Abstract
INTRODUCTION: Rapid inference of ancestral origin fromDNA evidence is critical in time-sensitive forensic investigations, particularly during the initial hours when crucial investigative decisions must be made. Although comprehensive analyses using multiple genetic markers provide thorough results, they often require significant processing time and resources. Y-chromosome short tandem repeats (Y-STRs) exhibit population-specific allelic distributions that facilitate rapid analysis, making them particularly valuable for initial screening in forensic contexts. METHODS: This study aims to enhance population classification accuracy using Y-STR profile analysis, with a particular focus on Northeast Asian populations that are often merged into a single group by commercial ancestry panels. We developed a machine learning architecture centered on an attention-based ensemble mechanism that incorporates three complementary algorithms: a One-vs-Rest Random Forest, XGBoost, and Logistic Regression, each configured to effectively manage imbalanced datasets. RESULTS: Utilizing only Y-STR data, the model achieved an overall accuracy of 80%-81% and demonstrated high stability. Notably, the model effectively processes imbalanced datasets, generating reliable outcomes for rapid ancestry assessment in time-critical investigations. DISCUSSION: By addressing a key limitation in commercial ancestry panels--their failure to differentiate among Northeast Asian subpopulations--this framework provides valuable preliminary guidance in forensic cases involving Asian individuals. Consequently, our approach enhances rapid screening capabilities, which can inform early-stage investigations while complementing subsequent, more comprehensive genetic analyses.