Related Experiment Video
Updated: Aug 6, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Integrated machine learning for cause-of-death classification and postmortem interval prediction: Liver and kidney
Jianghuan Lu1,2, Yuzhao Xu1,2, Siqi Chen1,2
1Key Laboratory of Tropical Translational Medicine of Ministry of Education, School of Basic Medical Sciences, Hainan Academy of Medical Sciences, Hainan Medical University, Haikou, China.
Purpose:
To assess whether liver and kidney metabolomics combined with machine learning can distinguish seawater drowning from postmortem submersion after CO2 euthanasia and estimate postmortem interval (PMI) under controlled conditions.
Methods:
Sixty male Sprague-Dawley rats were assigned to a seawater-drowning group or a postmortem-submersion group after CO2 euthanasia (30 per group). Liver and kidney tissues were collected at 0, 12, 24, 36, 48, and 72 h after death and analyzed by untargeted LC-MS/MS. Principal component analysis (PCA) and cross-validated orthogonal partial least squares discriminant analysis (OPLS-DA) were used to characterize global metabolic variation. Random Forest (RF), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), and Gradient Boosting Decision Tree (GBDT) were evaluated using repeated 10-fold cross-validation (five repeats; 50 folds in total). PMI-specific XGBoost models used 20 metabolites selected within each training fold by absolute Spearman correlation and were assessed by repeated 10-fold and leave-one-time-point-out cross-validation. SHAP analysis summarized feature contributions.
Results:
PCA showed that dominant metabolic variation was mainly associated with PMI, with substantial overlap between the two groups in the PC1-PC2 space. Cross-validated OPLS-DA identified group-associated structure in liver and kidney, with Q2(cum) values of 0.847 and 0.723, respectively. Mean classification AUCs ranged from 0.973 to 0.996 in liver and from 0.931 to 1.000 in kidney. Under repeated 10-fold cross-validation, the liver and kidney PMI models achieved MAEs of 4.78 and 4.05 h and R2 values of 0.823 and 0.892, respectively. Under leave-one-time-point-out cross-validation, MAEs increased to 13.52 and 13.64 h, while R2 values were 0.587 and 0.592. SHAP analysis showed partly different metabolite contribution patterns between the two organs.
Conclusion:
In this controlled rat model, liver and kidney metabolomic profiles combined with machine learning supported discrimination between seawater drowning and postmortem submersion after CO2 euthanasia and captured PMI-related metabolic changes. The classification models, PMI-specific XGBoost regression models, and SHAP analyses characterized organ-specific metabolite patterns associated with discrimination between the two modeled conditions and PMI prediction.