Linear and nonlinear analysis of multimodal physiological data for affective arousal recognition

Area of Science:

• Physiological computing
• Biomedical signal processing
• Human-computer interaction

Background:

• Accurate human arousal classification is crucial for adaptive systems.
• Peripheral bio-signals offer a non-invasive method for monitoring physiological states.
• Previous research has explored various signal processing techniques for arousal detection.

Purpose of the Study:

• To design a system for classifying human arousal into five stress levels.
• To evaluate the effectiveness of different feature extraction methods (linear and nonlinear) from bio-signals.
• To determine the optimal combination of physiological features for arousal classification.

Main Methods:

• Collected four peripheral bio-signals: photoplethysmography (PPG), galvanic skin response (GSR), and thorax and abdominal respiration (TR, AR).
• Induced five levels of mental stress using the Stroop test in 98 participants.
• Extracted statistical, frequency, geometrical, recurrence quantification analysis (RQA), and detrended fluctuation analysis (DFA) features.
• Classified arousal levels using a Naïve Bayes classifier.

Main Results:

• Classification accuracy reached 69.13% when combining linear and nonlinear features.
• The combined feature set yielded an average accuracy of 69.13%, an Intraclass Correlation Coefficient (ICC) of 88.12%, and an F1 score of 69.43%.
• Linear features alone achieved 58.45% accuracy, while nonlinear features achieved 57.1%.

Conclusions:

• Combining linear and nonlinear dynamic analysis methods enhances arousal level recognition accuracy.
• Future work should explore optimal weighting strategies for multimodal bio-signal integration.
• The findings contribute to developing more responsive and personalized human-computer interaction systems.