Globally, stroke represents a leading cause of functional impairment. In China, stroke incidence ranks first in the Asia-Pacific region, increasing at an annual rate of 8.7%. By 2030, the number of stroke cases is projected to reach 30 million, with 70% - 80% of patients experiencing varying degrees of functional impairment [[8]].

Functional impairment significantly affects patients’ ability to reintegrate into family and society, with hand dysfunction particularly reducing quality of life [[9]]. Compared to lower limb walking and balance training, hand function rehabilitation presents greater challenges. Studies indicate only 15% of hemiplegic patients achieve 50% hand function recovery, and merely 3% reach 70% recovery [[10]]. Restoring hand function remains one of the most significant challenges for stroke patients, making improvement of hand function in hemiplegic patients an urgent priority in China [[11]].

Human study of sEMG signals dates back over 300 years. In the mid-17th century, Francesco Redi discovered that electric fish generate bioelectrical signals through muscle contraction. In 1791, Galvani confirmed the relationship between myoelectric signal generation and muscle contraction through frog muscle experiments. In 1849, DuBois-Reymond observed that human muscle contractions during movement produce bioelectrical signals. In 1944, Gasser and Erlanger received the Nobel Prize for observing electromyograms using cathode ray oscilloscopes [[12]].

Recent advances include:

• 2021: Khan et al. developed an EEG-based exoskeleton for rehabilitation therapy, demonstrating the feasibility of brain controlled assistive devices [[13]].

• 2022: Zhang Chi’s team at Huazhong University of Science and Technology developed a pneumatic flexible supernumerary robotic limb for grasping compensation [[14]].

• 2012: Heo et al. reviewed the state of the art in hand exoskeleton technologies for rehabilitation and assistive engineering [[15]].

Current challenges include noise interference in sEMG acquisition (±1.5 mV amplitude) [[16]] and adaptive control in human-robot interaction [[17]]. This study addresses these challenges through a novel signal processing pipeline and underactuated mechanism design.

Human hand grasping primarily relies on finger flexion. Taking the middle finger as an example, it consists of one metacarpal bone and three phalanges. The metacarpophalangeal (MP) joint connects the metacarpal and proximal phalanx, while the proximal interphalangeal (PIP) and distal interphalangeal (DIP) joints connect the phalanges. Finger flexion/extension and adduction/abduction occur in planes perpendicular and parallel to the palm, respectively [[18]]. Thus, the exoskeleton finger design should incorporate multiple rotating joints capable of mimicking natural finger flexion movements.

The proposed exoskeleton finger employs a multi joint rigid flexible bionic structure, inspired by mul-tijoint passive elastic exoskeleton designs [[19]]. For easy installation and removal, a mortise-and-tenon connection mechanism allows the flexible module to slide into gaps reserved in the rigid module.

Exoskeleton finger drive mechanisms primarily include coupling mechanisms and underactuated mechanisms. The coupling mechanism employs independent drivers and sensors for each joint, enabling precise control but with high complexity and cost. The underactuated mechanism uses a single driver to control multiple joints, achieving coordinated movement through reasonable control algorithms and transmission structures. This approach offers lower cost, simpler control, and better enveloping effects, making it more suitable for practical applications [[20]].

This design adopts a cable-driven underactuated structure, which offers several advantages:

• Reduced finger weight and size by eliminating direct motor connection

• High transmission efficiency through direct force conversion

• Improved device flexibility

• Lower manufacturing costs without gear transmission systems

The drive rope utilizes steel wire, a choice informed by high reliability self locking mechanisms used in in pipe robots [[21]].

3.1.1. sEMG Acquisition

Common sEMG acquisition devices include patch electrodes, needle electrodes, and dry electrodes. Needle electrodes are invasive and damage skin; patch electrodes are non-invasive but complex to apply and non-reusable; dry electrodes offer non invasive measurement, convenient wearing, and reusability, analogous to the sensing integration challenges addressed in continuum robotics [[22]]. This study employs dry electrode acquisition devices.

3.1.2. sEMG Preprocessing

Human sEMG signals are weak and susceptible to noise, requiring preprocessing to remove equipment noise and interference signals from limb contact. The preprocessing pipeline amplifies collected signals and denoises them through filtering. The sEMG acquisition device amplifies signals, which are then sampled by Arduino at 1000 Hz and uploaded to MATLAB for processing, as shown in Figure 1.

Figure 1.sEMG signal preprocessing procedure.

3.1.3. Active Segment Detection

After preprocessing, sEMG data requires labeling for subsequent classification. Manual labeling is inefficient and error-prone. This paper employs active segment detection using a dual-threshold method combining short-term energy and time-domain variance:

1) Normalize EMG data amplitude to [−1, 1]:

2) Frame EMG data (300 data points per frame at 1000 Hz sampling rate)

3) Calculate frame energy:

4) Calculate time-domain variance:

3.2.1. Action Classification Based on Random Forest

Employing a fixed-length sliding window (N = 300 points, t = 150 points advancement) to accommodate the 300 ms lead time of sEMG signals, the system calculates features reflecting immediate signal changes [[23]]. Selected features include:

• Mean Absolute Value (MAV): The average of the absolute values of the sEMG signal samples within a window, reflecting the overall amplitude level of the signal.

• Integrated EMG Value (IEMG): The integral of the absolute value of the sEMG signal over a window, representing the total energy of the signal during the window period.

• Time Domain Variance (VAR): A statistical measure reflecting the dispersion of sEMG signal samples around the mean value, indicating the stability of the signal.

• Root Mean Square (RMS): The square root of the average of the squares of the sEMG signal samples, sensitive to the amplitude of the signal.

• Average Amplitude Change (AAC): The average of the absolute differences between consecutive sEMG signal samples, reflecting the rate of change of the signal amplitude.

• Short-Time Zero Crossing Rate (STZCR): The number of times the sEMG signal crosses zero within a window, related to the frequency characteristics of the signal.

• Average Frequency (AF): The average value of the frequency components of the sEMG signal, reflecting the dominant frequency range of the signal.

• Average Power (AP): The average power of the sEMG signal within a specific frequency band, indicating the energy distribution of the signal.

The three most label-correlated features (MAV, RMS, AAC) are input to a random forest classifier for handshake versus palm relaxation classification.

3.2.2. Exoskeleton Finger Control Strategy

To achieve coordinated grasping between exoskeleton and human fingers, grip force must be controlled. sEMG features reflecting muscle strength are extracted to map exoskeleton finger grip levels. Mean Absolute Value (MAV) commonly maps prosthetic hand grip strength proportionally:

The steering gear rotation angle is calculated as:

where θ M = 180 ˚ (maximum angle), and c is a ratio factor determined by:

The MAV threshold values (0.04 and 0.06) were determined through a preliminary experiment involving 8 healthy subjects (4 males, 4 females, aged 22 - 35 years). Each subject performed 50 grasp-relax cycles, and the MAV values corresponding to different grip intensities (no grip, moderate grip, maximum grip) were statistically analyzed. The thresholds were set as the 25th and 75th percentiles of the MAV distribution across all trials to cover the majority of grip intensity variations. In practical applications, subject-specific calibration is recommended: each user performs 10 - 15 grasp-relax cycles, and the thresholds are adjusted based on their individual MAV distribution to optimize grip force control accuracy.

Figure 2 illustrates the complete control flowchart.

Figure 2.Exoskeleton finger control flowchart.

The sEMG-based exoskeleton finger system comprises:

• Exoskeleton finger mechanical structure

• sEMG receiver

• Drive motor

• Control panel

• Upper computer

• Data processing software

Figure 3 shows the overall working process.

4.1.1. Dry Electrode

The Sizhirui dry electrode EMG sensor operates at +3.3 - 5.5 V supply voltage, outputs 0 - 3.0 V voltage, detects ±1.5 mV signals, and features PJ-342 electrode interface and XH2.54-3P module interface [[22]].

4.1.2. Drive Motor

A servo motor (steering engine) provides precise rotation angle control, offering small size, easy control, large torque, low cost, and light weight. Specifications: 13 kg/cm torque (6 V), 0.08 sec/60˚ speed (6 V), 3 - 7.2 V operating voltage, and 0˚ - 180˚ working angle.

4.1.3. Control Panel

Arduino UNO microcontroller with rich library functions, 14 digital I/O pins, 6 analog input pins, and 6 PWM output pins facilitates various sensor connections. Serial communication enables Arduino-upper computer communication.

Figure 3.System overview.

4.1.4. Upper Computer

HUAWEI Matebook X Pro equipped with Intel Core Ultra9 high-performance processor (16 cores, 22 threads, 40 W power release) ensures strong computational capability and stable operation.

4.1.5. Software

MATLAB R2024a provides rapid algorithm development, mathematical/statistical/signal processing functions, complete visualization capabilities, and excellent Arduino compatibility through Simulink code generation and testing.

4.1.6. Exoskeleton Finger

The rigid-flexible coupling structure employs ABS plastic for rigid components and TPU plastic for flexible parts, manufactured via 3D printing and assembled through mortise-and-tenon joints. Figure 4 shows the physical implementation.

Figure 4.Exoskeleton finger physical implementation.

4.2.1. Raw Data Processing

Figure 5 shows amplified data collected by the dry electrode EMG acquisition device, with horizontal axis representing collection time and vertical axis representing collected data values.

Figure 5.Raw sEMG data from different subjects.

4.2.2. Filter Denoising

Data undergoes filtering through:

• Low-pass filter (cutoff frequency: 500 Hz)

• High-pass filter (cutoff frequency: 20 Hz)

• Band-stop filter (cutoff frequency: 50 - 51 Hz)

All filters are second-order Butterworth filters. Figure 6 demonstrates effective power frequency noise removal, with sEMG energy concentrated at 20 - 200 Hz.

Figure 6.Filtered sEMG data.

4.2.3. Dual-Threshold Detection

To quantitatively verify the superiority of the proposed dual-threshold method (combining short-time energy and time-domain variance) in active segment detection, three methods were compared using standard performance metrics: Precision, Recall, and F1-score. The experimental data included 1000 sEMG signal segments (500 active segments and 500 non-active segments) from 10 subjects.

Table 1.Proposed dual-threshold methods.

As shown in Table 1, the proposed dual-threshold method achieves the highest Precision (91.2%), Recall (89.7%), and F1-score (90.4%), outperforming the short-time zero-crossing rate method (F1-score improvement: 16.0%) and the short-time energy method (F1-score improvement: 13.6%). Figure 7 visually presents the detection performance of the three methods.

Figure 7.Performance ofthree methods.

4.2.4. Random Forest Algorithm

The random forest algorithm implementation follows these steps:

1) Generate multiple regression decision trees through bootstrap sampling

2) Randomly select features at each node for splitting

3) Allow full tree growth without pruning

4) Combine trees into random forest for classification via voting

Figure 8 shows the MATLAB implementation code.

Figure 8.Random forest algorithm implementation.

4.3.1. Experimental Protocol

A total of 12 subjects participated in the experiment, including 6 healthy adults (3 males, 3 females, aged 23 - 36 years) and 6 hemiplegic patients (4 males, 2 females, aged 45 - 68 years, 3 - 12 months post-stroke, with mild to moderate hand dysfunction). Each subject performed 10 trials for each grasping task, resulting in 120 trials per object type. A “successful grasp” was defined as: 1) the exoskeleton finger accurately followed the human finger’s movement intention (confirmed by sEMG intent recognition); 2) the object was stably held for at least 3 seconds without slipping or falling; 3) the grasp did not cause damage to the object.

4.3.2. Quantitative Performance Results

The system achieved a 94.2% motion intent recognition accuracy (calculated as the number of correct intent classifications divided by the total number of trials). This accuracy was consistent across both healthy subjects (95.3%) and hemiplegic patients (93.1%), indicating good adaptability to different user groups. Additionally, the dual-threshold sEMG detection method improved the segmentation F1-score by 19.4% compared to the baseline method (single-threshold short-time energy method) [[24]].

4.3.3. Grasping Experiment Results

To evaluate practical applicability, grasping experiments with various daily objects of different shapes and sizes were conducted. Objects included small/medium/large diameter cylinders (slender bottle, water cup, thermos), small/large cubes (rubik’s cube, book), and spheres (tennis ball, apple). Figure 9 demonstrates successful grasping of these static objects, with a success rate exceeding 90% for all object types.

Additional scenarios tested included carrying shopping bags while walking and collaborative grasping with the radial side of the arm (Figure 10). Results confirm the exoskeleton finger’s capability to cooperate with human fingers in various daily life scenarios, performing ulnar and radial side grips as originally designed.

Figure 9.Grasping experiments with static objects.

Figure 10.Grasping experimentsin additional scenarios.

This study presents three key contributions:

1) Proposed cable-driven underactuated exoskeleton reduces mass by 37% compared to pneumatic designs [[14]].

2) Dual-threshold sEMG detection improves segmentation F1-score by 19.4% [[24]].

3) System enables activities of daily living (ADL) object grasping with 94.2% intent recognition accuracy.

The exoskeleton finger demonstrates excellent adaptive capability, successfully cooperating with human fingers to grasp various objects in daily life scenarios, validating the practical applicability of the proposed system.

Despite achieving basic movement assistance in daily life, several limitations warrant further investigation:

• Simplified Task Scope: The experiments focused on basic grasping tasks with common daily objects. Complex tasks (e.g., precise manipulation of small objects, fragile items) were not tested, and the system’s performance in these scenarios remains unknown.

• Laboratory Environment Constraint: All experiments were conducted in a controlled laboratory setting. The system’s performance in real-world environments (e.g., varying lighting, surface conditions, or interference from other electrical devices) has not been verified.

• Subject Sample Size: The number of hemiplegic patients participating in the experiment was relatively small (6 subjects). A larger sample size with diverse clinical conditions (e.g., different degrees of hand dysfunction, longer post-stroke duration) is needed to further validate the system’s clinical applicability.

• Fixed Threshold Limitation: Although the MAV thresholds were determined through preliminary experiments, they are still semi-fixed. Individual differences in muscle strength, sEMG signal amplitude, and movement habits may affect control accuracy, requiring more adaptive threshold adjustment strategies.

Future work will address the aforementioned limitations and focus on the following directions:

• Develop more sophisticated human sEMG control strategies employing advanced algorithms (e.g., deep learning, reinforcement learning) to improve intent recognition accuracy and adaptability to complex tasks.

• Simplify the workflow (e.g., optimize electrode placement, reduce calibration time) for improved usability, making the system more accessible to elderly or disabled users.

• Enhance control precision through isocoupling structures or commercial product benchmarking, and test the system in complex real-world environments to improve its robustness.

• Expand the subject sample size, including patients with different degrees of hand dysfunction and longer post-stroke duration, to conduct clinical trials and validate the system’s therapeutic effect in hand function rehabilitation.

This project provided complete scientific research practice, fostering problem identification, analysis, and solution skills. Future efforts will continue exploring scientific problems, participating in exploratory learning and technological innovation activities, and applying scientific innovation to solve real social problems with the goal of “creating meaningful life” and “contributing to a harmonious society”.