Comparison of automatic and physiologically-based feature selection methods for classifying physiological stress...

Area of Science:

• Physiological monitoring
• Machine learning applications
• Biomedical signal processing

Background:

• Physiological stress assessment is crucial for well-being.
• Heart rate variability (HRV) and pulse rate variability (PRV) are key indicators of autonomic nervous system activity.
• Distinguishing between different types of stress (e.g., postural vs. mental) using physiological signals is challenging.

Purpose of the Study:

• To evaluate machine learning algorithms for classifying physiological stress.
• To compare automatic feature selection (Akaike's criterion) with a physiology-based approach.
• To assess the effectiveness of HRV and PRV in stress classification.

Main Methods:

• Four machine learning algorithms (LDA, SVM, k-NN, Random Forest) were employed.
• Ten HRV and PRV indices from time, frequency, and information domains were analyzed.
• Data were collected from 127 healthy individuals under rest, postural, and mental stress conditions.

Main Results:

• Machine learning models could classify physiological stress, but differentiating between postural and mental stress proved difficult.
• Automatic feature selection using Akaike Information Criterion yielded better results than physiology-driven selection.
• PRV-based features demonstrated performance comparable to HRV-based features.

Conclusions:

• Specific HRV/PRV features relevant for stress classification were identified.
• The findings suggest PRV analysis is a viable option for stress monitoring with wearable devices.
• The study contributes to advancing stress assessment methodologies in clinical and real-world settings.