Exploring sex classification from earprints - A comparison of supervised machine learning algorithms and conventional...

Area of Science:

• Forensic Science
• Biometrics
• Machine Learning

Background:

• Biometric trait analysis is crucial for forensic identification.
• Traditional methods like Linear Discriminant Analysis (LDA) are established for sex classification.
• Machine learning (ML) offers advanced techniques for improved accuracy and robustness in biometric analysis.

Purpose of the Study:

• To compare the effectiveness of various ML classifiers against conventional LDA for sex classification using earprint data.
• To evaluate earprint morphometry and derived indices for their utility in sex determination.
• To benchmark performance using metrics like accuracy, PPV, NPV, F1 Score, and MCC.

Main Methods:

• Analysis of 12 earprint measurements and 4 derived indices from 351 individuals.
• Sex classification using traditional LDA and nine ML algorithms (Boosting, SVM, Decision Tree, KNN, Neural Network, Logistic/Multinomial Regression, ML-LDA, Naïve Bayes, Random Forest).
• Performance evaluation using standard classification metrics.

Main Results:

• Significant sex-based differences observed in earprint dimensions, with males generally having larger measurements.
• Conventional LDA achieved the highest accuracy (77.8%) and F1 score (0.82) on morphometric data.
• ML models (Boosting, ML-LDA, Decision Tree) showed comparable performance to LDA; Neural Networks underperformed significantly on both morphometric and index data.

Conclusions:

• Morphometric earprint data are superior to earprint indices for forensic sex classification.
• Traditional LDA remains a robust method, with ML classifiers like Boosting, Decision Trees, and Logistic Regression serving as viable alternatives.
• Neural Networks demonstrated poor performance, potentially due to overfitting and limited sample size, highlighting the need for careful model selection and data considerations.