Interpretable machine learning for individualized sex estimation from long bones.

Area of Science:

• Forensic Anthropology
• Bioinformatics
• Computational Biology

Background:

• Sex estimation is critical for identifying individuals from skeletal remains.
• Accurate sex estimation enhances the reliability of other biological profile analyses, such as age and stature.
• Current methods face challenges with interpretability, particularly artificial intelligence (AI) models.

Purpose of the Study:

• To investigate the application of machine learning (ML) techniques for sex estimation from skeletal remains.
• To address the 'black box' problem in AI by focusing on model interpretability.
• To evaluate the accuracy and transparency of different ML algorithms in forensic sex estimation.

Main Methods:

• Utilized a dataset of long bone measurements from 2,969 individuals.
• Evaluated 12 different machine learning algorithms.
• Employed iterative regression imputation for handling missing data and SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• Linear Discriminant Analysis (LDA) achieved the highest accuracy at 95.2%.
• SHAP values provided individualized insights into prediction factors, enhancing model transparency.
• Missing data significantly impacted prediction accuracy, highlighting the need for robust data handling.

Conclusions:

• Machine learning models can effectively balance accuracy and interpretability in forensic sex estimation.
• Interpretable AI models offer personalized, actionable insights crucial for legal and scientific standards.
• This research advances AI-driven methods for defensible, individualized evidence in forensic science.