Optimized XGBoost for Multimodal Affective State Classification Using In-Ear PPG and Behind-the-Ear EEG Signals.

Area of Science:

• Physiological computing
• Affective computing
• Wearable technology

Background:

• Emotion recognition using physiological signals is advancing.
• Traditional electroencephalography (EEG) and photoplethysmography (PPG) methods can be uncomfortable.
• Novel wearable device designs are needed for improved user comfort and data acquisition.

Purpose of the Study:

• To introduce a novel in-ear wearable device for simultaneous EEG and PPG signal capture.
• To enhance user comfort during physiological data collection for emotion recognition.
• To evaluate the effectiveness of combined physiological signals and machine learning for accurate emotion identification.

Main Methods:

• Collected EEG and PPG data from 21 participants across four emotional states (fear, happy, calm, sad).
• Preprocessed signals, extracted temporal and frequency domain features, and selected features using the ReliefF algorithm.
• Employed an XGBoost classifier with Bayesian hyperparameter tuning for emotion classification.

Main Results:

• Combined EEG and PPG features significantly outperformed individual modalities.
• The optimized XGBoost classifier achieved 97.58% accuracy, 97.57% precision, 97.57% recall, and 97.58% F1 score.
• The proposed method outperformed Support Vector Machine, Decision Tree, Random Forest, and K-Nearest Neighbor classifiers.

Conclusions:

• Multimodal physiological sensing from a comfortable in-ear device enhances emotion recognition.
• Optimized machine learning models are crucial for reliable emotion characterization.
• Findings have significant implications for mental health monitoring and human-computer interaction.