Neural Network Versus Stepwise Regression for Forensic Stature Estimation from Percutaneous Tibial Dimensions in South Sumatran Malay Adults

Abstract

Introduction: Stature estimation from skeletal elements underpins forensic biological profiling and is decisive in disaster victim identification, yet no validated osteometric standard exists for the South Sumatran Malay population, and stepwise regression assumes a linearity that skeletal growth biology does not obey.

Methods: 450 healthy South Sumatran Malay adults (225 male, 225 female) aged 20-50 years were enrolled at CMHC Research Center, Palembang, and reported per STROBE. Five percutaneous right-tibial dimensions were measured by one certified anthropologist under the Martin and Saller convention (all relative technical errors of measurement below 1.5%), and a stratified 70:30 partition withheld 135 observations for testing. Stepwise multiple linear regression (SMLR) was compared against a grid-search-optimised 5-64-32-16-1 multilayer perceptron artificial neural network (MLP-ANN).

Results: Sexual dimorphism was large for every dimension (Cohen's d 1.42-1.75; all p < 0.001). Percutaneous tibial length was the strongest single predictor (males r = 0.812, 95% CI 0.762-0.852). The calibration-corrected pooled regression reached R² = 0.742 with RMSE ±4.82 cm, whereas the network reached R² = 0.914 (95% CI 0.891-0.933) and RMSE ±2.78 cm — ΔR² = 0.172 (bootstrap 95% CI 0.141-0.203), a 23.2% gain in explained variance and a 42.3% reduction in error that narrows the 95% forensic identification window from ±9.45 cm to ±5.45 cm.

Conclusion: These first population-specific, artificial-intelligence-driven standards materially improve the biological profiling of incomplete human remains in Indonesian medicolegal and disaster victim identification practice.