Detecting Young Children's Deception Using Facial Expression Analysis With High Dimensional Statistical Methods.

Although deception is common in early childhood, accurately detecting deception remains a significant challenge. The present study aims to investigate the feasibility of using high-dimensional statistical methods to detect children's deception through facial expressions, with a focus on incorporating two time-dependent statistics: slope and fitness. We analyzed the facial expressions of 93 three- to six-year-old children during a guessing game. A machine vision algorithm (FACET) assessed the likelihood of nine basic facial expressions. We extracted three statistics (mean, slope and fitness) for each facial expression. Then, we used high-dimensional feature selection method to filter out irrelevant features. We adopted two parametric (logistic regression, linear discriminant analysis) and two non-parametric statistical methods (random forest, adaptive boosting) and tested both the full feature set ("full model") and the selected feature set ("2-step model"). The results showed adding time-dependent statistics (slope and fitness) improved performance, with the average balanced accuracy increasing from 47.2% to 62.0%. Feature selection further enhanced the model performance (from 60.8% to 63.3%). In addition, balanced accuracy provided a more conservative and reliable measure than the original accuracy (54.6% vs. 61.5%). These findings demonstrate the feasibility of using dynamic facial emotional features to detect deception in early childhood and highlight the potential of computational methods for advancing developmental research on emotion and social cognition.