ATR-FTIR spectroscopic characterization and interpretable machine learning for time since injury estimation in rat...

Accurate estimation of time since injury (TSI) remains a major challenge in forensic science. In this study, attenuated total reflectance Fourier transform infrared (ATR-FTIR) spectroscopy combined with chemometric and machine learning methods was used to characterize the temporal spectral evolution of scabs and to establish quantitative prediction models for TSI. ATR-FTIR spectra were collected from scab samples obtained from Sprague-Dawley rats at different post-injury time points. After spectral preprocessing, principal component analysis (PCA) was performed to evaluate overall spectral variation, and the predictive performance of multiple regression models was compared. Variable importance in projection (VIP) and Shapley additive explanations (SHAP) were further employed to identify the key spectral regions contributing to TSI prediction. The results showed that both protein-related regions and carbohydrate/nucleic acid-related regions exhibited distinct time-dependent changes. Partial least squares regression, used as the baseline model, demonstrated good robustness in the independent external validation set (R²P = 0.900, RMSEP = 15.291 h), whereas Ridge regression achieved the best overall predictive performance (R²CV = 0.922, RMSECV = 14.456 h; R²P = 0.924, RMSEP = 13.341 h). Both VIP and SHAP consistently highlighted the 1540-1520 and 1640-1620 cm^-1 intervals, with additional contributions from phosphate/carbohydrate-associated regions around 1050-1030 cm^-1. These findings demonstrate that ATR-FTIR spectroscopy, integrated with interpretable machine learning, provides a robust and objective analytical strategy for quantitative TSI estimation in forensic practice.