Reliability and Validity of Radiographic Comparisons for Positive Identification

Forensic scientists routinely use radiographic comparisons to generate positive identifications in medicolegal casework. However, identifications are subjective—each practitioner decides individually what anatomical points to use and how many constitute a valid identification. An objective and robust methodological approach to radiographic comparison is required to meet Daubert standards. We propose to address a gap in best practices in forensic anthropology by investigating: 1) observer reliability in radiographic identifications, and 2) predictive accuracy based on anatomical location and number of features. This work will improve best practices for generating positive identifications via radiographic comparisons by expanding sample sizes to produce powered statistical results and identifying recommended minimum experience levels for practitioners. Improving best practices is directly relevant to the daily operations of medicolegal offices by addressing their mission of identifying unknown decedents.

We will generate a large dataset from a contemporary sample of radiographic data via partnership of the University of Nevada, Las Vegas (UNLV); Clark County Office of the Coroner/Medical Examiner (CCOCME); and Forensic Anthropology Center at Texas State University (FACTS). We will use two sources to collect 700 sets of radiographs from diverse anatomical locations: 1) archived case files from CCOCME, and 2) whole-body donations at FACTS. Demographics of the sample will vary but will generally be diverse. We hypothesize that: 1) unique human skeletal morphological variation is sufficient to produce valid personal identifications, and the identification probability can be calculated, regardless of the anatomical location being used; and 2) experience and education affect the reliability of radiographic comparisons. Over two years, eight participants of varying levels of experience and education will each conduct 700 radiographic comparisons of image sets grouped by anatomical location to identify correct matches. Data will be analyzed using a generalized mixed-effects model to identify the factors that are most likely to predict an accurate identification. These data will be used to propose confidence intervals for different anatomical locations based on number of points of concordance and recommended minimum training.  In addition, a multivariate regression tree will be generated to identify which specific points of concordance contribute most to an accurate identification using “spatial” data collected from a custom-built framework on DICOM images. This work will generate a predictive model to assess the objective validity of identifications made from radiographic comparisons.