Comparative evaluation of metabolomic, microbial, and integrated features for distinguishing a rat model of heatstroke death from cervical-dislocation controls and estimating postmortem interval across organs.

OBJECTIVE: Postmortem diagnosis of heatstroke death and postmortem interval (PMI) estimation remain challenging in forensic practice. This study aimed to evaluate the value of untargeted metabolomics, postmortem microbial profiling, and integrated metabolite-genus features for distinguishing a rat model of heatstroke death from cervical-dislocation controls and estimating PMI across multiple organs.

METHODS: A controlled rat model of heatstroke death (HS) and a cervical-dislocation control group (C) were established. Brain, cerebellum, liver, kidney, and skeletal muscle samples were collected at multiple PMIs and analyzed by untargeted metabolomics and 16S rRNA sequencing. Organ-specific machine learning models were constructed using metabolomic, microbial, and integrated features. For HS/C classification, PMI-matched model-development and external-validation subsets were analyzed using random forest with repeated five-fold cross-validation, whereas PMI regression models were constructed within the HS group using XGBoost. Model performance was evaluated by cross-validation and independent external validation, and feature contribution by random forest importance and SHAP analysis.

RESULTS: Multi-organ metabolomic profiles differed between HS and C and varied with PMI, whereas postmortem microbial profiles mainly reflected organ-dependent and PMI-associated community succession. For classification, aggregated OOF AUCs ranged from 0.971 to 1.000 for metabolomics, 0.837-0.918 for microbial profiling, and 0.986-1.000 for integrated models. Integration provided limited and inconsistent numerical gains over metabolomics alone, and feature-contribution analysis indicated that the discriminative signal was dominated by metabolites. For PMI regression, liver, kidney, and skeletal muscle showed stable external validation performance across their longer available PMI ranges, whereas brain and cerebellum were evaluated over shorter sampling windows. Integrated regression models achieved external validation R 2 values of 0.749-0.979, and SHAP analysis again indicated predominantly metabolomic contributions.

CONCLUSION: Postmortem metabolomic and microbial profiles provided complementary but unequal predictive information in this controlled heatstroke model. Metabolomics was the principal information source for distinguishing HS from cervical-dislocation controls and estimating PMI, whereas microbial profiling mainly reflected PMI-associated succession. Under the present experimental setup, adding 16S-derived genus profiles provided limited additional value for classification and should be regarded mainly as a comparative or auxiliary component rather than a necessary routine layer.