Time, Frequency & Complexity Analysis for Recognizing Panic States from Physiologic Time-Series

Abstract

This paper presents results of analysis performed on a physiologic time-series dataset that was collected from a wearable ECG monitoring system worn by individuals who suffer from panic disorder. Models are constructed and evaluated for distinguishing between pathologic and non-pathologic states, including panic (during panic attack), pre-panic (preceding panic attack) and non-panic (outside panic attack window). The models presented use data fusion to combine both traditional time and frequency domain heart rate variability analysis together with nonlinear/complexity analysis. The best performing model is shown to be a random forest classifier that achieves an accuracy of 97.2% and 90.7% for recognizing states of panic and pre-panic, respectively. The models presented have application in pervasive and ubiquitous mobile and wearable health management systems.

Wearable-Based Affect Recognition—A Review

“…In addition, cross-cultural data collection would be very interesting in order to facilitate universal wearable-based affect recognition. Furthermore, as shown by Grünerbl et al [182] or Rubin et al [96], affect recognition systems can find application in clinical settings. However, in order to facilitate this type of application, datasets containing data from patients with specific health conditions are required.…”

“…The pre-/post-study questionnaires can be employed as additional features or to group the participants [82,100]. In contrast, the data gathered via EMAs is often used as a subjective ground truth [13,96].…”

“…Changes in low frequency (LF) band, ranged between 0.03 and 0.15 Hz, are mostly associated with the activity of the sympathetic nervous system (SNS). In contrast, the high frequency (HF) band, ranged from 0.15 to 0.4 Hz, is believed to reflect mostly the activity of the parasympathetic nervous system (PNS) [96]. Therefore, the LF/HF ratio quantizes is a descriptive feature indicating the influence of both, SNS and PNS, on the cardiac activity.…”

“…In addition to time and frequency domain-based features, non-linear features derived from ECG data were employed successfully wearable-based AR. Rubin et al [96], for instance, presents a detailed description of non-linear ECG features (e.g., maximal Lyapunov exponent, standard deviations (SD1 and SD2) along major axes of a Poincaré plot, the SD1/SD2 ratio, sample entropy, etc.). Moreover, Valenza et al [89], aiming to detect five levels of valence and arousal, compared the performance of linear and non-linear features.…”

“…Another class of features based on the cardiac cycle are referred to as geometrical features . An example is the triangular interpolation index [89,96,155]—A histogram of the RR intervals is computed, a triangular interpolation performed, and the baseline of the distribution is computed. Finally, the respiration is known to have an impact on the ECG signal.…”