Aplexy patient upper limb movement function assessment method based on surface electromyogram signals

Abstract

The invention belongs to the technical field of biomedical signal processing, and particularly relates to a surface electromyogram signal-based upper limb movement function evaluation method for a stroke patient. The method comprises the steps of subject selection, electrode arrangement and signal acquisition, experimental normal form design, surface electromyographic signal preprocessing, macroscopic electromyographic feature extraction and microscopic electromyographic feature extraction. The method comprises the following steps: collecting surface electromyogram signals under a plurality of specific gestures, extracting eight macroscopic features and two microcosmic features, and constructing a model for Brunnstrom stage classification and upper limb Fugl-Meyer scoring prediction by using a K nearest neighbor algorithm; through systematic comparison of an affected side, an uninjured side and a healthy control group of a stroke patient, accurate quantification of the degree of stroke dyskinesia is realized, and objectivity, fineness and clinical practicability of evaluation are improved; manual evaluation deviation and time consumption can be greatly reduced, and an objective and real-time evaluation basis is provided for formulating a personalized rehabilitation scheme.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal processing technology, specifically relating to a method for assessing upper limb motor function in stroke patients based on surface electromyography signals. Background Technology

[0002] Stroke is a common central nervous system disease and one of the leading causes of death and disability worldwide. According to the 2021 Global Burden of Disease Study, stroke ranks second among causes of death globally and third among combined causes of disability and death. The most common functional impairment after stroke is upper limb motor dysfunction, particularly impairment of fine motor skills in the hands, which severely impacts patients' daily living abilities and quality of life. Therefore, how to scientifically, objectively, and quantifiably assess upper limb, especially hand, functional impairment in stroke patients has become a research hotspot in the interdisciplinary fields of rehabilitation medicine, neuroscience, and human-computer interaction.

[0003] Currently, the Brunnstrom motor recovery staging system and the Upper Limb Fugl-Meyer Assessment (UE-FMA) are widely used in clinical practice to assess the recovery of motor function in stroke patients. While these scales have some clinical guiding significance, their assessment relies on the clinical physician's experience and judgment, exhibiting strong subjectivity and environmental dependence. Furthermore, they are difficult to use frequently, limiting their application in high-frequency, remote, and precise rehabilitation.

[0004] In recent years, surface electromyography (HD-sEMG) signal technology has been gradually introduced into stroke rehabilitation assessment due to its advantages such as high spatial resolution and non-invasiveness. This technology, by deploying a two-dimensional electrode array on the skin surface, can acquire real-time spatial-temporal distribution information of muscle activity, demonstrating strong neuromuscular activity monitoring capabilities. Existing studies have explored indicators such as muscle reconstruction (MU), muscle coordination, and spatial distribution symmetry based on HD-sEMG, showing preliminary application prospects in stroke motor function assessment.

[0005] Current research uses HD-sEMG signal extraction features to analyze changes in muscle activation patterns after stroke. For example, some studies have compared the spatial activation distribution of muscles in the affected and unaffected hands during active or passive movements in stroke patients, revealing heterogeneous recruitment patterns of spastic muscles; other studies have characterized muscle function at different stages of rehabilitation by calculating the synergistic patterns and time-series stability of EMG signals.

[0006] In addition, some studies have attempted to establish preliminary associations between HD-sEMG features and Brunnstrom or UE-FMA scores, and applied classification models (such as SVM and LDA) to determine functional levels, verifying the predictive ability of certain electromyographic features for clinical scores. However, these methods still have many shortcomings in practical applications.

[0007] Currently, analysis of neuromuscular activity changes in stroke patients based on HD-sEMG mainly focuses on the following aspects:

[0008] (1) Spatial and temporal distribution analysis of muscle activation: Xie et al. compared spastic hemiplegic patients with healthy controls and studied the spatial distribution and intensity differences of muscle activation during passive stretching and active contraction. They found that spastic muscles exhibited speed-dependent activation heterogeneity, indicating that the muscle control mechanism changed significantly after stroke [1].

[0009] (2) Analysis of bilateral muscle coordination and symmetry: Zhang et al. systematically evaluated the changes in neuromuscular activity in stroke patients in time and space, and pointed out that there were significant differences in muscle coordination and symmetry between the affected side and the healthy side, reflecting the damage of stroke to bilateral neuromuscular coordination ability[2].

[0010] (3) Research on the recruitment patterns of motor units (MUs): Xie et al. studied the temporal patterns and spatial heterogeneity of MU recruitment in spastic muscles under passive stretching, and found that its temporal variability decreased significantly, and that the activation distribution was correlated with the degree of spasticity [3].

[0011] (4) Quantification of motor unit distribution and co-activation pattern: Ruan et al. quantitatively analyzed the spatial distribution and co-activation of MU action potentials and found that stroke patients showed enhanced external muscle co-activation and abnormal MU distribution, which were closely related to the decline in independent finger control ability[4].

[0012] (5) Study on the difference in MU depth and force regulation mechanism: Liu et al. compared the spatial distribution and recruitment strategy of MU in the bilateral biceps brachii of patients with chronic stroke. The results showed that there was dysregulation in the MU recruitment process based on output force on the affected side, and the distribution of MU activation depth was significantly different from that on the healthy side [5].

[0013] Although current research has initially shown the potential of HD-sEMG in stroke assessment, the following limitations still exist:

[0014] (1) Single feature dimension and lack of multi-level fusion analysis: Most studies only extract macroscopic electromyographic features (such as root mean square RMS, mean frequency MNF, etc.) or microscopic features (such as MUAP waveform morphology), lacking the fusion of macroscopic and microscopic features, making it difficult to fully characterize the multidimensional manifestations of muscle dysfunction.

[0015] (2) Coarse grouping and lack of hierarchical comparison: Current studies focus on the comparison between the healthy group and the affected side, lacking systematic analysis between subgroups such as the healthy side and patients at different Brunnstrom stages, making it difficult to reflect the progression pattern and subtle differences of stroke dysfunction.

[0016] (3) Poor task generalization and neglect of gesture-specific differences: Some studies only use simple flexion and extension movements or isometric contractions, without in-depth analysis of the differences in muscle activation patterns induced by different gestures, which limits the role of HD-sEMG in fine function recognition and rehabilitation intervention guidance.

[0017] (4) It is out of touch with clinical assessment standards and has low clinical applicability: Most existing studies lack systematic comparative verification with clinical scores such as Brunnstrom or UE-FMA. The interpretability and clinical applicability of the assessment indicators are weak, making it difficult to use as a basis for adjusting rehabilitation programs.

[0018] Definitions of abbreviations and key terms

[0019] (1) HD-sEMG (High-Density Surface Electromyography) is a non-invasive detection technique that records muscle electrical activity by arranging a high-density electrode array on the skin surface. Compared with traditional sEMG, HD-sEMG provides higher spatial resolution and can be used to extract macroscopic (such as overall muscle group activation) and microscopic (such as single motor unit discharge) features of muscle activity. It is widely used in the diagnosis of neuromuscular diseases, rehabilitation assessment and human-computer interaction research.

[0020] (2) Post-stroke motor dysfunction: Brain tissue damage caused by stroke affects the cerebral cortex, subcortical structures or descending motor pathways, resulting in motor function impairment such as weakened muscle strength, poor motor coordination, slowed movement speed, abnormal muscle tone, and limited postural control. Common manifestations include hemiplegia, ataxia, muscle spasm and decreased motor control ability. This dysfunction directly affects the patient's daily living ability and is one of the core targets of rehabilitation treatment.

[0021] References

[0022] [1] T. Xie, Y. Leng, P. Xu, L. Li, and R. Song, “Mapping of spasticmuscle activity after stroke: difference between passive stretch and activecontraction,” Journal of NeuroEngineering and Rehabilitation, vol. 21, no. 1,p. 102, 2024.

[0023] [2]H. Zhang, B. Peng, Z. Chen, Y. Peng, X. Zhou, Y. Geng, and G. Li,“Characterizing upper limb motor dysfunction with temporal and spatialdistribution of muscle synergy extracted from high-density surfaceelectromyography,” Journal of Neural Engineering, vol. 21, no. 5, p. 056006,2024.

[0024] [3] T. Xie, Y. Leng, R. He, C. Wang, and R. Song, “Changes inspatiotemporal variability of muscular response under the influence of post-stroke spasticity,” IEEE Transactions on Neural Systems and RehabilitationEngineering, 2024.

[0025] [4] Y. Ruan, H. Shin, and X. Hu, “Quantifying muscle co-activationfor impaired finger independence in stroke survivors,” IEEE Transactions onBiomedical Engineering, 2024.

[0026] [5] Y. Liu, Y.-T. Chen, C. Zhang, P. Zhou, S. Li, and Y. Zhang, “Motor unit distribution and recruitment in spastic and non-spastic bilateral biceps brachii muscles of chronic stroke survivors,” Journal of neuralengineering, vol. 19, no. 4, p. 046047, 2022. Summary of the Invention

[0027] The purpose of this invention is to provide a method for assessing upper limb motor function in stroke patients based on surface electromyography (HD-sEMG) signals, aiming to solve the following technical problems existing in the assessment of motor dysfunction in stroke patients:

[0028] This invention proposes a method for assessing upper limb motor function in stroke patients based on surface electromyography (sEMG) signals. This method involves collecting HD-sEMG signals under multiple specific hand gestures, extracting eight macroscopic features and two microscopic features, and constructing a model using the K-nearest neighbor (KNN) algorithm for Brunnstrom staging and upper limb Fugl-Meyer score prediction. By systematically comparing the affected side, unaffected side, and healthy control group of stroke patients, the method achieves precise quantification of the degree of stroke motor function impairment, improving the objectivity, precision, and clinical applicability of the assessment, thereby providing a basis for developing personalized rehabilitation plans. Specifically, the method includes the following steps:

[0029] Step 1: Selecting Subjects

[0030] A number of participants were selected, including several stroke patients with varying degrees of bradykinesia (generally 10-20) and several healthy participants without hand motor dysfunction symptoms (generally 6-10), taking into account appropriate age groups and a suitable male-to-female ratio; all stroke patients exhibited significant difficulty in performing movements with the affected hand; detailed clinical information of the stroke patients was included (including gender, age, medical history, affected side, pathogenesis, Brunnstrom stage, and UE-FMAr score).

[0031] The Brunnstrom stage reflects the level of motor recovery after stroke; a higher score indicates stronger voluntary motor ability and better recovery. The UE-FMA is a subset of the Fugl-Meyer scale (Proximal Fugl-Meyer assessment scores predict clinically important upper limb improvement after 3 stroke rehabilitative interventions) and is specifically used to assess distal upper limb function, including shoulder, elbow, and finger function. The UE-FMA consists of 15 items, each scored on a 3-point scale: 0 = cannot complete, 1 = partially complete, 2 = fully complete; the maximum score is 30 points, and a higher score indicates better motor function recovery.

[0032] Step 2: Electrode Arrangement and Signal Acquisition

[0033] The sampling rate is set to 2048 Hz, gain to 150, and resolution to 16 bits. Specifically, the OT Bioelettronica Quattrocento system can be used. The electrode arrangement must ensure the spatial continuity and regional independence of the muscle group activity signal. Specifically, four 8×8 high-density electrode arrays are arranged in the forearm flexor and extensor muscle group area to form a 16×8 channel layout for high-resolution acquisition of the spatial distribution of the muscle group. The reference electrode is placed at the elbow tip, and the driving electrode is placed at the ulnar head.

[0034] Step 3: Experimental Paradigm Design

[0035] This invention designs seven different gesture actions (such as...) Figure 1 (As shown), these seven movements cover common hand motor functions in daily life and are key movement types that are often impaired in stroke patients; each movement involves different muscle groups, providing a comprehensive assessment of muscle control and strength in the hand and wrist; explanations of the seven hand gestures:

[0036] (1) Wrist flexion (WF):

[0037] Action description: Begin in a relaxed natural posture with palms facing up and wrists bent towards the palms, bringing the palms close to the forearms;

[0038] The main muscles activated are the flexor carpi radialis and the flexor carpi ulnaris.

[0039] Functional significance: This movement is used for wrist stability when grasping objects and is a representative activity of the wrist flexor muscles; in stroke patients, weakened flexion ability reflects impaired strength and coordination of the flexor muscles.

[0040] (2) Wrist Extension:

[0041] Action description: Slowly raise the back of your hand from a flexed position, extending your palm away from your forearm;

[0042] The main muscles activated are: extensor carpi radialis longus, extensor carpi radialis brevis, and extensor carpi ulnaris.

[0043] Functional significance: The stretching movement is an important counter-mechanism between grasping and relaxing. Weakness of the extensor muscles will lead to limited finger extension. This movement is used to assess the wrist joint's counter-control ability and muscle coordination, and is an important indicator of restoring the ability to switch between relaxation and grasping.

[0044] (3) Forearm pronation (WP):

[0045] Action description: Rotate the forearm from palm up to palm down, keeping the elbow joint fixed;

[0046] The main muscles activated are the pronator teres and the pronator quadratus.

[0047] Functional significance: Pronation is often used to place or manipulate objects and is an important gesture for assessing forearm rotational control. In stroke patients, pronation disorder can lead to limitations in daily activities (such as turning pages and using utensils).

[0048] (4) Forearm supination (WS):

[0049] Action description: The forearm rotates from a downward palm position to a palm-up position;

[0050] Primary muscles activated: supinator and biceps brachii.

[0051] Functional significance: Supination is an important basic movement for holding, grasping and picking up objects; stroke patients often exhibit limited supination, stiffness or inability to fully rotate, and this movement can be used to identify the degree of damage to the forearm rotator muscles.

[0052] (5) Hand Clenching (HC):

[0053] Action description: Start by straightening your fingers, gradually bend all your fingers to form a fist, and then slowly release.

[0054] The main muscles activated are the flexor digitorum superficialis and the flexor digitorum profundus.

[0055] Functional significance: Clenching the fist reflects overall grip strength and finger flexor coordination, and is a key movement for hand strength recovery after stroke; changes in electromyographic characteristics can reveal the degree of synchronous contraction of muscle groups and the activation pattern of motor units.

[0056] (6) Thumb-Index Finger Pinch (TIFP):

[0057] Action description: Bring the tips of your thumb and forefinger together to form a "pinch" motion, then relax;

[0058] Primary muscles activated: flexor pollicis longus and first dorsal interosseous muscle.

[0059] Functional significance: This action is used for fine grasping of small objects (such as pens and buttons) and is an important indicator for testing fine motor skills and fingertip coordination; stroke patients often experience unstable coordination between their thumb and index finger or are unable to make complete contact.

[0060] (7) Thumb–Index–Middle Finger Pinch:

[0061] Action description: The thumb simultaneously touches the tips of the index and middle fingers, forming a three-finger pinch, then releases;

[0062] The main muscles activated are: flexor pollicis longus, flexor digitorum superficialis, and intrinsic hand muscles (especially the lumbrical and interosseous muscles).

[0063] Functional significance: The three-finger pinch represents a complex coordinated movement used for grasping and manipulating medium-sized objects, such as picking up a key or writing; this movement reflects higher-order motor control ability and multi-muscle group coordination level, and is highly representative of the recovery of fine motor function in stroke rehabilitation;

[0064] These seven hand gestures, ranging from large muscle groups (wrist flexion, extension, and rotation) to fine motor skills (two-finger and three-finger pinching), form a functional hierarchy from coarse to fine, and from strength to coordination. Through HD-sEMG feature analysis of these movements, this invention can achieve: the distribution and localization of abnormal muscle group function (macroscopic features); the identification of changes in neuromuscular coordination and firing patterns (microscopic features); and thus, objectively assess and predict the overall and fine motor recovery level of stroke patients.

[0065] During the data collection process, the subjects sat in a comfortable position and performed seven hand gestures based on prompts displayed on the computer screen.

[0066] Step 4: HD-sEMG signal preprocessing

[0067] The original HD-sEMG signal was subjected to the following steps in sequence: Butterworth bandpass filtering (10–500 Hz) to remove drift and high-frequency noise; notch filtering (50 Hz and harmonics) to eliminate power frequency interference; inter-channel mean correction and normalization; time alignment and segmentation to synchronize the test trigger point.

[0068] Step 5: Extraction of macroscopic electromyographic features

[0069] The following eight categories of macroscopic electromyographic features were extracted from the preprocessed HD-sEMG signals to reflect the overall activity and functional status of the muscles:

[0070] (1) Complexity and variability characteristics: including mean amplitude variation (AAC) and difference mean square standard deviation variation (DASDV), used to measure the dynamic variability of muscle activity;

[0071] , (1)

[0072] , (2)

[0073] Where x(i) represents the i-th sample in the signal, and N is the length of the signal. The standard deviation represents the absolute difference of consecutive samples;

[0074] (2) Muscle strength characteristics: including mean absolute value (MAV), mean amplitude (ASM) and root mean square (RMS), used to assess the strength and endurance of muscle contraction;

[0075] , (3)

[0076] (4)

[0077] , (5)

[0078] in, This represents the absolute value of the signal at sampling point i, while This indicates the exponent applied to that value; for samples within the middle 50% of the signal length, the exponent is 0.5; for samples outside this range, the exponent is 0.75.

[0079] (3) Energy distribution and amplitude variation characteristics: including modified Hamming window (MHW) and modified trapezoidal window (MTW) ​​characteristics, reflecting the temporal variation trend of muscle contraction energy;

[0080] , (6)

[0081]

[0082] , (7)

[0083] ; in, and These represent the start and end indices of the current window segment, respectively. It is the Hamming window located at position j. Here, N is the value of the signal at position j, and N is the length of the window. This represents the trapezoidal window located at position j. It is a parameter that defines the width of the trapezoid;

[0084] (4) Coordination and synchronization characteristics: including inter-channel synchrony (Sync), assessing the coordinated control ability between muscle groups;

[0085] (8); Where λ1 is the largest eigenvalue of the local electromyographic signal covariance matrix, representing the variance captured by the first principal component, and... This is the total variance captured by all N principal components;

[0086] The innovation of this step lies in forming a quantitative description of temporal energy changes and muscle coordination by using custom weighted window functions (MHW, MTW) and inter-channel synchronicity indicators (Sync).

[0087] Step 6: Extraction of microscopic electromyographic features

[0088] The FastICA-based motor unit decomposition method is used to extract micro-control features at the neural level, specifically including (e.g.) Figure 2 (as shown)

[0089] (1) Data splicing processing: When splicing multiple repeated signals under the same gesture, a rest segment with 100 sampling points is added to prevent decomposition errors caused by false peaks;

[0090] (2) MU decomposition processing: The HD-sEMG signal is subjected to channel delay extension (extension order L=8) and then covariance whitening processing is performed to improve the independence of the source signal;

[0091] (3) FastICA decomposition: The q-non-Gaussian objective function (q=1.5) is used to improve the robustness of the algorithm to abnormal noise and separate the discharge signals of multiple motion units;

[0092] (4) MU detection and screening: K-means clustering is used to identify the discharge time of MUs and the Silhouette coefficient (SIL) is calculated to evaluate the clustering quality. If the overlap rate of the discharge sequence is >50% within ±1 ms, the one with higher SIL is retained. The screening condition is SIL≥0.75. If there are less than 6, the highest 6 are selected. All MUs are manually checked for the rationality of their morphology and discharge frequency.

[0093] Based on the effective MU discharge sequence, the following microscopic neural drive features are extracted:

[0094] (1) Number of MUs (NMUs): Reflects the number of motor units activated and is used to measure the degree of muscle recruitment;

[0095] (2) Average discharge frequency (AFR): defined as the average discharge frequency of MU per unit time, used to assess the intensity of neural drive.

[0096] The innovation of this step lies in the introduction of splicing smoothing + q non-Gaussian FastICA decomposition + SIL screening criteria, which significantly improves the accuracy and stability of MU decomposition and can quantitatively reflect the degree of damage at the neural drive level in stroke patients.

[0097] Based on the above method, the present invention also provides an upper limb motor function assessment system for stroke patients based on surface electromyography (EMG) signals. This assessment system comprises six functional modules: a subject selection module, an electrode placement and signal acquisition module, an experimental paradigm design module, HD-sEMG signal preprocessing, macroscopic EMG feature extraction, and microscopic EMG feature extraction module; each of the six functional modules performs one of the six steps of the method.

[0098] The main technical features and functional advantages of this invention are as follows:

[0099] (1) To fully reflect the muscle function status of the affected and unaffected hands of stroke patients and healthy controls.

[0100] This invention collects high-density surface electromyography (HD-sEMG) signals from the affected side, unaffected side, and healthy control group of stroke patients under seven hand gestures (WF, WE, WP, WS, HC, TIFP, TIMFP). Eight macroscopic features (AAC, DASDV, MAV, ASM, MHW, MTW, RMS, Sync) and two microscopic features (NMUs, AFR) are extracted. The differences among the three groups are analyzed using the nonparametric Kruskal-Wallis test combined with Bonferroni-Holm correction. The results show that, in terms of macroscopic features, the healthy group generally has higher values ​​on the affected side than the stroke patient, while the affected side has lower values ​​on the unaffected side, reflecting a decline in muscle contraction ability and insufficient energy output caused by stroke. In terms of microscopic features, NMUs are significantly elevated on the affected side, while AFR is decreased, suggesting increased motor unit recruitment and decreased firing rate, revealing the degenerative pattern of neuromuscular control. This method enables quantifiable identification of muscle function status among different groups, providing an objective basis for determining the lesion side and functional impairment site in stroke patients, and providing a characteristic-level physiological basis for subsequent staging assessment and rehabilitation intervention.

[0101] (2) Multi-stage rehabilitation grading assessment to achieve automated classification of the stroke exercise recovery stage.

[0102] In the second verification scheme, this invention constructs a K-nearest neighbor (KNN) classification model based on extracted macroscopic and microscopic features to distinguish between patients' Brunnstrom stages (stages 4, 5, and 6) and health status. Results show that the classification accuracy rates for gestures such as WE, TIMFP, and TIFP reach 92.08%, 87.82%, and 87.03%, respectively, effectively distinguishing different stages of motor recovery. Furthermore, misclassifications are mainly concentrated in adjacent stages, consistent with the continuity of patient functional recovery. These results demonstrate that this invention can automatically grade and assess patients' motor function recovery stages without relying on subjective scoring, significantly reducing the bias and time-consuming nature of manual assessments and providing rehabilitation physicians with objective and real-time functional assessment data. For example, the system can identify the phased changes in patients' electromyographic characteristics in real time during rehabilitation training, thereby automatically indicating the timing for stage transitions or adjustments to training intensity.

[0103] (3) Based on electromyographic characteristics, quantitative prediction of motor function can be achieved to realize accurate assessment of the recovery level of fine motor function and personalized rehabilitation guidance.

[0104] This invention utilizes a KNN regression model to fit the characteristics of the patient's affected side under various hand gestures to the upper limb Fugl-Meyer (UE-FMA) score, establishing a mapping model from electromyographic features to clinical scales. Regression results show that the TIMFP, HC, and TIFP hand gestures exhibit the best predictive performance, with R² values ​​of 0.86, 0.74, and 0.73, and RMSE values ​​of 3.26, 4.42, and 4.56, respectively, accurately reflecting subtle changes in the patient's upper limb motor function. Through this model, the system can directly predict the patient's UE-FMA score based on electromyographic signals, automating rehabilitation assessment. For example, when the system detects that the patient's characteristic improvement under a specific hand gesture reaches the corresponding stage threshold, it can automatically update the rehabilitation level and suggest targeted training (such as strengthening thumb fingertip control or wrist extension movements), thereby achieving personalized and dynamic adjustments to the rehabilitation plan. Attached Figure Description

[0105] Figure 1 It demonstrates seven different hand gestures.

[0106] Figure 2 This is a flowchart for decomposing the motion unit.

[0107] Figure 3 Box plots show the distribution of macroscopic and microscopic features for seven gestures in three groups of subjects: the dominant side for healthy subjects, the affected side for stroke patients, and the unaffected side. Figure (ah) shows eight macroscopic features, and (ij) shows two microscopic features. Significance levels are labeled as follows: * p < 0.05, ** p < 0.01, *** p < 0.001. It is assumed that there is a significant one-sided advantage in the feature values, with macroscopic and microscopic AFR features following the trend: healthy group > affected side > unaffected side; while microscopic NMUs features show the opposite trend: healthy group < affected side < unaffected side.

[0108] Figure 4 The classification accuracy is represented by different gestures; the error bars represent the standard deviation of the data.

[0109] Figure 5 The values ​​of R and RMSE are for different gestures.

[0110] Figure 6 Scatter plots and residual analysis of the three gestures for optimal performance. Detailed Implementation

[0111] The invention will be further described below with specific examples and accompanying drawings.

[0112] Step 1: Subject Preparation

[0113] Nineteen participants were recruited, including 11 stroke patients with varying degrees of bradykinesia (10 males and 1 female, aged 27 to 75 years, mean age 54.27 ± 16.01 years) and 8 healthy participants without hand motor dysfunction symptoms (5 males and 3 females, aged 30 to 66 years, mean age 53.88 ± 10.84 years). All stroke patients exhibited significant difficulty in performing movements with the affected hand. Detailed clinical information of the stroke patients (including sex, age, medical history, affected side, pathogenesis, Brunnstrom stage, and UE-FMAr score) is shown in the table below.

[0114] ; The Brunnstrom stage reflects the level of motor recovery after stroke; a higher score indicates stronger voluntary motor function and better recovery. The UE-FMA is a subset of the Fugl-Meyer Assessment, specifically designed to assess distal upper limb function, including shoulder, elbow, and finger function. The UE-FMA consists of 15 items, each scored on a 3-point scale: 0 = unable to complete, 1 = partially complete, 2 = fully complete; the maximum score is 30 points, with higher scores indicating better motor function recovery. All participants signed informed consent forms before the experiment, and this study was approved by the Ethics Committee of Shanghai Yangzhi Rehabilitation Hospital (Approval No.: YZ 2024-140).

[0115] Step 2: Determination of Electrode Arrangement and Sampling Scheme

[0116] The OT Bioelettronica Quattrocento system was used with a sampling rate of 2048 Hz, a gain of 150, and a resolution of 16 bits. Four 8×8 high-density electrode arrays (10 mm electrode spacing, elliptical electrodes, model ELSCH064NM1) were arranged in the forearm flexor and extensor muscle group region, forming a 16×8 channel layout for high-resolution acquisition of the spatial distribution of muscle groups. The reference electrode was placed at the elbow tip, and the drive electrode was placed at the ulnar head. This layout ensures the spatial continuity and regional independence of muscle activity signals.

[0117] Step 3: Sampling according to the designed experimental paradigm.

[0118] Before the experiment, participants practice each gesture repeatedly until they can perform it accurately. Gesture cues are presented one by one in a random order, with each action repeated in two blocks. Each block contains three dynamic tasks: transitioning from a relaxed state to the designated gesture; and one maintenance task: maintaining stillness after completing the action. The execution time is dynamically adjusted based on participants' abilities. Generally, dynamic tasks must last at least 1 second, and maintenance tasks at least 3 seconds. A 2-second rest period is provided between each trial, and a 5-second rest period is provided between every two task blocks to prevent muscle fatigue. Each trial begins with an audible alert and a synchronized trigger signal to mark the trial time for later segmented analysis of electromyographic data. For stroke patients, each participant completed 84 dynamic trials (2 sides × 7 gestures × 2 test blocks × 3 repetitions) and 28 maintenance trials (2 sides × 7 gestures × 2 test blocks × 1 repetition) on both the affected and unaffected sides. For healthy participants, data were collected only on the dominant side (both right-handed), totaling 42 dynamic trials (7 gestures × 2 test blocks × 3 repetitions) and 14 maintenance trials (7 gestures × 2 test blocks × 1 repetition). During the trials, if a participant failed to complete or incorrectly performed a gesture, they could notify the experimental assistant, who would then use the "Retry" button on the control interface to allow the participant to re-perform the trial.

[0119] Step 4: HD-sEMG signal preprocessing

[0120] The original HD-sEMG signal was subjected to the following steps in sequence: Butterworth bandpass filtering (10–500 Hz) to remove drift and high-frequency noise; notch filtering (50 Hz and harmonics) to eliminate power frequency interference; inter-channel mean correction and normalization; time alignment and segmentation to synchronize the test trigger point.

[0121] Step 5: Extract macroscopic electromyographic features, specifically:

[0122] (1) Complexity and variability characteristics: including mean amplitude variation (AAC) and difference mean square standard deviation variation (DASDV).

[0123] (2) Muscle strength characteristics: including mean absolute value (MAV), mean amplitude (ASM) and root mean square (RMS).

[0124] (3) Energy distribution and amplitude variation characteristics: including modified Hamming window (MHW) and modified trapezoidal window (MTW) ​​characteristics;

[0125] (4) Coordination and synchronization characteristics: including inter-channel synchronization (Sync).

[0126] Step 6: Extract microscopic electromyographic features. The specific process is as follows:

[0127] The FastICA-based motor unit decomposition method is used to extract micro-control features at the neural level, including:

[0128] (1) Data splicing processing: When splicing multiple repeated signals under the same gesture, a rest segment with 100 sampling points is added to prevent decomposition errors caused by false peaks;

[0129] (2) MU decomposition processing: The HD-sEMG signal is subjected to channel delay extension (extension order L=8) and then covariance whitening processing is performed to improve the independence of the source signal;

[0130] (3) FastICA decomposition: The q-non-Gaussian objective function (q=1.5) is used to improve the robustness of the algorithm to abnormal noise and separate the discharge signals of multiple motion units.

[0131] (4) MU detection and screening: K-means clustering was used to identify the discharge time of MUs, and the Silhouette coefficient (SIL) was calculated to evaluate the clustering quality. If the overlap rate of the discharge sequences within ±1 ms is >50%, the ones with higher SIL are retained. The screening condition is SIL ≥ 0.75, and if there are less than 6, the highest 6 are selected. All MUs were manually checked for the rationality of their morphology and discharge frequency.

[0132] Based on the effective MU discharge sequence, the extracted microscopic neural drive features include:

[0133] (1) Number of MUs (NMUs);

[0134] (2) Average discharge frequency (AFR).

[0135] Step 7: Assessment

[0136] To verify the effectiveness of the macroscopic and microscopic features extracted in this invention in assessing motor dysfunction after stroke, three assessment and verification schemes were designed. This assessment scheme is based on 10 extracted features (8 macroscopic features and 2 microscopic features), combined with the K-Nearest Neighbor (KNN) algorithm to achieve classification and regression modeling, forming a multidimensional assessment system for functional differentiation, stage determination, and recovery prediction.

[0137] Validation Scheme 1: Difference Analysis of Electromyographic Patterns Among Multiple Populations

[0138] Objective: To analyze the differences in electromyographic characteristics of the unaffected hand of healthy subjects, the unaffected hand of patients, and the affected hand of patients under different hand gestures, and to reveal the neuromuscular control ability and functional degeneration patterns of different hand gesture muscle groups in stroke.

[0139] Implementation steps:

[0140] (1) Calculate the mean and standard deviation of eight macro features and two micro features for the seven gestures (WF, WE, WP, WS, HC, TIFP, TIMFP) of the three groups of samples;

[0141] (2) For each feature under each gesture, the Kruskal-Wallis nonparametric rank-sum test was used to compare the distribution differences of the three groups of samples;

[0142] (3) To control the false positive error caused by multiple comparisons, the Bonferroni–Holm method was used for significance correction;

[0143] (4) If the adjusted significance level p < 0.05, the feature is determined to have statistical differences among the three groups and is marked as a key feature with discriminative power.

[0144] Output results: Output box plots of feature distribution for each gesture, and label the feature dimensions with significant differences to form a set of feature indicators for distinguishing muscle function differences between healthy and stroke patients (affected side and unaffected side).

[0145] Results: Through comparative analysis of multiple gestures and multidimensional features, the functional impairment patterns of different gesture muscle groups after stroke can be revealed, and the abnormal distribution of electromyographic features on the affected side can be quantitatively identified, providing basic feature support for subsequent functional recovery assessment.

[0146] Verification Scheme 2: Discriminant Analysis of the Exercise Recovery Phase

[0147] Objective: To verify the effectiveness of the extracted macroscopic and microscopic features in distinguishing different stages of motor recovery in stroke patients (Brunnstrom stages 4, 5, and 6) and in healthy individuals, thereby assessing the sensitivity and stage discrimination ability of electromyographic features to the degree of motor recovery.

[0148] Implementation steps:

[0149] (1) Collect electromyographic signals of the affected hand of stroke patients and the unaffected hand of healthy controls under seven hand gestures (WF, WE, WP, WS, HC, TIFP, TIMFP);

[0150] (2) Extract 10 standardized macroscopic and microscopic features and label them according to the Brunnstrom stage or health status of the subjects;

[0151] (3) Construct a classification model based on the K-nearest neighbor algorithm (KNN) and use leave-one-out cross-validation strategy for training and testing;

[0152] (4) Calculate the classification accuracy and confusion matrix for each gesture, and analyze the discrimination performance of each gesture;

[0153] (5) Mark gestures with high classification accuracy and low misjudgment rate as key actions that are most sensitive to the recovery phase.

[0154] Output results: Output the classification accuracy and confusion matrix results for each gesture, identify the most discriminative gesture, and thus achieve reliable judgment of the patient's motor recovery stage.

[0155] Technical benefits: This solution uses classification modeling based on the fusion of multiple gesture features to objectively distinguish different stages of motor recovery, revealing the evolution of electromyographic characteristics from gross motor to fine motor recovery, and providing quantitative support for rehabilitation staging assessment.

[0156] Verification Plan 3: Predictive Analysis of Upper Limb Motor Function Scores

[0157] Objective: To evaluate the predictive ability of extracted macroscopic and microscopic features on Upper Limb Motor Function Assessment Scale (UE-FMA) scores, and to achieve a quantitative assessment of the degree of recovery of fine motor control in stroke patients.

[0158] Implementation steps:

[0159] (1) Summarize the characteristic data of the affected hand of stroke patients under seven gestures and label the corresponding UE-FMA scores; the data of the unaffected hand of the healthy control group are labeled with a full score of 30.

[0160] (2) Construct a regression model based on the K-Nearest Neighbor (KNN) algorithm, input 10 standardized features, and output the predicted UE-FMA score;

[0161] (3) The leave-one-out cross-validation strategy was used for model training and testing, and the coefficient of determination (R²) and root mean square error (RMSE) were calculated.

[0162] (4) Compare the regression performance of different gestures and select the key gestures with the highest prediction accuracy;

[0163] (5) If each gesture can achieve high prediction accuracy, then further calculate the specificity and sensitivity statistics to evaluate the model stability.

[0164] Output results: Output the R² and RMSE values ​​for each gesture, label the optimal predicted gesture and corresponding feature combination, and generate the UE-FMA score prediction result map.

[0165] The results showed that, in terms of macroscopic characteristics, the overall level of the healthy group was higher on the affected side than that of stroke patients, while the affected side was lower than that of the healthy side, reflecting the decline in muscle contraction ability and insufficient energy output caused by stroke. In terms of microscopic characteristics, NMUs were significantly elevated on the affected side, while AFR was decreased, suggesting increased motor unit recruitment and decreased firing rate, revealing the degenerative pattern of neuromuscular control. This method can achieve quantifiable identification of muscle function status among different groups, providing an objective basis for determining the lesion side and functional impairment site in stroke patients, and providing a characteristic-level physiological basis for subsequent staging assessment and rehabilitation intervention (e.g., Figure 3 (As shown).

[0166] (2) Multi-stage rehabilitation grading assessment to achieve automated classification of the stroke exercise recovery stage.

[0167] In the second verification scheme, this invention constructs a K-nearest neighbor (KNN) classification model based on extracted macroscopic and microscopic features to distinguish between patients' Brunnstrom stages (stages 4, 5, and 6) and health status. Results show that the classification accuracy rates for gestures such as WE, TIMFP, and TIFP reach 92.08%, 87.82%, and 87.03%, respectively, effectively distinguishing different stages of motor recovery. Furthermore, misclassifications are mainly concentrated in adjacent stages, consistent with the continuity of patients' functional recovery (e.g., ...). Figure 4 (As shown). This result demonstrates that the present invention can automatically grade and assess the recovery stages of a patient's motor function without relying on subjective scoring, significantly reducing the bias and time-consuming nature of manual assessments and providing rehabilitation physicians with objective and real-time functional assessment data. For example, the system can identify the phased changes in a patient's electromyographic characteristics in real time during rehabilitation training, thereby automatically indicating the timing for phase transitions or adjustments to training intensity.

[0168] (3) Based on electromyographic characteristics, quantitative prediction of motor function is achieved to realize accurate assessment of the recovery level of fine motor function and personalized rehabilitation guidance.

[0169] In the third validation scheme, this invention utilizes a KNN regression model to fit the characteristics of the patient's affected side under various gestures to the upper limb Fugl-Meyer (UE-FMA) score, establishing a mapping model from electromyographic features to clinical scales. Regression results show that the TIMFP, HC, and TIFP gestures exhibit the best predictive performance, with R² values ​​of 0.86, 0.74, and 0.73, respectively. Figure 5 As shown), the RMSE values ​​were 3.26, 4.42, and 4.56 respectively (as shown). Figure 6As shown in the figure, it can accurately reflect subtle changes in the patient's upper limb motor function. Through this model, the system can directly predict the patient's UE-FMA score based on electromyography signals, thereby automating rehabilitation assessment. For example, when the system detects that the patient's characteristic improvement under a specific gesture reaches the corresponding stage threshold, it can automatically update the rehabilitation level and suggest targeted training (such as strengthening thumb fingertip control or wrist extension movements), thereby achieving personalized and dynamic adjustment of the rehabilitation plan.

Claims

1. A stroke patient upper limb motor function assessment method based on surface electromyographic signals, characterized by, The method comprises the following steps: Step 1: selecting subjects Step 2: electrode arrangement and signal acquisition Step 3: experimental paradigm design Step 4: HD-sEMG signal preprocessing Step 5: feature extraction Step 6: model construction Step 7: model evaluation Step 8: clinical application ​ ​ ​ ​ ​ ​ ​ ​ ​ The original HD-sEMG signal is sequentially subjected to: Butterworth band-pass filtering to remove drift and high-frequency noise; notch filtering to eliminate power frequency interference; inter-channel mean correction and standardization; time alignment and segmentation to synchronize the test trigger point; Step 5: Macro-electromyographic feature extraction The following eight types of macro-electromyographic features are extracted from the preprocessed HD-sEMG signal to reflect the overall activity and functional status of the muscle: The eight types of macro-electromyographic features are: (1) Complexity and variability features: including average amplitude change (AAC) and differential average standard deviation change (DASDV), which are used to measure the dynamic fluctuation of muscle activity; (2) Muscle strength features: including mean absolute value (MAV), average spectral amplitude (ASM), and root mean square (RMS), which are used to assess the strength and endurance of muscle contraction; (3) Energy distribution and amplitude change features: including modified Hamming window (MHW) and modified trapezoidal window (MTW) features, which reflect the trend of muscle contraction energy over time; (4) Coordination and synchronization features: including inter-channel synchronization (Sync), which evaluates the synergistic control ability between muscle groups; Step 6: Micro-electromyographic feature extraction Based on the effective MU discharge sequence, the FastICA motor unit decomposition method is used to extract neural-level micro-control features, including: (1) MU number (NMUs): reflecting the number of activated motor units, used to measure the degree of muscle recruitment; (2) Average discharge frequency (AFR): the average discharge frequency of MUs per unit time, used to assess the strength of neural drive.

2. The stroke patient upper limb motor function assessment method according to claim 1, characterized in that, Among the seven different gestures described in Step 3: (1) Wrist flexion (WF), action description: starting from a natural relaxed posture, palm up, wrist flexion towards the palm side, making the palm close to the forearm; main activated muscles: radial flexor carpi, ulnar flexor carpi; (2) Wrist extension (WE), action description: slowly lift the back of the hand from the flexion position, making the palm away from the forearm direction; main activated muscles: radial extensor carpi longus, radial extensor carpi brevis, ulnar extensor carpi; (3) Forearm pronation (WP), action description: forearm from palm up to palm down, keep elbow joint fixed; main activated muscles: pronator teres, pronator quadratus; (4) Forearm supination (WS), action description: forearm from palm down to palm up; main activated muscles: supinator, biceps brachii; (5) Hand clenched (HC), action description: from finger extension, gradually bend all fingers to form a fist, then slowly release; main activated muscles: flexor digitorum superficialis, flexor digitorum profundus; (6) Thumb and index finger pinch (TIFP), action description: thumb and index finger tips are close to each other to form a "pinch" action, then relax; main activated muscles: flexor pollicis longus, first dorsal interosseous muscle; (7) Thumb, index finger and middle finger pinch (TIMFP), action description: thumb simultaneously contacts index finger and middle finger tips to form a three-finger pinch, then relax; main activated muscles: flexor pollicis longus, flexor digitorum superficialis, intrinsic hand muscles.

3. The stroke patient upper limb motor function assessment method according to claim 2, characterized in that, Among the eight types of macro-electromyographic features extracted in Step 5: (1) The calculation formula of average amplitude change (AAC) is: ， （1） Wherein, x(i) represents the i-th sample in the signal, N is the length of the signal; (2) The calculation formula of differential average square standard deviation variation (DASDV) is: ， （2） Wherein, std represents the standard deviation of the absolute difference value of the continuous samples; (3) The calculation formula of mean absolute value (MAV) is: ， （3） Wherein, |x(i)| represents the absolute value of the signal at the sampling point i; (4) The calculation formula of average spectral amplitude (ASM) is: （4） Wherein, exp(i) represents the exponent applied to the value; for the samples within the range of 50% of the signal length, the exponent is 0.5; for the samples outside the range, the exponent is 0.75; (5) The calculation formula of root mean square (RMS) is: ， （5） (6) The calculation formula of modified Hamming window (MHW) is: ， （6） where beginIdx and endIdx represent the start and end indices of the current window segment, respectively, W H (j ) is the Hamming window at position j, sig(j ) is the value of the signal at position j, and N is the length of the window. (7) The calculation formula of modified trapezoidal window (MTW) is: ， （7） ； where W T (j ) represents a trapezoidal window located at position j, and a is a parameter defining the width of the trapezoid. (8) The calculation formula of inter-channel synchronization (Sync) is: （8） where λ1is the largest eigenvalue of the local EMG signal covariance matrix, which represents the variance captured by the first principal component, and is the total variance captured by all N principal components.

4. The stroke patient upper limb motor function assessment method according to claim 1, characterized in that, The effective MU discharge sequence in step 6 is obtained after the following processing: (1) Data splicing processing: 100 sampling point rest segments are added to the splicing of multiple repeated signals under the same gesture to prevent decomposition errors caused by false peaks; (2) MU decomposition processing: channel time delay expansion is performed on the HD-sEMG signal, the expansion order L=8, and then covariance whitening processing is performed to improve the independence of the source signal; (3) FastICA decomposition: the q-non-Gaussianity objective function is used to improve the robustness of the algorithm to abnormal noise, and multiple motor unit discharge signals are separated; (4) MU detection and screening: K-means clustering is used to identify the MU discharge time, and the Silhouette coefficient (SIL) is calculated to evaluate the clustering quality; if the coincidence rate of the discharge sequence within ±1 ms is >50%, the one with higher SIL is retained; the screening condition is SIL≥0.75, if there are less than 6, the top 6 are taken; all MUs are manually checked for their shape and discharge frequency rationality.

5. A system for evaluating the motor function of the upper extremity of a stroke patient based on the evaluation method according to one of claims 1 to 4, characterized in that It includes six functional modules: subject selection module, electrode placement and signal acquisition module, experimental paradigm design module, HD-sEMG signal preprocessing, macroscopic electromyographic feature extraction, and microscopic electromyographic feature extraction module; The six functional modules respectively perform the operations of the six steps of the wind patient upper limb motor function evaluation method.

Images