Artificially intelligent scoring and classification engine for forensic identification. (January 2020)

• Record Type: Journal Article
• Authors: Siino, Viviane; Sears, Christopher
• Is Part Of: Forensic science international. Volume 44(2020)
• Publication Date: 2020-01
• Subjects: Pedigree analysis -- Genetic kinship analysis -- Likelihood ratio -- Bayesian network analysis -- Machine learning cascade; Forensic genetics -- Periodicals; Génétique légale -- Périodiques; Forensic genetics; Electronic journals; Periodicals; 614.1
• DOI: 10.1016/j.fsigen.2019.102162
• Languages: English
• ISSNs: 1872-4973

Abstract

Highlights: We demonstrate the first application of artificial intelligence to genetic kinship analysis. The AI significantly and reproducibly outperforms a human interpreter in classifying true matches. This method quantifies the confidence underlying likelihood ratios across the spectrum of pedigrees. Not limited by amount of genetic information or number of missing persons or presence of mutations or consanguinity. Abstract: Despite advances in genotyping technologies, traditional kinship analysis tools utilized in forensic identification have seen limited evolution and lack measures of accuracy. Here, we leverage artificial intelligence (AI) and extend the Elston-Stewart algorithm to deliver a method that provides an unprecedented level of flexibility to matching individuals with pedigrees by likelihood ratio. We designed an AI that utilizes a prediction cascade based on gradient descent logistic regression which allows for iterative solution of multi missing person scenarios. Furthermore, the AI can quantify the confidence underlying likelihood ratios across the spectrum of pedigrees, regardless of the amount of genetic information available and the number of missing persons. The algorithm accommodates an arbitrary number of generations and ancestral relationships, including multiple marriages, mutations, and consanguinity. We demonstrate that a properly trained AI significantly and reproducibly outperforms a human interpreter. We discuss published limitations of existing tools and demonstrate that they are not amenable to the size and complexity of this study. This novel method significantly improves the trade-off between sensitivity and specificity beyond the limits of traditional kinship analysis tools and introduces opportunities beyond the field of forensic genetics.