Feature Extraction and Selection for Pain Recognition Using Peripheral Physiological Signals.

Area of Science:

• Physiological computing
• Affective computing
• Biomedical signal processing

Background:

• Feature selection is critical for robust pattern recognition systems, especially in emotion classification.
• Machine learning methods can be unreliable with small datasets, potentially leading to suboptimal feature choices.
• Inconsistent optimal feature sets in pain-based emotion classification literature necessitate a generalized approach.

Purpose of the Study:

• To conduct a deep exploration of pain-based emotion classification using physiological signals.
• To develop an interpretable and generalizable feature selection protocol to address literature inconsistencies.
• To visualize and understand the physiological feature space for pain-induced emotions.

Main Methods:

• Explored 155 time and frequency domain features from electromyogram (EMG), skin conductance level (SCL), and electrocardiogram (ECG) signals.
• Applied an exhaustive and interpretable feature selection protocol.
• Utilized a topological feature space visualization (Mapper) and statistical interaction analysis.

Main Results:

• Identified five main functional feature groups: signal amplitude/power, frequency, nonlinear complexity, unique, and connecting.
• Visualized associations between features, synthesizing results from prior literature.
• Observed that EMG and SCL features can functionally substitute for ECG features in pain emotion classification.

Conclusions:

• A generalized feature set for pain-based emotion classification was established through a robust protocol.
• Topological visualization revealed key physiological information sources and functional feature groupings.
• EMG and SCL offer viable alternatives to ECG, informing future research in sensing, feature design, and deep learning for emotion recognition.