Hilbert- and wavelet-derived cardiorespiratory features for machine learning classification of anxiety-inducing stimuli

Abstract Anxiety disorders are associated with measurable alterations in cardiorespiratory physiology that can be captured using wearable biosensing systems. This study investigates the classification of anxiety-inducing stimulus conditions using synchronized electrocardiography (ECG) and respiration (RSP) signals combined with Hilbert- and wavelet-derived signal features. Data were collected from 18 participants during anxiety-inducing and non-anxiety-inducing video conditions. A multimodal feature extraction framework incorporating heart rate variability, respiration rate variability, instantaneous amplitude, circular phase, instantaneous frequency, and wavelet descriptors was developed and evaluated using multiple machine learning classifiers. The primary evaluation used participant-independent nested leave-one-subject-out validation. On the original cohort, combined ECG and RSP features yielded a balanced accuracy of 0.532, close to chance level, whereas on the independent WESAD stress-versus-baseline dataset the identical pipeline achieved a balanced accuracy of 0.862, showing that the feature representation retained discriminative information across unseen individuals in a related affective task. In an exploratory within-dataset analysis, in which windows from the same participant could appear in both partitions, the Extra Trees classifier reached an accuracy of 0.968 and a compact subset of eight Hilbert- and wavelet-derived features reached an accuracy of 0.964 with a Matthews correlation coefficient of 0.894; these window-level values are within-dataset estimates rather than measures of generalization to unseen participants. Reliable generalization to unseen individuals and to measured anxiety requires larger cohorts with validated state-anxiety measurements.