Research On Multi-dimensional Sparse Time Series Prediction And Prescription Prediction For Clinical Data

Medical artificial intelligence has long become a hot field of global attention.Multi-faceted research on complex clinical data can assist doctors in diagnosis and treatment to promote the development of medical intelligence.Clinical time series data is limited by the influence of collection scenarios,and has the characteristics of missing values,irregular time intervals,and sparseness,which makes the research of real clinical time series data a great challenge,and more advanced models need to be built.Prescription runs through China’s the long-standing TCM theoretical system,and it is very challenging to use algorithms to learn complex information and simulate TCM prescribing.In this regard,this paper starts with the real clinical diagnosis and treatment process of new crown patients,takes symptoms as the correlation point,and constructs features to characterize patients in two dimensions: numerical and textual,and conducts research in two aspects.(1)For the clinical time series data with missing,sparse and irregular sampling,the experimental comparison and analysis of the three advanced models(Mtans,GRU-D,PLSTM)time series modeling performance are better than the traditional time series model.Firstly,a data set of physical and chemical indicators of patients with new coronary pneumonia was constructed,and two research tasks were carried out based on the multiple detection values of patients: death outcome classification task and interpolation task.On classification tasks,the three advanced models outperform traditional missing value processing methods,namely GRU-mean and GRU-forward.The AUC value of the Mtans model reaches 0.96,which is 19% and 21% higher than GRU-mean and GRU-forward,respectively.In addition,a classification task based on index blocks is further designed.After being divided into blocks according to the meaning of physical and chemical indexes,different index blocks show different contributions to the characteristics of patients.On the interpolation task,the mean squared error value of the Mtans model is lower than that of the RNN＿VAE baseline method.And the experimental comparison is also carried out on the public dataset Physio Net.(2)Aiming at the prediction of TCM prescriptions for specific patient populations,a multi-label classification prescription prediction model(Text＿Attention Seq2 Seq,TASeq2Seq)integrated with attention mechanism is proposed.First,a dataset of symptom prescriptions of new crown patients was constructed.Based on the Seq2 Seq model architecture,an attention mechanism is added to the encoder and decoder respectively,and the correlation between symptoms and between symptoms and drugs is considered to further improve the classification effect of the model.Compared with multi-label classification methods such as Text＿RNN,the TASeq2 Seq model achieves better results,and the F1 value and jaccard value are 24% and 22% higher than the multi-label decision tree,respectively.By further integrating the duration of symptoms to build more abundant features,the TASeq2 Seq model has significantly improved in terms of accuracy,F1 value and other indicators.At the same time,experimental comparisons are also performed on the chronic lung disease dataset(COPD).