Artificial Intelligence Versus Stepwise Regression for Stature Estimation from Tibial Dimensions: A Forensic Osteometric Study in South Sumatran Malay Adults

Abstract

Introduction: Stature estimation from skeletal elements constitutes a foundational component of forensic biological profiling, critically supporting disaster victim identification in disaster-prone nations such as Indonesia. Traditional Stepwise Multiple Linear Regression (SMLR), while widely employed, is constrained by linearity assumptions that inadequately model the complex, multidimensional osteometric biology of population-specific cohorts.

Methods: This cross-sectional study enrolled 450 healthy adult South Sumatran Malay participants (225 males, 225 females), aged 20–50 years, from Palembang and surrounding regencies. Five percutaneous tibial measurements were acquired under standardized protocols by a single trained anthropologist. A 70:30 stratified train-test split yielded 315 training and 135 test observations. Predictive performance of SMLR was rigorously compared against an optimized three-hidden-layer Multilayer Perceptron Artificial Neural Network (MLP-ANN).

Result: Significant sexual dimorphism was demonstrated across all variables (independent samples t-test, p < 0.001). Percutaneous Tibial Length (PTL) was the strongest individual stature predictor (males: r = 0.812; females: r = 0.795). The best SMLR pooled model (PTL + PDB + DDB) achieved R-squared = 0.742 and RMSE = ±4.82 cm. The MLP-ANN substantially outperformed SMLR across all subgroups, achieving a pooled R-squared of 0.914 and RMSE of ±2.78 cm-representing a 23.2% improvement in R-squared and a 42.3% reduction in prediction error.

Conclusion: These population-specific AI-driven standards offer forensic practitioners in the Indonesian medicolegal context a markedly more reliable tool for biological profiling of incomplete human remains.