Emotion classification using gait biomechanics and machine learning.

BACKGROUND: Emerging research suggests that emotions alter the way people walk, suggesting gait data may serve as a potential source for emotion recognition. Compared to traditional modalities (i.e., facial expressions, speech), gait-based emotion detection may have advantages, including reduced susceptibility to deliberate manipulation. RESEARCH QUESTION: Is it feasible to recognize emotional states using 3D gait biomechanics and machine learning? METHODS: Fifteen healthy young adults participated in this study, performing gait trials while recalling autobiographical memories to elicit five target emotions: anger, sadness, joy, fear, and neutral. Gait biomechanics were recorded using a 3D optoelectronic motion capture system, and 155 biomechanical variables were extracted for analysis. Five machine learning algorithms, K-Nearest Neighbors, Logistic Regression, Random Forest, and Multi-layer Perceptron, and eXtreme Gradient Boosting (XGBoost), were evaluated using Leave-One-Participant-Out cross-validation and Synthetic Minority Over-sampling Technique to handle class imbalance. RESULTS: Machine learning models classified emotional states (anger, sadness, joy, fear) with accuracy higher than chance (59 % vs. 25 %). XGBoost showed the highest performance (59 % accuracy) using the top 20 biomechanical variables ranked by a decision tree entropy index. Among the emotions, sadness was detected most accurately (66 %). SIGNIFICANCE: Our findings demonstrate that 3D gait analysis combined with machine learning holds promise for an alternative way of emotion recognition. This study may provide foundational evidence supporting the development of tools for the early detection of emotional fluctuations in mental health conditions such as bipolar disorder.