Predicting Facial Soft Tissue Thickness from Skeletal Malocclusion Class and Statistically Engineered Features Using Machine Learning in Cambodian Adults

Abstract

This paper presents a two-experiment study aimed at predicting facial soft tissue thickness (FSTT) at 11 anatomical landmarks in Cambodian adults based on skeletal malocclusion class and sex. Experiment 1 (preliminary) assesses three multi-output regressors—Linear Regression (LR), Random Forest (RF), and Gradient Boosting (GB)—utilizing four baseline features and 70 engineered features derived from landmarks across seven categories of statistical transformations. Baseline models achieve a test mean absolute error (MAE) of 1.506–1.527 mm (R² = −0.002 to 0.026), whereas engineered models attain near-perfect R² (≥0.987), identified as structural target leakage: z-score and class-deviation features deterministically encode the prediction targets. Experiment 2 (systematic three-phase) addresses the leakage issue through a leakage-free feature design. Phase A reinstates the baseline (MAE 1.427–1.453 mm); Phase B broadens inputs to 16 valid class-sex interaction features across three output strategies; Phase C assesses Regressor Chains through four landmark orderings (anatomical, reverse, variance-based, and data-driven optimal) for a total of 12 configurations. Neither feature engineering (∆MAE = −0.004 mm avg.) nor chain ordering (detrimental to RF and GB; neutral for LR) produces significant enhancement. LR attains the optimal outcome in both experiments (Phase A test MAE = 1.4266 mm, R² = 0.047). Kruskal-Wallis analysis indicates Stomion and Labrale superius as the most discriminative landmarks for class differentiation. Engineered interaction features (Class_ordinal_Sex: 15.9%; ClassIII_Male: 13.1%) predominate in feature importance in Experiment 2, yet yield no predictive advantage. These results validate a performance ceiling dictated by significant intra-class FSTT variability, irrespective of feature engineering method or output structure.