Multimodal Assessment of Biomechanical and Neuromuscular Characteristics of Zygomaticus Major in Healthy Participants

Abstract

Background:

The midfacial plane overlying the zygomaticus major (ZM) muscle undergoes unique compressive/tensile forces and displacements that vary with emotional expressions. In cases of facial paralysis, surgeons attempt to recapitulate these dynamics with muscle transplantation. New technologies attempting to simulate these movements require a holistic understanding of the biomechanical and neuromuscular characteristics of the midface and lower face.

Methods:

We recruited 33 healthy participants (12 men, 21 women) and measured physical movements, forces, and surface electromyography (sEMG) characteristics of the midface associated with smiling and puckering expressions. We assessed measures of central tendency and quantified characteristic neuromuscular features. We also trained machine learning classifiers to accurately predict muscular loading based on the sEMG features.

Results:

The midfacial ZM plane experiences a maximum contraction of 9.8% ± 3.1% during smiling while exerting a force of 0.93 ± 0.48 N. Conversely, it experiences a maximum extension of 9.2% ± 3.4% and a tension of 0.47 ± 0.23 N during puckering. Furthermore, we observed the emergence of characteristic neuromuscular features associated with varying degrees of biomechanical loading, which allowed us to successfully train machine classifiers to differentiate between facial expressions with up to 86% accuracy.

Conclusions:

We have successfully performed multimodal assessment of the physical and functional parameters (displacement, force, and sEMG patterns) associated with the midfacial ZM plane. Our results present a benchmark for the quantitative assessment of the midface and lower face, which could provide metrics for functional rehabilitation or even training data for technologies that may attempt to recapitulate facial animation.

Takeaways

Question: Can multimodal assessment of the midfacial zygomaticus major (ZM) plane establish quantitative benchmarks to inform rehabilitation and the development of future facial prosthetics?

Findings: In 33 healthy participants, we measured displacement, force, and surface electromyography of the midfacial ZM plane during various facial expressions. We identified distinct biomechanical and neuromuscular signatures under varying loads and trained machine learning models to classify expressions with up to 86% accuracy.

Meaning: Multimodal quantification of ZM dynamics offers foundational data to inform rehabilitation strategies and support the development of responsive facial prosthetics.

INTRODUCTION

Facial expressions are a result of a complex interplay of numerous muscles and nerves in the facial neuromuscular system.¹ Of these, the zygomaticus major (ZM) is one of the most important facial muscles, as it is responsible for elevating the corners of the mouth, leading to smiling and laughing expressions.² The subconscious activity of the ZM is often taken for granted until its function is affected by a disease, trauma, or iatrogenic intervention. Such disfigurements can cause psychosocial distress, withdrawal from social activities, negative body image, and low mood.^3–5 Plastic surgeons regularly interact with the neuromuscular anatomy of the face through both aesthetic and reconstructive surgery, including the administration of neuromodulators, facial aesthetic surgery, craniofacial surgery, head and neck oncological surgery, and facial paralysis surgery.^6,7 Yet, a comprehensive understanding of the biomechanical and neuromuscular properties of the midface and lower face is lacking.

Attempts to decipher the physical and functional characteristics of this anatomy have been limited by their single-dimensional approaches that quantify either contraction patterns, biomechanical outputs, or neuromuscular activity. Examples include noninvasive imaging techniques for measuring dynamics of facial movements^8–11; direct measurement of force outputs in healthy participants and patients with unilateral chronic facial palsy^12,13; and surface electromyography (sEMG)–based neuromuscular recordings in healthy participants and those with diseases.^14–17 Although each of these studies established safe and noninvasive methods for applying adhesive sensors to the face to measure muscle response, they all either had relatively small cohorts (<10) or had a homogenous group of participants (eg, all men), limiting the generalizability of their results.

Therefore, we argue that objective data related to both the biomechanical and neuromuscular activity of ZM are suboptimal and scarce because no prior study has described the biomechanics and neuromuscular patterns of the ZM plane under varying loads in a large cohort. Thus, by quantifying these metrics under varying loading conditions, we have established a benchmark for assessing physiological and functional integrity. We have achieved this by developing a multimodel quantitative database of biomechanical and neuromuscular activation patterns of the ZM plane by cataloging displacements, forces, and sEMG patterns associated with standard contraction and tensile movements in healthy participants. We also demonstrate that it is possible to leverage machine learning algorithms to accurately classify facial expressions from simple quantitative sEMG features of ZM activation. Such insights are crucial for advancing medical diagnostics, informing rehabilitation strategies, improving cosmetic outcomes, and enabling precise interventions tailored to individual neuromuscular profiles.

MATERIALS AND METHODS

Study Design

For this study, we recruited 33 healthy adults (12 men and 21 women; average age = 30.9 y; age range = 19–68 y). Participants with prior facial injuries or pathologies were excluded. During the study, force and EMG sensors were placed along the plane overlying the ZM on one side of the participants’ faces. They were subsequently asked to perform smiling and puckering actions (5 movements with 5-s hold for each movement). The smiling and puckering actions were further distinguished between “regular” and “maximum,” with regular defined as “what feels natural to the participant” and “maximum” referring to the maximal voluntary action the participant could perform subjectively. To minimize motion artifacts, we asked each participant to consciously avoid unnecessary movement of their heads and bodies. All participants who completed the study received a $25 ClinCard. This study protocol was reviewed and approved by the institutional review board of the University of Pennsylvania (protocol no. 853043).

Displacement Recording and Analysis

For static displacement measurements, the participants were asked to hold smiles or puckers while the distance between the central tragus and the lateral commissure was measured (Fig. 1). The central tragus was chosen given its stationary position and natural landmark for visual identification of bifacial width. Displacement and normalized displacement were calculated as follows:

Fig. 1.

Measurement setup. Schematics of the location of the ZM (A) and the experimental setup for force and sEMG measurements (B).

${\displaystyle \begin{array}{c}\Delta d=d_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}-d_{\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{t}}.\end{array}}$ (1)

${\displaystyle \begin{array}{c}\mathrm{n}\mathrm{o}\mathrm{r}\mathrm{m}.\Delta d=\frac{\Delta d}{d_{\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{t}}}*100\end{array}.}$ (2)

For static measurements, data from 32 of 33 participants are presented because, for 1 participant, the data were incorrectly recorded.

For dynamic displacement analysis, black markers were placed on the face over the zygomaticus origin and the lateral commissure to facilitate motion tracking. We selected the external locations of these markers based on a recent study describing the anatomy of the ZM muscle.¹⁸ The participants were asked to perform smiling and puckering actions (5 movements with 5-s hold for each movement) while being video-recorded from the front using a camera (iPhone, Apple). Dynamic displacement analysis was performed using Kinovea, where a unique coordinate system was established for each participant, and the scale was set using the ruler taped to the participant’s forehead. A vertical line was then drawn from the middle of the eyebrows down the nose to capture the natural tilt of the head and set the horizontal and vertical axes. Using this calibrated coordinate system, the movement of the black markers was tracked in horizontal and vertical directions. Horizontal and vertical distances were used to calculate the vector distances using the Pythagorean theorem. The individual displacements were then calculated according to equations (1) and (2). For dynamic measurements, data from 32 of 33 participants are presented, where both markers were distinctly visible and properly tracked.

Force Measurements and Data Analysis

Force measurements were performed using a custom assembly with a LSB205 S-Beam Load Cell (Futek). Before placing the load cell assembly over the ZM, the skin was gently cleaned with an alcohol wipe. The anchor points of the load cell assembly were placed over the lateral commissure and the zygomaticus origin using medical tape and a double-sided adhesive (Fig. 1). Force data were acquired at 100 Hz using the SENSIT Test and Measurement software (Futek).

For force measurements, smiling and puckering data from 31 of 33 and 30 of 33 participants, respectively, are presented. Participants with mechanical failure of the adhesive between the load cell and the face during measurements were excluded.

sEMG Measurements

We recorded sEMG activity in a bipolar configuration using a Shimmer 3 EXG Unit (Shimmer) and gelled Natus electrodes. Before placing the electrodes on the participant, their skin was gently cleaned with an alcohol wipe. A pair of gelled electrodes were placed over the muscle belly of the ZM, and a reference electrode was placed over the bony mastoid (Fig. 1B). The sEMG data were acquired at 512 Hz using the Consensys software (Shimmer).

For sEMG measurements, smiling and puckering data from 31 of 33 and 30 of 33 participants, respectively, are presented. Participants with detached sEMG electrodes during the measurements were excluded.

sEMG Data Analysis

Raw sEMG data were filtered using a high-pass, second-order Butterworth filter with a cutoff frequency of 10 Hz, along with iterative notch filters at 60 Hz and harmonics to minimize noise artifacts. The root mean square of the filtered sEMG data (EMG_RMS) was calculated over 200 milliseconds with a 50-milliseconds sliding window. EMG_RMS activity was normalized for each participant using the maximum EMG_RMS recorded across all actions for that participant. Normalized EMG_RMS activity for each action and the baseline activity without any actions were then segmented.

Power spectral density (PSD) estimate was calculated from the raw sEMG data with only a high-pass, second-order Butterworth filter with a cutoff frequency of 10 Hz. Welch PSD estimates for each action were then computed using the built-in MATLAB function pwelch(). Noise in the PSD due to 60 Hz interference and its harmonics was suppressed using interpolation. The average PSD for each expression for a given participant was computed from the PSD of each corresponding action. The average PSD of baseline activity was computed as the average of PSD activity from baseline segments during smiling and puckering data acquisition sessions. Two-sample Kolmogorov–Smirnov tests were conducted using the average PSD for each expression.

The mean and median frequencies were averaged across the mean and median frequencies of the PSD of each action. The total band power was computed as the area under the PSD curve for each action and averaged across all actions for a given expression.

Classifying Facial Expressions

The following features were extracted from the sEMG recordings for each regular and maximum smiling and puckering action: maximum EMG_RMS, zero-crossing rate, mean frequency of PSD, median frequency of PSD, and total band power of the PSD. Three different kinds of classifiers (decision tree, quadratic discriminant, and linear support vector machine) were trained using these features with 5-fold cross-validation. Briefly, the data were randomly divided into 5 equal parts at the subject level. Each classifier was trained and validated 5 times, each being used once, as the validation set and the remaining 4 folds as the training set.^19–21 The average prediction accuracy was calculated across the 5 folds.

Statistical Analysis

All statistical analyses were performed using MATLAB (MathWorks). Statistical significance was determined using a 1-way analysis of variance with post hoc Tukey testing. Statistical significance was established at a P value of less than 0.05.

RESULTS

Contraction and Extension of ZM

The ZM originates at the zygomatic bone and inserts diagonally into the modiolus.¹⁸ To directly study the biomechanics of the facial plane overlying the ZM, we focused our attention on the smiling and puckering actions, as these actions involve direct contraction and extension with minimal shear stresses. Measuring static expressions, we observed that the contraction during regular and maximum smiling was 0.6 ± 0.3 cm and 1.1 ± 0.3 cm, respectively (Table 1). On the other hand, regular and maximum puckering led to the extension of the plane by 0.5 ± 0.3 cm and 1.0 ± 0.4 cm, respectively (Table 1). Analysis of dynamic actions revealed that when smiling or puckering, the plane undergoes greater displacements in the horizontal direction than in the vertical direction (Table 1).

Table 1.

Average Displacements Under Active and Dynamic Conditions

Action	Static		Dynamic
	$\Delta d$, cm	norm.∆d, %	norm.∆d-horizontal, %	norm.∆d-vertical, %	norm.∆d-vector, %
Regular smile	−0.6 ± 0.3	−5.5 ± 2.7	−7.0 ± 4.0	−5.5 ± 3.8	−4.9 ± 3.7
Maximum smile	−1.1 ± 0.3	−9.8 ± 3.1	−11.3 ± 6.5	−7.9 ± 5.0	−7.4 ± 5.2
Regular pucker	+0.5 ± 0.3	+4.9 ± 2.7	+6.6 ± 3.4	+4.4 ± 3.2	+3.9 ± 2.3
Maximum pucker	+1.0 ± 0.4	+9.2 ± 3.4	+8.3 ± 3.8	+7.6 ± 5.7	+5.0 ± 3.2

Data presented as mean ± SD (n = 32 participants with 4–5 movements for each action for dynamic conditions). + and – signs indicate extension and contraction during the movement, respectively.

Contractile and Tensile Forces by the ZM Midfacial Plane

To accurately measure the forces produced, we placed a custom load cell between the commissure and a fixed point on the zygoma where the ZM originates (Fig. 1B). Because smiling requires contraction of the ZM, contractile (negative) forces are produced, whereas puckering causes extension of the ZM, leading to tensile (positive) forces being experienced by the muscle (Fig. 2). Intuitively, as the intensity of smiling or puckering increased, the absolute magnitude of the force measured also increased (Fig. 2). Specifically, “regular” and “maximum” smiling produced –0.52 ± 0.31 N and –0.93 ± 0.48 N of force, respectively (Table 2). On the other hand, “regular” and “maximum” puckering produced +0.27 ± 0.15 N and +0.47 ± 0.23 N of force, respectively.

Fig. 2.

Force measured at ZM plane. Forces at the ZM plane during regular (light) and maximum (dark) actions for a representative participant (A) and the distribution across all participants (B).

Table 2.

Average Force for Different Facial Expressions

Action	Force, N
Regular smile	−0.52 ± 0.31
Maximum smile	−0.93 ± 0.48
Regular pucker	+0.27 ± 0.15
Maximum pucker	+0.47 ± 0.23

Data presented as mean ± SD (n = 31 for smiling and 30 for puckering, 4–5 movements per action).

Neuromuscular Activity at the ZM Midfacial Plane

Using surface electrodes overlying the muscle belly of the ZM, we recorded the neuromuscular activity during the smiling and puckering actions (Fig. 3). The magnitude of EMG_RMS for maximum smiling was significantly greater than that for regular smiling because maximum smiling requires greater activation of the muscle (Fig. 3). We observed the emergence of sEMG activity during regular and maximum puckering (Fig. 3). However, the magnitude of these sEMG signals was significantly lower than those observed for regular and maximum smiling expressions (Fig. 3). Mathematical deconstruction of the EMG signals into a function of spectral components is a powerful technique for identifying and isolating unique EMG signatures.²² Therefore, we applied the Welch PSD estimate to calculate the average spectral response for each facial expression and participant (Fig. 3). As expected, the magnitude of the PSD response for each action was greater than baseline (ie, no facial movement was performed). Furthermore, for most of the participants, the PSD signature of each facial action was statistically different from the others (Fig. 3).

Fig. 3.

Representation of sEMG activity at the ZM plane. Normalized EMG_RMS during regular (light) and maximum (dark) actions for a representative participant (A) and distribution across all participants (B). The sEMG PSD for a representative participant (C) and distribution of participants with significantly different sEMG PSD (D).

Classifying Facial Expressions Using sEMG

We trained a machine learning algorithm using supervised learning paradigms to automatically classify facial expressions (smiling versus puckering) based on sEMG activity at the ZM midfacial plane. For this task, a total of 5 features were extracted from the EMG data: 2 from the time domain (maximum EMG_RMS and zero-crossing rate) and 3 from the frequency domain (median frequency, mean frequency, and total power of the PSD estimates) (Fig. 4). Three distinct classifiers were then trained using a 5-fold cross-validation scheme, and the average classification accuracy was calculated. The average classification accuracies were 78%, 86%, and 86% for decision tree, quadratic discriminant, and linear support vector machine, respectively (Fig. 4). However, there was no statistical difference between the classification accuracy of the 3 classifiers. These results show that a generalizable machine learning algorithm can be trained to automatically classify facial movements with high accuracy using just simple features of the sEMG.

Fig. 4.

Predicting facial expressions. A, Schematic of the pipeline for predicting facial expressions. B, Prediction accuracies for different machine learning models (results are presented as mean ± SD, n = 5 folds for cross-validation).

DISCUSSION

The goal of this study was to perform a multimodal assessment of the biomechanical and neuromuscular characteristics of the midface and lower face overlying the ZM in healthy participants. We achieved this by mapping the displacement, the compressive and tensile forces experienced, and the neuromuscular activity during isotonic contraction (smiling) and extension (puckering). A comprehensive atlas of quantified metrics (displacement, force, and EMG) is of vital importance to the understanding and tracking of rehabilitation in patients with facial neuromuscular-related pathologies, as well as for designing technologies attempting to recapitulate midface and lower face movement.

A recent study evaluated the contraction of the ZM during smiling using functional magnetic resonance imaging in 34 young individuals and determined an average linear contraction of the muscle of 14.4%.² In our study, we tracked the contraction and extension of the midfacial place overlying that ZM via static and dynamic measurements of the distance between the commissure and a fixed point on the zygoma. This provides us with an accurate representation of the displacement of the ZM and its connective tissues during various facial expressions. Our results indicate a maximum contraction of 9.8% ± 3.1% during smiling and a maximum extension of 9.2% ± 3.4% during puckering. Our measurements account for the cumulative movement of ZM, corresponding soft tissue, and the overlying skin of the midface. Therefore, these measurements are observably lower than the direct measurements of ZM contraction obtained through extensive magnetic resonance imaging. We have also proposed a simple, fast, and cost-effective approach for tracking the displacement of the skin using video recording and straightforward data analysis. This technique allowed us to track dynamic movements of the region of interest and calculate relative displacements. It should be noted that the overall displacements observed along the chosen vector were lower for dynamic contractions than for static contractions. This is attributed to the fact that (1) static displacements were measured along the true vector of the ZM in 3-dimensional space, whereas dynamic measurements were projections onto a front-facing 2-dimensional plane, and that (2) our static measurements compared the distance between commissure and tragus as opposed to that between the commissure to zygoma. This second choice was made because the front-facing camera often lost sight of the tragus in individuals who moved their heads.

In-depth quantification of the compressive forces generated by the facial musculature and the tension experienced during various facial expressions is helpful for achieving natural functional outcomes, especially after surgical intervention.^13,23 We recorded maximum contractile forces of 0.93 ± 0.48 N for smiling and tension of 0.47 ± 0.23 N for puckering. The large variations in forces measured for each expression are attributed to the diversity in the participant population. These results are in line with previous attempts to determine the forces generated during smiling in healthy participants, which have reported approximate contractile forces of 1.36 and 1.92 N.^12,13 Furthermore, these prior studies had significantly smaller cohorts, which may have skewed the force distributions toward individuals with “maximal” facial expressions.

When muscles are activated via neuromuscular junctions, depolarization of the muscle fibers occurs. This leads to the generation of compound muscle action potentials and contraction.^24,25 This can be noninvasively recorded using sEMG,²⁶ which is a long-established technique for examining the neuromuscular activation patterns with well-developed standards for instrumentation and techniques.^27–30 The neuromuscular activity of facial muscles during a variety of facial movements has been well studied using high-density sEMG.^29,31 We use a simple and clinically feasible approach to quantifying the neuromuscular activity of the ZM by using a bipolar electromyography scheme with commercial electrodes. Because smiling involves direct contraction of the ZM, it leads to the largest magnitude sEMG signals. The emergence of sEMG signals during puckering, when the ZM is not directly activated, can be attributed to the crosstalk due to the proximity of the orbicularis oris and other nearby muscles.^32–34 Through our recordings, we were therefore able to quantify the characteristic neuromuscular features in the temporal and frequency domains associated with biomechanical loading of the ZM plane.

By quantifying the displacement, forces generated/experienced, and electrophysiological activity of the midface and lower face in healthy participants, we have created an atlas that can serve as approximate benchmarks for tracking rehabilitation from reconstructive surgery for facial reanimation.³⁵ Our pilot study demonstrated the breadth of quantitative features related to facial neuromuscular activity that can be obtained from a small number of facial movements and recording techniques. Our results also lay the foundation for implementing data-driven rehabilitation where the sEMG can be noninvasively recorded and used to track the progression of pathology or outcomes of rehabilitation in patients with facial paralysis. In terms of measuring outcomes, these quantitative metrics may also complement future patient-reported outcome measures, including those already validated, such as SCAR-Q,³⁶ BREAST-Q,³⁷ and BODY-Q.³⁸ These have become central to modern plastic surgery outcome assessments. As technology continues to advance, integrating objective tools such as sEMG with patient-reported outcome measures could enable more comprehensive pre- and postoperative evaluations and support future data-driven evolutions in clinical decision-making and rehabilitation protocols. Another application is to improve postoperative assessments of peripheral nerves after aesthetic procedures such as rhytidectomy, given the risks associated with nerve dysfunction and injuries.^39–42 Our multimodal recording techniques can provide a holistic map of the functionality of peripheral nerve injuries, leading to early identification and accurate diagnosis.

As the field of neuroprosthetics evolves, having baseline data from healthy participants to serve as inputs for emerging technologies will be absolutely critical. Our study provided the methodology for multimodal quantitative assessment of neuromuscular activity as well as data associated with ZM biomechanics in healthy participants. Machine learning tools can significantly enhance clinical diagnosis and decision support.⁴³ In the case of facial neuromuscular pathologies, accurate identification of expressions and emotions can be extremely valuable for neurological and psychological assessments.⁴⁴ Here, we demonstrate that by leveraging simple machine learning algorithms and sEMG signals alone, we can classify the intended facial movements as smiling or puckering with more than 85% accuracy. Effectively incorporating machine learning and artificial intelligence into medical devices will require a database of quantitative training and validation data. Our ability to link a small sample of objective data with emotional intent in a simple way opens the door for a wide range of future applications.

CONCLUSIONS

In this study, we have quantified the biomechanical and neuromuscular activity of the facial plane overlying the ZM in healthy participants, thus establishing benchmark metrics for tracking rehabilitation from reconstructive surgery and training data for future smart neuroprosthetics. Our pilot study demonstrated the breadth of objective data related to facial neuromuscular activity that can be obtained from a small number of facial movements and recording mechanisms.

DISCLOSURES

The authors have no financial interest to declare in relation to the content of this article. This work was (in part) supported by Edwin and Fannie Gray Hall Center for Human Appearance Research and Education Fund (N.M., L.S.L.). This work was also supported by the Center for Precision Engineering for Health from the University of Pennsylvania School of Engineering and Applied Sciences (C.S., F.V., N.M.) and from the National Institutes of Health (award no. R01AR081062, F.V.).