A study of grading strategy on postmortem intervals (PMIs) estimation from hypothermia deaths based on non-targeted metabolomics and Machine learning algorithms.

Postmortem intervals (PMIs) estimation in hypothermia has long been a challenge in forensic practice. Recently, many studies have linked postmortem metabolic changes with animal models, providing a potential strategy for estimating PMIs. In the present study, we profiled the non-targeted metabolomics of iliopsoas muscle samples from hypothermia death rats at different PMIs within 48 h by Ultra Performance Liquid Chromatography-Tandem Mass Spectrometry (UPLC-MS/MS). Six different machine learning algorithms, including K-nearest neighbor (KNN), decision tree (DT), support vector machine (SVM), random forest (RF), logistic regression (LR), and Gaussian naive Bayes (GNB), were used based on 48 differential metabolites, which predict the first layer PMIs (0 h, 4-8 h, 12-24 h, 36-48 h). Models were obtained through splitting data, training models, cross-validation, and parameter tuning. The assessment results of prediction models indicate that curtain models, including KNN, LR, SVM, RF, have better performance. The confusion matrices prove three models, KNN, SVM, and LR, only produce one error in testing samples. The second PMIs layer was analyzed using receiver operating characteristic (ROC) curve and statistic analysis based on the differential metabolites, which allowed the identification of biomarkers that could distinguish between the different finer points in time. Finally, a strategy of grading diagnosis is proposed to assess PMIs in cases of death due to hypothermia.