Forensic Anthropology & Forensic Pathology | Session 2 from the 2026 NIJ Forensic Science Research and Development Webinar Series

Rachel Wendt: Good afternoon and welcome to the 2026 NIJ Forensic Science Research and Development Webinar Series. We are excited for you to join us. My name is Rachel Wendt and I will be moderating today's session. I am a program manager at the National Institute of Justice, or NIJ. NIJ is the research, evaluation and technology agency of the US Department of Justice and is dedicated to improving knowledge and understanding of crime and justice issues through science.

Today's webinar, Forensic Anthropology and Forensic Pathology will explore research advancing human identification, skeletal trauma analysis, and death investigation. Presentations will feature innovations in skeletal analysis, decomposition studies, time since death estimation, and multidisciplinary approaches to complex medicolegal investigations.  

Next slide.

Before we begin, let's review some important information. To adjust your audio settings, click on the down arrow button next to the grayed-out microphone to adjust as needed. Questions can be submitted at any time during the presentations. There will be a Q&A session after each presentation. To submit your questions, click the Q&A icon at the top of your screen.

Closed captioning is available. To access, click the more button with three dots. Select Language and Speech and then select Show Live Captions. A recording of this webinar will be available at a later date on NIJ's website. Certificates of attendance are not being provided for today's event.  

With that, I would like to introduce today's presenters. Rachel Buckley is a PhD candidate in mechanical engineering at the University of Utah. Dr. Andrea Zurek-Ost is a postdoctoral scholar in the Department of Anthropology at Michigan State University. Dr. Jessica Metcalf is a professor in the Department of Animal Sciences at Colorado State University. Dr. Christopher Godin is an assistant professor in the Department of Anthropology and Middle Eastern Cultures at Mississippi State University. Again, there will be a Q&A session after each presentation, but we may not be able to address all questions. We appreciate your understanding, and we'll do our best to respond to as many questions as possible.  

Next slide.

And with that let's hear from Rachel Buckley.

Rachel Buckley: Okay. Thank you all for being here today. I will be talking about collagen damage in the retinal vasculature from failure and subfailure loading. To break that down for you a little bit further, I'll be talking about collagen molecules, which are a structural component of blood vessels. And I'll be talking specifically about blood vessels in the retina and how they're damaged during different amounts of loading on the retina. And I'll be talking about this specifically in the context of retinal hemorrhage in pediatric eyes. And so here on the screen you can see two images of adult retinas. On the left is a normal uninjured retina and on the right is an injured retina with several different retinal hemorrhages. And the black arrows just indicate different types of retinal hemorrhages, differentiated by size and shape and perhaps even depth within the retina.

Next slide.  

First, though, I'd like to just talk about the retina in general. It's the innermost layer of the eye, sandwiched between the choroidal scleral complex and the vitreous. The vitreous is that gel-like substance that forms the inside of the eye, and the retina itself is formed of many different complex cellular layers that help send sight information to the brain.

But today, we're most interested in the highly vascularized layers that are susceptible to injury in the retina. Next. It just so happens that in pediatric eyes specifically, the retinal hemorrhage can occur in up to 78% of abusive head trauma cases, but also in up to 15% of accidental head trauma cases. So in regards to retinal hemorrhage, it can be difficult to discern between abusive and accidental head trauma cases. And this is exacerbated by a lack of reliable witnesses, but also by a lack of data, experimental data linking forces during head trauma to damage to the retinal vasculature. Next.

And so, we really want to ask the question exactly what forces are involved in head trauma and how do they cause failure to the retinal vasculature that leads to these hemorrhages? One dominant theory in the field has arisen, and that's the theory of traction forces causing failure in the vasculature.  

Next slide.  

So, to illustrate traction forces, if you imagine a head experiencing repetitive back and forth motion, the eye would experience this as well. And perhaps that vitreous gel-like substance would pull on the retina in a perpendicular motion, making these traction forces and transmitting them to the retinal vasculature. And perhaps this would damage the weak points in the retinal vasculature,. perhaps branch points, and this might cause the bleeding seen inretinal hemorrhage. Next. And so, this though is still an untested, experimentally untested theory. And we really would love to dive into testing traction forces. But they're quite complex and difficult to do right away. And we want to also define failure mechanics in the retinal vasculature. So, in order to tackle this complex problem, we really need to approach it from a much more fundamental perspective. And so, from a mechanics perspective this fundamental approach is going to be tensile tests.

And so, we will subject the retina to tensile forces to really seek to define failure mechanics in the retinal vasculature and provide a foundation for how the retinal vessels are damaged due to forces, so that future research can build on that and understand how traction forces contribute to damage the retinal vasculature to provide a launching point for future research about retinal hemorrhage and pediatric eyes.

Next slide.

And so, one of the main tools that we will use to define failure in the retinal vasculature are collagen hybridizing peptides, a really unique tool. So, in case you're unfamiliar, collagen molecules present in collagenous tissues like tendon and ligament, but also all blood vessels in the body, have three triple helices, and they're organized like the three strands of a rope, as seen in the bottom left.

Now, when any collagenous tissue becomes damaged due to injury or illness, they start to unravel like the three strands of a rope, and CHPs have been uniquely engineered to mimic the triple helical structure of collagen so that they don't bind to an intact collagen molecule, but do bind to an unraveled collagen molecule. Next. So that if we attach a fluorophore to CHPs, we can report on where collagen damage has occurred. Next.

So, CHPs have been used in a wide variety of tissues, including sheep cerebral blood vessels. So, here you can actually see an image of an unstretched, uninjured sheep cerebral blood vessel. There is like a minimal amount of CHP staining in this control image, indicating that there is some collagen turnover just naturally present in a healthy tissue. Next. But with increasing amounts of overstretch of the tissue and mechanical damage, we see an increasing amount of CHP staining, indicating an increasing amount of collagen damage.

And so, we hope to see a similar trend in our blood vessels in the retina. But before our study, CHPs hadn't yet been used in the retina at all.  

Next slide.  

So with that, we'll take a look at the objective of the study, which was to develop a method to evaluate damage in the immature sheep retinal microvasculature. Now of course, we're primarily interested in the pediatric eye, but it is actually quite difficult to obtain immature sheep eyes and in the future, it will be very difficult to obtain pediatric human eyes. And so, our first key question is actually to ask, can we use adult sheep eyes as a substitute or alongside immature sheep eyes? Next, and we will look at this in the context of both our mechanics and our collagen damage metrics. Next, our second key question for the study is going to be can CHP be used in the retinal microvasculature in general? Will it work? And if it does, can it be used in a more sensitive way to detect a difference between failure, catastrophic failure and more subfailure levels of damage?  

Next slide.  

So, to do this, we took eyes from on campus sheep studies, and we carefully excised the retina, and we took two samples from each retina, and those remained paired for the rest of our analysis. Each sample was taken from a major visible retinal blood vessel, and one sample remained unstretched as a control, and the other sample was stretched in uniaxial tensile testing. That sample that was going to be stretched was safely transferred to our materials testing system.  

Next slide.  

It was stretched at two millimeters per minute either to failure, which was complete separation of the tissue, or to varying displacements prior to failure, which were different subfailure strains. And this was before any separation of the tissue.  

Next slide. Sorry. Next slide.  

And immediately after testing, all damaged and control samples were stained with CHP conjugated with a green dye. Then they were mounted on slides with a cover slip. And so here is a chart showing exactly how many samples we used in this study. And I would like to point out that we did not do any subfailure testing on our immature eyes. And this is because we really wanted to focus on getting as high of a quality of comparison at the failure level between our adult and our immature eyes as possible. Because like I shared before, immature eyes were a scarce resource for this study.  

Next slide.  

And so after mounting, we were able to do confocal imaging where we took multiple images of each sample through the thickness, obtaining a Z stack in different locations.

Next slide.  

So, to analyze these images, it was quite a complex process to quantify the collagen damage from these Z stacks. So the first step was to compress the Z stacks using a maximum pixel projection. Next. And then we segmented out the background leaving only the primary vessel. And so, in order to tell from the tested images exactly how much of that green signal is collagen damage, and not just natural collagen turnover like you might see in the control image, we developed a threshold based on the control.

Next. This threshold was two and a half times the mean pixel intensity. And when applied to the tested image, you can see that any pixel that was higher in value than the threshold is labeled in red. And so initially we were just going to pick out one image that represented the tested sample and whichever one was the brightest to represent the worst-case scenario of damage in this sample.

But we noticed an interesting trend, and that was that even if one image wasn't necessarily the brightest, there were some images that had a lot of staining, and then there were some images that had not very many pixels, but perhaps they were really bright. So, we asked the following questions. Next. Which images had the most pixels above the threshold regardless of brightness?

Which images had the brightest pixels above the threshold regardless of how many pixels pass the threshold? And then which images didn't have the brightest or the most pixels? And based on these questions, next, we developed two metrics to measure collagen damage instead of just one. A quantity metric and a brightness metric. Next, our quantity metric was pixel density, or the number of pixels that surpassed the threshold divided by the number of pixels in that image that were segmented.

And then our brightness metric was the normalized mean pixel intensity, or the average value of the pixels that surpass the threshold divided by the value of the threshold itself. So, these are just two ways of saying how many pixels are there and how bright are those pixels.  

Next slide.  

So now I'll talk through how we analyzed the mechanics.

Of course, we had our raw force and displacement values, but we can't necessarily compare those from sample to sample because they can tend to vary based on geometry. So, to measure geometry, next, we took a look at our gauge length or the grip-to-grip length of the sample and we took a look at our vessel cross-sectional area, estimated from the width of any vessel that spanned the grip-to-grip length of our samples. Next. So, taking a look at our vessel cross-sectional areas, we saw that of course in our adult group they didn't differ between those that were stretched to failure and those that were stretched to subfailure displacements. However, as expected in our immature eyes, they tended to be smaller vessels compared to our adult eyes. So this will come into play when we normalize our forces by our vessel cross-sectional areas.

In addition, we’ll, so when we normalize forces by cross-sectional areas, we get stress. And then we normalize our displacements by that gauge length to obtain strain. Next. And so, our peak stress is the maximum stress of the test regardless of whether it went to failure or subfailure. The peak strain is the strain at which that peak stress occurred, and then our maximum modulus is the maximum linear slope of that stress-strain curve. So now understanding how we got our metrics we can go ahead and take a look at the results.  

Next slide.  

So within our failure group between our immature eyes comparing to our adult eyes we're reminded that the vessels were smaller in the immature eyes, which resulted in a peak stress that was similar between adult eyes and immature eyes.

Similarly, there were no statistical differences between adult and immature eyes for our maximum modulus, which was that slope, or for our peak strains between adult and immature eyes. So, we conclude that immature eyes are similar to adult eyes for mechanics. Now for CHP results. Next. We can take a look at some representative CHP images, and we don't really see much difference here. Next. And the data really tell the same story. So, for our normalized mean pixel intensity or our brightness metric, there were no statistical significances, no statistical statistically significant differences between our adult and our immature group. However, there was an increase in pixel density between adults and our immature group. Next. But when I took a look at these three samples that were really high in our immature group, I noticed something interesting.

In all of our immature samples, there was really bright background staining from CHP, and that was largely taken care of by segmenting out that background in our initial analysis step. But for these three samples, that bright background was also behind our large vessels and bled through when we took the maximum pixel projection. And so, we believe that this artificially increased the number of pixels that pass the threshold and artificially increased the pixel density for these three samples.

And so, while this is a statistically significant finding, we don't believe that this is necessarily a meaningful difference between the adult and immature eyes for our pixel density. So, we still concluded that immature eyes are similar to adult eyes in terms of our CHP metrics, particularly with regards to brightness.  

Next slide.

So now we can take a look at our adult group and when we compare our failure and our subfailure mechanical results, we're really just looking here to establish the fact that, hey, we designed our subfailure tests well. Our peak stresses were lower, our peak strains were lower, which means that our subfailure group experienced lower amounts of mechanical stresses and strains, which is what we wanted. You will notice that the mechanical modulus does not differ between failure and subfailure groups and all this means is that for our subfailure group, we did reach the linear region of the stress-strain curve. And this isn't necessarily a good or a bad thing. It just means that we reached a certain part of the test for those samples. Next. So, looking at some representative images for our CHP results. Next. We can compare that to the data and see that there was a reduction in the normalized mean pixel intensity and the pixel density.

So, both our brightness and our quantity metrics between failure and subfailure groups. And this is a really exciting outcome for two reasons. One, we were able to obtain some signal in CHP samples. It was a possible outcome that we weren't able to see any staining in our CHP groups, because either the CHP molecules didn't accumulate enough, or perhaps we didn't damage our samples enough.

And so, this was exciting to achieve any staining at all. But it was also exciting that we were able to have metrics that were able to discern a difference between our failure and our subfailure amounts of damage. It's possible that, you know, with some amount of mechanical damage to the samples, damage is damage, and you can't necessarily tell the difference, but these metrics tell a different story, that you can differentiate between failure and subfailure amounts of mechanical damage to the collagen molecules.

And so, this really bodes well for the future. Perhaps in the future, we can try to correlate different amounts of collagen damage to different mechanical metrics. And we hope to do this in our future analyses.  

Next slide.  

So now we can go ahead and take a look at our key questions. When we asked if adult sheep eyes can be used as a substitute for immature lamb eyes, next, we saw that mechanical outcomes were not different. Next. We saw that there weren't any meaningful differences in collagen damage. So, we conclude that yes, adult sheep eyes can be used alongside immature lamb eyes. Next. We also asked if CHP can be used to identify damage in the retinal microvasculature. Next. Not only can CHP be used in general, but it can also be sensitive enough to determine a difference between failure and subfailure amounts of damage, which is very exciting.

Next slide.  

So now we're left with some data linking retinal mechanics to pediatric head trauma.  

Next.  

In our future directions can help us launch into traction testing. Perhaps we'll even look at repetitive overload to see if an accumulation of damage occurs with cyclic overload in the retina, and this will help us someday move from benchtop to animal models, and even into the future to help us establish a foundation to understand how head trauma leads to damage in the retinal vasculature, which can lead to retinal hemorrhages.  

Next slide.

So with that, I would like to thank my lab. Special thanks to the labs on campus that donated eyes to the study and thanks to the National Institute of Justice for funding this work.

And with that, I'll take any questions.

Rachel Wendt: Thank you. Rachel. Go ahead and submit your questions and you can continue to submit them throughout the Q&A session. I'll start with a question about applicability of this staining technique. Is there potential for it to be used in other capacities, like diagnostics of different types of disease and injury, or perhaps for other types of chronic trauma, like athletes who are dealing with CTE?

Rachel Buckley: So, yes. CHPs are widely applicable as long as there is collagen in the tissue that is of interest, then it's applicable. And as long as you can have, you know ex vivo postmortem tissue. So, it's main limitation is that it cannot currently be used in vivo though the cutting edge of research with CHPs is perhaps currently in that stage. But as long as you can excise your tissue of choice and as long as it has enough collagen, then you can look and see if those collagen molecules have been damaged. It can be, you know, if you're looking at osteoarthritis, if you're looking at rheumatoid arthritis or other systemic diseases, even if you're looking at, you know, traumatic brain injury, perhaps you're looking at the collagen and the trabeculae of the brain you could perhaps apply it there. You asked specifically about CTE. I'm not familiar with that. But if you tell me a little bit more about that, I could speak to that as well, perhaps.  

Rachel Wendt: Yeah, I honestly forget what the acronym stands for, but it's a brain disease that they're finding in a lot of NFL players and athletes, just from that really chronic head trauma that they are experiencing.

So, I think you addressed that with your answer in general.  

Rachel Buckley: Yeah. Yeah. It would depend on the area of the brain that's being affected and whether or not it has collagen. And again, like it depends on how the collagen is being affected. If the disease doesn't affect the collagen, CHP would actually help you tell that story. If you don't see staining, then that mechanism for that disease likely doesn't affect the collagen.

Rachel Wendt: Got it. Thank you. And why were sheep eyes chosen as a model for representing human eyes?  

Rachel Buckley: Yeah. So, sheep eyes and pig eyes are actually both good choices to represent human eyes because they have a very similar structure in the retina and in the vitreous. And actually, the interface between the vitreous and the retina is also important in these studies and that is also similar to humans. Of course, primates are also a very good choice. But sheep and pigs both are better choices in terms of just being conservative with our research, sheep are a slightly better choice than pigs in this case, because pigs tend to have retinal blood vessels that sit on top of the retina instead of more within the retina like humans and so sheep tend to have those more embedded retinal blood vessels like humans.

Rachel Wendt: Great. And our last question, what's the future trajectory for your research?  

Rachel Buckley: Yeah. So, I'm really excited that I've begun doing some of that traction testing, because we were able to establish that foundation of collagen damage from the very simple tensile tests. I'm able to translate that knowledge into the more physiological relevant traction tests that are perpendicular to the retina and even do that with some cyclic overload.

So, looking at if you have multiple applications of subfailure levels of forces, does that introduce an accumulation of damage, which of course is applicable if you have maybe a shaken baby scenario

Rachel Wendt: Great. Well thank you Rachel. Yeah. Pleasure. We will now hear from Dr. Andrea Zurek-Ost.

Dr. Andrea Zurek-Ost: All right. Good afternoon, everyone. My name is Andrea Zurek-Ost, and I'm a postdoctoral scholar in the anthropology department at Michigan State University, working on the NIJ-funded initiative MOSAIC, which I have the privilege of speaking about today. MOSAIC stands for unifying methods of sex, stature, affinity, and age for identification through computational standardization. And I just want to up top acknowledge the MOSAIC team.

Our various institutions are shown here on the slide, as this is very much a collaborative and multi-institutional effort. And so, the goal of MOSAIC is to model the biological profile, age, sex, stature and population affinity in a way that reflects human skeletal variation. So as an integrated and standardized multivariate system. And the ultimate goal with the development of MOSAIC is to improve the identification process for unidentified individuals in United States medicolegal context.

This project is ongoing as a multi-year grant, so I'll be presenting preliminary work here today, including the rationale behind the creation of MOSAIC, our data collection plan, and our process to develop a large matched skeletal data set and the use of Bayesian network analysis in modeling biological variables. And then I'll talk about some of the plans we have for the future.

Next slide please.

So, I want to start with a quick review of the biological profile. As I'm sure most folks here are very familiar with, identification efforts in forensic anthropology rely on estimating four core components of the biological profile. Age, sex, stature, and population affinity. And these biological profile parameters structure how we interpret skeletal remains as forensic anthropologists, how we narrow missing persons lists, and how we communicate conclusions to medicolegal stakeholders.

The estimates of these biological profile parameters are foundational to casework, but they are also probabilistic, interdependent, and sensitive to both biological and methodological assumptions. And biologically, there are relationships between them. So, for example, we know that sexual dimorphism varies across populations. An analyst often doesn't think about every skeletal observation in complete isolation. So, we contextualize our observations based on what we know about individual, and more broadly, skeletal variation.

However, there is little formalized guidance on how to incorporate these relationships into our analytical and reporting frameworks.  

Next slide.

Our traditional analytical approaches tend to treat biological profile parameters as independent factors, the methods for which are typically developed, validated, and applied in isolation. So, in a traditional workflow, we might estimate sex from the pelvis, stature from the long bones, population affinity from the cranium, and age from morphological features across the skeleton.  

Each of these methods has its own analytical assumptions, error structures, and reference samples, which then must be distilled by the analyst into the description of a single individual. And in that process, we often come up with a set of estimates that is internally inconsistent and difficult to reconcile probabilistically. And further, when these estimates don't agree, there isn't a robust, standardized framework for determining how they should be weighted or reconciled by the analyst.  

Next slide.  

So, all of this was the impetus for MOSAIC. MOSAIC approaches the creation of the biological profile by moving the integration of those disparate components into the analytical model itself, rather than leaving that integration primarily to the analyst at the end of the process.

And part of what we wanted to investigate with this grant isn't just, for example, which variables or traits are best for estimating sex, we also wanted to evaluate how observations from different anatomical regions may relate to one another across biological profile parameters. So, for that, we're applying machine learning and artificial intelligence in a way that can handle complex multivariate biological data. And MOSAIC allows us to model nonlinear relationships, integrate multiple data types, and also explicitly evaluate uncertainty across biological parameters. Our plan is for MOSAIC to be a dynamic, modular program where, in the future, we can implement not just more data, but also expand the data categories that are represented.

Next slide.

MOSAIC is built around three primary components. So, the first is the matched skeletal data set, which provides the empirical foundation for examining relationships between different types of skeletal observations.

So, by matched I mean that we have observations for all four parameters of the biological profile for the same individuals, rather than having separate data sets for sex, stature, age, and population affinity. And the second primary component of MOSAIC is the Bayesian network and model, which provides the analytical framework for modeling those relationships. And the third component of MOSAIC is the biological profile that results from that framework, where observations are used to generate estimations for age, skeletal sex, stature, and population affinity. And we want the creation of those biological profile estimates to be transparent and testable, and also user-friendly for practitioners and defensible in a post-Daubert world.  

Next slide please.

So, as you can imagine, this is a big undertaking and it requires a lot of coordination between 12 project personnel across five universities. We are fortunate in that our PIs for this project are all board-certified forensic anthropologists and experts in various skeletal and dental regions that we're utilizing from MOSAIC. And we saw it as crucial for the foundations of this program that we brought everyone together to ensure that we were all operating at the same baseline.

So about one year ago, at the inception of the grant, the MOSAIC team convened for our project onboarding at Texas State University for two weeks of intensive training. And during that time, we laid the foundations for our data collection procedures and standardized our recording and scoring protocols. And this period also included assessments of inter-observer error between all analysts with regard to our scoring and recording so that we had a measure of quality control for our data.

So, if we identified during that training period that there was a specific trait or method that we as a team had lower agreement on, we revisited training in that area and then reevaluated our agreement scores. And since that time, our team has taken those skills from the onboarding session into the phase we’re in the process of finishing up right now, the data collection phase.

So really, the bulk of the work we've done so far has been devoted to assembling the large matched skeletal data set, so that's mostly what I'm going to be discussing today.  

Next slide please.  

To date, our efforts have focused on establishing both proficient and consistent data collection across sites. So, our team has now traveled for the better part of a year across the United States to visit eight documented and donated skeletal collections.

And to date, we have collected data for 379 skeletal variables from 2,467 individuals, and we still have a couple of data collection trips planned for later this year. Our goal is to have all of these skeletal variables from across the body represented for the same individuals in this matched data set. So, on this slide, you'll see the types of traits we're currently collecting and you'll see that one thing we are not trying to do with MOSAIC is reinvent the wheel. We know that there are well-established and well-validated biological profile estimation methods that exist. So rather than create a new massive suite of skeletal traits to collect, we want to integrate preexisting traits into the model. And you'll see here that the data encompass both metric and morphological variables from across the entire skeleton, including cranial, postcranial, and dental traits.

Next slide.  

And while this is a fairly large data set, its value really lies in the fact that all of these observations are matched. And this again provides that empirical foundation for investigating how these variables relate to one another. So, you can see here on the slide the collections we visited so far. We've focused our efforts not just on which skeletal collections have the largest numbers of individuals, but also with an attempt to achieve good demographic coverage across the United States.

So, we have, for example, individuals from the Midwest, the Southeast, the Southwest, and also individuals from Hawai’i in our sample to account for skeletal variation in different demographic groups. And we've also tried to balance our sample in terms of males and females, and also to collect data from young as well as older skeletal donors. While the MOSAIC team has been responsible for data collection at these locations, we also have data that's been contributed to the grant.

So, most of our contributed data comes from our PIs, but we also have colleagues who have contributed data who've been trained by the MOSAIC project PIs. And we hope to continue to expand our sample and our coverage across these variables and the biological profile parameters with collaborators in the future, so we can better reflect the overall demographic composition of the United States and ultimately better fit our forensic casework.

Next slide please.

So, to accommodate these types of data where we have both continuous, such as stature and categorical, such as sex variables, and also where we have relationships between many variables rather than simply one predictor and one outcome. We are using a Bayesian network analytical framework, which you can see an example of here. So, this is being built by Mike Kenyhercz, our project biostatistician. And at its most basic level, a Bayesian network is a graphical model in which nodes represent variables and edges represent conditional relationships between those variables. So, MOSAIC will use Bayes theorem to update the probability of the target biological profile as different skeletal observations are entered and the network structure for the current proof of concept was informed by our biological knowledge while the parameters of the model were estimated using matched data. So, that included for this version of the model, morphology, craniometrics and age factors. And the idea is to provide a formalized framework for relationships that we already understand, at least conceptually, to exist within human skeletal variation, while also allowing those relationships to be quantified through probabilistic inference.

One major goal of the project is for this to not function as a black box, so we want transparency and reporting to be key parts of the analytical outcomes so that MOSAIC can be confidently applied to case reports.  

Next slide please.

So, what you're seeing on this slide for example, is just a representation of some of the variables we've collected and the initial connections between them. The next phase of our work is to understand these relationships in the context of what we know about skeletal variation and biology. And again, to produce estimates of all four components of the biological profile that are conditioned on the same underlying set of observations and relationships within the network.

So, the network that exists right now is based on a small subset of the sample and the variables and it was created with 335 individuals and 224 variables. However, the sample in that initial creation stage lacked some coverage, so it was mostly older individuals who had a Euro American background, and that was prior to any of our big data collection trips.

So, this that I'm showing you now reflects the initial stage of development, and we don't yet have finalized predictive outputs or a complete algorithm or model. But that said, just last week we finished a big push for data collection. So, we're going to be incorporating over 2,000 more individuals into that original model and recalibrating it. And then we'll be expanding our 224 variables to 379 total variables.

Next slide please.  

So, what will MOSAIC actually look like for a practitioner? The idea with MOSAIC is that the program will function as both a data collection interface, where observations are recorded and stored, and also as an analytical interface where data are incorporated into the model. So, we are working on the creation of a user manual and a trait guide that would assist practitioners in assessing a skeleton.

So, practitioners would, with the help of this user manual, collect data from the entire skeleton that includes skeletal and dental traits, both metric and morphological. And one of the goals with MOSAIC is for it to be both accessible and affordable for practitioners with minimal expensive equipment required to use it, so you really wouldn't need more than an osteometric board and a pair of sliding and spreading calipers.

So, the analysts will collect this suite of skeletal and dental traits, which takes one of our analysts around an hour to an hour and a half for one skeleton. Then the analyst will enter their observations into the model. That evidence will propagate through the network, with the understanding that one part of the skeleton can contribute probabilistically to multiple biological profile parameters.

So essentially, MOSAIC will take the probabilistic relationships from the matched reference data set in which those relevant classifications are already known, and then it will perform simulation-based inference to explore the range of biological profiles that are compatible with those data, and this will allow the model to identify the most probable configurations of age, sex, stature, and population affinity given all the available evidence, and the output will not be a single deterministic estimate. So, for example, a simple classification of male or female, but rather a set of probabilistic estimates that characterize the posterior uncertainty associated with each biological profile parameter., And these distributions are summarized using credible intervals, providing explicit characterization of any uncertainty encountered by MOSAIC. And so, in many ways, this mirrors expert reasoning by calibrating confidence about analyses according to a specific set of observations.

However, MOSAIC formalizes that reasoning mathematically, providing transparent results to help an analyst communicate both accuracy and uncertainty to medicolegal stakeholders. And one more thing about the outputs. If anyone is familiar with FORDISC 4 there is a downloadable summary page that's produced by that program, and we'd like to have something similar with MOSAIC that's user friendly, easy to interpret, and that would be able to be incorporated directly into forensic casework bench notes.

Next slide please.

So, there's one more thing I'd like to take a moment to discuss. And that's that of course we want MOSAIC to be compatible with the realities of forensic casework. And one of those realities is that analysts frequently work with incomplete or damaged skeletal remains. So, in a traditional workflow, the absence of a particular skeletal element or observation might mean that a particular method just simply can't be applied, and that leaves the analyst to make decisions about which alternative method to use. But within this framework we're building, an unavailable observation can simply remain unobserved, and the model will still generate posterior distributions based on the traits that are available. So, with this approach, we recognize that missing data are data themselves. So, for example, with fewer observations, uncertainty will generally increase. And in this way, we're ensuring that missingness will be explicitly integrated into the analyses where posterior distributions can become broader when the model has less available evidence for a particular individual.

Next slide.  

So, MOSAIC is progressing in phases. We're moving through data collection, model development, testing and refinement. And the work for MOSAIC officially started in July of 2025. And this past year of the grant was really onboarding and training, building our recording forms and our data collection protocols. And then, of course, traveling to skeletal collections to gather data to build our matched data set.

So going into this next phase, we now have a matched data set with great coverage of both metric and morphological data for the cranium, postcranium and dentition for nearly 2,500 individuals from across the United States. And we also have the foundations of the model and are developing the program interface in collaboration with programmers, software developers and user experience experts at the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign. And our goal for the next year is to continue to build on this work. So, we're aiming to have a testable beta by the upcoming American Academy of Forensic Sciences meeting in February. And we'll also use that space as an opportunity to solicit beta testers for the program. And we anticipate that formal validation from other practitioners will be a very important step moving forward.

And that will include assessments of accuracy, identification of potential sources of bias, and then tests for how the model performs under varying conditions of missing data. And one other thing I want to mention is that we also have multiple graduate students who are supported by this grant for their graduate research, so their research will investigate more closely some of the relationships between different variable types such as cranial morphological, dental morphological, and demographic data, so that we can continue to refine our probabilistic predictive models of the biological profile.

And right now, we're also in the process of preparing two manuscripts for a submission to the Forensic Anthropology Journal to introduce MOSAIC. So, the first of these papers will be the conceptual and methodological rationalization for the project. And then the second paper will be a proof of concept of the program itself. And ultimately the goal is to have the final deployment of the program by December of 2027. And our plan is for that program to be freely accessible to practitioners. And we're also in conversation about presenting a workshop at future American Academy of Forensic Sciences annual meetings to facilitate training in the data collection process, as well as the use of the program itself.  

Next slide please.  

So, today I've just presented the rationale, foundational framework and anticipated timeline for MOSAIC at this first stage of the project's development. We have substantial work ahead of us in the next phase, but we have an exceptional team as well as a proof of concept demonstrating the incorporation of empirically-derived covariation across the skeleton into a model that we believe can closely reflect the realities and factors of forensic casework. And the overall goal of this project is ambitious. MOSAIC represents a different paradigm for thinking broadly about how skeletal evidence is understood and integrated and ultimately operationalized within forensic anthropology. But with these foundational components in place, and with continued testing and feedback from the broader forensic anthropology community, it's our goal that MOSAIC will be a driving force in streamlining and standardizing the identification process, and ultimately improving the identification of unidentified individuals within United States medicolegal contexts. Thank you.  

And I just have some acknowledgments on the next slide.

I just want to take the time to acknowledge personnel at each of the donated and documented skeletal collections we visited so far, as well as our data contributors. So, we are grateful to all of our collaborators and of course, to the NIJ for funding this research and for the opportunity to discuss it with all of you today. Thank you, and I'm happy to take any questions.

Rachel Wendt: Thank you, Andrea. We do have a few questions. How is MOSAIC’s capacity to incorporate new and emerging technologies? Is that something that's going to evolve over time?  

Dr. Andrea Zurek-Ost: Yeah, that's a fantastic question. So, this is something we've talked about a lot with other collaborators. So, we first gave our initial introduction to MOSAIC at the previous American Academy of Forensic Sciences meeting in February, and we had several people approach us after who said, I'm working on X research project with Y type of data, is it possible to incorporate these data into MOSAIC down the line. And so that's something that we're very interested in and that we're very focused on is being able to have sort of a modular approach so that we can if someone, someone, for example, has offered to donate histological data to us or share those data with us, and we'd like to have a, a sort of model within MOSAIC that can incorporate those types of data.

So, we anticipate really that this will be a sort of, or we hope that it will be a disciplinary-wide effort to kind of pull together all of these different emerging technologies under a single umbrella. And along with that, one of the considerations that we've talked about with our software designers and developers is this idea of being able to constantly update the program without it crashing on people.

And so, to be able to incorporate updates easily and seamlessly so that practitioners won't have to, to worry about sort of what version they're using and, and these types of things. So, we would like all of that to be a major focus of the project moving forward. And we also, I should speak a little bit to that, just with regard to things like subadult sex estimation and age estimation, other parameters of the biological profile with subadult individuals. This is something else we've been really focused on. So, we think that there is sort of an arbitrary divide in some ways between subadult and adult individuals because many of these traits persist throughout life, Right, and so not all of them may be useful at all ages for the biological profile. But this is something that we are really interested in investigating in the future is also incorporating subadult data into MOSAIC broadly.

So, I hope that answers your question.  

Rachel Wendt: It does. Thank you. We have a couple more questions. Can you speak to the methods anticipated for reporting posterior probabilities?  

Dr. Andrea Zurek-Ost: Yeah. So, I think that's a fantastic question. So, you know, I mentioned just very briefly right now that we've talked about likelihoods, we've talked about attaching credible intervals to every parameter of the biological profile. But as the model is still in development and just full transparency, my role on the model or my role on the project has been primarily data collection and organizing and coordinating the data collection trips. I would want to refer or pass that question along to our programming and modeling team or the components of the team because I don't want to misrepresent what may be produced ultimately, but I'm very happy to pass those questions along to our statisticians and our programmers.

If anyone wants to email any specific questions about the model and what the outputs will look like to that email that was on the slide. So, and again, we've just finished that data collection push last week. So, we're really hoping to be able to get those data incorporated quickly so that we can start to test what those what those distributions might look like and what will be most useful to forensic anthropologists for their reporting.

Rachel Wendt: Thank you. A couple more. Do you anticipate MOSAIC being, sorry, MOSAIC data being applied to anthropological disciplines outside of unidentified human remains identification?  

Dr. Andrea Zurek-Ost: If this question, I'm not exactly certain what this question is referring to, but if this person is maybe asking about broader applications to bioarchaeology, maybe studies of demography in the past instead of identifications of modern individuals, right. I would say that that's absolutely something that we would hope to see MOSAIC tested within the future. So, the MOSAIC sample of course, is United States-based at this point. But we have several international collaborators in both forensic and bioarchaeological contexts who have already written to us to ask about testing the model with their samples. And so that's definitely a direction that we also want to be able to apply MOSAIC in.  

Rachel Wendt: Thank you. Does MOSAIC have the capacity to capture taphonomic data and anything else that might impact analysis?  

Dr. Andrea Zurek-Ost: This is a really great question. This is something that we've talked a little bit to folks at, at various other institutions about is the idea of maybe because this will be a data collection interface, maybe incorporating some additional data collection variables that aren't just related to the biological profile. What we have talked about so far that we would like to sort of concretely apply moving forward, is that we recognize that taphonomic agents do often fragment bones, they will reduce the number of observable skeletal traits that are able to be incorporated into any given method, Right, and so, we would really welcome any testing. And we have plans to do testing of this type as well that would incorporate different taphonomic patterns into our data set.

And so maybe have a better understanding of how different taphonomic agents and different taphonomic patterns can influence the model's performance with respect to missingness, potential bias that would be produced in the model, etc. So yeah, if that's what that individual was referring to, that is definitely something that we're thinking about moving forward that we would really like to incorporate into the testing phase.

Rachel Wendt: Great. One last question. What areas of the of your data so far are you looking to expand on? Like what's lacking in terms of balance and availability?  

Dr. Andrea Zurek-Ost: Yeah, that's an excellent question. So, we are working on attempting to balance the sample from a sort of variety of different directions. So, the first one is ensuring good demographic coverage, right, across all of these biological profile parameters.

And so that's really been a push for us in the last couple of months with this data collection process. And I think we know that we have enough at this point to build out a model, but we would really like to, especially if we have beta testers in the future who are interested in testing the model on samples or individuals who are from demographic backgrounds that may be less well-represented in donated skeletal context, because of course, that's where our data have come from so far, we would really welcome expanding in that direction so that we can better represent all individuals. But of course, especially those who would typically show up in our forensic casework, Right, and we, so that's one direction we're expanding is looking at the demographic coverage and looking at increasing that. And we're also looking at incorporation of different traits. We have, for example, less dental data at this point than other types of data, for example.

So, we just went to the donated skeletal collection in Hawai’i in May at the University of Hawai’i, and the teeth there were by far sort of the best preserved we've seen so far on any of our data collection trips from those donated individuals. And if there are other individuals that have maybe at their universities or at their institutions that have dental data, good teeth that are not absent or fragmented, that would be fantastic to bolster our sample. And because those are data that we just have, we have less well represented in our data set at present. So yeah.  

Rachel Wendt: Amazing. Thank you so much. Thank you. Next, we'll move to Dr. Jessica Metcalf.

Dr. Jessica Metcalf: Hi there. Great. So, I'm excited to talk to you today about how microbes can be useful in forensic sciences.  

Next slide.

So, a core research program in my lab is we study the microbial ecology of carrion and human cadaver decomposition. And so, we study this both in in an ecology and also in a forensic supplied framework. And we do this by studying the microbiome which are the genes and molecules of microbes in a given environment.  

Next slide.  

And so, similar to insects, a succession of microbes is associated with decomposing bodies after death. And that's something that we can leverage potentially as evidence if we have a good understanding of the microbial ecology. Next.

So, with the support of OJP funding, we have been studying and uncovering the succession of microbes during decomposition over the last about 15 years. And I want to highlight sort of the major milestones that we have overcome and sort of been building on before getting to the project that I'm going to talk about today, because all this work is builds on previous work. And so, in 2013, we first started this research with mouse models. And you can see the little mouse there where we had to answer the first really basic question of when animals, including humans, decompose, is there a predictable succession of microbes? Is it repeatable across individuals? So here we started with a lab, a laboratory research with mouse where we controlled everything minus time.

So, we controlled environment. There were no insects. We controlled host genetics and diet. And what we found was that we had big changes in the microbial community during decomposition. And across individuals, those changes were almost exactly the same. So, it was very impressive. And so, as our next step, we then wanted to study how different environments and also outdoors with human donors, using some of those facilities that we heard about in the last talk.

So, this is one of the ways we can use those donors before they become skeletal collections. And what we found is that, again, we were able to uncover using microbiome data this really predictable succession of microbes. It was a bit different between mice and humans, outdoors and indoors. But overall, we could see within study that we had a very predictable set of microbes that changed. They had major changes during decomposition that allowed us to use a random forest machine learning model to build one that predicted the postmortem interval. And so, the next two papers that I wanted to highlight were published in 2024 and 2026. And I'm going to go in a little bit more detail into those in the next few slides, because it really provides the foundation for the study that I'm going to talk about today.

So next slide.

So, our 2024 paper, we were able to work with three different human taphonomy facilities. Colorado Mesa University, University of Tennessee Knoxville and Sam Houston State University.  

Next slide.

So, we were able to collect skin and soil samples daily for 21 days for three bodies per facility across four seasons. So, a total of 36 bodies. And from that we generated a suite of different types of microbiome data. And we're not going to go too deep into that because I know, I think there's a lot of non-microbiome people here in the audience. But we have different ways to look at the taxonomic profile, the gene profile and the molecules that they're making. Next.

And so, as I mentioned before, one of the ways that we then use these data is we can train a random forest model on the microbial abundances or the gene abundances or the metabolite abundances, and how those change over time. Given that we know the PMI, right, for these donor bodies. And so, with those 21 days of daily sampling, we're able to build these models.

And so, these on this on this slide, I show you the different omic types. So, the different types of data and how accurate they were for predicting PMI. And this is, so again this is random forest data. On the y axis we have the cross-validation error. So that means that when we're training the model we always withhold a subset of the data and then use that to test, to estimate an error and our errors are in accumulated degree day with a base temperature of zero. And so, you can see here that we have very similar errors across the different data types. So, it's 16 S data, 18 S data and various different types of metagenomics data. And then finally metabolomics data. And for some of those data sets we only had certain sample types. But the amplicon data worked very well. So the 16 S. And it gave us a a cross-validation error of about 50 which is around three days chronological time. Next.

And so, kind of looking under the hood how does this work. So, if we get from our model the most important feature. So those are the microbial abundances that if you take them out they increase error the most. So, they're the ones that are really making the model accurate. And you can see some examples here. So Helcococcus seattlensis is our most in this data set is our most important feature. And what you can see is that the relative abundance over time in ADD base 0 is very similar across individuals and across facilities. And that's why this is such a powerful data point or data type to help inform the model and you can see that in several others. And in some cases, we have a little bit of a different pattern for Colorado Mason University, which is in a different climate region than the other two facilities. I forgot to mention that earlier.

Okay. Next.

And so, if we look at this from more of an ecology sort of standpoint, I'm not going to go through, there's a lot of methods that are behind this network. This network uses all the data types that we generated. And so, you'll see genes, you'll see metabolites, you'll see microbes. But the really important thing is that many of those important features are part of this really tight and complex ecological network where there's lots of cross-feeding, there's a lot of ecology going on. Next.

And so, who are these taxa? And so, one of the things we did in this paper is we search for our sort of key decomposer taxas, what we call them across a lot of big data sets like the American Gut Project, the Earth Microbiome Project, and in some cases, we could find them, such as with acinetobacter. We could find them in some of these data sets. But for the most part, as you can see in black here, a lot of these taxa were not in any major data set. But where we did find them, Next, is on insects. And so, we were able to go using some of our own blowfly insect data as well as others in literature. We downloaded those data, and we were able to find our sort of key decomposers really closely related to or associated with insects. So, this was a big finding and sort of becomes very important. Okay.  

Next slide.

So, this brings us to the 2026 paper that's also really relevant to what we're doing now. So, this one was the PI was Sibyl Bucheli at Sam Houston State University. And in this project, which was led by our PhD student Victoria Nieciecki, at Sam Houston State University. Next. We were able to study decomposition in indoor versus outdoor scenarios. So, for indoor we used sheds like the common sort of tough shed that you see that you can get from Home Depot or some other sort of store. So, these were not these were not air conditioned. This was a very simple, sort of straightforward projects. We had 12 cadavers that we placed indoors over several seasons, and then we had 15 comparative ones that we placed outdoors. And again, we kept the same sampling as daily sampling over 21 days of skin and soil samples from which we generated microbiome data. Next.

Okay, so not surprisingly to any entomologist, the maggot emergence was delayed in the enclosed shelter. So, when we ran this experiment, we did not know that the bulk of our, the microbes that were important for our model and that become really abundant during decomposition are associated insects. So, this wasn't an insect exclusion experiment, but we did thankfully track insects.

So, we knew when we recorded when flies were present or maggots emerged. And so, this is a taxa bar plot. And you don't have to read this too much. Just kind of let your eyes pass to it. But what you can see is that the maggot emergence on the left for indoor was 11.2 days on average, and you kind of see this increase in relative abundance of certain taxa that looks somewhat associated with that. Right? And spoiler alert, those are the taxa associated with bacteria that are important in our model. Next.  

So not surprisingly, if we take that Burcham et al. model from 2024, right, which is trained on 36 outdoor decomposing bodies, and we use the samples that we collect from this study as test samples. So not training the model with them and getting a cross validation, but actually as test samples, we get a higher error for indoor than outdoor. Makes a lot of sense. Tracks everything we've been figuring out. And you can see as an example here, some of the relative abundances of the important taxa and how some of them differ indoor versus outdoor. Next.

So, our conclusion here, which we've sort of come to over the last two years, is that insect information is going to be really important for predicting the PMI of indoor bodies. So, either we're going to have to have two different models, one for indoor, one for outdoor, or we're going to have to include insect data somehow in the model as a feature. Next.

So that brings us to what we've been working on with our NIJ project expansion and validation of our microbiome database for predicting the postmortem interval. And this project was a really big, coordinated project that we started a few years ago, maybe three years ago, and we worked with every available human taphonomy facility that that we were able to.

And so, from some of these facilities, we wanted to add to the training data set. So, in particular FROST and REST. REST is now closed. Those were from a different climate region. So, we wanted to improve our that's where the expansion is. We wanted to improve our models to include data from bodies from another, another climate region. And then we also wanted to include samples, more samples that were pure test samples. So they weren't from bodies that were ever used in training or data set.  

And if you go to the next slide.

What you can see here, Tr means training. Those are the number of cadavers. And again, this was daily sampling for skin and soil for 21 days. So, in a place like FIRS at CMU, we have eight cadavers that were held out as test samples and a total of 12 cadavers that are part of the training set. And this includes the samples from the Burcham et. al data and also from the Nieciecki et. al the last those two, because we're now expanding the dataset and we're using all the samples we have. And so, there are some locations like Texas State University and Western Carolina, where we only have test samples. So we've never, so our training data set does not include any cadavers ever placed at those. And that really helps us understand how robust our model is, how generalizable our model is. Next.

Yes. So, there are the totals. We had 62 outdoor bodies for training. Only 12 indoor bodies. That's going to become important. And then 33 outdoor bodies for independent test samples. And most of those bodies we collected like four samples from not the full 21. Next.

And so, here's what some of those results look like. So here we have mean absolute error again in ADD. One of the things that's different here is this is for test samples. So, it's not a cross-validation within training error. So, this is much more robust. And you can see that we have an average around 50 which is very close to what we had. Oh, overall MAE is 64. But it's very close to what we had with the cross-validation error in the Burcham et. al study where we didn't have as many test samples. And you can look across facility, we don't have, what's really interesting is so the facilities FARF, FOREST, both of those are not represented in our training data set. Most of those samples from those cadavers predict really well. So that's a really nice result that we have like pretty robust, a pretty robust model built on a whole lot of samples. Next.

Okay. So just some sort of ground truth thing checking that we don't have anything major change. We'd look at the most important taxa for this newer model that's built on much more samples. And again, we see really similar taxa to what we saw with the Burcham et. al. So, no big surprises there. Next.

And just to kind of show you, again, pulling that network out of the Burcham et. al study, I kind of highlighted here, not comprehensively, but just a few of the important features that, again, we know are associated with, well, that were, these are important features from this current model. We see them again in this network that is ecological, that has an ecological meaning, right, like the taxa are doing something together. Next. And we see that most of these are insect associated, okay. Next.

Okay. So now I'm getting to the part of like of what are we doing right now. So of course, we pulled out as best as we could from all of our research projects, the notes we did have metadata columns about insects. And we added them as a feature in the model. So similar to a microbial abundance, we don't have to just have microbial abundances. We can have other environmental data in the models. And the insect data did not affect the error. So, it didn't get worse, didn't make it worse. Didn’t make it better. But if you recall, all of our test samples are outdoor. And so that's probably why. What we need are test samples that are indoor or something where insects are restricted like burial to understand how well this insect metadata will help us overcome the differences that we see indoor, outdoor. It's not really indoor outdoor, it's really about insects being restricted. And so, that remains to be better tested. Although we're fairly confident that insect metadata should, for indoor or insect restricted samples, improve the model. Next. So, given what we have at this time, what does affect the error?

So, if you recall a few slides back, we had some dots with really high errors. Like, there's some samples and what that often is, it's a really late ADD sample getting predicted to be really early ADD. And so that could mean one of two things or probably more than that. But I can think of 1 or 2 things.

So, one our training set. So here, we show you for the bodies that predicted the least, that had the highest mean absolute errors of our test samples. You can see that where they really had a problem predicting was in the later ADDs. And so, one area is that we don't have, we have a lot more training data for the 300 and below ADD. And so it may be that there's just not enough training data. The other option is that we don't have like the microbial community change in skin and soil slows down at those ADDs, and we can't really use microbial abundances as well to sort of parse between a 400 or 450 ADD.

So yeah. So, you can see some of these that have errors at 400 and 500 that are like 300. That means they were basically, they were being predicted at the wrong end of decomposition. So that's very informative for us to help us understand where we can improve the model or you know, where the model is going to work the best and what are the, you know, what are the cases where we shouldn't trust the model as much. Next.

So, for our future work, if I've done a good job, it shouldn't be surprising. So, I think we will need to build up our indoor training and testing samples and potentially also build up our late, later ADD training samples. We need to be very careful in using insect and collecting evidence of insects if they've been present or not in any of our future research.

We also have been working on the modeling quite extensively with Rob Knight's lab, where we've been testing for, we've used random forest machine learning models for a really long time, and they tend to work really well with microbiome datasets. It's not just for predicting PMI, they're used in other things, but we have been testing whether transformer models may also work well. Next.

So, with that, I just want to acknowledge our many partners and the funding that we've had for this research. And so first of all, we'd like to thank the donors. These are extremely large data sets that we are fortunate and to be able to generate. And all of our collaborators at those facilities and the labs that have helped us, several, many NIJ, several NIJ grants that allowed this, this cumulative body of work to continue to grow as we, as we work towards a model that can be, we can start to have medical examiners and coroners and other medicolegal folks begin to, to test out. So, with that, thank you for listening and I'm happy to take any questions.  

Rachel Wendt: Thank you, Jessica. So go ahead and ask your questions in the Q&A box. And I will start us off with a question. So, for a variable like body size or BMI, which could cause some decompositional variance, is that a potential variable for the microbiome or are insects going to behave the same regardless?

Dr. Jessica Metcalf: Yeah. So, I think we will always have an error, right. We're never going to really get it. You know right now our error is about 50 to 65 ADD. And what is that error caused by? And, it's probably accumulation of a lot of these variables like BMI, you know, possibly medications, we don't know. And so, we have, there's different groups that have studied BMI. We did one study with pigs and piglets, and we didn't find any major difference between the microbial community and its changes with like one kilogram versus 20 kilogram. But that wasn't really BMI. That was more just like size, mass. And so, I don't think it's going to be a major influence. But sure, it could certainly be one of the many things that's cumulatively adding to the, to the error that we probably can't completely overcome.

Rachel Wendt: Thank you. You guys have addressed seasonality, and you mentioned burial as another environment that needs to be explored. Are there other things like a semi -submerged aquatic “burial” that you are interested in trying to get data from?  

Dr. Jessica Metcalf: Yeah, I, I am interested in aquatic. I mean, I think the aquatic environment is going to be so different that, that's not one that I've delved into yet. I'm sort of trying to get a handle on all the variables around terrestrial, terrestrial, and sort of then now terrestrial scenarios where insects are going to be restricted or delayed and the seasonality can have something to do with that, obviously.  

Whether it's too hot for insects or too cold for insects could certainly cause, you know, contribute to some, some errors. So having and remember in our PMI models we do have accumulated degree dose, so temperature is part of the information that's happening here. And so, it's possible the models will be able to figure that out. We will see.

Rachel Wendt: Great. Is there a potential utility for the microbiome to help in a case where a body has been moved after death?  

Dr. Jessica Metcalf: Yes. Great question. So, if the body is for example, so, one scenario we tested back when I was doing the mouse decomposition work many years ago is from one of the experiments. I removed the body and I and I sampled the soil for I don't remember if it was like up to 30 days or something after removing the bodies and you still got that sort of soil signature. So, there's a signature that remains that microbial community doesn't in the soil. Those changes, they don't just disappear. Right. Partly probably because the nutrients are in the soil at that point. And so there certainly can be signatures in the soil, for example, if a body has been moved elsewhere or if you have very decomposed remains and there's no evidence in the soil that decomposer microbes have been present, you know, I think a lack of insect evidence would also point to that.

And it could probably work in the reverse, some of the reverse scenarios as well. You certainly, I think, could tell some indoor versus outdoor scenarios from, from skin, from a decomposing body. Right. Because if it's outdoors at a certain time of year, it certainly should have had insects and those insect-associated microbes, I call a little mobile ecology unit, you know, sort of dumped on to the body.

Yeah. So, I think there's various clues beyond PMI that we could, that we could use from the microbes.

Rachel Wendt: Awesome. Well thank you. And we're going to go to our final presenter Dr. Christopher Goden.

Dr. Christopher Goden: Hello, hopefully y'all can hear me. I believe that I'm going to now share my screen.

Yeah, that's not it.

We're going to do this all right. Can everybody see okay? Somebody can confirm from NIJ that y'all can see my slides. Everything looks good. Fabulous. Thank you so much. So yes, I am Chris Goden. I am now an assistant professor at Mississippi State University. I also serve as one of two forensic anthropologists for the state of Mississippi. The other one's Doctor Jesse Goliath, also here at Mississippi State.

And so, this research is actually going to go over a couple of different NIJ grants that we got and a couple of different papers that were published out of my doctoral institution, the Ohio State University, more specifically the Injury Biomechanics Research Center. So first, I'd like to acknowledge my co-author and my co-investigators on these grants for their contributions to this work.

As I said, these are going to be covering some work from two grants from the NIJ. The first was in 2019, and the second was a continuation in 2023. But all of the data is still related back to the tibia. So, we combined them for some of these analyses. The two papers that I'm going to be giving data from are shown here. Those QR codes will allow anybody in the audience to go ahead and scan. And they can download those to read at their own, whenever they would like to. The one in red is my article from I believe it was published back in May. The one in green is one that just came out, I think like last week in Early View.

So, those are both giving different approaches to human variation but all still from the same project. So, if you notice any differences with like the sample size, it's just because we've impacted more since then. So, to start us off, trauma in forensic anthropology is really one of the most frequent services that we’re asked to provide. Specifically, we're seeing an increased utilization, and we're seeing a lot of instances where we know who our decedent is, and they are wanting, actually, the anthropologist to go in and talk about that skeletal trauma that we're seeing there.

And so, when we're collecting these skeletal trauma data, we're trying to ascertain what might have happened to this individual that led to their death, and that’s specifically what the coroner and medical examiner's office usually are requesting to know. And that has a huge investigative impact. Obviously, as forensic anthropologists, we don't determine cause and manner of death, but our findings can inform that forensic narrative for prosecutors, death investigators, medical examiners and coroners, and ultimately that can influence some court case outcomes.

So, there's a huge, huge burden here on the anthropologist to make sure that when we're collecting these data, we are doing so in a proper way, and we're interpreting things carefully. Right. But what is this data that we're collecting actually mean? It's all contributing to our interpretation of trauma. And right now in the field, it's kind of relying on some assumptions. It’s really experience based.

We have a lot of good data in the field about we know this trauma from a car crash. We know these trauma events happen from similar instances that we might see in a medical examiner setting, but we don't have as much experimental data that exists. We have some non-human analogs, and we have a little bit of human data, but not enough.

But we also see when we do think about those human impact studies, it's highly variable. Some different research out of Ohio State and research out of Michigan State have demonstrated that there's a lot of variability in those fracture outcomes. So, it leads us to the question, first and foremost of how accurate are our trauma interpretations? And we really want to know does human variation influence trauma outcomes?

So, for our leg experiment and impact experiments that we did at Ohio State, we've tested over 100 isolated biomechanically fresh human tibia. And by biomechanically fresh I mean that they were excised, kept frozen until we were going to use them for the impact studies. And then they were thawed to get back to normal temperature and then impact. All of the tibia were kept 100% fresh. Nothing was embalmed. And this is still an ongoing interdisciplinary project. It's not just forensic anthropologists, it's also engineers as well. And this experiment that we're going to go over today is a simplified bending scenario that's replicating essentially a pedestrian-vehicle impact, specifically that bumper and I'll show you in the experimental videos here in a moment. And some justification for this type of loading scenario is that it's very common. There were over 70,000, sorry, almost 70,000 pedestrians that were injured referenced in a recent study. And we most commonly see tibia fractures resulting from this impact. And more specifically, because this is a bending scenario, we're expecting that to result in a wedge or a butterfly fracture, if you're familiar with biomechanics of the of bending. And so, one of the first things we're going to look at before we get into the nitty gritty of the two articles that I'm talking about today, is we want to actually just look at what the fracture outcomes in these human tibiae are when we're loading them to failure and bending.

So, for our methodology, we utilized postmortem human subjects as a fancy way to say, our anatomical donors, which we were lucky to have through the Ohio State University College of Medicine. And first we started with an actual physical examination of those remains to make sure that they're going to fit within the parameters of this project. We want to make sure that when we're testing these bones, that they are representative of a generalized population rather than somebody that's maybe super very osteoporotic, etc..

So, we got to see those remains at first and then also had a ton of access to records, age at death, some medical history, cause of death, their stature, body habitus, etc. but we went a step further. We went and did full body CT scans of these donors so that that way we could go and actually look at the skeleton before doing anything. Get a sense for any antemortem trauma that might be going on. Also, any perimortem trauma and as well as pathologies. If we determine that that individual was free of any existing trauma specifically to the tibia and/or pathologies, that bone was then excised and received its own isolated CT that we’ll go over later. And, also standard osteometric measurements were taken using the measurements outlined by Langley et. al in DCP 2.0.

So, for our methodology of our experimental impact, you can see here our distribution in this bar chart of the donors that we utilized for this study. Again, this is a growing data set. So, we're always going to be adding a little bit more. In general, we're only looking at adults aged 22 at the low end. And on our high end, I believe it's 102 years and we have a pretty decently matched male to female ratio here. Obviously, it skews a little bit older, but we have a sizable amount of younger and middle-aged individuals as well. We've partitioned them off just for ease of visibility here in these different age brackets. This is going to change a little bit as we go through, so I'll just point that out as we get to it. To get ready for this experiment, again, all bones were kept biomechanically fresh and they were wrapped in saline-soaked gauze during all these different procedures to prepare the bones for impact. We started by isolating down to getting the bone potted in these rigid pots shown here in teal, and that's just a fast cast resin to help stabilize the bone when it's undergoing impact. We put those at the 20 and 80% sites. The epiphyses were removed to facilitate orientation within our experimental testing device, and those bones were simultaneously impacted by two impactors at the 40% and 60% sites of the total diaphyseal length in four point bending in either a lateral to medial or posterior anterior condition at 6m/s or 2m/s. You might be asking why we have a couple of different loading scenarios going on here because we had two separate grants.  

The first grant was exclusively looking at loading conditions, and then the second grant added in a couple of different things that we started looking at. Our largest condition is going to be lateral-medial six meters per second. So, I'll point that out as we start switching around looking at different groups later. This is a high-speed video example of that bending test. You can see those impactors coming down. This video was filmed at 20,000 frames per second, and you can see that it's resulting in that expected butterfly or wedge fracture right at the mid shaft of the bone. But play one more time. There's a couple more videos throughout the rest of the presentation.

So, after the bones were impacted, we went through using the AOOTA Trauma and Dislocation Compendium to classify and analyze any of the skeletal trauma that was a result of this test. So, this procedure is hierarchical. So, there's a type and a group. We're just going to focus on the overarching fracture types which wedge or butterfly fracture. It's that triangular fracture that you see a simple fracture.

So, any type of oblique or transverse fracture would be classified in that. And then of course multi-fragmentary which are going to be more complex fragmented bones. Because we did start seeing the tibiae had more than one fracture that could be classified, we classified each fracture independently. So, for example, this one has a wedge and an oblique fracture. So, we ended up tabulating a separate fracture count variable. So, in this instance, since we have a wedge and a simple fracture, that's two fractures.

So, for our expected outcome the wedge since this is a bending scenario 64% of the sample produced a wedge fracture. And that was 69 tibiae in this particular setup. But we also had 24 of them that ended up having a simple fracture and 15 of them that were multi-fragmentary. And that's almost 40% of the entire sample that did not result in a wedge fracture. So that's kind of interesting when we think about that because that's kind of sure we see that the bulk of the sample has that expectation. But when we start looking at divvying up, okay, what was the breakdown of wedges in the different loading scenarios that didn't explain any of these deviations. We also briefly started to look at sex and age trends, but again, we weren't seeing those trends come out clean because atypical patterns that simple or that multi fragmentary outcome were evident across all demographics and loading conditions. And so yes, while we had 64% that did follow the expectation, almost 40% did not follow expectation. And we don't know exactly why yet. And that's an interdisciplinary concern, right? In forensic anthropology, we want to be sure that when we're documenting trauma and they're asking us what happened, if we're going to say what maybe a bending scenario is producing, or if we're going to interpret bending from that wedge, we want to make sure that we're accurate in that, right?

On the flip side, for medical context, we want to know potentially what types of scenarios are leading to different injuries. So, we interdisciplinary we want to figure out why are these deviations happening. So, we need to investigate the influence of intrinsic factors and what they are doing to our fracture outcomes. And so, just to kind of put that knowledge gap a little bit more cleanly, we have extrinsic factors direction loading rate that we know influence injury outcomes.

And that's why we can determine the difference between projectile and blunt trauma on bones because bone is viscoelastic. It's going to change how it's responding based on that loading rate especially. But we don't really fully understand yet how intrinsic factors, height, weight, osteometrics, and bone shape collectively, how those are influencing injury outcomes. So, by looking at those things, our question is becoming can we better understand why we're seeing variation in fracture outcomes? And maybe then we can start saying, okay, based on certain metrics, somebody might be more predisposed to a less severe or a more severe outcome. And so, we're looking at what is the actual contribution of those intrinsic factors. When we look at the research from Harden et al. in 2026, that article that just came out a couple of weeks ago, I won't spoil all of it for you, but we did some really complex analyses.  

The statistics were led primarily by Dr. Stull and then interpreted by the primary author, Dr. Harden. So, when we're looking at age, we saw some weak trends. We look at sex. We did see that females tended to have fewer fractures. When we look at stature overall, we started to see that it was influencing some fracture classification significantly. And similarly with BMI, we saw that that was also influencing our fracture classification significantly.

And generally, we saw that overweight and obese individuals had a higher propensity for fragmentary fractures. And when we look at some more complex FMAD factor maps, we can see that there are tibiae that are grouping together based on their age, based on their BMI category, based on sex. And also, if we look at the fracture classifications, tibiae are also kind of starting to group together into those nice clusters of fragmentary or wedge or simple.

And so, this is setting us up really nicely to start looking at whether bone geometry is influencing these things, because we're seeing in these maps that bone geometry, body composition and sex are covarying with our fracture morphology. So, when we move into looking at just our shape data that we're going to talk about, for my dissertation work, we wanted to simplify our approach, and we just wanted to control our loading condition.

So, we had just the tibiae that were loaded lateral to medial six meters per second. That's like 15mph approximately for anybody that doesn't want to do that conversion in their head. So, because of that we only had 60 tibiae that we looked at in this particular study. The sample breakdown is provided here in those bar charts. You can see that we're still relatively equal males and females.

And we have a pretty good breakdown between young and old. For this analysis, we had a young end of 24, high end of 102 years. Young is going to be classified as anybody 24 to 60 and then old 60 to 102. For our injury outcomes at outcome variable, we started to look at fracture classifications and fracture count. But we quickly realized and if anybody's stat savvy here, you know that this can create some issues when you have unbalanced data.

So, we had 60 observations for 60 tibiae for fracture count. But when we started looking at fracture type, we had 101. And that's because certain fractures were co-occurring with other fractures more often than others. So, we started to think back to the literature, and if you're familiar with the work of Symes, and if you were familiar with the work of Labbé, you know that there is kind of a focus on modeling whole total-bone injury. And because fracture classifications and fracture count are differentially capturing our bone injury, if we have a simple fracture versus one count, that makes sense. But if we have a multi fragmentary fracture that's still classified as potentially one fracture because it's one classified fracture. So, we started to look differently at how we could view this. And we developed a complexity index of low, medium, and high.

And since I only have like ten-ish minutes left, we can't really get into the intricacies of that. So, just think about it as our low severity or low complexity is going to be less numerous, less severe fractures. And on that continuum, the high complexity fractures more numerous, more severe. So, we took our tibiae before they were ready for impact and we got our individual CT scans. Those were rendered as 3D meshes and made watertight models in 3D slicer. In order to make sure that we were referencing our actual physical bones that were impacted during the studies, we took Faro scans of them before impact, and those were all aligned so that we could make sure that we had them representative of how they were impacted in that standard loading scenario.

So, this posterior surface is looking at us with the lateral surface being at the superior end. The analogous CT scan model was then aligned to that Faro model, so that we could make sure that we're doing analysis on the shape of the bone that was explicitly used in that bending scenario, and that includes cropping it down to make sure it's just that diaphyseal segment. In order to quantify shape, we went through and established surface correspondence by putting different landmarks throughout the length of the bone in standardized locations shown here in green. And then we did a dense landmark placement to get really in-depth data on that surface of that bone. Those landmarks were then projected to the rest of the sample, and variation was quantified in both an unscaled and a scaled for length scenario, so that we're taking that big parameter of shape difference out of the equation so that we can see okay. Are there any new traits that are identified in scale that weren't identified in the unscaled shape analysis? In order to interpret those, we went through a principal component analysis where that's resulting in principal components. Think of those as continuous variables because that's what it's providing to us. It’s a combination of all the data there. But then also, we can make visual exemplars of those shape traits.

We can calculate what a negative extreme looks like to a positive extreme. And then there's also going to be a certain amount of variation up to 100% that is captured by that specific trait. So, this one, for example, is showing a change in that bowing of the diaphysis capturing 88.2% of the variation. This is just an example. So, we'll go through those shape traits in just a second. But when we look at the proportion of total variance found per trait, you can see that it's dominated in UPC1 in gray by that PC1. And so that's going to be length. But then when we look at our SPCs, our scale principal components in purple, they're more distributed. But what was interesting was when we compared them and did a correlation matrix, a lot of them were very highly correlated.

And so, this is essentially demonstrating that we are identifying mostly the same traits after removing that size variable, just the proportions of variance are redistributed. So, we're capturing the same traits across the different analyzes. For our statistical approach, we took both those unscaled and scaled principal components, and we put them into multinomial logistic regression models in R. And we used our fracture classification or fracture count variables as our outcome. And we also wanted to look at sex interaction effects for sex and age. And then we went through and calculated predicted probabilities to make it a little bit easier than using odds ratios. And so predicted the probability of low, medium and high or one, two and three fractures at the negative extreme shown in blue, and the positive extreme of that shape shown in red.

Each of those columns are going to add up to 100%. Going into our actual injury outcome results, you can see when we look at our fracture count models compared to our injury complexity model, that distribution is changing, so we have a lot of things moving around, things that were originally classified as one fracture might have been moved into a medium or a high complexity. And while we don't have time to go through every nitty gritty detail today, that's for the paper. We'll go through a couple of different examples. So, this is the second unscaled principal component, and we're looking at complexity as the outcome variable. The shape trait captured here is 3.8% of the total variation. And it's looking at the proximal half of that bone where that negative extreme in blue has a more curved proximal diaphysis. And then the positive extreme in red has a more straight proximal half of that diaphysis. In general, when we look at that positive extreme, they were more likely to experience low complexity fractures. And that's interesting when we look at the distribution of the predicted probabilities. For low complexity fractures, that negative extreme was predicted at 3%. And then that positive extreme was predicted at 99% to have low probability of, sorry, a probability of, a low complexity fracture. When we look at a different trait, though, SPC10, that's looking at this torsional distal half. So that negative extreme is having a more twisted distal end. And our positive extreme is having a straighter distal half of that bone, even though it's only accounting for 1.4% of the variation. There is a big trend again here where our positive side, our positive shape trait is more likely to be associated with a low complexity fracture or a medium complexity fracture.

It's 52% versus 41%, whereas our negative extreme, that blue is only a 2% probability of having a low complexity fracture. So, we see that if you have that distal torsion, you're more likely to have a high complexity fracture at 60% probability. And when we look at our sex interaction effects this was a very interesting trend. So, for UPC3 that's going to be width of the bone as you can see here, negative extreme a wider diaphysis positive extreme narrower diaphysis. We predicted different probabilities for males and females at those extremes of those different traits. And what was an interesting trend here was that similar bone shapes were experiencing different outcomes when we looked at sex.

So, for example, if we look at medium complexity, those females at that low extreme had almost an 80% predicted probability compared to 5% at the positive extreme. For our males, we saw the inverse, 7% predicted medium complexity versus 89% for that medium complexity at that positive extreme. So, they're in direct contrast of one another. And so, in summary, bone shape and intrinsic variation in general is showing relationships with injury. And so similar shapes have those different injuries between sexes and relationships. Also in that paper, we were seeing that they were becoming more complex with age changes. And that kind of makes sense, giving our known differences with age and sex, how that contributes to bone biology. But also, another interesting thing to consider is that low variance traits identified here still had significant relationships with outcomes. So even though low variance is kind of the idea that maybe it was a rarer outcome in the sample, it was a rarer trait in the sample. This may represent individualized anatomy, and it may demonstrate that we're predisposing certain individuals towards a less or more severe outcome based on individual bone traits. And so, the other finding was that our complexity index was a little bit more intuitive than fracture count. But I'll save that for you guys to look at in the paper. For our limitations, we didn't look at any cross-sectional geometry, but you can probably infer based on some of the results that maybe that’s having a role here. So, we're planning on looking at that for future work. And we're also doing histological analyses in future work to flesh out all the data as well.

So, the big overall here is that different fractures are able to be produced in the same loading scenario. Bone shape and intrinsic variation warrant consideration when we're interpreting trauma. So, I'd like to conclude by thanking the anatomical donors for their generous gifts, without which we would not be able to do any of this research. And I'd also like to thank the National Institute of Justice for funding this work.

And I'd also like to again shout out my co-investigators on this work, and especially my doctoral advisors, Dr. Harden and Dr. Agnew, for all their contributions to this work and helping develop some of the stuff that we presented on today. So, with that, I would be happy to take any questions after I figure out how to go back to my screen and stop sharing.

Sorry for that matrix thing that was happening there.  

Rachel Wendt: Thank you. We do have a few questions. So, the first one I'll ask you is how do fracture patterns differ based on fat and/or muscle density?  

Dr. Christopher Goden: Oh, great question. We are starting to look at that. And I can't give any results yet because we haven't fully fleshed it out. But we're actively investigating that right now. That's something that we're working on as part of that second NIJ grant. And I'm literally working through some of that data right now. We're at a little bit of a delay since I've moved institutions, and I'm waiting on all my stuff to get in so I can start getting all of my measurements in, but we're actively looking at that data.

Great question.  

Rachel Wendt: It's really exciting to hear. Next question, since you had access to medical records, were factors like pathology, disease or medications controlled for.  

Dr. Christopher Goden: We could not control for medications because we weren't provided that explicit information, but we did go through and make sure that nobody had any type of pathology to the bone. Again, if anybody has familiarity with some of those records, it's at worst entirely inaccurate and at best just limited.

So, we can't get super in-depth with what, and that's also protected information, so we can't get super in-depth as to cause and manner of death for that kind of stuff. But you can know that none of the tibiae in this sample were pathologic or traumatic in any way. We made sure to control for that.  

Rachel Wendt: Awesome. Next question. Could this methodology be applied to irregular or flat bones, or would a different approach be necessary and what potential obstacles would be anticipated?

Dr. Christopher Goden: Great question. So, because this didn't utilize any if you're familiar with type one, two and three landmarks, the tibia is mostly type two and type three landmarks. Especially when we cut off those epiphyses, we got rid of a lot of rich information to create an analogous form to different tibs. So what we ended up having to do was go with that dense landmark approach. And after some trial and error, I'm toying with the idea of developing a separate tech note for this. We found that it worked very well, and somebody has developed a similar protocol to look at juvenile humeri in an anatomical sciences article, I think it was anatomical record. It was a very similar approach. But we have a couple of other bones that we have similar research going on with for in particular the rib or the ulna if you're familiar with Dr. Agnew's work.

So we're exploring moving on to different bones as well. I think that it would just again, be kind of a trial and error to see, okay, did this method work on this bone? Why or why not? And then of course, with any methodology you have to probably tweak it a little bit. So, with my anchoring landmarks, in order to get that, we were projecting our dense landmarks to the right places. It was based on standardized locations along the bone, so proximal and distal end 60, 50, 40% sites based on total diaphysis around the bone. So anterior, posterior, medial and lateral. For a long bone, that might work great. For a rib, you probably would have to figure out, okay, what are we going to do instead? And so, we're looking at that as well. But I think it's a great opportunity to start looking even for other questions, not even related to trauma, just for quantifying human variation. Yeah.  

Rachel Wendt: So, you mentioned the removal of the epiphyses. Do you think there is any potential that that could have impacted the way that those fractures are appearing?  

Dr. Christopher Goden: That's a great question because everything was rigidly potted, I don't believe so. When we were taking them off, we sawed them off. We didn't like break them off or anything like that. And it was done after they were potted. So, it was stable within those pots. So, nothing was happening beforehand. We didn't cut off the epiphyses and then fix them in those pots. It was after the fact.

So, that would be minimized. But at the same time, I do think it would be great to have a comparative sample where it is the entire bone that's impacted as well. That obviously introduces some new issues to control for in the actual experimental setup. I know that Mari Isa and collaborators of hers at Michigan State, they've looked at a similar type of loading scenario.

I think it was three point bending in the femur and they did use the entire bone. And so, I think that in future work, it would be great to collaborate with a team like that who have done that type of work. So yeah, future work should investigate that for sure.  

Rachel Wendt: Well, great. That was our last question, so thank you Doctor Goden.

Thank you. That is a wrap for the Forensic Anthropology and Forensic Pathology sessions. Thank you for dedicating your afternoon to today's webinar. If there are any follow up questions, please contact [email protected].

We look forward to your participation in future events.