Metabolomic profiling and machine learning application for discriminating fatal hypothermia in a rat model.

Abstract

Cause of death discrimination of fatal hypothermia has been a challenge in forensic practice. In the present study, we profiled the non-targeted metabolomic of iliopsoas muscle samples from rats of the fatal hypothermia group (n = 48), hypoxia group (n = 20), control group(n = 22) by ultra performance liquid chromatography tandem mass spectrometry (UPLC-MS/MS). A total of 63 differential metabolites were identified between the hypothermia group and the hypoxia group, while 60 differential metabolites were identified between the hypothermia group and the control group. Six different machine learning algorithms, including k-nearest neighbor (KNN), decision tree(DT), support vector machine (SVM), random forest (RF), logistic regression(LR), gaussian naive bayes(GNB), were used to compare the efficiency of the prediction methods. We obtained good models through splitting data, training data, and preliminary prediction models screening. The poorly performing models were subjected to cross-validation and parameter tuning, resulting in optimized models. The results of all models employed to test data of between control group and hypothermia group shows that the all models performed well. Compared the hypothermia group and hypoxia group, LR, SVM, KNN,GNB have better prediction ability. In the practical application of forensic investigations into fatal hypothermia-related deaths, any of the top-performing models presented in this study can be considered for future application. Additionally, our screening identified key metabolic biomarkers and pathways, which suggest potential correlative changes at the mRNA and protein levels. This study provides new insights into the determination of causes of death in low-temperature fatalities. The machine learning codes will be made publicly available after the paper is accepted.