Intelligent analysis system based on MWS-RF-SDAE:A novel analysis method in forensic science for body fluids and interferents identification by Raman spectra

Raman spectroscopy has emerged as a groundbreaking tool for body fluid analysis in forensic science, providing a unique approach to evidence examination. However, its practical application is often hampered by challenges such as noise interference, baseline drift, and interferents, which can impede accurate identification. To address these limitations, this study introduces an innovative intelligent analysis system that combines Moving Window Smoothing (MWS), Random Forest (RF), and Stacked Denoising Autoencoder (SDAE)—collectively referred to as MWS-RF-SDAE—to significantly improve the recognition accuracy of body fluids and interfering substances. A robust Raman spectroscopy database was established, containing spectra from six types of body fluids and 31 common interfering substances typically encountered at crime scenes. The Backpropagation (BP) classifier was utilized as an evaluation metric to gauge system performance. Among six preprocessing methods tested, MWS was identified as the most effective in enhancing spectral data quality. Furthermore, the study investigated various feature selection techniques, demonstrating that the RF method effectively minimized feature redundancy and increased BP prediction accuracy to 88.4%. Nevertheless, challenges remained in differentiating body fluids from interfering substances due to spectral similarities and instrumental noise. To address these issues, an SDAE prediction model was developed and rigorously assessed using three critical metrics: Accuracy, Recall, and F1 score. The results revealed outstanding performance, with the MWS-RF-SDAE method achieving a prediction accuracy of 99.85%, effectively resolving the issue of low accuracy caused by interfering substances. The system’s practical applicability was further confirmed through six samples in real-world cases, underscoring its reliability for forensic uses. A key feature of the system is its capability to calculate prediction confidence rates for unknown samples, allowing for targeted validation of low-confidence results. This functionality not only streamlines forensic investigations but also offers robust support for crime scene analysis. By integrating these advanced techniques, this study marks a significant advancement toward the intelligentization and automation of forensic science, with the potential to revolutionize the field by enhancing both the accuracy and efficiency of body fluid identification.