Modeling Physiological Data with Deep Belief Networks.

Area of Science:

• Physiological computing
• Machine learning
• Affective computing

Background:

• Traditional physiological data analysis relies on manual feature engineering, which is time-consuming and lacks scalability.
• Automated feature extraction is crucial for efficient and effective modeling of complex physiological signals.

Purpose of the Study:

• To develop and evaluate a Deep Belief Network (DBN) system for unsupervised feature extraction from raw physiological data.
• To assess the DBN system's performance in predicting arousal, valence, and liking levels.
• To compare the DBN-derived features against expert-designed features.

Main Methods:

• Utilized a Deep Belief Network (DBN) architecture for unsupervised feature learning from 4-channel raw physiological data.
• Developed three classifiers based on the learned features to predict arousal, valence, and liking.
• Employed Gaussian Naïve Bayes with expert-designed features as a benchmark for comparison.

Main Results:

• The DBN system achieved classification accuracies of 60.9% for arousal, 51.2% for valence, and 68.4% for liking.
• The performance of the DBN-based features was comparable to state-of-the-art expert-designed features.
• Demonstrated the effectiveness of unsupervised learning for physiological feature extraction.

Conclusions:

• Deep Belief Networks (DBNs) offer an effective approach for automatic feature extraction from raw physiological data.
• DBNs provide a scalable and efficient alternative to manual feature engineering in affective computing.
• The learned features enable accurate prediction of emotional states from physiological signals.