Anthropological estimates of the post-mortem interval (PMI) or the time since an individual died depend on understanding the morphological features of the body present at the time of examination. Such changes may include skin color, bloating, or mummification that are assumed to occur sequentially from the time of death until skeletonization, broadly indicating how long a person may have been deceased. However, there are no standards or even agreed-upon stages in which these morphological changes are observed, given the number of factors influencing human decomposition over time. This lack of standards makes the observer's reliability of morphological decomposition traits and stages tenuous and an important research subject. Since photography is widely used in forensic cases and is an affordable and replicable method of capturing the morphological changes in human decomposition, pairing two-dimensional photographs of human decomposition with the growing field of artificial intelligence (AI) may have the ability to decrease observer error and bias in estimating PMI and produce an automated process for human decomposition assessment.
Over 20,000 photographs of human decomposition from the University of Tennessee, Knoxville Anthropological Research Facility (ARF) digital catalog were categorized and utilized as training data for AI machine learning algorithms. These algorithms reflected AI's ability to accurately and repetitively categorize human decomposition utilizing two published methods and one novel method. AI performed well, correctly identifying states of decomposition using three different methods, as evidenced by average MaF1 scores above 0.8 and precision score averages of more than 85%. Results indicate that AI can produce an automated and accessible method of accurately and repetitively categorizing human decomposition. Additional raters with diverse experience and skills and an increased and diverse sample size are needed to develop AI training and performance further.
(Publisher abstract provided.)