A Research On Rapid And Accurate Analysis Technology Of Trauma Metabolomics Based On Machine Learning

Purpose:The research aimed to conduct a complete analysis process of trauma clinical data and proton nuclear magnetic resonance(~1H-NMR)metabolomics data based on data science.By a variety of machine learning(ML)algorithms,we created a systematic solution to each step in the whole process including missing data processing,model evaluation,screening biomarkers,data dimensionality reduction and establishing complication prediction models.Materials and methods:For missing data processing,a complete clinical data set of 83 patients who had major abdominal surgery was collected and we generated 5%,10%and 15%data missing randomly.Various interpolation methods were used to missing data processing,including probabilistic principal component analysis(PPCA),multiple interpolation(MI)and so on.Mean square error(MSE)was a metric to evaluate different methods.This process was repeated 100 times to get the mean MSE(m-MSE).Then a retrospective clinical study was conducted.We collected a set clinical data of 117 traumatic brain injury patients with 40 clinical indicators and established 23 survival prediction models.Receiver operating characteristics(ROC),area under ROC curve(AUC),accuracy,sensitivity and specificity were used as performance metrics.After that,we conducted a prospective cohort study involving 50patients who were undergoing major abdominal surgery.Clinical data and ~1H-NMR metabolomics samples were collected before,1 day after and 3 days after operation.To explore the metabolic differences between patients with complications and patients without complications.,support vector machine-recursive feature elimination(SVM-ref)was used to screen biomarkers and clinical data dimensionality reduction.Finally,the postoperative complication prediction model was established based on Cubic SVM.Results:In the study of missing data processing,the m-MSE of PPCA was lowest in 100 tests,which is the optimal methods.In the 23 survival prediction models,AUC of all ML models ranged from 86.3%to 94%.AUC of logistic regression(LR)was83%,and accuracy was 88%.Of these,Cubic SVM,Quadratic SVM,Linear SVM performed significantly better than LR.Finally,there were significant differences in amino acid metabolism and steroid metabolism between patients with or without complications.The AUC of the postoperative complications prediction model based on Cubic SVM was 0.82 and the prediction performance is satisfactory.Conclusion:We constructed a data analysis process that can be used for trauma clinical and metabolomics studies,including PPCA for missing data processing,establishing survival prediction models based on SVM and exploring metabolic mechanism by combining with ~1H-NMR metabolomics.