Research On Method Of Mining Drug Patterns And Prognosis Prediction For Critically Ill Patients

The electronic medical records(EMR)are composed of various data generated by patients after admission,including amounts of clinical diagnosis data,which can indicate the patients’ condition in hospital.By mining the information under the data,it can provide support for clinical decisions.However,the EMR has problems such as poor data quality and strong data sparsity so that it is difficult to be used directly in machine learning.In this paper,the e ICU Collaborative Research Database composed of critically ill patients can be used to establish a model which can provide assistant in diagnosis and treatment for critically ill patients.In this paper,a prognosis prediction method based on drug pattern mining for critically ill patients is established,which includes multiple algorithms such as electronic medical record data preprocessing,drug pattern mining,feature selection technology,prognosis prediction model,and model interpretability analysis.The main works of this paper are divided into four parts:(1)the clinical data of patients can be divided into explicit features and implicit features.Then,we introduce a series of algorithms for the preprocessing of the explicit features,which make the conversion between the explicit features and explicit features variable,so that it can be used for the establishment of the subsequent prognosis prediction model;(2)the algorithm of drug pattern mining for critically ill patients are proposed,which provide a tool of converting the drug data into the drug pattern variables.Moreover,the generated drug pattern variables provide the data support for the subsequent model establishment;(3)explore the establishment of a prognosis prediction model with drug pattern mining for critically ill patients.Based on the patients’ explicit features mentioned above and implicit features obtained through the drug pattern mining algorithm,a prognosis prediction model based on drug pattern mining is established.Through feature selection technology,the key features related to patients’ prognosis are extracted.Logistic regression,Random forest,XGBoost and Cat Boost algorithms are used to build the prediction models.(4)In order to better demonstrate the prognosis prediction model proposed in this paper,the interpretability analysis of the proposed prediction model is carried out with SHAP.In this paper,we use the EMR of critically ill patients in the e ICU database to verify the proposed prediction model and evaluate the effect.Finally,the main contributions of this paper are divided into five parts:(1)The existence of data imbalance has a negative impact on the prediction model,and the SMOTEENN algorithm can reduce its influence and improve the effect of the model;(2)Then,19 features which is identified by the feature selection technology can be used to establish the prognosis prediction model through comparative analysis;(3)The experimental results show that the excellent prediction model is the prediction model constructed based on the XGBoost algorithm,and its AUC is 0.75;(4)The patient prognosis prediction model based on drug pattern mining proposed in this paper also has a good performance on the validation set,which proves that the proposed prognostic prediction model can provide the support for clinical decision in reality;(5)The interpretability analysis of the prediction model is carried out with SHAP,and the analysis is carried out from the overall level and the individual level.It is concluded that the “apachescore” is the most relevant feature of the prognosis of the patient.