Investigating lightweight and interpretable machine learning models for efficient and explainable stress detection.

Area of Science:

• Computational intelligence
• Biomedical signal processing
• Machine learning for healthcare

Background:

• Prolonged stress negatively impacts mental and physical health.
• Heart rate variability (HRV) is a key indicator for stress measurement.
• Accurate stress detection using limited HRV features with machine learning (ML) is challenging.

Purpose of the Study:

• To develop computationally efficient, lightweight ML models for stress detection using minimal HRV features.
• To enable real-time stress monitoring suitable for Internet of Things (IoT) deployment.
• To evaluate model performance and interpretability for practical applications.

Main Methods:

• Utilized the SWELL-KW dataset for model training and evaluation.
• Implemented efficient feature selection and hyper-parameter tuning for ML models.
• Developed and compared lightweight models, including k-nearest neighbors (k-NN) and Decision Tree.

Main Results:

• Lightweight models achieved competitive accuracy with reduced computational demands.
• The k-NN algorithm demonstrated superior performance, reaching 99.3% accuracy with only three HRV features.
• The best k-NN model maintained 99.26% accuracy on an NVIDIA Jetson Orin Nano edge device, training in 31 seconds.

Conclusions:

• Lightweight ML models, particularly k-NN, are effective for accurate and efficient stress detection from HRV.
• The proposed approach is suitable for real-time stress monitoring in resource-constrained IoT environments.
• Local interpretable model-agnostic explanations enhance the understanding of ML-based stress detection.