Machine learning approaches for stature estimation from cranial computed tomography: Performance comparison with linear regression in a Northeastern Thai population

Background/Objective Stature estimation from skeletal remains is a fundamental component of biological profile construction in forensic anthropology. When long bones are unavailable, the cranium represents a viable alternative source. This study compared linear regression (LR), support vector regression (SVR), and random forest (RF) for stature estimation from cranial computed tomography (CT) measurements in Northeastern Thai adults. Methods A total of 240 subjects (120 males, 120 females) aged 20–94 years were included. Eight cranial dimensions were measured from CT reconstructions, and models were developed for combined, male, and female groups using seven cranial measurements selected on the basis of significant correlations with stature, with all models validated using leave-one-out cross-validation. Results Seven of eight cranial measurements correlated significantly with stature in the combined sample. RF achieved the highest predictive accuracy (R 2 = 0.424, RMSE = 6.975 cm), followed by LR (R 2 = 0.414, SEE = 7.109 cm) and SVR (R 2 = 0.382, RMSE = 7.227 cm), with RF producing the narrowest limits of agreement in Bland–Altman analysis. In the male subsample, LR outperformed both ML models (R 2 = 0.314, SEE = 6.206 cm). Female-specific models demonstrated limited predictive capacity across all approaches (R 2 ≤ 0.063). Conclusion Cranial CT measurements provide a viable basis for stature estimation in this population. RF offers superior performance in the combined sample, while LR remains a practical and interpretable alternative. The combined model is recommended when biological sex has not been established.