The Establishment Of Prediction Models For Successfully Placing Self-propelled Nasojejunal Tubes

[Objective]Nutritional support is an indispensable part of comprehensive treatment for critically ill patients. Generally, patients prefer enteral nutrition, and tube feeding is its main method. The use of nasogastric tube may easily cause complications related to enteral nutrition, such as gastro-esophageal reflux, aspiration and aspiration pneumonia, due to high residue caused by insufficiency of gastric motility. Comparatively, the use of nasojejunal tube seldom causes these complications because insufficiency of enteral motility is rather scarce. Bedside pernasal placement of nasojejunal tube is commonly accepted by healthcare providers, patients and their dependents because it is safe, cheap and minimally invasive. However, its high technical requirements and low success rate restrict its wide application. Recently, a new type of nasojejunal tube called self-propelled nasojejunal feeding tube (SNT) is invented. Studies find that the success rate is relatively high in its bedside placement, but rather low in post-pyloric placement for critically ill patients. This study, by using data mining technique, aims to construct two decision tree and one binary logistic regression prediction models as well as to test their predictive performance. The objectives of this study are:(1) to investigate whether it is feasible to construct a prediction model and predict the success rate of SNT placement by using data mining technique;(2) to investigate whether it is feasible to predict the success rate of SNT placement by constructing decision tree and binary logistic regression prediction models;(3) to compare and evaluate the performance of the decision tree and binary logistic regression prediction models foresaid. Through the study of these prediction models, We can take the initiative to choose the critically ill patients who may success in placing SNT. Thereby we can achieve the purpose of improving SNT placement success rate.[Methods]In this study, prediction models on the success of SNT placement will be constructed by applying data mining technique, the process of which is completed according to Cross-Industry Standard Process for Data Mining (CRISP-DM).Business understanding:It is known in preliminary study that the success rate of SNT placement is low because of several factors. Data mining is used to classify critically ill patients and to predict the success rate of SNT placement in a certain patient based on this classification. SPSS21.0and Rapidminer6.0.001is used for data mining.Data understanding:In this study,609patients receiving bedside pernasal SNT intubation between August2005and November2013in the intensive care unit of Guangdong General Hospital (GGH),(Guangdong Academy of Medical Sciences, GAM)are included. This primary dataset is divided in two parts. One part is searched and collected from the GGH(GAM) hospital information system, presenting as Microsoft Excel2010spreadsheets (.xlsx). Another part is collected from case reports of the study "The prospective, multi-center, randomized controlled clinical trial of improving the success rate of post-pyloric placement of spiralnasaljejunal tubes in critically ill patients by prokinetic drugs, PROMOTE". Inclusion and exclusion criteria as well as variables and variable names are also defined in this stage.Data preparation:Data is included, excluded, processed, and confirmed.(1) Inclusion. Primary dataset is analyzed in a medical professional perspective, and variables that affect the success rate of intubation are included for further exclusion.(2) Exclusion. Primarily included data is checked to discover and exclude the discorded, depleted and repeated items.(3) Processing. Variables are converted according to clinical standard or common sense; variables that can be processed together are integrated; quantitative variables are discretized as categorical variables.(4) Confirmation. Retrospective data of491patients receiving SNT placement between August2005and December2011are defined as a training cohort; Prospective data of118patients receiving SNT post-pyloric placement between April2012and November2013, also included in the study "The prospective, multi-center, randomized controlled clinical trial of improving the success rate of post-pyloric placement of spiralnasaljejunal tubes in critically ill patients by prokinetic drugs, PROMOTE", are defined as a testing cohort.Modelling:(1) The training cohort is applied to construct a CHAID-based decision tree in SPSS21.0. The growth method is CHAID algorithm and set parameters by maximum tree depth:3; significance of splitting nodes:0.01; significance of merging categories:0.05; calculated using the Pearson chi-square statistic; maximum number of iterations:100; minimum change in expected cell frequencies:0.05.(2) The training cohort is applied in the decision tree module of RapidMiner6.0.001and a RapidMiner decision tree is generated under the condition of Gini Index. The parameters set by minimal size for split:4; minimal leaf size:2; minimal gain:0.1;maximum tree depth:5; confidence:0.25.(3) The training cohort is applied to construct a Logistic regression model in SPSS21.0. Training cohort is classified based on placement outcome. Categorical variables are χ2tested. Quantitative variables are t tested. P<0.10is considered significant. Significant variables are regarded as single impact factor and Wald method is applied to include those larger than0.05and exclude those larger than0.10for binary logistic regression model construction.Evaluation:Testing cohort is used to test the performance of the three prediction models. Predictive success rates in different models are used to generate ROC curve in MedCalc12.7.8, and the optimal operating points of ROC curve and their correspondent sensibility and specificity are confirmed with Youden index maximum method; AUC(area under the curve) is z tested as a criterion to evaluate the predictive performance of models. P<0.05is considered significant.Deployment:It is expected to use these decision trees models to create a flowchart for critically ill patients and help them decide whether to undergo nasojejunal tube placement. Compliance of healthcare staff and operability of the process can also be evaluated.[Results]1. In CHAID-based decision tree, high success rate of SNT placement for critically ill patients is related to two criteria:(1) if using prokinetic agent and APACHE II score<20then SNT placement will be succeed,(2) if using prokinetic agent, APACHE II score≥20and having EN by nasogastric tube before SNT placement then SNT placement will be succeed.2. In RapidMiner decision tree, high success rate of SNT placement for critically ill patients is related to three criteria:(1) if using prokinetic agent and APACHE II score≤10then SNT placement will be succeed,(2) if using of prokinetic agent (excluding domperidone) and APACHE II score ranged from11to19then SNT placement will be succeed,(3) if using prokinetic agent, APACHE II score≥20, patients with no digestive system disease and having EN before SNT placement then SNT placement will be succeed.3.Logistic regression model indicates that success rate of tube placement is brought down by risk factors including the use of sedations, the disuse of prokinetic agent, high APACHE II score and digestive system disease. It can be noted in the two decision tree models and the logistic regression model that prokinetic agents and APACHE II score are two more important factors that affects the success rate of SNT placement.4.The AUC of the three models are:0.687(95%CI:0.595-0.769) for CHAID-based decision tree model,0.723(95%CI:0.633-0.802) for RapidMiner decision tree model, and0.762(95%CI:0.674-0.835) for Logistic regression model, respectively. By using Youden index maximum method, the correspondent sensibility and specificity of the optimal operating points in ROC curve are: CHAID-based decision tree model:sensibility:80.0%(95%CI:65.4%-90.4%), specificity:60.3%(95%CI:48.1%-71.5%); RapidMiner decision tree model: sensibility:64.4%(95%CI:48.8%-78.1%), specificity:76.7%(95%CI:65.4%-85.8%); Logistic regression model:sensibility:71.1%(95%CI:55.7%-83.6%), specificity:74.0%(95%CI:62.4%-83.5%). The results of AUC z test are as followed:Difference between CHAID-based decision tree model and Logistic regression model is significant (z=2.089, P=0.037); while RapidMiner decision tree model is not statistically different form logistic regression model (z=0.929, P=0.353). Results show that the predictability of the logistic regression model is better than that of the CHAID-based decision tree model; while the RapidMiner decision tree model has the same predictability as the logistic regression model. The AUCs in both RapidMiner decision tree model and logistic regression model are larger than0.700, indicating relatively good predictability; while the AUC of CHAID-based decision tree model is0.687(<0.700), showing a defect in predictability.[Conclusion]This study applies data mining technique to optimize SNT placement selection in critically ill patients. Under the framework of CRISP-DM,609patients receiving bedside pernasal SNT intubation between2005and2013in the intensive care unit of Guangdong General Hospital (GGH),(Guangdong Academy of Medical Sciences, GAM) are included and used to construct decision tree and logistic regression prediction models on the success rate of SNT placement. It is testified that the three models (CHAID-based decision tree, RapidMiner decision tree and logistic regression) are relatively accurate in predicting the success rate of SNT placement. Therefore, Data mining technique proves to be feasible for constructing such models. By analyzing the results of the study, it is found that criteria of successful placement drawn from decision tree models are easier to understand and apply. Compared to the complex regression coefficients and regression equations in logistic regression model, the result of decision tree is presented in an easier way. Hence, decision tree model is much convenient to be applied. With good predictability and convenience, decision tree model is a referable guideline for critically ill patients who actively requires SNT placement. In conclusion, it can be further confirmed from the analysis of the three models that APACHE II score and the use of prokinetic agents are two more important factors affecting the success rate of SNT placement.