D. Puglisi, K. Montelius, E. Tavaglione, S. Tsubota, B. Fabbri, J. Eriksson, I. Shtepliuk
Linköping University,
Sweden
Keywords: MOS sensors, electronic nose, classification models, volatile organic compounds
Summary:
In this work, we propose an innovative dog-inspired multi-sensor platform capable of detecting and classifying unique odor signatures released by human remains as a form of forensic evidence. Our device, equipped with 32 metal oxide semiconductor (MOS) chemiresistive sensor elements, was applied to three different scenarios, demonstrating the potential of its application in forensic medicine. Understanding the volatile organic compound (VOC) profile released during the early post-mortem period (0-72 h interval) is essential for applications in search and rescue operations to rapidly locate victims [1]. Gas chromatography-mass spectrometry (GC-MS), commonly used to analyze odor profiles in the laboratory, requires multiple data processing steps and is not suitable for in situ use and rapid response [1,2]. Specially trained dogs, so called cadaver detection dogs (CDDs), have been successfully used in forensic investigations since 1888 [3] to locate missing people, identify human remains or even detect traces of human decomposition odors. However, although CDDs are considered the most rapid and efficient tool for the search and detection of human remains, from a juridical perspective, their results cannot be used as physical evidence in court. Here, we show that our proposed machine learning (ML) enhanced gas sensor technology is a promising solution that can help investigators overcome existing constraints in the use of CDDs [4-6]. In the first case (CASE I), we analyzed samples from deceased and living individuals to demonstrate the ability of our model to distinguish between post-mortem and ante-mortem samples as a critical factor for determining the time and circumstances of death. In the second case (CASE II), we compared animal (pig) and human samples as a key discriminant in complex forensic scenarios, particularly in the detection and identification of human remains. In the third case (CASE III), we studied the aging process of pig meat to understand the dynamics of organic tissue decomposition over time. The Optimizable Ensemble was found to be the most effective model for all classification tasks. The obtained results show that our model was able to distinguish between post-mortem and ante-mortem human biosamples with highest accuracy of 99.5%, sensitivity of 99.9 %, and specificity of 98.0%. A majority-voting mechanism was implemented to finalize classifier decisions on the status of samples. Furthermore, our model, based on a smartly designed feature matrix, was capable of accurately classifying animal and human samples, achieving 97.2% accuracy. Finally, using pig meat samples as a model, we demonstrate the capacity for precise post-mortem interval (PMI) estimation down to a single day in the early stages of decomposition (1, 2, or 3 days) and to a time range during later stages (4-11 or 18-32 days). The study also introduces a cascading algorithm that integrates three trained classifiers to determine the PMI. This work underscores the promising role of integrating volatilome analysis, advanced ML techniques, and e-nose technology to advance forensic sciences capabilities, and has the potential to be used in other relevant focus areas where binary classification models are suitable.