Rapid, Non-Destructive Bone Identification with Handheld Near-Infrared Spectroscopy

A recent study published in the journal Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy highlighted how handheld near-infrared (NIR) spectroscopy, combined with artificial neural networks (ANNs), could potentially change the way forensic scientists distinguish between human and animal bones, especially in real-world scenarios involving fragmented or degraded remains (1). This study, which was led by Johannes Dominikus Pallua at the Medizinische Universität Innsbruck, demonstrates that portable spectroscopic tools can deliver high classification accuracy while preserving skeletal integrity.

Handheld instrumentation is becoming increasingly used in forensic analysis, and it has allowed researchers to expedite their analysis on-site. Given that most spectroscopic methods are laboratory-based, this trend is crucial for forensic scientists who need more practical ways to do their work in field settings (2). In their study, the researchers investigated whether recent advances in miniaturized NIR spectroscopy, paired with modern machine learning approaches, could offer a rapid, non-destructive alternative suitable for on-site use.

As part of their experimental procedure, the research team analyzed 225 femoral bone samples, including both human and animal specimens, using a handheld NIR spectrometer. Unlike benchtop instruments, handheld devices typically operate over narrower spectral ranges and at lower resolution, raising concerns about their ability to support reliable species classification (1,2). To address this, the authors generated data sets acquired specifically with compact instruments, rather than relying on data collected from laboratory-scale systems.

Through the use of standard normal variate (SNV) transformation, the spectral data were preprocessed before being analyzed with both binary and multi-class ANN models (1). Both ANN models achieved different results. For the binary model, it demonstrated consistently high performance, with a final evaluation on a strictly held-out test set (representing 10% of all spectra), yielding an accuracy of 96.3% (1). Additional performance metrics were equally robust, including a precision of 97.6%, recall of 94.3%, specificity of 98.1%, and an F1-score of 95.9% (1).

Notably, the correct classification rates observed in the independent test set fell within, or even slightly above, the interquartile ranges established during cross-validation. However, the authors observed marked differences in variability between classes (1). Animal samples were classified more consistently, whereas predictions for human samples showed greater spread across validation runs (1). According to the researchers, this variability likely reflects a combination of factors. Some of these factors included imbalanced sample sizes, differences in post-mortem interval (PMI) distributions among human bones, and the broader species diversity within the animal class (1).

To better understand the underlying spectral structure, the team also applied principal component analysis (PCA). The resulting score plots revealed clear clustering patterns among species and highlighted the influence of post-mortem intervals on human bone spectra (1). These findings underscore the sensitivity of NIR spectroscopy not only to species-specific chemical signatures, but also to time-dependent changes associated with decomposition (1).

Unlike the binary classification model, the multi-class ANN struggled to perform reliably. The recall and F1 scores indicated limited success in resolving closely related species, a limitation with important forensic implications. As an example, cattle samples were frequently confused with human bones, with a median misclassification rate of 45.8% (1). However, it was noted that this issue was far less pronounced in the independent test set, where the corresponding rate dropped to 9.4%, suggesting that broader training data and improved representation may help mitigate such risks (1).

Despite these challenges, the study demonstrates the utility of handheld NIR spectroscopy for typical forensic tool. Its portability, speed, and non-destructive nature make it particularly attractive for on-site investigations, where rapid triage decisions are often required (1). The authors argue that, even in its current form, the binary classification approach is well suited for detecting human bones, which remains the primary forensic priority in many contexts (1).

Looking ahead, Pallua and colleagues outline several avenues for further refinement. They wrote how expanding the spectral range of handheld devices, incorporating a wider variety of species and bone types, and testing samples under different environmental and preservation conditions could all enhance model performance (1).

“Future studies should explore a broader range of species, consider the effects of additional variables such as bone preservation methods, implementation of more diverse bone sources and validate findings using more extensive datasets,” the authors wrote in their study (1).

References

1. Weisleitner, K.; Wöss, C.; Kampik, L. et al. Rapid Forensic Differentiation of Human and Animal Bones Using Handheld Near-infrared Spectroscopy and Deep Learning. Spectrochimica Acta Part A. Mol. Biomol. Spectrosc. 2026, 344 Part 1, 126657. DOI: 10.1016/j.saa.2025.126657
2. Zbrog, M. How Portable Instruments are Changing Forensic Investigations. Forensics Colleges. Available at: https://www.forensicscolleges.com/blog/handheld-devices-in-forensics (accessed 2026-01-30).