Facial micro-movements as a proxy of increasingly erratic heart rate variability while experiencing pressure pain

Figures

• 

  Figure 1. Different tasks in an assay to measure the effects of pressure pain. (A) Resting task where the person sits and is instructed to remain as still as possible. (B) Drawing task where the person draws a line connecting digits and letters in a specified order (1 A, 2 B, etc.) from beginning to end. (C) Pointing task, where the person is asked to point to a target (the full loop of forward motion to the target and backward motion toward rest). The second backward segment occurs spontaneously, without instruction, while the first motion toward the target is instructed. (D) The grooved peg task where the person picks up a grooved peg and aims it at the grooved hole on the board to insert it in the appropriate orientation. This task requires orienting the grooved peg to match the grooved hole, similar to opening a door by inserting a key into the lock. (E) Experimental assay to probe the control condition vs. the pain condition. For each task (using the pointing task as an example), the person performs the task normally, without any pain (control case). A blood pressure cuff is then placed on the non-dominant arm (the arm that is not performing the task) and inflated to 200 mmHg for the…

• 

  Figure 2. Data acquisition tools. (A) A webcam collects video of the participant for the full duration of the task. (B) OpenFace software is used to estimate the facial grid from the video of one of the authors (Dr. Elsayed), along with the 3D gaze and head orientation, saving action units responsible for the micro-motions of the facial muscles. (C) Trigeminal subdivision based on the facial nerves innervating three main regions: V1 (ophthalmic), V2 (maxillary), and V3 (mandibular). The facial grid is subdivided into V1 (red), V2 (blue), and V3 (black) subregions to study the stochasticity of the micro-motions. (D) Blood pressure cuff used to evoke the sensation of pressure pain in the non-performing arm. (E) The MC10 Biostamp sensors and phone used to collect time series data by registering different biorhythms. (F) Locations on the body where the Biostamp sensors were placed to co-register data from the biceps, anterior thigh, and heart (image acquired via registered study on the MC100 Biostamp-n-Point Cloud system, now acquired by Medidata).

• 

  Figure 3. Analytical pipeline of facial biorhythms. (A) Face grid mapped (indexed) across regions: ophthalmic V1, maxillary V2, and mandibular V3. (B) Sample XY-position traces of one grid point (19 in (A)) recorded for 35 s. (C) Smoothed trajectories. (D) Scalar speed profiles from 26 sites in region V1. (E) Absolute speed deviations from the empirically estimated mean (from the red peaks). (F) Locally scaling allometric effects using Equation 1, due to anatomical disparities in lengths, impacting speed ranges across participants. (G) Full micro-movement spikes preserving the frame of each original speed peak. (H) Frequency histogram of the MMS peaks. (I) Gamma parameter plane used to represent the empirically estimated Gamma PDFs as the shape and scale parameters with 95% confidence intervals for each parameter. (J) Parameter space spanned by the Gamma NSR (scale parameter) of each PDF thus estimated for each face region. (K) Parameter plane defined by the norms of the Gamma NSR vector spanned by the V1, V2, and V3 NSRs and the corresponding Gamma PDF skewness. (L) Empirical Gamma PDFs for each region and condition (control vs. pain).

• 

  Figure 4. An analytical pipeline for pre-processing the cardiac data. (A) Upon filtering noise and detrending the ECG data, the R-peaks are localized, and in (B), the temporal distances between the peaks are obtained. This produces the RR-peaks series data or inter-beat interval times series (IBI). (C) The temporal domain is assessed by obtaining the Poincaré plot, a shifted version of the time series one step forward along the y-axis and the current time version along the x-axis. The scatter is fitted by an ellipse, and the principal axes are examined as SD1 (variability associated with the parasympathetic system) and SD2 (variability associated with the sympathetic system). These parameters are then examined in relation to noise regimes from the empirically estimated micro-movement spikes (micro-peaks) distributions. (D) In this step, we perform frequency domain analysis using power spectral decomposition techniques and obtain the low- and high-frequency bands for examination of the sympathetic and parasympathetic regimes associated with lower and higher frequencies, respectively. A magnitude scalogram can then be used to visualize the data as frequency over time with power information…

• 

  Figure 5. The continuous gamma family of probability distributions fits the empirical IBI data well in a maximum likelihood estimation sense, with 95% confidence. (A) Control and pain conditions at rest, (B) Drawing, (C) Pointing, and (D) Peg tasks. Each (Гshape, Гscale) point on the gamma parameter plane is plotted to reveal a self-emerging pattern; the log–log scatter aligns according to a tight linear negative correlation: as the distribution becomes more symmetric (with increasing shape values), scale (dispersion) values decrease. The scale is the noise-to-signal ratio (NSR, variance over the mean). (E) The pairwise comparison of the NSR according to the rank-sum test reveals statistically significant differences (p < 0.01) for all tasks during the pain state, relative to the absolute difference between pain and control states for the resting task. (F–H) The parameter plane represents the difference in NSR vs. the difference in skewness for each task relative to the resting task, in both the control (red) and pain (blue) blocks.

• 

  Figure 6. Gamma parameters for facial micro-movements across all regions V1, V2, V3 for all 4 tasks (resting, drawing, pointing, and peg) and for control (red dots) and pain (blue dots). Each dot represents a participant’s empirically estimated (Γshape, Γscale) measurement, spanning a family of PDFs for each participant. Insets show the gamma PDFs from the empirically estimated shape and scale parameters. Scatters are well fit by exponential curves, and as in the heart IBI case of Figure 5, the log–log representation is well fit by a tight linear negative correlation whereby as the gamma NSR decreases, the gamma shape increases (see Supplementary Table 2).

• 

  Figure 7. Statistical tests of the Gamma parameters for facial micro-movements. (A) Control condition (pre-pain) shows statistically significant differences in the Gamma NSR parameter according to the non-parametric Kruskal–Wallis test for the ophthalmic area V1 across all tasks. The MATLAB Multcompare test reveals p < 0.01 for all pairwise comparisons relative to the resting task. (B) During the pain condition, these differences are accentuated. (C) Stats table for comparisons A and B and for the control vs. pain comparison according to the Friedman’s test. All comparisons yield p < 10−10. (D) Pairwise comparisons of NSR rank-sum test in ophthalmic area V1, (E) maxillary area V2 and (F) mandibular area V3.

• 

  Figure 8. Different representations of the face data capture shifts in gamma families of probability distributions from control to pain conditions and across the different tasks. (A) Parameter space spanned by V1, V2, V3 gamma NSR with control (red dots) and pain (blue dots), where each dot represents a participant, and each parameter space denotes the state of the face during the task. (B) The Euclidean norm of the parameters vector (V1, V2, V3) is obtained for both the NSR and the skewness, with the scalar quantities represented on a parameter plane spanned by the norm ГNSR and the ГSkewness. Notice the different shifting patterns (see details in the results section).

• 

  Figure 9. Individualized probability distribution families across participants are compared pairwise using the EMD and grouped into self-emerging clusters. (A) Matrix of EMD values denoting the normalized distance (color bar) for each participant in the V1 area (the most significantly impacted by tasks and pain) – PRE refers to the control (before pain condition) and POST refers to the pain condition. (B) Tree clustering of the participants into 8 groups according to the EMD values in (A), obtained from the distributions of gamma NSR parameters empirically estimated. Cluster composition of control vs. pain PDF percentages is detailed in the main text.

• 

  Figure 10. Linear correlations between the difference in gamma NSR shifting from control resting to the pain condition for each task. (A) Resting task. (B) Drawing task. (C) Pointing task. (D) Peg task. Supplementary Table 3 reports the correlation values and the goodness of fit.

• 

  Figure 11. Measures of transfer entropy during the control session for all 4 tasks and across participants. (A) TE sorted in increasing levels within the sample of participants (used as an individualized threshold) for the case where the past heart IBI & face activity predicts the current face activity. Superimposed for each participant is the TE curve (denoted face → IBI) where past heart IBI & face activity predict current heart-IBI activity. The plot with the star highlights the peg task where participants with lower values of TE (in IBI → face) express higher TE (in face → IBI) for lower thresholds and lower TE (in face → IBI) for higher thresholds (of IBI → face TE). (B) Same analysis as in (A) but inverting the process by sorting the face → IBI TE signal across participants.

• 

  Figure 12. Measures of transfer entropy during the pain session for all 4 tasks and across participants. (A) Identified pattern in the Peg Task is consistently observed during the pain state in other tasks. (B) Results from the reverse analysis also show marked differences from the control case (peg task highlighted for comparison purposes).