lower than the LSTM+Attention model with the highest accuracy,but 71% less in training time.