FTIR Technique Combined With Machine Learning Algorithms For The Identification Of Electrocution Death And Post-mortem Electrocution In Rats

Objective:The use of Fourier Transform Infrared Spectroscopy（FTIR）combined with machine learning（ML）algorithm for feature selection and classification model recognition,and the study of the biochemical differences of cardiac and lung tissues of rat electrocution death and post-mortem electrocution are expected to provide objective diagnosis as a supplementary tool in forensic identification.Methods:1.Animal experiment:80 SPF-grade healthy Sprague-Dawley（SD）rats,male and female,weighing 220g±10g,8 weeks old,were randomly divided into electrocution death group,immediate post-mortem electrocution group,30min post-mortem electrocution group and control group,20 rats in each group.In the electrocution death group,the rats were anaesthetized and electroshocked with the metal clips of the homemade electroshock device attached to the left forelimb and right hindlimb,and electroshocked with the power supply（220V,50Hz）for 2 min;in the immediate post-mortem electrocution group,The rats were anaesthetised and killed by breaking their necks and then electrocuted for 2 min in the same way as in the electrocution death group;in the 30 min post-mortem electrocution group,the rats in the electroshock group were anaesthetised and killed by breaking their necks 30 min after death and then electrocuted for 2 min after death in the same way as in the electrocution death group;in the control group,the rats were anaesthetized and executed by breaking their necks 30min after death.All groups were dissected 30 min after death and cardiac and lung tissues were removed.2.Hematoxylin-Eosin（HE）staining:The histopathological changes of the heart and lung were observed by HE staining in each group of rats.3.Routine data pre-processing:FTIR spectra were collected and pre-processed with smoothing and noise reduction,Standard Normal Variation（SNV）and first-order derivation.4.Feature selection:Three algorithms are established for the routine preprocessing data,namely,the Relief algorithm,the Random Forest（RF）algorithm,and the Support Vector Machine Recursive Feature Elimination（SVM-RFE）algorithm.5.Establishment of machine learning classification models:Establish five models:Back Propagation（BP）neural networks,Extreme Learning Machine（ELM）,Support Vector Machine（SVM）,Naive Bayes Model（NBM）,and RF to carry out classification and recognition work on samples that have not undergone feature selection and processed by three feature selection methods.6.Biochemical analysis:The feature selection method with the highest recognition rate of the classifier is used to intercept the number of feature waves and analyse the biochemical information.Results:1.Animal modeling results:The rats in the electrocution death group had erected hair,tense muscles and tremors in the trunk and extremities after the shock.The rats in the immediate post-mortem electrocution group still showed muscle spasm and tremors after electrocution,but the symptoms were less severe than those in the electrocution death group.The rats in the 30min post-mortem electrocution group showed muscle tension and contraction in the limbs after electroshock,which were less severe than those in the electrocution death and immediate post-mortem groups.2.HE staining results:Cardiac tissue:No pathological changes were seen in the control group;myocardial fibers in the electrocution death group were disordered and wavy;myocardial cells in the electrocution death group showed vague and irregular transverse structures;the cytoplasm showed intense eosinophilia,and the nuclei were fixed;some of the myocardial cells in the immediate post-mortem electrocution group showed vague transverse structures,some of the cytoplasms showed intense eosinophilia,and some of the nuclei were fixed.An increase in cell eosinophilia was occasionally seen in the 30 min post-mortem electrocution group.Lung tissue:No pathological changes were seen in the control group.In the electrocution death group,the alveolar structure was significantly damaged,with thickening,widening and extensive rupture of the alveolar wall,extensive haemorrhage in the alveolar cavity and dilated capillaries.In the immediate post-mortem electrocution group,some of the alveolar walls were ruptured,thickened and widened,and the capillaries in the alveolar wall were dilated and bruised.Some of the alveolar walls were thickened and widened in the 30 min post-mortem electrocution group.3.Feature selection results:The experimental results show that the classification accuracy of both cardiac and lung tissue spectral samples after feature selection is much higher than that of the original samples without feature selection,and the RF feature selection algorithm has the highest classification effect overall.The RF feature selection algorithm has the highest classification accuracy of 92.44%and 96.32%for heart and lung tissues,respectively,after the NBM model.The results of feature selection of sample wavelengths by the RF algorithm showed that the cardiac tissue variables that contributed more to the differentiation of the electrocution death group from the immediate post-mortem electrocution group,the 30 min post-mortem electrocution group and the control group were mainly distributed around 1066cm^-1,1231 cm^-1,1456 cm^-1,1535 cm^-1,1656 cm^-1,and the lung tissue variables were mainly distributed around 966 cm^-1,1078cm^-1,1234cm^-1,1542cm^-1,1654cm^-1.Among the variables in cardiac tissue,the wave number absorption peaks of 1066 cm^-1,1231 cm^-1,1456 cm^-1,1535cm^-1,and 1656 cm^-1 were higher in the electrocution death group compared to the immediate postmortem electrocution group,30 min postmortem electrocution group and the control group,respectively,with statistically significant differences（P<0.05）.Among the variables of lung tissue,the wave number absorption peaks of 966 cm^-1,1078cm^-1,1234cm^-1,1542cm^-1,and1654cm^-1 were elevated in the electrocution death group compared to the immediate post-mortem electrocution group,the 30 min postmortem electrocution group and the control group,respectively,and the differences were statistically significant（P<0.05）.4.The results of the biocomponent analysis:the variables that contributed more to the differentiation of the electrocution death group from the immediate postmortem electrocution group,the 30 min postmortem electrocution group and the control group were mainly distributed around 1066 cm^-1,1231 cm^-1,1456 cm^-1,1535 cm^-1,1656 cm^-1,indicating that electric shock caused changes in the cardiac tissue nucleic acids,amidases I and II and other relevant biological components were changed.The variables in lung tissue were mainly distributed around 966 cm^-1,1078 cm^-1,1234 cm^-1,1542 cm^-1,and 1654 cm^-1,indicating that the electric shock caused protein phosphorylation in lung tissue and changes in biological components related to nucleic acids,amidases I and II.Conclusion:1.The FTIR technique combined with machine learning methods can visualise the biochemical differences between rat cardiac and lung tissues,thus providing new ideas and methods for the identification of electrocution death and post-mortem electrocution.2.The feature selection algorithm improved the classification and recognition effect,and the best recognition effect was achieved by the RF algorithm in this study.Among the five machine learning models built after feature selection based on this algorithm,the NBM model had the highest accuracy in distinguishing electrocution death from post-mortem electrocution.