Adjusted binary classification (ABC) model in forensic science: an example on sex classification from handprint dimensions (CROSBI ID 290009)
Prilog u časopisu | izvorni znanstveni rad | međunarodna recenzija
Podaci o odgovornosti
Jerković, Ivan ; Kolić, Andrea ; Kružić, Ivana ; Anđelinović, Šimun ; Bašić, Željana
engleski
Adjusted binary classification (ABC) model in forensic science: an example on sex classification from handprint dimensions
Binary classification techniques are commonly used in forensic examination to test if a specimen belongs to a particular group and base the expert opinion on the questioned evidence. However, most of the currently used methods do not achieve sufficient accuracy due to the ignoring of the specimens classified in the overlapping area. To address the issue, we proposed a novel Adjusted binary classification (ABC) algorithm that automatically adjusts posterior probabilities to reach classification accuracy and positive/negative predicted values (PPV, NPV) of 95%. In the presented example, we used three handprint measurements from 160 participants (80 males and 80 females) to develop models that would classify sex from their dimensions. The sample was split into the training/cross-validated (70%) and testing sample (30%). We developed four classification models using linear discriminant analysis (LDA) by employing traditional single cut-off values and ABC approach that for each group provides a specific posterior probability cut-off threshold. In the cross- validated sample, the accuracy of traditional models was 78.7-92.5%, while PPVs/NPVs ranged between 78.2 and 93%. ABC models provided 95% accuracy, PPV, and NPV, and could classify 35.5-88.1% of specimens. In the testing sample, ABC models achieved accuracy of 97.3-100%, PPV/NPV 95.4-100%, and could be applied to 29.1-87.5% of specimens. The study demonstrated that the ABC approach could adjust classification models to reach predefined values of accuracy, PPV, and NPV. Therefore, it could be an efficient tool for conducting a binary classification in forensic settings and minimizing the possibilities of incorrect classifications.
binary classification ; discriminant analysis ; posterior probability ; forensic science ; handprints
nije evidentirano
nije evidentirano
nije evidentirano
nije evidentirano
nije evidentirano
nije evidentirano
Podaci o izdanju
320
2021.
110709
8
objavljeno
0379-0738
1872-6283
10.1016/j.forsciint.2021.110709
Povezanost rada
Kognitivna znanost (prirodne, tehničke, biomedicina i zdravstvo, društvene i humanističke znanosti), Sigurnosne i obrambene znanosti, Temeljne medicinske znanosti