Determining the extent to which large animal food resources influenced our evolution relies on the ability to accurately identify early hominins as agents of bone modification. Analyses of bone fracture patterns form a substantial body of research into early hominin subsistence patterns and has major implications for early human evolution. Yet bone breakage patterns are still largely described and analyzed using categorical data that do not always reveal which agents broke the bones. Current quantitative methods offer limited, localized information about breakage patterns. When using three-dimensional (3D) models of bone fragments, more data can be rapidly extracted and processed which invites opportunities to apply machine learning as a possible avenue for such taphonomic research.

This paper investigates how state of the art machine learning algorithms perform when classifying the agent of breakage of bone fragments. In our experiment, cervid long bones were broken by humans using hammerstone and anvil, or by spotted hyenas (Crocuta crocuta). Then, 3D surface meshes of the subsequent bone fragments were created using CT scans and the batch artifact scanning protocol (Yezzi-Woodley et al., 2022). We introduce new feature sets, using computational tools such as the virtual goniometer (Yezzi-Woodley et al., 2021) that provide more detailed information about each bone fragment, extracted in a fully replicable manner using semi-automatic and automatic image processing tools. Our analysis achieved an average mean accuracy of 77% across tests. Overcoming issues of equifinality would strengthen these analyses.

Early research using fracture patterns relied on descriptions of observed break patterns (Dart, 1957, 1959b; Martin, 1910). Since then, researchers have established various criteria for distinguishing different types of breakage (Bunn, 1983; Johnson, 1985; Karr and Outram, 2012a, 2012b, 2015; Morlan, 1984; Outram, 2001, 2002; Pickering, 2002; Villa and Mahieu, 1991). Quantitative assessments of breakage angles began when Capaldo and Blumenschine (1994) developed a method for analyzing notches to differentiate fragments produced by anthropogenic and carnivore breakage (see also, Galán et al., 2009). Alcántara-García et al. (2006) began measuring single fracture angles on breaks in the middle of each fracture edge with a handheld goniometer though De Juana and Dominguez-Rodrigo (2011) and Coil et al. (2017) found that fracture angles can vary based on species and skeletal element. This variation might also be attributed to the fact that a single measurement was taken per break and might not sufficiently represent the break. Most of these methods lack a measurable level of precision and accuracy and are not sufficient (O’Neill et al., 2020; Yezzi-Woodley et al., 2021) for addressing more complex and nuanced questions about site formation processes that are subject to equifinality. The challenges presented by equifinality are exacerbated by concerns over inter- and intra-analyst error and intense disagreement among research groups about the validity of methods employed (e.g., Domínguez-Rodrigo et al., 2017b, 2019; Harris et al., 2017; James and Thompson, 2015; Merritt et al., 2019).

The application of sophisticated modern tools capable of extracting more replicable information about breakage patterns could resolve long-standing debates over early hominin subsistence patterns at important paleoanthropological sites such as Dikika (Domínguez-Rodrigo et al., 2010, 2011, 2012; McPherron et al., 2010; Thompson et al., 2015) and FLK 22 (see Pante et al., 2012, 2015; Domínguez-Rodrigo et al., 2014; Parkinson, 2018, and citations contained therein). Researchers have developed and are continuing to develop various methods for analyzing bone surface modifications and fracture patterns through approaches such as geometric morphometrics (e.g., Arriaza et al., 2017; Courtenay et al., 2019a, 2019b, 2019c; Maté-González et al., 2019a; Otárola-Castillo et al., 2018, 2023; Palomeque-González et al., 2017; Yravedra et al., 2017, 2018), confocal profilometry (e.g., Braun et al., 2016; Gümrükçu and Pante, 2018; Pante et al., 2017; Schmidt et al., 2012), and other digital data extraction methods (e.g., Bello et al., 2011; Macdonald et al., 2022; O’Neill et al., 2020; Yezzi-Woodley et al., 2021). Many of these new methods rely on digital imaging, in particular 3D scanning, which has become a prominent avenue of research within the field (e.g., Maté-González et al., 2019b; Yezzi-Woodley et al., 2022). This enables powerful computational tools for data analysis such as machine learning to be employed in bone modification studies (Arriaza et al., 2021; Byeon et al., 2019; Cifuentes-Alcobendas and Domínguez-Rodrigo, 2019; Courtenay et al., 2019a, 2019b, 2020; Domínguez-Rodrigo, 2019; Domínguez-Rodrigo and Baquedano, 2018; Domínguez-Rodrigo et al., 2017a, 2021; Jiménez-García et al., 2020a, 2020b; Moclán et al., 2019, 2020; Pizarro-Monzo and Domínguez-Rodrigo, 2020). Most of these applications of machine learning have focused on discriminating bone surface modifications, rather than fracture pattern analysis. Additionally, there has been limited input on the proper implementation of machine learning by trained computer scientists and machine learning experts, leading to inadvertent misuses and spurious results with falsely elevated accuracy rates (Calder et al., 2022; Holcom et al., 2022; McPherron et al., 2022).

Machine learning is a powerful tool that has the potential for producing fundamental advances in the analysis of bone fracture patterns. Here we introduce a method that can be easily replicated by independent research teams and, as we will demonstrate, holds great potential for deciphering when and where hominins first began exploiting bone marrow and clarifying its importance and influence on their evolution.