Machine Learning-Based Stress Detection Using Robust HRV Features and Person-Specific Normalization

This paper presents an artificial intelligence (AI) system for the electrocardiogram (ECG)-based detection of psychological stress, developed as part of a prototypical digital assistance system for nurses. The AI system employs the following key steps: After initial preprocessing including ECG windowing and heartbeat detection, 70 time, frequency, and time-frequency domain features are extracted from each window’s heartbeats. All features are descriptors of heart rate variability (HRV), a well-known stress indicator. They are primarily calculated from sophisticated estimates of NN intervals, NN interval differences, respiration rates, and autonomic nervous system (ANS) activity. Novelties of this step include an increased noise robustness of estimated respiration rates and the derivation of ANS activity features from the integral pulse frequency modulation model of heartbeat generation. Furthermore, the entire feature extraction process is explicitly designed for robustness against irregularities like beat detection errors and ectopic beats, which could severely distort HRV analysis. All features are normalized with a novel, person-specific approach based on short baseline data sections. This lightweight personalization serves to increase classification performance while avoiding costly alternatives like person-specific modeling, and actively prevents normalization data leakage. Finally, stress detection is performed using a machine learning model of type random forest, gradient boosting machine, or multilayer perceptron. Hyperparameters of all steps were rigorously tuned using Bayesian optimization and manual search on the popular datasets WESAD and SWELL-KW as well as their union. The resulting stress detection performances under leave-one-subject-out cross-validation consistently surpassed the comparable state of the art.