Evaluating machine learning and neural network architectures for forensic sex estimation using mandibular ramus and...

Estimating a biological profile, such as sex, is a fundamental step in forensic identification when primary identifiers are unavailable for direct individual comparison. In forensic scenarios involving advanced decay, specific taphonomic alterations, or midfacial blunt force impacts, the mandibular ramus serves as a valuable anatomical marker due to its distinct sexual dimorphism and thick cortical structure, making it more resilient to fragmentation than other, more fragile facial bones. Despite its utility, the application of machine learning (ML) and artificial neural networks (ANN) to evaluate mandibular measurements in Indonesian populations remains limited by sample size constraints and predictive variability. This study analyzed three continuous vertical parameters (condylar-ramus height, projective ramus height, and coronoid-ramus height) and one categorical morphological variable (sigmoid notch) from 1,000 digital panoramic radiographs of Indonesian individuals, employing ten ML algorithms and three ANN architectures. A tenfold nested cross-validation approach was adopted to validate each model and ensure unbiased performance estimation. The optimized ANN and Logistic Regression frameworks exhibited the highest overall discriminative power (AUC > 0.99), while the Random Forest algorithm achieved the peak classification accuracy (97.10%). SHAP analysis confirmed that these vertical parameters collectively contribute to model performance, with condylar-ramus height identified as the most influential feature, which aligns with known patterns of male vertical mandibular growth. This study confirms that integrating these dimensions with computational models creates a reliable tool for forensic sex estimation. However, as this research relied on a localized dataset, further multicenter studies involving diverse populations are necessary to establish broader forensic standards.