Optimizing Probability Threshold of Convolution Neural Network to Improve HRV-based Acute Stress Detection...

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Health Informatics

Background:

• Stress significantly impacts emotional and physical health, necessitating accurate detection and management.
• Heart Rate Variability (HRV) analysis, often derived from Electrocardiogram (ECG) signals, offers physiological markers for stress.
• Convolution Neural Networks (CNNs) show potential for automated stress detection by analyzing complex physiological data.

Purpose of the Study:

• To enhance the performance of CNNs for acute stress detection.
• To investigate the impact of probability threshold optimization on CNN-based stress detection accuracy.
• To improve the balance of classification between stress and rest states using HRV data.

Main Methods:

• A two-step training approach was proposed for CNN models.
• The core method involved optimizing the probability threshold of the CNN classifier.
• Analysis focused on HRV signals derived from ECG to detect acute stress.

Main Results:

• The average error rate in stress detection was significantly reduced from 17.3 ± 9.2% to 9.2 ± 5.7%.
• Probability threshold optimization led to more balanced classification results between stress and rest conditions.
• The proposed method improved CNN performance without requiring additional training data.

Conclusions:

• Optimizing the CNN probability threshold is an effective strategy for improving acute stress detection using HRV.
• This simple yet effective method enhances the practical applicability of HRV-based stress monitoring systems.
• The findings contribute to more reliable and balanced stress detection in real-world health applications.