J Forensic Leg Med. 2026 Apr 15;120:103134. doi: 10.1016/j.jflm.2026.103134. Online ahead of print.
ABSTRACT
Sex classification using biometric traits is vital in forensic identification. While traditional methods like Linear Discriminant Analysis (LDA) have been widely used, recent advances in supervised machine learning (ML) offer potential improvements in accuracy and robustness. This study investigates the effectiveness of several ML classifiers compared to conventional LDA for sex classification based on earprint morphometry and indices. A dataset comprising 423 individuals was examined to assess sexual dimorphism in earprint morphometry; however, for sex classification modelling, only 351 individuals with complete discriminating variables were included. Twelve distinct earprint measurements were analyzed, and four derived earprint indices were computed. Sex classification was performed using traditional Linear Discriminant Analysis (LDA) alongside nine machine learning algorithms: Boosting, Support Vector Machine (SVM), Decision Tree, K-Nearest Neighbors (KNN), Neural Network, Logistic and Multinomial Regression, Linear Discriminant Analysis, Naïve Bayes, and Random Forest. Performance was benchmarked using Classification Accuracy, Positive Predictive Value (PPV), Negative Predictive Value (NPV), F1 Score, and Matthews Correlation Coefficient (MCC). Statistically significant sex-based differences (p < 0.05) were observed in the majority of earprint dimensions, with males typically exhibiting larger values. Moreover, conventional LDA achieved the highest overall accuracy (77.8%) and F1 score (0.82) on morphometric data, closely followed by Boosting, ML-LDA, and Decision Tree models (∼75.7%). In contrast, Neural Networks performed poorly (accuracy = 32.9%). KNN and Logistic Regression performed best on earprint indices (accuracy = 74.3% and 72.9%, respectively), while Neural Networks again underperformed (accuracy = 41.4%). MCC scores confirmed model reliability, with LDA and ML classifiers outperforming Neural Networks across both datasets. Morphometric earprint data outperform earprint indices in forensic sex classification. Traditional LDA remains robust, but ML models such as Boosting, Decision Trees, and Logistic Regression offer comparable alternatives. Neural Networks showed poor performance, likely due to overfitting and limited sample size.
PMID:42001636 | DOI:10.1016/j.jflm.2026.103134
