A Novel Analytical Strategy Based On Tandem Machine Learning Model And Metabolites In Wound Age Estimation

ObjectiveIn the present study,differential metabolites which could be used to estimate rat skeletal muscle contusion wound age would be identified with UPLC-HRMS.Besides,a tandem machine learning model for estimating wound age more accurately would be established by comparing the performance of four supervised machine learning algorithms concerning Logistic Regression,support vector machine,random forest and multi-layers perceptron.MethodsTwelve wound age groups were set as 4-hour blocks from 4 hours to 48 hours and one control group was set,13 groups as total and 9 rats per group.Then,UPLC-HRMS was applied for skeletal muscle specimens’ metabolomic profile detection.VIP > 1 with and FDR-adjusted P < 0.05 were combined for differential variables selection.Metabolite identification for differential variables was based on accurate mass and mass fragmentation pattern spectra against MS-MS spectra of metabolites available on mz Cloud database.The metabolites identified above were applied to construct a series of machine learning models.At first,the first-layer time period labels were determined by OPLS-DA and then four machine learning algorithms were established and compared with cross-validation,and the model with the highest accuracy was selected as the first-layer model.Then the secondlayer time periods labels were determined with OPLS-DA based on the first-layer labels.After cross-validation and model comparison,the models with the highest accuracy were also selected as second-layer models.At last,we constructed tandem model with python coding and tested its performance with another 13 unseen samples.ResultsWith metabolomic analysis,43 metabolites varing with wound age were identified in the present study,and their biological functions focus on primary bile acid biosynthesis,aminoacyl-t RNA biosynthesis,phenylalanine,tyrosine and tryptophan biosynthesis,biosynthesis of unsaturated fatty acids,and arachidonic acid metabolism,etc.As a result,the first-level model discriminating wound age into 4-12h(group Ⅰ),16-32h(group Ⅱ)and 36-48h(group Ⅲ)consisted of one MLP model and its validation accuracy was 93.9%.The second-level model consisted of three MLP models discriminating group Ⅰ into 4h,8h,and 12 h,group Ⅱ into 16-20 h and 24-32 h and group Ⅲ into 36-40 h and 44-48 h group and their validation accuracy was 100%,92.8%,and 90.9%,respectively.Finally,we constructed tandem model with python coding and tested its performance with another 13 unseen samples and the test accuracy was 61.5%.ConclusionIn the present study,43 metabolites were identified using UPLC-HRMS analysis,and a tandem machine learning model of MLP-MLP was established to refer the interval time.The accuracy for external-validation data was 61.5%.The time window of wound age estimation narrowed into a minimum of 4 hours and a maximum of 12 hours and demonstrating a useful measure for future practical work of wound age estimation.