Interpretable machine learning for forensic post-mortem interval prediction: SHapley Additive exPlanations analysis of corneal ATR-FTIR spectral features

The precise estimation of the post-mortem interval (PMI) has always been a core challenge in forensic pathology. Due to the significant environmental interference and the lack of objective quantitative methods of traditional morphological methods, this study collected 130 rat corneal samples and used Attenuated Total Reflectance Fourier-Transform Infrared (ATR-FTIR) spectroscopy technology combined with machine learning algorithms to systematically characterize the spectral evolution characteristics of corneal tissues under different PMIs. The study constructed and compared 10 regression models including ElasticNet, Ridge, Lasso, PLSR, GPR, SVR, GBDT, XGBoost, RF, and MLP. The results showed that the ElasticNet model demonstrated the best prediction performance and generalization ability, with a cross-validation coefficient of determination CV-R² of 0.93, a test set R² of 0.89, and corresponding CV-MAE and Test-MAE of 4.71 h and 5.27 h, respectively. Moreover, to address the “black box” problem of the algorithm, the SHapley Additive exPlanations (SHAP) framework was further introduced for interpretive analysis, and 20 key spectral features were selected. The analysis indicated that the attenuation of the vibration intensity of the carbohydrate skeleton (C–O–C) and the accumulation of characteristic peaks of metabolites were the core bases driving the model's decision-making, which was highly consistent with the biochemical mechanism of “component degradation – product accumulation” in the post-mortem cornea. This study established for the first time a corneal PMI prediction method based on ATR-FTIR spectroscopy and machine learning combined with SHAP interpretive analysis, providing a rapid, non-destructive, and interpretable new approach for the inference of death time in forensic medicine.