Classification of Common Relationships Based on Short Tandem Repeat Profiles Using Data Mining

Korean J Leg Med 2019;43:97-105

Department of Statistics, Korea University, Seoul, Korea,

Product Development HQ,  Dong-A ST, Seoul, Korea

Department of Forensic Medicine, Seoul National University College of Medicine, Seoul, Korea

Forensic Science Division 2, Supreme Prosecutor’s Office, Seoul, Korea

E-mail: jael@korea.ac.kr

We reviewed past studies on the identification of familial relationships using 22 short tandem repeat markers. As a result, we can obtain a high discrimination power and a relatively accurate cut-off value in parent-child and full sibling relationships. However, in the case of pairs of uncle-nephew or cousin, we found a limit of low discrimination power of the likelihood ratio (LR) method. Therefore, we compare the LR ranking method and data mining techniques (e.g., logistic regression, linear discriminant analysis, diagonal linear discriminant analysis, diagonal quadratic discriminant analysis, K-nearest neighbor, classification and regression trees, support vector machines, random forest [RF], and penalized multivariate analysis) that can be applied to identify familial relationships, and provide a guideline for choosing the most appropriate model under a given situation. RF, one of the data mining techniques, was found to be more accurate than other methods. The accuracy of RF is 99.99% for parentchild, 99.44% for full siblings, 90.34% for uncle-nephew, and 79.69% for first cousins.