Postmortem interval (PMI) estimation is a necessary but often difficult task that must completed during a death investigation. The level of difficulty rises as time since death increases, especially with the case of skeletonized remains (long PMI). While challenging, a reliable PMI estimate may be of great importance for investigative direction and cost-savings (e.g. suspect identification, tailoring missing persons searches, non-forensic remains exclusion). Long PMI can be estimated by assessing changes in the organic content of bone (i.e. collagen), which degrades and is lost as the PMI lengthens. Visible-near infrared (VNIR) spectroscopy is one method that can be used for analyzing organic constituents, including proteins, in solid specimens. A 2013 preliminary investigation using a limited number of human cortical bone samples suggested that VNIR spectroscopy could provide a fast, reliable technique for assessing PMI in human skeletal remains. Clear separation was noted between "forensic" and "archaeological" specimen spectra within the near-infrared (NIR) bands. The goal of this research was to develop reliable multivariate classification models that could assign skeletal remains to appropriate PMI classes (e.g. "forensic" and "non-forensic"), based on NIR spectra collected from human cortical bone. Working with a large set of cortical samples (n=341), absorbance spectra were collected with an ASD/PANalytical LabSpec® 4 full range spectrometer. Sample spectra were then randomly assigned to training and test sets, where training set spectra were used to build internally cross-validated models in Camo Unscrambler® X 10.4; external validations of the models were then performed on test set spectra. Selected model algorithms included soft independent modeling of class analogy (SIMCA), linear discriminant analysis on principal components (LDA-PCA), and partial least squares discriminant analysis (PLSDA); an application of support vector machines on principal components (SVM-PCA) was attempted as well. Multivariate classification models were built using both raw and transformed spectra (standard normal variate, Savitzky-Golay) that were collected from the longitudinally cut cortical surfaces (Set A models) and the superficial cortical surface following light grinding (Set B models). SIMCA models were consistently the poorest performers, as were many of the SVM-PCA models; LDA-PCA models were generally the best performers for these data. Transformed-spectra model classification accuracies were generally the same or lower than corresponding raw spectral models. Set A models out-performed Set B counterparts in most cases; Set B models often yielded lower classification accuracy for older forensic and non-forensic spectra. A limited number of Set B transformed-spectra models out-performed the raw model counterparts, suggesting that these transformations may be removing scattering-related noise, leading to improvements in model accuracy. This study suggests that NIR spectroscopy may represent a reliable technique for assessing the PMI of unknown human skeletal remains. Future work will require identifying new sources of remains with established extended PMI values. Broadening the number of spectra collected from older forensic samples would allow for the determination of how many narrower potential PMI classes can be discriminated within the forensic time-frame.

No description available.