Forensic analysts have long sought precision when determining time of death. While on crime scene investigation television shows, the presence of insects always seems to reveal when a person died, there are many elements to account for, and the probable date may still not be accurate. Insects arrive days after death if at all (e.g., if the body is found indoors or after burial), and the stage of insect activity is influenced by temperature, weather conditions, seasonal variation, geographic location and other factors. All this makes it difficult to estimate the postmortem interval (PMI) of a body discovered an unknown time after death. One way to make estimating PMI less subjective would be to have calibrated molecular markers that are easy to sample and are not altered by environmental variabilities.
Bacterial communities called microbiomes have been frequently in the news. The influence of these microbes encompass living creatures and the environment. Not surprisingly, research has focused on the influence of microbiomes on humans. For example, changes in gut microbiome seem to affect human health. Intriguingly, microbiomes may also be a key to determining time of death. The National Institute of Justice (NIJ) has funded several projects focused on the forensic applications of microbiomes. One focus involves the necrobiome, the community of organisms found on or around decomposing remains. These microbes could be used as an indicator of PMI when investigating human remains. Recent research published in PLOS ONE examined the bacterial communities found in human ears and noses after death and how they changed over time. The researchers were interested in developing an algorithm using the data they collected to estimate of time of death.
Samples were collected from the ears and noses of four cadavers every 2–3 days over several months at the Anthropological Research Facility (ARF) at the University of Tennessee at Knoxville, colloquially known as The Body Farm. In addition, single samples were taken from 17 other cadavers. All samples had DNA isolated, PCR amplified and microbes identified using 16S next generation metagenome sequencing. The bacterial species were classified from the sequence data, and each sequence was assigned to its genus, family, order, class, phylum and kingdom. For the computer analysis and learning, the days since death were calculated as accumulated degree days (ADD), taking into account time and temperature variations because the cadavers were subject to an uncontrolled outdoor environment.
How does the microbe diversity change over time? Correlation between time and species, genus or phylum was analyzed for ear, nose and both samples using Pearson’s correlation coefficient. In fact, bacterial diversity decreased over time and there was negative correlation with ADD for ear; a smaller positive correlation with nose. These correlations were seen over species, order and phylum.
The data were taken, split into two, a training set to fit the data and a testing set used to predict, and used to create a model using a regressor with a choice of hyperparameters. Seven regressors were used in a grid-search algorithm, and resulting the model from each regressor is evaluated to select the model with the best cross-validation score (lowest error) on the training set. After scoring and ranking each model, the model is retrained on the training set and then the testing set applied. The error rate from this test offered an independent validation of the model with an unbiased accuracy estimate. Each model was then compared to a dummy regressor. This dummy or mean model simply memorized the mean ADD found in the training samples.
For the nose data, generated from 69 viable samples, the dummy model outperformed the ten lowest-error models based on curated or uncurated data over various taxa (species, genus, order, phylum or class). However, for the ear data based on 83 samples, the ten ear models from curated or uncurated taxa data outperformed the dummy model. There were 67 samples from the four cadavers that had good sequencing reads from both ear and nose samples. Despite the smaller data set, all of the top ten models for the dual samples surpassed the dummy model by a wide margin.
In addition, Johnson et al. were interested in exploring which combination of dataset and regressor may give the best overall generalization based on our data, eschewing accuracy and using the entire dataset for model selection. Through testing, researchers found that simpler algorithms (e.g., linear regression) performed well when ranked for validation error and phylum taxon worked well in several of the models.
While the top-performing models suggest that using all the data sequenced from cadaver ears and noses produced the best predictive algorithm, researchers were interested to know which taxa might be more informative. To do this, three different feature selections, a method to select the best independent variables for predicting a dependent variable, were performed. Each method produced different features, suggesting that the data set as whole was more important than any particular taxon. There were two or three models that had agreement on some phyla with the data set only encompassing 52 phyla. While some phyla may have predictive value, the best models were constructed from the entire data set and not restricted to any one taxon.
Using the full data set of nose and ear sequences, Johnson et al. were able to develop a model through computer learning from an algorithm that can identify time of death plus or minus 55 ADD or about two days in the Tennessee summer. This is a proof-of-concept that integrates 16S RNA sequences from microbes swabbed from the ears and noses of decomposing bodies with a predictive algorithm to determine time-of-death with more accuracy than current methods. However, the sample size is small and only from a single geographic location. Are there differences in the necrobiome in other locations? Are there other sites on the body that could benefit the model? Exploring the value of the necrobiome is still in the early stages, but the results from this research suggest it is a worthwhile pursuit and may give forensic investigators another tool for their arsenal.
Johnson, H.R., Trinidad, D.D., Guzman, S., Khan, Z., Parziale, J.V., DeBruyn, J.M. and Lents, N.H. (2016) A machine learning approach for using the postmortem skin microbiome to estimate the postmortem interval. PLOS ONE 11, e0167370. doi: 10.1371/journal.pone.0167370