A Machine Learning Approach for Using the Postmortem Skin Microbiome to Estimate the Postmortem Interval

Abstract

Research on the human microbiome, the microbiota that live in, on, and around the human person, has revolutionized our understanding of the complex interactions between microbial life and human health and disease. The microbiome may also provide a valuable tool in forensic death investigations by helping to reveal the postmortem interval (PMI) of a decedent that is discovered after an unknown amount of time since death. Current methods of estimating PMI for cadavers discovered in uncontrolled, unstudied environments have substantial limitations, some of which may be overcome through the use of microbial indicators. In this project, we sampled the microbiomes of decomposing human cadavers, focusing on the skin microbiota found in the nasal and ear canals. We then developed several models of statistical regression to establish an algorithm for predicting the PMI of microbial samples. We found that the complete data set, rather than a curated list of indicator species, was preferred for training the regressor. We further found that genus and family, rather than species, are the most informative taxonomic levels. Finally, we developed a k-nearest- neighbor regressor, tuned with the entire data set from all nasal and ear samples, that predicts the PMI of unknown samples with an average error of ±55 accumulated degree days (ADD). This study outlines a machine learning approach for the use of necrobiome data in the prediction of the PMI and thereby provides a successful proof-of- concept that skin microbiota is a promising tool in forensic death investigations.

Emerging Priorities for Microbiome Research

“…The gut microbiome can be collected from toilets, and individuals from diverse geographic locations can be differentiated both by specific microbial sequence signatures (Yooseph et al, 2015), and by 16S rRNA-based taxa composition (Yatsunenko et al, 2012). Moreover, postmortem intervals can be determined by the microbiomes' turnover in the host's decomposition (Damann et al, 2015), and in different body locations (Damann et al, 2015;Hauther et al, 2015;Johnson et al, 2016).…”

The Mathematically Describable ILD Patterns

“…The 2 nd order Gaussian function with six-coefficients was presented (1), and the fit options for the Gaussian models parameters and goodness of fit statistics were also explained in detail, Appendix. The 2 nd order Gaussian functions used in the model fitted all seven data cluster sets.…”

“…2. The closer the values of SSE and RMSE are to zero the better the fit, in contrast to a higher value for the R 2 and Adj.R 2 approaching one being a better fit (Appendix) for the 1 st order polynomial model of the coefficients (a 1 , b 1 , c 1 , a 2 , b 2 , and c 2 ), Table 3.…”

“…The non-linearity variations of "σ" values in seven pdf models are found to be inconclusive for this work. The reasons for an inconclusive result was; (i) the linearity test was not based on only "µ" parameters, and (ii) The "σ" and "µ" parameters were both data dependent, meaning that these parameters (from a normal pdf in (3)) could also be expressed as part of a cluster where the coefficients (a 1 , b 1 , and c 1 ) of the 1 st order Gaussian function (2) are not data dependent. These coefficients are varied and only one out of the seven ILD data clusters were selected.…”

“…There have been many modelling applications in machine learning, and considering to derive an algorithm from biological data sets; for recently examples, using the taxonomic levels of microbiomes information that helps to predict the time of death of the human cadaver [1] or predicting relationship trouble from vocal patterns in couples' speech [2], or developing hearing devices [3]. The patterns used for this study also came from the biological data sets, in next subsection.…”