sEMG technology provides a non-invasive and highly time-resolution detection method for the assessment of facial nerve function by capturing the bioelectrical activities generated during muscle contraction. Its working principle is based on the spatiotemporal superposition effect of muscle fiber action potentials. When the muscles innervated by the facial nerve (such as the zygomatic major muscle and orbicularis oculi muscle) contract, the action potentials generated by the motor units are conducted along the muscle fibers, forming a detectable potential difference on the skin surface. The sEMG electrode extracts these weak signals through differential amplification technology. After filtering to remove noise, they are transformed into time-frequency domain parameters reflecting the degree of muscle activation.

The core advantage of this technology lies in its direct association with the functional state of the facial nerve muscles. Studies have shown that the root mean square value of sEMG signals on the affected side of patients with FP is 37–62% lower than that on the healthy side, and it is significantly negatively correlated with the House-Brackmann classification (r = −0.84, p < 0.001) (Cui et al., 2020). In addition, sEMG spectrum analysis can reveal the pathological characteristics of neuropathy: The median frequency of patients with acute FP decreased by 21% compared with the healthy population, while patients in the chronic phase showed progressive loss of high-frequency components (>200 Hz). These indicators provide quantitative basis for differentiating the nature of the lesion, such as nerve disuse vs. axonal rupture (Franz et al., 2024). In clinical applications, sEMG technology has broken through the positioning of a single diagnostic tool. By synchronically collecting the electromyographic activities of the bilateral masseter muscles and frontal muscles, the facial symmetry index can be quantified, providing precise target localization for botulinum toxin injection therapy combined with exercise. Furthermore, combined with 3D motion capture data, sEMG signals can analyze the neural control strategies of complex expression movements, such as the cooperative activation mode of the zygomatic major muscle and the orbicularis oculi muscle in the smiling movement, providing biofeedback for the expression reconstruction training of patients with FP (Petrides et al., 2023).

sEMG collects the summed potentials of superficial muscle groups through surface electrodes, with a low spatial resolution and being susceptible to interference from factors such as subcutaneous fat. However, it can collect dynamic electromyographic signals in real time, and its accuracy depends on the signal processing algorithm and electrode layout. It is mainly used in auxiliary fields such as rehabilitation assessment. Needle EMG inserts needle electrodes into the muscle, achieving a spatial resolution of millimeters, which can precisely locate deep or small area muscle lesions and accurately determine the nature of nerve-muscle lesions. It is recognized as the “gold standard” for the diagnosis of neuromuscular diseases, but it cannot directly monitor dynamic functions. The differences between sEMG and needle EMG in clinical practical applications are shown in Table 1.

3D motion capture technology tracks the movement trajectories of facial marker points with sub-millimeter accuracy through optical or inertial sensor systems, and builds digital expression models. The optical system uses high-speed infrared cameras to capture the position changes of reflective marking points, with a sampling frequency of up to 200 Hz and a spatial resolution better than 0.1 mm. The inertial measurement unit directly measures the motion parameters of facial tissues through a miniature gyroscope and an accelerometer. The multimodal fusion system developed by the Mayo Clinic synchronously deploys 7 optical marker points and 4 distributed inertial measurement units, and combines the Kalman filtering algorithm to achieve spatio-temporal alignment of data, keeping the motion trajectory reconstruction error within 0.3 mm (Nguyen et al., 2022).

This technology reveals the functional characteristics of facial expression muscles by quantifying parameters such as displacement, velocity, acceleration and joint Angle. Clinical studies have shown that in the smiling movement of patients with FP, the angular displacement of the affected side is reduced by 41% compared with the healthy side, the peak speed of the blinking movement decreases by 57%, and the contraction acceleration of the zygomatic major muscle is significantly correlated with the conduction velocity of the facial nerve (Zhu et al., 2023). The facial nerve function map constructed based on kinematic parameters can dynamically evaluate the movement patterns of 12 basic expressions and discover minor functional disorders that cannot be recognized by traditional scales. In addition, this technology provides a quantitative basis for individualized rehabilitation programs. By correcting abnormal movement patterns through biofeedback training and combining it with the immersive rehabilitation system developed with VR technology, the naturalness score of patients’ expressions has been significantly improved, and the treatment compliance has increased to 89% (Qidwai et al., 2019).

AI has significantly enhanced the accuracy and objectivity of FP diagnosis and treatment through deep learning models and multimodal data fusion technology. In the field of image and signal processing, convolutional neural networks demonstrate powerful feature extraction capabilities. For example, the FaceNet model improved based on the ResNet-50 architecture achieves an accuracy rate of 95.1% in the classification diagnosis of FP by introducing an attention mechanism to focus on key facial areas, and the diagnosis time is compressed to within 0.5 s (Feng et al., 2025). Recurrent neural networks excel at handling temporal biological signals, such as sEMG data stream analysis. Their Long Short-Term memory network structure can capture the temporal dependence of muscle activation patterns, increasing the accuracy of conjugated motion prediction to 89% (Zhuang et al., 2021).

Generative Adversarial Network (GAN) has opened up a new path for the diagnosis and treatment of FP. This technology can synthesize realistic facial expression images through adversarial training between the generator and the discriminator. For example, the Cycle GAN model generates a “virtual repair” image of the affected side expression by learning the expression mapping relationship between healthy people and patients with FP, enabling doctors to visually compare the symmetry of expressions before and after treatment. In more advanced applications, conditional GAN can dynamically generate corresponding expressions based on sEMG signals, providing a visual template for biofeedback training for patients with complete FP (DeBord et al., 2023). Furthermore, GAN has demonstrated value in data augmentation. By generating diverse pathological expression samples, it has expanded the training dataset size of deep learning models by 10 times, alleviating the bottleneck of scarce medical image annotation data.

The multimodal data fusion algorithm has further unleashed the clinical potential of AI. Feature-level fusion integrates sEMG spectral features, 3D motion trajectory parameters and patient self-rating scale data through deep neural networks to construct a joint feature space. Experiments show that the AUC value of the fused feature vector in differentiating central and peripheral FP reaches 0.94, which is 21% higher than that of a single mode. Decision-level fusion adopts Bayesian networks or D-S evidence theory to integrate the results of different evaluation tools. For example, by combining the sEMG signal improvement rate, motion capture parameters and the prediction results of the AI model, a probability map of facial nerve function recovery can be generated, providing a quantitative basis for the formulation of individualized rehabilitation plans. This multi-dimensional data fusion strategy enables the diagnosis and treatment of FP to shift from the assessment of a single indicator to comprehensive decision-making based on biomechanics, electrophysiology and clinical phenotypes.

Although sEMG technology can reflect muscle activity, the signals are prone to interference and have limited spatial resolution. 3D motion capture technology can accurately record movements, but the data processing is complex and the cost is high. AI assisted analysis technology relies on high-quality data and the models have poor interpretability. A single technology is difficult to comprehensively and accurately solve complex problems in practical applications. The integration of sEMG, 3D motion capture, and AI assisted analysis technology significantly enhances the accuracy, efficiency, and personalization level of facial paralysis recognition and assessment. Through the collaborative effect of electrophysiological signals, spatial motion data, and intelligent algorithms, the diagnostic accuracy of traditional methods (single sEMG or 3D Motion Capture) ranges from 70 to 80%, while with the integration of AI, it has improved to 95.1% (an increase of 25.1 percentage points). Combined with this technology, the false alarm rate of automatic assessment of facial paralysis grades can be reduced from 8 to 2%, the detection time can be shortened from several minutes in traditional methods to within 0.5 s (a speedup of approximately 90%), meeting the real-time requirements of clinical practice. Moreover, combined with this technology, the anxiety score of patients can be reduced by 37%, the efficiency of monitoring the rehabilitation progress can be increased by 50%, and personalized plans can reduce the repetitive work of doctors.

Previous studies on diseases such as FP or human movement often focused on a single aspect of assessment, such as only paying attention to muscle electrical activity or movement trajectories. The combination of these three techniques enables a comprehensive assessment from multiple dimensions, including muscle electrical activity, movement learning, disease diagnosis, and treatment effect evaluation. Moreover, by integrating these three techniques, individual differences such as muscle characteristics and movement patterns of patients can be fully considered, allowing for the formulation of personalized rehabilitation plans for each patient. Most importantly, the combination of these three techniques supports real-time data collection and analysis, enabling real-time tracking of disease progression and movement effects, and providing a basis for timely adjustment of treatment plans.

sEMG technology provides highly sensitive biomarkers for the early diagnosis of FP by capturing the bioelectrical activities of facial muscles. Taking the Biomask system as an example, this wearable device integrates an 8-channel flexible microneedle electrode array and an infrared temperature sensor, which can synchronously collect electromyographic signals and local blood perfusion data of key muscle groups such as the zygomatic major muscle and orbicularis oculi muscle (Cui et al., 2021). During the acute phase of Bell’s FP (within 72 h of onset), the Biomask system analyzed the time-frequency characteristics of sEMG signals through machine learning algorithms and successfully advanced the diagnostic window to 6.3 h after the appearance of symptoms, which was 52 h shorter than the traditional clinical assessment (Ryu et al., 2018).

The analysis of signal characteristics reveals the dynamic evolution law of facial nerve injury. The sEMG spectrum of patients with acute FP shows a “low-frequency migration” phenomenon. The median frequency is 21% lower than that of the healthy population, and it is strongly negatively correlated with the results of the facial nerve excitability test (r = −0.82). The combined time-frequency analysis further revealed that with the progression of the disease course, the high-frequency components (>200 Hz) of the electromyographic signal were progressively lost, and this change was particularly significant in patients with complete FP. Compared with traditional needle EMG, the Biomask system uses non-invasive microneedle electrodes, avoiding muscle twitch artifacts caused by acupuncture pain and improving the stability of the signal baseline by 67% (Demeco et al., 2021). In addition, its portable design supports home dynamic monitoring. In a multi-center study, continuous 72-h sEMG recording successfully captured the nocturnal associated motor episodes of 83% of patients, while the missed diagnosis rate of traditional single EMG detection was as high as 61% (Watts et al., 2020). This paradigm shift from “single-point assessment” to “continuous monitoring” has bought a valuable time window for the early intervention of FP.

3D motion capture technology quantifies the kinematic characteristics of facial expression muscles, providing an objective biomechanical basis for the clinical classification of FP. In the differentiation between central (supranuclear) and peripheral (subnuclear) FP, this technique reveals the differences in the trajectory characteristics of the two lesions: In the smiling movements of patients with central FP, the retention rate of the upward lift amplitude of the affected corner of the mouth can reach 67%, but there is a significant “synchronous contraction delay” phenomenon, which is related to the abnormal motor command conduction caused by cortical brainstem tract injury. However, patients with peripheral FP show a loss of symmetry in angular mouth displacement, and the contraction acceleration of the zygomatic major muscle is linearly correlated with the conduction velocity of the facial nerve (r = 0.76). By classifying the motion trajectory parameters through the Support Vector Machine algorithm, the automatic identification of two types of FP can be achieved, with an accuracy rate of 89.3%, which is 24% higher than that of the traditional neural localization diagnostic method (Tran et al., 2023).

This technology shows unique advantages in combined motion detection and postoperative functional assessment. The dynamic capture system can continuously record facial movement data for 72 h. Through spatio-temporal clustering analysis, it was found that 32% of patients with FP had minor conjunctival movements at rest. This abnormal pattern was not significantly correlated with the House-Brackmann scale classification, but indicated a poor prognosis. In the postoperative monitoring of facial nerve transplantation, the “functional recovery index” constructed by 3D motion capture combined with sEMG data can quantitatively evaluate the quality of nerve regeneration. Clinical studies have shown that for patients with this index >0.75 at 6 months after surgery, the improvement in the facial symmetry index was 2.3 times that of the low-index group, and the incidence of associated movement decreased by 58% (Zhao et al., 2020). In more advanced applications, the multimodal fusion of motion capture data and the fiber bundle tracking results of diffusion tensor imaging enables the visualization accuracy of the facial nerve regeneration path to reach 0.5 mm, providing precise anatomical navigation for secondary repair surgeries (Gupta et al., 2013).

AI has achieved the precision and individualization of FP diagnosis and treatment by constructing a deep learning framework. Models based on facial key point detection usually adopt a two-stage architecture: Firstly, 468 facial anatomical landmark points are located through the improved YOLOv7 algorithm, covering subtle areas such as the space between the eyebrows, nasolabial folds, and orbicularis oris muscle, with a detection accuracy of 0.8 pixels; Subsequently, the Graph Convolutional Network extracts the spatio-temporal features of the key points and fuses the dynamic information of the expression and action sequences. For instance, the FaceAI system achieved a differential accuracy rate of 96.4% for Bell’s FP and Hunt syndrome, which was 28% higher than that of experienced neurologists (Kong et al., 2024). The innovation of this model lies in the introduction of the self-attention mechanism, which enables the dynamic adjustment of the topological relationship weights between key points, thereby capturing minor functional disorders that cannot be identified by traditional scales.

In terms of performance evaluation, the FaceAI system demonstrates outstanding diagnostic efficiency. On the independent test set, its accuracy rate and recall rate reached 95.1 and 92.3%, respectively. It was significantly superior to the traditional House-Brackmann scale and Sunnybrook scale. Especially in the diagnosis of grade III - IV FP, the model synchronously optimizes the grading and localization tasks through a multi-task learning strategy, increasing the Kappa value to 0.89, which is much higher than the consistency among human evaluators (0.77). The ablation experiment confirmed that after fusing the sEMG spectral characteristics and motion capture parameters, the detection sensitivity of the model for linked band motion increased from 68 to 89%. In terms of clinical decision support, this system can generate individualized treatment suggestions. Based on the abnormal patterns of the patient’s facial movement trajectory, it automatically recommends the intensity of rehabilitation training, reducing the treatment plan formulation time by 73%. In a multicenter study in a tertiary hospital, the improvement in the facial symmetry index of patients who adopted AI suggestions was 1.9 times higher than that of the conventional treatment group, and the recurrence rate of associated movement was reduced by 54% (Zhang et al., 2024; Kimura et al., 2025).

In the current diagnosis and treatment of facial paralysis, sEMG and traditional needle-type EMG have the highest level of evidence strength (Level I), and should be used as the core diagnostic and monitoring tools; MRI and 3D motion capture technology need to be selected carefully in combination with clinical scenarios (Level II-III); although AI-assisted analysis has a promising future, it requires further verification (Level III). The key findings and evidence strength of each technology in the diagnosis and treatment of facial paralysis are shown in Table 2.

Intelligent sensing technology, through multimodal data fusion and dynamic monitoring, has constructed a quantitative index system for the therapeutic effect evaluation of FP. The assessment system based on sEMG developed by Tsinghua University uses skin-friendly and breathable PU membrane electrodes to cover core muscle groups such as the frontal muscle, zygomatic major muscle, and orbicularis oculi muscle, and combines a wireless transmission module to collect signals in real time. This system accurately quantifies the differences in muscle activity between the affected side and the healthy side by analyzing the standard deviation of movement (MSD) and the correlation of signal energy. Clinical data show that the MSD of the affected side muscles in patients with FP is 41% lower than that of the healthy side, while the difference in bilateral MSD between healthy people is only 7%. The monitoring sensitivity of this system for treatment response reaches 92%, and it can capture the tiny electrophysiological changes in the early stage of nerve regeneration, providing a real-time basis for adjusting the treatment plan (Frigerio et al., 2015).

3D motion capture technology further enriches the dimensions of therapeutic effect evaluation. The multimodal system of the Mayo Clinic tracks facial movement trajectories with an accuracy of 0.3 mm through 7 optical marker points and 4 distributed IMUs. This system can extract parameters such as displacement, velocity, acceleration and joint Angle, and quantify the recovery of facial expression muscle function. For example, in the smiling movement of patients with FP, the displacement of the corner of the mouth on the affected side is 40% less than that on the healthy side, while after treatment, this indicator improves at a rate of 8% per week. Dynamic assessment also revealed associated movement patterns that could not be identified by traditional scales. For instance, 28% of patients experienced involuntary contractions of the nasolabial folds in the early stage of rehabilitation. The correction of such abnormal movement trajectories is closely related to long-term prognosis (Nguyen et al., 2022). The feasibility analysis of the main assessment plan for facial paralysis is shown in Table 3.

Combined with the mixed model of patient self-rating scales, a comprehensive assessment from physiological signals to quality of life has been achieved. A multicenter study from Busan National University in South Korea showed that the integrated traditional Chinese and Western medicine treatment group improved the EQ-5D-5L health score by 15% compared with the conventional treatment group, and the improvement rate of facial sEMG signals was significantly positively correlated with the quality of life score (Goo et al., 2025). This correlation continued to strengthen during the 6-month follow-up, suggesting that the recovery of physiological functions is the basis for the improvement of quality of life. Furthermore, there is a strong correlation between the motion capture parameters and the social function dimension in the FACIAL-QOL scale. For every 10% increase in facial motion symmetry, the patient’s social confidence score increases by 0.8 points accordingly.

Long-term follow-up studies have verified the prognostic predictive value of intelligent sensing technology. A team conducted a 6-month follow-up on 85 patients with FP and found that the correlation of sEMG signal energy increased from 0.51 at the beginning of treatment to 0.73, and this index was linearly related to the recovery of facial nerve conduction velocity. Motion capture data show that the symmetry index of the facial movement trajectory of patients improves at a rate of 12% per month. For patients with an symmetry index of the facial>0.85 after 3 months of treatment, the risk of recurrence is reduced by 67%. What is more worthy of attention is that the AI model has constructed an efficacy trend prediction algorithm by integrating sEMG spectral features, motion capture parameters and patients’ self-evaluation data. The prediction error of this model for the rehabilitation period of patients with complete FP is only ±1.2 weeks, which is three times more accurate than the traditional empirical judgment method, providing a scientific basis for the dynamic optimization of individualized treatment plans (Miller et al., 2021). The clinical diagnosis and treatment of facial paralysis require the integration of multimodal technologies to achieve precise diagnosis, dynamic monitoring, and personalized prognosis assessment (Kafle and Thakur, 2021; Shi et al., 2024; Kishimoto-Urata et al., 2023). The analysis of the current status and future development direction of clinical diagnosis and treatment techniques for facial paralysis is presented in Table 4.

By integrating the millimeter-level accuracy of 3D motion capture, a “holographic map” of facial muscle activity can be constructed. For example, by covering key muscle groups such as the frontal muscles and orbicularis oculi with dense electrode arrays, simultaneous capture of muscle electrical activity and movement trajectories can be achieved, solving the problems of signal interference and blurred positioning in traditional methods (Zhu et al., 2022). Using deep learning algorithms (such as convolutional neural networks, graph neural networks) to perform real-time fusion analysis of multi-modal data and establish a dynamic model of the facial nerve-muscle system. For example, by training AI to identify differences in facial movement patterns between healthy individuals and patients with facial paralysis, automatic classification of the causes (such as viral infection, trauma, or tumor compression) can be achieved (Petrides et al., 2022).

By integrating facial expression analysis technology, through AI monitoring of subtle changes such as blink frequency and symmetry of the corners of the mouth, a facial paralysis risk prediction model is constructed (Gaber et al., 2022). The combination of sEMG and nerve conduction velocity detection is utilized to analyze the electrophysiological characteristics of the facial nerve in real time. AI can identify early signals of nerve damage by comparing normal and abnormal EMG data, thus securing the golden time for treatment (Zhang et al., 2022). In the future, environmental factors (such as cold stimulation) and patient medical history can be further integrated to improve the accuracy of warning.

An AI-based biofeedback training platform is developed. Through 3D motion capture, real-time monitoring of patients’ facial movements is conducted, and the training intensity is adjusted in combination with sEMG data (Fattah et al., 2014). Studies have shown that high-resolution sEMG can quantify the effect of biofeedback training on facial muscle coordination. In the future, combined with virtual reality technology, immersive rehabilitation scenarios can be designed to enhance patient engagement (Dusseldorp et al., 2018). Exploring the combination of sEMG with ultrasound imaging, brain-computer interface (BCI), and achieving more precise neuromuscular control is also possible. For example, through ultrasound imaging to verify the source of muscle activity recorded by sEMG, combined with BCI technology, help severely affected patients recover facial expression functions.

Promote the miniaturization and wirelessization of sEMG and 3D motion capture technology, and develop rehabilitation monitoring devices suitable for home use. For instance, the combination of flexible electrode patches and smartphone apps enables real-time upload of patients’ daily rehabilitation data, facilitating remote adjustment of treatment plans by doctors (Zimmermann et al., 2019). Collaborate with medical institutions, research institutes, and enterprises to establish technical specifications for facial paralysis assessment and efficacy evaluation standards. For example, clarify operational details such as sEMG electrode layout and 3D motion capture marker point settings, to enhance the comparability of research results and the reliability of clinical application (Rao et al., 2025). In the future, with continuous technological breakthroughs, this multimodal integration solution is expected to achieve the “precise diagnosis - personalized treatment - dynamic assessment” full process management of facial paralysis, significantly improving patients’ quality of life and providing a new technical paradigm for the field of neurorehabilitation.