Artificial neural networks reconstruct missing perikymata in worn teeth
Dental evolutionary studies in hominins are key to understanding how our ancestors and close fossil relatives grew from the early stages of embryogenesis into adults. In a sense, teeth are like an airplane’s ‘black box’ as they record important variables for assessing developmental timing, enabling comparisons within and between populations, species, and genera. The ability to discern this type of nuanced information is embedded in the nature of how tooth enamel and dentin form: incrementally and over years. This incremental growth leaves chronological indicators in the histological structure of enamel, visible on the crown surface as perikymata. These structures are used in the process of reconstructing the rate and timing of tooth formation. Unfortunately, the developmentally earliest growth lines in lateral enamel are quickly lost to wear once the tooth crown erupts. We developed a method to reconstruct these earliest, missing perilymata from worn teeth through knowledge of the later-developed, visible perikymata for all tooth types (incisors, canines, premolars, and molars) using a modern human dataset. Building on our previous research using polynomial regressions, here we describe an artificial neural networks (ANN) method. This new ANN method mostly predicts within 2 counts the number of perikymata present in each of the first three deciles of the crown height for all tooth types. Our ANN method for estimating perikymata lost through wear has two immediate benefits: more accurate values can be produced and worn teeth can be included in dental research. This tool is available on the open-source platform R within the package teethR released under GPL v3.0 license, enabling other researchers the opportunity to expand their datasets for studies of periodicity in histological growth, dental development, and evolution.