Artificial intelligence in forensic genetics: a multi-database bibliometric and science-mapping analysis (2003-2026).

Abstract

Purpose: Artificial intelligence, machine learning and computational prediction are entering forensic genetic workflows, from probabilistic DNA-profile interpretation to the prediction of externally visible characteristics, biogeographical ancestry and age. This study maps the structure of that literature and interprets it in terms of translational maturity, to support research prioritisation and validation agendas.

Methods: Two complementary retrieval arms were executed between 10 and 13 September 2026: a venue-anchored search of 80 forensic and genetics journals derived from the Scopus Source List, and a topic-only probe with no source restriction, both applied to Scopus and Web of Science. The probe was introduced to measure what a source-restricted strategy costs in recall, and was incorporated into the design once more than half of the additional records it returned proved eligible. All 728 unique records retrieved were screened for eligibility by two independent raters, with agreement computed on the independent codings (Cohen kappa 0.708 to 0.806). Analyses were computed on three corpora: performance indicators on all documents, conceptual structure on the records carrying index terms, and citation structure on Web of Science coverage.

Results: The corpus comprises 520 documents from 108 sources and 1,908 authors, with 15.63% annual growth over complete years and 23.86 citations per document; four documents in five were published from 2017 onwards. The literature is organised around a single publishing nucleus accounting for 39.0% of documents and forming the entire first Bradford zone, alongside a dispersed tail of 100 sources that is more recent and substantially less cited; a quarter of the eligible literature lies outside the venue-anchored frame. Index-term co-occurrence analysis resolves into three thematic neighbourhoods spanning marker-based profiling and probabilistic interpretation, epigenetic age estimation, and sequencing with body-fluid identification. The number of clusters depends on the resolution parameter and on the composition of the corpus, and is reported with a formal stability analysis rather than as a finding. Contemporary neural paradigms are largely absent, and the two-arm design shows this absence to be a property of the field rather than of where it publishes.

Conclusion: Artificial intelligence has entered forensic genetics along distinct trajectories that differ in validation maturity, operational implementation and evidential admissibility. Validation conducted prospectively and across laboratories, uncertainty expressed in evidential terms, and reasoning that can be contested in court, rather than predictive performance alone, determine readiness for casework.