Meta-Heuristic Feature Optimization for Classifying Multimodal HRV Datasets for Stress Prediction: An Explainable Approach Towards Ethical AI

Abstract Heart Rate Variability (HRV) is intimately associated with stress and can serve as a valuable indicator of the individual's stress level. HRV is the variation in the length of time between heartbeats. It demonstrates the adaptability and versatility of the autonomic nervous system, which regulates the body's stress response. It's a measure of the autonomic nervous system's activity, specifically the ratio of the sympathetic and parasympathetic branches. HRV has been the subject of an extensive investigation regarding its association with stress and can provide insightful information regarding the physiological response to stress in a particular individual. Lower HRV is associated with higher stress levels, while higher HRV indicates better stress resilience and adaptability. The HRV parameters can be classified as - time-domain, frequency-domain, and non-linear. Parameters associated with HRV can be employed to assess individuals' health and observe the effects of interventions such as exercise, stress reduction, and medication. Research in the field of artificial intelligence is ongoing that attempts to classify stress based on HRV data. HRV, which is associated with stress and physiological health, has received attention as a potential component to incorporate into models that classify and predict stress levels accurately. The primary goal of this study is to develop accurate, reliable, explainable, and ethical AI models that can classify stress levels based on HRV data. In this regard, we have considered the well-known multimodal SWELL knowledge work (SWELL-KW) dataset that represents the following three stress conditions - no stress, time pressure, and interruption. We have explored feature selection and dimensionality reduction techniques to extract relevant features from HRV signals and enhance classification accuracy. We have used various machine-learning techniques (e.g., traditional and ensemble methods) to analyze imbalanced and balanced HRV data and classify stress levels. We have used different oversampling techniques, such as Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic (ADASYN) to generate synthetic samples from the minority class. We have explored the Local Interpretable Model-Agnostic Explanations (LIME) to explain the model classifications. As the HRV features are non-linear, the genetic algorithm-based feature selection and followed by, model classification with Random Forest Classifier have produced the highest classification result on both the imbalanced and balanced datasets. The optimized feature set can be beneficial to design and develop a stress management system. Moreover, it's crucial to recognize that while HRV is a valuable non-invasive indicator of stress levels, its interpretation should be combined with other subjective and objective measures of stress in order to comprehensively understand the individual's stress response. Therefore, the monitoring of HRV may help individuals assess the effectiveness of stress management techniques and interventions. Furthermore, we have introduced the concept of domain ontology to represent the obtained knowledge.