Cerebral stroke patient lower limb motion intention prediction method and system based on deep learning

Abstract

The invention belongs to the field of motion prediction, and provides a cerebral apoplexy patient lower limb motion intention prediction method and system based on deep learning, and the technical scheme is as follows: obtaining a surface electromyogram signal before motion and an inertial measurement unit signal; obtaining a motion intention recognition result by combining the surface electromyogram signal before motion, an inertial measurement unit and the trained motion intention recognition model; the construction process of the motion intention recognition model comprises the following steps: extracting a plurality of discriminative features from a surface electromyogram signal before motion according to a preset discriminative rule; re-weighting the importance of the discriminative features in combination with a channel attention mechanism, and capturing key features related to biomechanics of each channel according to the importance degree; extracting a time sensing feature from the inertial measurement unit signal; and fusing the key features related to biomechanics and the time perception features of each channel to obtain fused features, and converting the fused features into motion probabilities. And accurate recognition of the motion intention is realized.

Description

Technical Field

[0001] The present invention belongs to the field of motion prediction, and in particular relates to a method and system for predicting lower limb movement intentions of stroke patients based on deep learning. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Most existing lower-limb exoskeleton systems for gait rehabilitation in stroke patients rely on a passive control paradigm, enforcing predetermined motion trajectories. This rigid approach often leads to fatigue and psychological resistance in patients during rehabilitation training. Therefore, developing intention-driven adaptive control strategies that can accurately decode user movement intentions and dynamically adjust exoskeleton assistance has become a key research frontier.

[0004] Surface electromyography (sEMG) and inertial measurement units (IMUs) have become important research areas for intention recognition in the field of exoskeleton human-robot interaction. sEMG can detect neuromuscular activation 30-150 milliseconds before movement onset and is therefore widely used for movement intention recognition. IMUs have attracted attention for their robustness and cost-effectiveness in capturing kinematic signals. However, these methods rely on a single modality, which can result in poor recognition robustness.

[0005] In recent years, the integration of IMUs and EMG has attracted increasing attention. However, these studies have primarily focused on motion recognition and joint angle prediction during or after movement. No research has explored the use of pre-movement signals to predict movement intentions for active exoskeleton control, which is not conducive to identifying the patient's movement intentions. Summary of the Invention

[0006] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides a method and system for predicting lower limb movement intention of stroke patients based on deep learning, which can effectively decode the pre-movement intention of patients with impaired neurological function and enhance patients' active participation in rehabilitation training.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A first aspect of the present invention provides a method for predicting lower limb movement intention in stroke patients based on deep learning, comprising the following steps:

[0009] Acquire surface electromyographic signals and inertial measurement unit signals before movement;

[0010] The movement intention recognition result is obtained by combining the surface electromyography signal before movement, the inertial measurement unit and the trained movement intention recognition model. The construction process of the movement intention recognition model includes:

[0011] Extracting multiple discriminative features from the pre-exercise surface electromyographic signal according to preset discriminative rules;

[0012] Combined with the channel attention mechanism, the importance of discriminative features is re-weighted to capture the key features related to biomechanics of each channel according to their importance.

[0013] Extracting time-aware features from inertial measurement unit signals;

[0014] The biomechanically related key features and time perception features of each channel are fused to obtain fused features, which are then converted into motion probabilities.

[0015] Furthermore, after obtaining the surface electromyographic signal before movement and the inertial measurement unit signal, the surface electromyographic signal before movement and the inertial measurement unit signal are preprocessed including signal filtering, signal extraction and data segmentation.

[0016] Furthermore, the time domain discriminative features include: mean absolute value, waveform length, corrected waveform length, slope sign change, zero crossing point, integrated EMG value, simple square integral, mean, variance, standard deviation, root mean square, Wilson amplitude, EMG percentage, skewness, kurtosis, standard deviation of absolute difference, corrected mean absolute value 1, corrected mean absolute value 2, mean absolute value slope, V order, logarithmic energy detector, mean amplitude change, maximum value, minimum value, crest factor, absolute deviation, and energy entropy.

[0017] Furthermore, the frequency domain discriminant features include: mean frequency, median frequency, autoregressive model coefficient, mean power, total power, spectral moment ratio, peak frequency, power spectrum ratio, center frequency variance, and mean power spectral density.

[0018] Furthermore, the channel attention mechanism is combined to reweight the importance of the discriminative features, and captures the key features related to biomechanics of each channel according to the degree of importance, including:

[0019] The extracted discriminative features are projected into the latent space through a linear layer to obtain the projected features;

[0020] The projected features are subjected to global average pooling and maximum pooling to generate channel attention weights;

[0021] The channel attention weights generated by global average pooling and maximum pooling are passed through the corresponding multi-layer perceptron to obtain the importance weights after discriminative feature adjustment. The optimized features are compressed through a linear layer with Dropout regularization to obtain the key biomechanical features related to each channel.

[0022] Furthermore, time perception features are extracted from the inertial measurement unit signal, including:

[0023] The IMU signal is passed through a 1D convolutional layer to extract the local motion pattern and obtain a one-dimensional IMU feature.

[0024] The extracted one-dimensional IMU features are downsampled using one-dimensional maximum pooling to obtain compressed IMU features;

[0025] The compressed IMU features are passed through the LSTM layer to further capture the temporal information of the downsampled feature sequence;

[0026] The time series features output by LSTM are passed through the attention module, and the adaptive weight of each time step is dynamically calculated based on the attention pooling mechanism to obtain weighted time series information. The time-aware features are obtained by weighted summation of the weighted time series information.

[0027] A second aspect of the present invention provides a system for predicting lower limb movement intention in stroke patients based on deep learning, comprising:

[0028] A signal acquisition unit, which is used to obtain surface electromyographic signals and inertial measurement unit signals before movement;

[0029] The intention prediction module is used to combine the surface electromyography signal before movement, the inertial measurement unit, and the trained movement intention recognition model to obtain the movement intention recognition result. The construction process of the movement intention recognition model includes:

[0030] Extracting multiple discriminative features from the pre-exercise surface electromyographic signal according to preset discriminative rules;

[0031] Combined with the channel attention mechanism, the importance of discriminative features is re-weighted to capture the key features related to biomechanics of each channel according to their importance.

[0032] Extracting time-aware features from inertial measurement unit signals;

[0033] The biomechanically related key features and time perception features of each channel are fused to obtain fused features, which are then converted into motion probabilities.

[0034] A third aspect of the present invention provides a computer-readable storage medium.

[0035] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for predicting lower limb movement intention of stroke patients based on deep learning as described above.

[0036] A fourth aspect of the present invention provides a computer device.

[0037] A computer device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for predicting lower limb movement intention of stroke patients based on deep learning as described above are implemented.

[0038] A fifth aspect of the present invention provides a program product.

[0039] A program product, which is a computer program product, includes a computer program. When the computer program is executed by a processor, it implements the steps in the method for predicting lower limb movement intention of stroke patients based on deep learning as described above.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention combines the surface electromyography signal before movement and the inertial measurement unit signal, designs a dual-channel adaptive channel attention module, combines the discriminative features of the extracted sEMG signal, fuses the two extracted signals, accurately captures the key features related to movement intention, and realizes accurate recognition of movement intention.

[0042] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0044] Figure 1 This is a flow chart of a method for predicting lower limb movement intention in stroke patients based on deep learning provided by an embodiment of the present invention;

[0045] Figure 2 This is the electromyographic signal extraction provided by an embodiment of the present invention; wherein (a) is the right leg signal, and (b) is the left leg signal;

[0046] Figure 3 This is a DCAF-Net network structure diagram provided by an embodiment of the present invention;

[0047] Figure 4 is a patient prediction confusion matrix provided by an embodiment of the present invention; wherein (a) is a left leg prediction confusion matrix, and (b) is a right leg prediction confusion matrix;

[0048] Figure 5 is the ablation experiment result provided by the embodiment of the present invention;

[0049] Figure 6It is the dual-modality and single-modality prediction accuracy provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0051] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0052] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0053] Existing IMU and EMG fusion methods mainly focus on motion recognition and joint angle prediction during or after movement. No research has yet explored the use of pre-movement signals to predict movement intentions to achieve active control of the exoskeleton, which is not conducive to identifying the patient's movement intentions.

[0054] This paper designs a deep learning-based method for predicting lower limb movement intention in stroke patients. The paper proposes a dual-channel attention fusion network (DCAF-Net) that synergistically integrates pre-motor surface electromyography (sEMG) and inertial measurement unit (IMU) data to predict lower limb movement intention in stroke patients. First, a dual-channel adaptive channel attention module is designed to extract key features from 48 time- and frequency-domain features extracted from bilateral gastrocnemius sEMG signals. Second, an IMU encoder combining a convolutional neural network (CNN) and an attention-based long short-term memory (Attention-LSTM) network is designed to decode spatiotemporal movement patterns. Third, sEMG and IMU features are fused through feature concatenation to accurately identify movement intention. Experimental validation was conducted on three healthy subjects and eight stroke patients. The method achieved an average prediction accuracy of 97.19% in the patient group, demonstrating its effectiveness and providing a feasible solution for intention-driven adaptive control in clinical rehabilitation exoskeletons.

[0055] Example 1

[0056] like Figure 1 As shown, this embodiment provides a method for predicting lower limb movement intention of stroke patients based on deep learning, comprising the following steps:

[0057] Step 1: Acquire pre-exercise surface electromyography (sEMG) signals and inertial measurement unit (IMU) data. In this embodiment, the pre-exercise surface electromyography signals are collected using a surface electromyography module, such as a NeuroHUB wireless surface electromyography module. The module includes an sEMG channel, a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer. The sEMG signal and IMU data are collected at a set sampling frequency. For example, the device collects sEMG signals at a sampling rate of 2,000 Hz and IMU data at a sampling rate of 200 Hz. Considering the key role of the gastrocnemius muscle in maintaining postural balance and driving movement, the device is placed on the gastrocnemius muscles of both legs to collect sEMG signals and IMU data. The gastrocnemius muscle of the left leg collects the sEMG signal and IMU data of the first channel, and the gastrocnemius muscle of the right leg collects the sEMG signal and IMU data of the second channel. Finally, the sEMG signal and IMU data of the two channels are obtained.

[0058] Step 2: Preprocess the acquired signal;

[0059] Specifically, the acquired signal preprocessing includes signal filtering, signal extraction and data segmentation.

[0060] Raw sEMG signals contain both muscle electrical activity and unavoidable noise, such as electrical noise from the amplification system. Therefore, filtering is necessary to remove baseline noise before further analysis. In this example, a Butterworth bandpass filter (with a frequency range of 10 Hz to 500 Hz) is used to reduce the impact of high-frequency noise, and a 50 Hz Butterworth notch filter is used to suppress power line interference.

[0061] For IMU data, vertical ground reaction acceleration was used. Because IMU data is susceptible to high-frequency noise, a fourth-order Butterworth low-pass filter was applied to the acceleration signal to remove high-frequency noise.

[0062] Since the goal of this invention is to predict the participant's movement intention to enable active exoskeleton control, and since patients cannot complete walking movements while wearing the exoskeleton, the experiment extracts sEMG and IMU data from a set time before heel lift-off, such as 300 milliseconds. This 300-millisecond data window is used to predict the walking leg (left or right) and stride length (short or long).

[0063] This embodiment adopts a heel-off detection method based on vertical acceleration change, by setting the acceleration threshold (1m / s 2 ) accurately identifies the moment when the heel leaves the ground, thereby providing a time reference for the prediction of movement intention. The schematic diagram of the data extraction process is shown in Figure 2Among them, (a) is the right leg signal, (b) is the left leg signal, the blue part represents the sEMG signal, the yellow part represents the IMU signal, and the red part represents the sEMG signal before movement used for analysis.

[0064] The present invention collected 50 small steps with the right leg, small steps with the left leg, large steps with the right leg, and large steps with the left leg for each experimenter, for a total of 200 steps. Furthermore, the 300ms electromyographic signal and IMU data before each step were extracted. In order to further improve the generalization ability of the prediction, each 300ms electromyographic signal and IMU data was segmented with a window size of 200ms and a window step size of 50ms to obtain three 200ms windows. 600 electromyographic signal and IMU data windows were extracted from the 200 steps data for prediction. Furthermore, these 600 windows were randomly divided into training set and test set according to the ratio of 8:2.

[0065] Step 3: Extract discriminative features based on the preprocessed sEMG signal, including time domain and frequency domain;

[0066] In this embodiment, in order to capture the neuromuscular activation pattern, energy distribution and transient dynamic characteristics, discriminative features are screened from the time domain and frequency domain representations;

[0067] Specifically include:

[0068] Step 301: extracting time domain features based on the preprocessed sEMG signal, and filtering to obtain time domain discriminative features according to preset rules;

[0069] In this embodiment, the time domain discriminative features after the final screening include: mean absolute value, waveform length, corrected waveform length, slope sign change, zero crossing point, integrated electromyographic value, simple square integral, mean, variance, standard deviation, root mean square, Wilson amplitude, electromyographic percentage, skewness, kurtosis, absolute difference standard deviation, corrected mean absolute value 1, corrected mean absolute value 2, mean absolute value slope, V order, logarithmic energy detector, mean amplitude change, maximum value, minimum value, peak factor, absolute deviation, and energy entropy.

[0070] Specific screening rules include:

[0071] Average absolute value: reflects the average signal strength and is directly related to muscle contraction force. Differences in force between the left and right legs will cause changes in the average absolute value.

[0072] Waveform length and modified waveform length: Accumulates signal changes. Large strides or differences in left and right leg movement patterns increase the waveform length value. The modified version suppresses noise interference and improves robustness.

[0073] Slope Sign Change: Detects how often the signal slope changes direction. Different muscle activation rates during the start of a step in the left and right legs can cause a slope sign change.

[0074] Zero crossing point: The number of times the signal crosses zero. The high-frequency adjustment before each leg steps will result in different zero crossing values for the left and right legs.

[0075] Integrated EMG value: represents the total energy output. The integrated EMG value of large strides or unilateral leg-dominated movements is different for the other legs.

[0076] Simple Square Integration: Sum of signal squares, emphasizing high-intensity transients and capturing explosive differences in stride speed.

[0077] Mean: The mean reflects overall strength. The difference in the mean strength of the left and right legs can help distinguish the dominant leg.

[0078] Variance and standard deviation: Variance / standard deviation characterizes signal fluctuations. Differences in force stability between the left and right legs can lead to different variances in the electromyographic signals.

[0079] Root mean square: It is related to the strength of muscle contraction. The stronger the muscle contraction of the swinging leg, the larger the root mean square will be.

[0080] Wilson Amplitude: Counts the number of times the amplitude exceeds the threshold. The amplitude threshold will be exceeded differently for large-stride movements and small-stride movements.

[0081] EMG percentage: The percentage of statistical signals exceeding the threshold. The percentage of large-stride movements and small-stride movements exceeding the amplitude threshold will be different.

[0082] Skewness and kurtosis: Skewness measures the symmetry of the signal distribution, while kurtosis detects peak characteristics. The swing leg EMG signal is more asymmetrical, and the explosive power at the moment of stepping (high kurtosis) is higher.

[0083] Absolute difference standard deviation: The standard deviation of the difference between adjacent sampling points. The absolute standard deviation of irregular gaits will be different.

[0084] Corrected mean absolute value 1 and corrected mean absolute value 2: The segmented weighted mean absolute value is calculated to suppress noise interference and more robustly extract the force difference between the left and right legs.

[0085] Mean absolute value slope: The average slope within the signal window. The slope of the swing leg will increase sharply.

[0086] V-order: Based on the order statistics of signal changes, it captures the dynamic pattern differences of muscle activity during stepping.

[0087] Logarithmic energy detector: emphasizes the energy representation after logarithmic transformation. The amplitude fluctuation of the swing leg signal is large, and the logarithmic energy detector will be lower.

[0088] Average amplitude change: The average value of the amplitude change of adjacent sampling points. The amplitude change will be different when the step amplitude is different.

[0089] Maximum value and minimum value: the amplitude of the signal extreme points. The maximum and minimum values of the electromyographic signal are strongly correlated with muscle force.

[0090] Crest Factor: The ratio of peak value to root mean square value, which can detect sudden force.

[0091] Absolute deviation: The average deviation of the signal from the mean. The absolute deviation of the swinging leg will be greater than that of the non-swinging leg.

[0092] Energy entropy: Quantifies signal complexity based on the entropy of the energy distribution. The entropy of the swinging leg will be different from that of the non-swinging leg.

[0093] Step 302: extract frequency domain features from the pre-processed sEMG signal, and filter according to preset rules to obtain frequency domain discriminant features;

[0094] In this embodiment, the final filtered frequency domain discriminant features include: mean frequency, median frequency, autoregressive model coefficient, mean power, total power, spectral moment ratio, peak frequency, power spectrum ratio, center frequency variance, and mean power spectral density. The filtering rules include: mean frequency and median frequency: the mean frequency is the average value of the spectral energy distribution, and the median frequency is the median of the spectral energy. Changing the step size will change the spectral distribution.

[0095] Autoregressive model coefficients: Use the autoregressive model to fit the spectrum characteristics and extract the coefficients. The differences in spectrum patterns in different gait stages can be distinguished by the autoregressive model coefficients.

[0096] Average power and total power: Average power is the mean of the spectrum, and total power is the total energy. The power will be different for different stride movements.

[0097] Spectral moment ratio: quantifies spectral distortion. The spectral moment ratio of a patient will be different from that of a normal person when walking.

[0098] Peak frequency: The frequency corresponding to the maximum energy in the spectrum. The dominant peak frequency will be different when different legs take steps.

[0099] Power spectrum ratio: the ratio of high-frequency energy. The ratio of high-frequency energy of the swinging leg is different from that of the non-swinging leg.

[0100] Center frequency variance: The variance of the center frequency of the spectrum, which describes the discrete range of the spectrum energy distribution. The center frequency variance will be different when the step size is different.

[0101] Average power spectral density: Based on the power spectral density estimation of the Welch method, the frequency band energy is extracted. The energy captured is different when the amplitude is different.

[0102] Ultimately, a comprehensive set of 48 discriminative features were extracted from the EMG signals.

[0103] Step 4: Combine sEMG, IMU data and the trained movement intention recognition model to obtain the movement intention recognition result;

[0104] like Figure 3 As shown, the specific steps include:

[0105] Step 401: Re-weight the importance of discriminative features by combining the channel attention mechanism, and capture the key features related to biomechanics of each channel according to the degree of importance;

[0106] Each sEMG channel corresponds to a sub-network. The first channel sEMG signal is processed by the first sub-network, and the second channel sEMG signal is processed by the second sub-network. Specifically, the following steps are included:

[0107] Step 4011: Project the 48 discriminative features extracted in step 3 into the latent space through a linear layer to obtain projected features;

[0108] In this embodiment, 48 features are linearly projected into a 128-dimensional latent space.

[0109] Step 4012: The projected features are subjected to global average pooling and maximum pooling to generate channel attention weights, which are expressed as:

[0110]

[0111] Among them, f avg is the output of the average pooling layer, f max is the output of the maximum pooling layer, AvgPool(·) is the global average pooling, MaxPool(·) is the maximum pooling, f is the input feature map, is the number of output channels. Step 4013: The channel attention weights generated by global average pooling and maximum pooling are passed through the corresponding multi-layer perceptron to obtain the adjusted importance weights of the discriminative features. Then, the optimized features are compressed into a 64-dimensional representation through a linear layer with Dropout regularization, which is expressed as:

[0112]

[0113] f attn =f⊙w (4),

[0114] Among them, w is the weight coefficient matrix, σ is the Sigmoid function, F MLP Represents a two-layer MLP (Multi-layer Perceptron) with a GELU activation function. This mechanism ensures that the network can focus on the most discriminative EMG features while suppressing noise and artifacts.

[0115] The purpose of this step is to reduce the influence of non-discriminative muscle activation patterns. The channel attention mechanism dynamically reweights the importance of features. This module combines global average pooling and maximum pooling to generate channel attention weights, thereby more accurately capturing key features related to biomechanics.

[0116] Step 402: extracting time perception features from the inertial measurement unit signal;

[0117] The specific steps include:

[0118] Step 4021: Extract local motion patterns from the IMU signal through a 1D convolutional layer.

[0119] The original IMU data is passed through a one-dimensional convolution layer with a convolution kernel size of 3 to preliminarily extract the IMU features.

[0120] Step 4022: Perform temporal downsampling through maximum pooling.

[0121] The extracted one-dimensional IMU features are downsampled using one-dimensional maximum pooling, compressing the time series IMU sequence size from 40 to 20 and reducing model parameters. The maximum pooling layer is defined as follows:

[0122]

[0123] Among them, f max is the output of the maximum pooling layer, MaxPool(·) is the maximum pooling, f is the input feature map, is the number of output channels.

[0124] Step 4023: The compressed IMU features are passed through the LSTM layer to further capture the temporal information of the downsampled feature sequence;

[0125] Attention-based pooling is defined as follows:

[0126]

[0127]

[0128] Among them, α t is the weight coefficient matrix, c is the output of the attention layer, is the hidden state of LSTM at time step t, W h ,v,b is a learnable parameter. This method ensures that the network can dynamically focus on the most informative time steps, thereby improving the robustness to timing changes.

[0129] Step 4024: The time series features output by the LSTM are passed through the attention module, and the adaptive weight of each time step is dynamically calculated based on the attention pooling mechanism to obtain weighted time series information and highlight the key motion stages; finally, the IMU representation obtains the time perception features by weighted summation of the weighted time series information.

[0130] Step 403: Fusing the biomechanical key features and time perception features of each channel to obtain fused features, and converting the fused features into motion probabilities;

[0131] The network combines sEMG features (64 dimensions per branch) and IMU (32 dimensions) into a unified 160-dimensional representation by concatenation. A lightweight classifier (composed of linear layers with layer normalization and GELU activation) converts the fused features into motion probabilities.

[0132] This architecture effectively preserves modality-specific information while promoting implicit cross-modal interactions in deeper layers, thereby achieving an optimal balance between computational efficiency and discriminative performance. The classification process is formally defined as follows:

[0133] P(y|x)=Softmax(W2·GELU(W1[f L ;f R ;f IMU ])) (7),

[0134] in,

[0135] In order to verify the beneficial effects of the present invention, 11 participants (8 stroke patients and 3 healthy controls) were recruited in this experiment.

[0136] The experimental protocol stipulated that each participant complete 200 gait cycles during data collection. Using sEMG signal window segmentation, 600 signal windows were extracted from each participant. The dataset was split into training and test sets in an 8:2 ratio, resulting in 480 training samples and 120 test samples for each participant.

[0137] As shown in Table 1, the classification results demonstrate the performance of DCAF-Net across all participants, including eight stroke patients (subjects 1-8) and three healthy individuals (subjects 9-11). In the stroke patient group, the model achieved an average accuracy of 96.2%, with four subjects (subjects 1, 2, 5, and 6) achieving 100% accuracy. The overall average accuracy for the stroke patient group was 97.19%. In comparison, the healthy control group performed slightly lower, with an average accuracy of 93.56%. These results strongly demonstrate the effectiveness of DCAF-Net for predicting movement intentions in both stroke patients and healthy individuals.

[0138] Table 1 DCAF-Net classification performance

[0139]

[0140]

[0141] To investigate the impact of different affected limbs on model predictions, Figure 4 The confusion matrices of stroke patients with affected left and right legs are shown, where (a) is the confusion matrix of stroke patients with affected left legs, and (b) is the confusion matrix of stroke patients with affected right legs. Figure 4 As shown in the figure, the confusion matrices for left- and right-leg patients exhibit similar patterns, with similar proportions and types of misclassifications. This indicates that the difference in affected limbs did not significantly affect the model's predictive performance, demonstrating that the model exhibits consistent robustness across patients with different affected limbs.

[0142] Figure 5 The ablation experiment results of DCAF-Net are shown. Figure 5 As shown in the figure, when the model switches from a dual-branch structure to a single-branch structure, its ability to extract features from electromyographic (sEMG) signals decreases significantly. Specifically, the average accuracy of motion intention prediction decreases by 1.65% in healthy individuals and 4.12% in stroke patients. In addition, when the attention mechanism in the electromyographic feature extraction module and the IMU feature extraction module is removed, the average accuracy of motion intention prediction decreases by 0.94% in healthy individuals and 1.62% in stroke patients. These results verify the effectiveness of the dual-branch structure and attention mechanism in DCAF-Net.

[0143] In order to investigate the role of each modality in the motion intention prediction task, Figure 6 The prediction accuracy of single modality (IMU only and EMG only) and dual-modality fusion in 11 subjects (including 8 stroke patients and 3 healthy individuals) was demonstrated. The results showed that dual-modality fusion achieved higher accuracy in all subjects compared to any single modality. Specifically, when using only IMU data, the average motion prediction accuracy of stroke patients and healthy individuals was 57.39% and 56.11%, respectively; while when using only EMG data, the accuracy was 68.95% and 65.82%, respectively. In contrast, the fusion of IMU and EMG data significantly improved the average accuracy of stroke patients to 96.96% and the accuracy of healthy individuals to 93.56%, highlighting the complementary advantages of multi-sensor fusion.

[0144] Example 2

[0145] This embodiment provides a deep learning-based system for predicting lower limb movement intention in stroke patients, including:

[0146] A signal acquisition unit, which is used to obtain surface electromyographic signals and inertial measurement unit signals before movement;

[0147] The intention prediction module is used to combine the surface electromyography signal before movement, the inertial measurement unit, and the trained movement intention recognition model to obtain the movement intention recognition result. The construction process of the movement intention recognition model includes:

[0148] Extracting multiple discriminative features from the pre-exercise surface electromyographic signal according to preset discriminative rules;

[0149] Combined with the channel attention mechanism, the importance of discriminative features is re-weighted to capture the key features related to biomechanics of each channel according to their importance.

[0150] Extracting time-aware features from inertial measurement unit signals;

[0151] The key biomechanically relevant characteristics of each channel

[0152] It should be noted that the specific implementation method of the deep learning-based lower limb movement intention prediction system for stroke patients in an embodiment of the present invention is similar to the specific implementation method of the deep learning-based lower limb movement intention prediction method for stroke patients in an embodiment of the present invention. Please refer to the description of the method part for details. In order to reduce redundancy, it will not be repeated here.

[0153] Example 3

[0154] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps in the method for predicting lower limb movement intention of stroke patients based on deep learning as described above are implemented.

[0155] Example 4

[0156] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for predicting lower limb movement intention of stroke patients based on deep learning as described above are implemented.

[0157] Example 5

[0158] This embodiment provides a program product, which is a computer program product, including a computer program. When the computer program is executed by a processor, it implements the steps in the method for predicting lower limb movement intention of stroke patients based on deep learning as described above.

[0159] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage and optical storage, etc.) containing computer-usable program code.

[0160] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0161] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0163] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0164] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for predicting lower limb movement intention in stroke patients based on deep learning, characterized by: The steps include: Acquire surface electromyographic signals and inertial measurement unit signals before movement; The movement intention recognition result is obtained by combining the surface electromyography signal before movement, the inertial measurement unit and the trained movement intention recognition model; The construction process of the motion intention recognition model includes: Extracting multiple discriminative features from the pre-exercise surface electromyographic signal according to preset discriminative rules; Combined with the channel attention mechanism, the importance of discriminative features is re-weighted to capture the key features related to biomechanics of each channel according to their importance. Extracting time-aware features from inertial measurement unit signals; The biomechanically related key features and time perception features of each channel are fused to obtain fused features, which are then converted into motion probabilities.

2. The method for predicting lower limb movement intention of stroke patients based on deep learning as claimed in claim 1, characterized in that: After obtaining the surface electromyographic signal before movement and the inertial measurement unit signal, the method includes performing signal filtering, signal extraction and data segmentation preprocessing on the surface electromyographic signal before movement and the inertial measurement unit signal.

3. The method for predicting lower limb movement intention of stroke patients based on deep learning as claimed in claim 1, characterized in that: The time domain discriminative features include: mean absolute value, waveform length, corrected waveform length, slope sign change, zero crossing point, integrated EMG value, simple square integral, mean, variance, standard deviation, root mean square, Wilson amplitude, EMG percentage, skewness, kurtosis, standard deviation of absolute difference, corrected mean absolute value 1, corrected mean absolute value 2, mean absolute value slope, V order, logarithmic energy detector, mean amplitude change, maximum value, minimum value, crest factor, absolute deviation, and energy entropy.

4. The method for predicting lower limb movement intention of stroke patients based on deep learning as claimed in claim 1, characterized in that: Frequency domain discriminant features include: mean frequency, median frequency, autoregressive model coefficient, mean power, total power, spectral moment ratio, peak frequency, power spectrum ratio, center frequency variance, and mean power spectral density.

5. The method for predicting lower limb movement intention of stroke patients based on deep learning as claimed in claim 1, characterized in that: The channel attention mechanism is combined to reweight the importance of discriminative features, and captures the key features related to biomechanics of each channel according to the degree of importance, including: The extracted discriminative features are projected into the latent space through a linear layer to obtain the projected features; The projected features are subjected to global average pooling and maximum pooling to generate channel attention weights; The channel attention weights generated by global average pooling and maximum pooling are passed through the corresponding multi-layer perceptron to obtain the importance weights after discriminative feature adjustment. The optimized features are compressed through a linear layer with Dropout regularization to obtain the key biomechanical features related to each channel.

6. The method for predicting lower limb movement intention of stroke patients based on deep learning as claimed in claim 1, characterized in that: Extract time-aware features from inertial measurement unit signals, including: The IMU signal is passed through a 1D convolutional layer to extract the local motion pattern and obtain a one-dimensional IMU feature. The extracted one-dimensional IMU features are downsampled using one-dimensional maximum pooling to obtain compressed IMU features; The compressed IMU features are passed through the LSTM layer to further capture the temporal information of the downsampled feature sequence; The time series features output by LSTM are passed through the attention module, and the adaptive weight of each time step is dynamically calculated based on the attention pooling mechanism to obtain weighted time series information. The time-aware features are obtained by weighted summation of the weighted time series information.

7. A deep learning-based prediction system for lower limb movement intention in stroke patients, characterized by: include: A signal acquisition unit, which is used to obtain surface electromyographic signals and inertial measurement unit signals before movement; The intention prediction module is used to combine the surface electromyography signal before movement, the inertial measurement unit and the trained movement intention recognition model to obtain the movement intention recognition result; The construction process of the motion intention recognition model includes: Extracting multiple discriminative features from the pre-exercise surface electromyographic signal according to preset discriminative rules; Combined with the channel attention mechanism, the importance of discriminative features is re-weighted to capture the key features related to biomechanics of each channel according to their importance. Extracting time-aware features from inertial measurement unit signals; The biomechanically related key features and time perception features of each channel are fused to obtain fused features, which are then converted into motion probabilities.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method for predicting lower limb movement intention of stroke patients based on deep learning are implemented.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps in the method for predicting lower limb movement intention of stroke patients based on deep learning are implemented.

10. A program product, wherein the program product is a computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting lower limb movement intention of stroke patients based on deep learning are implemented as described in any one of claims 1 to 6.

Images