An ATR-FTIR spectroscopy strategy coupled with ensemble learning for skin contusion age estimation under a 24 h Postmortem interval.

Accurate estimation of wound age is a critical issue in forensic practice. However, injured tissues in real-world cases are often affected by postmortem changes, which may reduce the reliability of conventional laboratory methods. This study established a skin contusion model in SD rats, collected skin tissue samples under different wound ages (0-48 h) and postmortem intervals (0 h and 24 h), utilized ATR-FTIR to obtain molecular spectral information, and combined it with 10 machine learning algorithms to construct predictive models. The results demonstrated that 24 h postmortem changes affected the spectral characteristics of skin contusion tissues, resulting in decreased predictive accuracy of wound age estimation models. To reduce the interference caused by postmortem alterations, a prediction framework integrating genetic algorithm (GA)-based feature selection and ensemble learning strategies was further developed. The Average ensemble model composed of GA-LightGBM and GA-SVR exhibited satisfactory predictive performance, achieving CV-R2 = 0.89 and CV-RMSE = 5.11 h in five-fold cross-validation, and Test-R2 = 0.85 and Test-RMSE = 5.97 h on the independent test set. SHAP analysis revealed that protein-related features gradually decreased with increasing wound age, whereas lipid- and phosphate-related features gradually increased. Ultimately, this study developed a predictive model capable of mitigating the effects of 24 h postmortem changes and enabling reliable estimation of skin contusion wound age.