Paired foot pressure data-based preliminary screening method for sarcopenia

Abstract

The invention belongs to the field of medical health, and provides a preliminary sarcopenia screening method based on paired foot pressure data, which comprises the following steps: firstly, acquiring foot pressure signals of left and right feet of a subject in a walking process to obtain corresponding foot pressure characteristics of the left and right feet; fusing by adopting a multiplicative combination mode to obtain total pressure characteristics of each walking stage in each walking period; then, a GMM model is adopted to create a plantar pressure mode, plantar pressure mode features are obtained, attention mechanism fusion is adopted to obtain single-cycle fusion pressure mode features and multi-cycle fusion pressure mode features, and time sequence features are obtained through learning of an improved GRU module; and finally, sending the corresponding multi-cycle fusion pressure mode features and the time sequence features into a multi-layer perceptron, and outputting the prediction probability of the sarcopenia. The method fully considers the independence of the left and right foot pressure signal features, introduces the joint relation of the two features, more comprehensively learns the gait features, and effectively improves the accuracy of primary screening of sarcopenia.

Description

Technical Field

[0001] This invention belongs to the field of medical and health care and relates to sarcopenia screening technology, specifically providing a preliminary sarcopenia screening method based on paired foot pressure data. Background Technology

[0002] Sarcopenia (or muscle loss) is an age-related disorder characterized by a decline in skeletal muscle mass, muscle strength, or bodily function. It is more common in older adults and increases the risk of falls, fractures, disability, hospitalization, and even death.

[0003] Current sarcopenia diagnosis mainly relies on the "gold standard" for diagnosis, such as DXA (dual-energy X-ray absorptiometry), CT (magnetic resonance imaging), ultrasound and other related image data. As artificial intelligence has demonstrated good processing capabilities for image data, AI-based sarcopenia-assisted diagnosis technologies have been proposed one after another. These technologies are based on the "gold standard" image data and involve everything from delineating sarcopenia-related key areas (ROIs) to acquiring image features. For example, Chinese patent document CN 115376681 A discloses an AI-based automatic screening method for sarcopenia, proposing an image segmentation method based on U-Net, which can not only indicate muscle area but also provide more effective information for intervention, such as muscle density and muscle-to-fat ratio. Another example is a method for constructing an automatic diagnostic model for sarcopenia disclosed in Chinese patent document CN 117095813 A, which proposes a scheme for automatic diagnosis using ultrasound images, converting ultrasound images to the frequency domain, and extracting features from the images through four-layer feature convolution. However, these methods all rely on "gold standard" image data, which often requires expensive examination equipment, making them impractical for initial screening in communities.

[0004] To address the aforementioned issues, Chinese patent document CN 117426770 A discloses a method and system for constructing a gait recognition model based on plantar pressure characteristics. This method performs gait recognition analysis on patient walking images to screen for sarcopenia, achieving an accuracy rate of up to 90% using machine learning SVM and LSTM models. However, this approach lacks discussion on the related connections between the left and right feet, thus affecting the accuracy of gait recognition. Furthermore, compared to image-based gait recognition, using pressure sensors for gait recognition is more feasible. Gait recognition based on plantar pressure can be applied to the auxiliary screening of degenerative diseases, such as early screening for sarcopenia. Therefore, this invention provides a preliminary sarcopenia screening method based on paired foot pressure data. Summary of the Invention

[0005] The purpose of this invention is to provide a preliminary screening method for sarcopenia based on paired foot pressure data. This method fully considers the independence of the pressure signal characteristics of the left and right feet while introducing the joint relationship between the two, so as to learn gait characteristics more comprehensively and effectively improve the accuracy of the preliminary screening for sarcopenia.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A preliminary screening method for sarcopenia based on paired foot pressure data, characterized by comprising the following steps:

[0008] Step 1: Collect the pressure signals of the left foot and right foot during the subject's walking process, and extract features after time alignment to obtain the left foot pressure features and right foot pressure features of each walking stage in each walking cycle.

[0009] Step 2: Use a multiplicative combination method to fuse the pressure characteristics of the left and right feet to obtain the overall pressure characteristics of each walking stage within each walking cycle;

[0010] Step 3: Use the GMM model to create plantar pressure patterns for each walking stage within each walking cycle, and obtain the corresponding plantar pressure pattern features.

[0011] Step 4: Use an attention mechanism to fuse the plantar pressure pattern features of each walking stage within a single walking cycle to obtain the single-cycle fused pressure pattern features of a single walking cycle; then use an attention mechanism to fuse the single-cycle fused pressure pattern features of multiple walking cycles to obtain the multi-cycle fused pressure pattern features.

[0012] Step 5: Based on the single-cycle fusion pressure mode characteristics and the multi-cycle fusion pressure mode characteristics, the improved GRU module is used to learn in time sequence to obtain the time sequence characteristics of the corresponding walking cycle.

[0013] Step 6: Feed the multi-cycle fusion stress pattern features corresponding to all walking cycles and the temporal features of the last walking cycle into a multilayer perceptron (MLP layer) to output the predicted probability of sarcopenia.

[0014] Furthermore, in step 1, the pressure signal sig of the left foot... l With right foot pressure signal sig r The data collection process is as follows:

[0015] The pressure signals (sig) of the left foot generated by the subject during a complete walking process were collected using an insole equipped with N pressure sensors. l With right foot pressure signal sig rThe system captured pressure signals from both feet for C walking cycles. Each walking cycle c was divided into S walking stages, and the pressure signal from the left foot for each walking stage s within walking cycle c was obtained. Pressure signal from the right foot c=1,2,…,C, s=1,2,…,S.

[0016] Furthermore, the walking cycle includes the following stages: initial contact, load response, intermediate standing, final standing, pre-swing, initial swing, intermediate swing, and end swing. The walking stages are divided by pressure extreme value detection.

[0017] Furthermore, in step 1, a 1D CNN network is used as the foot pressure feature extractor for feature extraction. For the walking phase s in the walking cycle c, the left foot pressure feature... Pressure characteristics of the right foot Represented as: CNN stands for 1D CNN network.

[0018] Furthermore, in step 2, the overall pressure characteristics are expressed as follows:

[0019]

[0020] Among them, P c,s This represents the overall pressure characteristics of the walking phase s within the walking cycle c. and W represents the pressure characteristics of the left foot and the right foot, respectively. l W r b is the learning matrix corresponding to the pressure features of the left and right feet. l ,b r This is the bias vector during the learning process.

[0021] Furthermore, in step 3, the plantar pressure pattern characteristics of walking phase s within walking cycle c. Represented as:

[0022]

[0023] Among them, P c,s This represents the overall pressure characteristics of the walking phase s within the walking cycle c.

[0024] Furthermore, in step 4, an attention mechanism is used to obtain fused pressure pattern features based on the plantar pressure pattern characteristics of the first s walking phases within the current walking cycle c. Represented as:

[0025]

[0026] in, The plantar pressure pattern characteristics of the walking phase s in the walking cycle c, a s W represents the attentional characteristics of each walking phase within the walking cycle. a W is the attention learning matrix during the walking cycle. f With b f For the transformation weights and biases of the pressure pattern characteristics within the walking cycle, A s Att represents the attention score for each walking phase within the walking cycle, and Att represents the attention mechanism.

[0027] The plantar pressure pattern characteristics of the last walking phase S of the current walking cycle c are taken as the single-cycle fusion pressure pattern characteristics F of walking cycle c. c :

[0028] For the single-cycle fusion stress pattern features of the first c walking cycles, an attention mechanism is used to obtain the multi-cycle fusion stress pattern features. Represented as:

[0029]

[0030] Among them, g c Attentional characteristics for each walking cycle; W b For attention learning matrices with multiple walking cycles, W F With b F Transformation weights and biases for single-cycle fusion pressure mode characteristics, γ c The attention score for each walking cycle.

[0031] Furthermore, in step 5, the time series features are represented as follows:

[0032]

[0033] r c =σ(W rm m c-1 +W rF F c +b rr )

[0034] m c =tanh(W mF F c +W mm (r c ⊙m c-1 )+b mm )

[0035] Among them, z c The temporal characteristics of the walking period c; W zm To update the weights for the features, W zFUpdate the weights of features for single-cycle fusion stress pattern features. To update the feature weights of multi-period fusion stress mode features, b zz Update the bias for the features; W rm For the feature forgetting weight, W rF b represents the forgetting weights for single-cycle fusion pressure pattern features. rr For forgetting bias; W mF W is the memory update weight for single-cycle fusion pressure mode features. mm To update the weights of the previous memory layer, b mm This is for memory bias.

[0036] Furthermore, in step 6, the predicted probability of sarcopenia is expressed as:

[0037]

[0038] in, z is the predictive probability of sarcopenia. C The temporal characteristics of the last walking cycle. The numbers represent the multi-cycle fusion pressure pattern features corresponding to each walking cycle, and MLP represents the multilayer perceptron (MLP) layer.

[0039] Furthermore, in step 6, the multilayer perceptron (MLP) consists of a linear layer and an activation function.

[0040] Based on the above technical solution, the beneficial effects of the present invention are as follows:

[0041] This invention provides a preliminary sarcopenia screening method based on paired foot pressure data. It constructs a joint pressure matrix for the left and right feet, combining pressure features from both feet to obtain the correlation between an individual's pressure features and walking patterns related to the left and right feet. Simultaneously, it uses clustering to derive pressure patterns observed during periodic walking, revealing the phased characteristics of walking. Furthermore, it utilizes the dynamic changes in these pressure patterns to obtain periodic information about the patient's walking, and further leverages the characteristics of the walking cycle to acquire temporal features during the subject's walking process. Finally, it completes the preliminary sarcopenia screening based on pressure pattern and temporal features, effectively improving the accuracy of initial sarcopenia screening. More importantly, this invention proposes a novel GRU module that, during the acquisition of dynamic feature changes, can obtain both information about the changing pressure features themselves and the global information coexisting with the changing pressure features, thereby better learning dynamic foot pressure features for use in sarcopenia screening tasks. Attached Figure Description

[0042] Figure 1This is a flowchart illustrating the preliminary screening method for sarcopenia based on paired foot pressure data in this invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0044] This embodiment provides a preliminary screening method for sarcopenia based on paired foot pressure data, such as... Figure 1 As shown, the specific steps include:

[0045] Step 1: Collect foot pressure data and preprocess it;

[0046] Pressure signals (sig) on ​​the left foot of the subject were collected during walking using a specially designed insole equipped with a pressure sensor. l and right foot pressure signal sig r And perform time alignment;

[0047] Specifically, this invention requires that during data collection, the subject wears insoles equipped with N pressure sensors on each of their left and right feet and walks for a duration of at least T. While preserving as much pressure signal as possible from a single complete walking process, the shortest possible time is extracted. The extracted walking time is the time taken for all subjects to complete C walking cycles. Within one cycle, human walking can be divided into S stages, such as initial contact, load response, intermediate standing, final standing (standing stage), pre-swing, initial swing, intermediate swing, and final swing (swing stage). The duration of each walking stage s in each walking cycle c is T. c,s Total walking time The initial moment of signal interception is when the pressure signal intensity first reaches 10% of the maximum intensity of the process, so as to ensure that the signal time of the left and right feet is consistent during walking;

[0048] After acquiring the walking process of the left and right feet for a duration of T, the walking duration T of each phase s in all cycles is determined using pressure extreme value detection. c,s The signal was then segmented to obtain the left and right foot pressure signals for each walking phase s in a single walking cycle c. The pressure extreme value detection is as follows: Based on the walking performance, a walking cycle is estimated in advance. Then, the first extreme (large or small) value point with discrimination is selected in the pressure signal during the walking process. Subsequent extreme (large or small) value points with discrimination are selected from the pressure signal changes around the estimated cycle of the extreme value point.

[0049] Through the above process, the left and right foot pressure signals of each walking phase s in a single walking cycle c are obtained. Next, the pressure data needs to be converted into a unified pressure feature space for easier subsequent processing. Specifically, this invention uses a 1D CNN as the left and right foot pressure feature extractor, first inputting the single-walking stage pressure signals from all sensors. (T s The signal sequence length of the walking phase s is used to obtain information at different levels hidden in the pressure signal. Finally, the pressure features of the left and right feet are obtained.

[0050] That is, the corresponding left foot pressure features are obtained after processing through a 1D CNN network. and right foot pressure characteristics

[0051]

[0052] Step 2: Combine the pressure characteristics of the left and right feet;

[0053] After obtaining the pressure characteristics of the left and right feet in step 1, the data is fused using a multiplicative combination method, specifically as follows:

[0054]

[0055] The overall pressure matrix during the walking phase can be obtained using equation (2). Among them, P c,s W represents the overall pressure matrix for the walking phase s within a single walking cycle c. l W r b is the learning matrix corresponding to the left and right foot pressure features. l ,b r is the bias vector in the learning process; Equation (2) projects the pressure features of the left and right feet from their respective feature spaces to a shared feature space, and then uses a multiplicative combination to fuse the features of the two into a total pressure matrix;

[0056] Step 3: Create a stress pattern;

[0057] The overall pressure matrix P during the walking phase is obtained through step 2. c,s The data distribution characteristics were analyzed, and clustering methods were used to aggregate the features. This invention uses the GMM model for this purpose. Foot pressure pattern features were created for walking phase s within the walking cycle c. Represented as:

[0058]

[0059] Equation (3) is the feature aggregation result of the pressure matrix, which summarizes the pressure characteristics during the walking process and reflects the probability distribution of the pressure matrix in a walking stage;

[0060] Step 4: Integrate multi-period stress patterns;

[0061] Steps 1 to 3 can yield a pressure pattern during the walking phase. To acquire dynamic pressure features, it is necessary to integrate the temporal and feature space information of pressure patterns from multiple walking stages. First, pressure patterns of each walking stage within a cycle are fused. By fusing within a cycle, dynamic pressure features of a single cycle can be obtained. Then, pressure patterns from multiple cycles are further fused.

[0062] For the current walking cycle c, the plantar pressure pattern of the current walking stage s is the local feature at this time. The plantar pressure patterns of walking stage s and all previous walking stages are the global time-segment features, which reflect the commonality of pressure pattern distribution across all time periods. By extracting the common features of pressure patterns in each walking stage, the individual characteristics of patients can be better represented, thereby improving the accuracy of model judgment.

[0063] Specifically, an attention mechanism is used to fuse the plantar pressure pattern features of the first s walking phases within a walking cycle c. Obtain fusion pressure mode characteristics Represented as:

[0064]

[0065]

[0066] Among them, a s Attentional characteristics at each stage of the walking cycle; W a This is the attention learning matrix within the walking cycle, which is adjusted during training; W f With b f The transformation weights and biases of the stress pattern features within the walking cycle are used to project the stress pattern features of a single walking phase into the attention space; A s Att represents the attention score for each walking stage within the walking cycle. It is used to control the contribution of each walking stage feature to the walking cycle feature, so that the fusion of stress pattern features can obtain stress pattern information from each walking stage, thereby extracting the overall features of stress pattern in a single walking cycle more comprehensively; Att represents the attention mechanism.

[0067] For the current walking cycle c, the last walking stage S of the above process is used as the fused pressure mode feature for that cycle. After obtaining the fused pressure mode features of a single-walk cycle, multi-cycle pressure mode features are obtained. By employing the same attention mechanism to fuse the single-cycle fusion stress pattern features of the first c walking cycles, multi-cycle fusion stress pattern features are obtained. Represented as:

[0068]

[0069] Among them, g c Attentional characteristics for each walking cycle; W b For attention learning matrices with multiple walking cycles, W F With b F Transformation weights and biases for single-cycle fusion pressure mode characteristics, γ c Attention scores for each walking cycle;

[0070] Step 5: Obtain feature time series information;

[0071] After step 4, the fusion pressure mode features are acquired periodically; for the fusion pressure mode over all C walk cycles, the improved GRU is used to update the corresponding periodic temporal features z. c It includes the characteristics of this cycle and the changes of previously learned cycles over time.

[0072]

[0073] r c =σ(W rm m c-1 +W rF F c +b rr )#(11)

[0074] m c =tanh(W mF F c +W mm (r c ⊙m c-1 )+b mm )#(12)

[0075] All W in equations (10)(11)(12) * Both represent learnable weight matrices, b * Both represent bias vectors, and σ(·) represents the Sigmoid function;

[0076] Equation (10) is the feature update gate, which includes: feature update weight W zm Single-cycle fusion pressure mode feature update weight W zF Multi-cycle fusion pressure mode feature update weight WzF With feature update bias b zz The feature update gate combines single-period features, multi-period features, and the temporal features of the previous state obtained from GRU.

[0077] Equation (11) is the feature forgetting gate, which includes: feature forgetting weight W rm Single-cycle fusion pressure pattern feature forgetting weight W rF With forgetting bias b rr By using a feature forgetting gate, some irrelevant information is selected for forgetting, thus ensuring the temporal relevance of features during the GRU learning process.

[0078] Equation (12) is a memory gate, which includes: single-cycle fusion pressure pattern feature update weight W mF The weight W for updating the memory of the previous layer mm With memory bias b mm By using memory gates to save intermediate temporal feature states, the features of this iteration of learning are saved, as well as the state features of the previous iteration of learning.

[0079] Equations (10) to (12) can be used to obtain the evolution information of stress pattern features in time series and capture the long-range dependency information of its time series data. Compared with the traditional GRU unit, the stress pattern features of the walking cycle that need to be learned are directly added to the update gate, which can more clearly obtain the stress pattern distribution information at this time. At the same time, common comprehensive stress pattern features related to all stress patterns are added, and the individual walking pattern features of the patient are optimized and adjusted. In addition, single-cycle stress pattern information is added to the forget gate and the memory gate to ensure that the features learned by the entire GRU unit module are strongly correlated with the individual walking pattern features of the patient.

[0080] Step 6: Preliminary prediction of sarcopenia;

[0081] The GRU in step 5 learned the characteristics of each walking cycle. Temporal features z c ,Will and the temporal characteristics z of the last walking cycle C The data is fed into the MLP layer of a multilayer perceptron to obtain the probability of a final diagnosis of sarcopenia. Specifically, it is expressed as follows:

[0082]

[0083] The multilayer perceptron (MLP) layer is specifically composed of a linear layer and an activation function.

[0084] This invention uses a binary cross-entropy loss function for training and selects the Adam optimizer to update the model parameters until the trained model loss converges or reaches 1000 rounds, at which point training ends; the training uses data from a total of K patients, y i This is the true label value for whether patient i has sarcopenia. The probability of the model predicting sarcopenia is 1 if it is sarcopenia and 0 otherwise; the binary cross-entropy loss function is specifically expressed as:

[0085]

[0086] Based on this, the present invention realizes the conversion from pressure signal data to sarcopenia classification and completes the preliminary screening of sarcopenia based on paired foot pressure data.

[0087] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A preliminary screening method for sarcopenia based on paired foot pressure data, characterized in that, Includes the following steps: Step 1: Collect the pressure signals of the left foot and right foot during the subject's walking process, and extract features after time alignment to obtain the left foot pressure features and right foot pressure features of each walking stage in each walking cycle. Step 2: Use a multiplicative combination method to fuse the pressure characteristics of the left and right feet to obtain the overall pressure characteristics of each walking stage within each walking cycle; Step 3: Use the GMM model to create plantar pressure patterns for each walking stage within each walking cycle, and obtain the corresponding plantar pressure pattern features. Step 4: Use an attention mechanism to fuse the plantar pressure pattern features of each walking stage within a single walking cycle to obtain the single-cycle fused pressure pattern features of a single walking cycle. Then, an attention mechanism is used to fuse the single-cycle fusion stress pattern features of multiple walking cycles to obtain multi-cycle fusion stress pattern features. Step 5: Based on the single-cycle fusion pressure mode characteristics and the multi-cycle fusion pressure mode characteristics, the improved GRU module is used to learn in time sequence to obtain the time sequence characteristics of the corresponding walking cycle. Step 6: Feed the multi-cycle fusion stress pattern features corresponding to all walking cycles and the temporal features of the last walking cycle into a multilayer perceptron and output the predicted probability of sarcopenia.

2. The method for preliminary screening of sarcopenia based on paired foot pressure data according to claim 1, characterized in that, In step 1, the pressure signal sig of the left foot. l With right foot pressure signal sig r The data collection process is as follows: The pressure signals (sig) of the left foot generated by the subject during a complete walking process were collected using an insole equipped with N pressure sensors. l With right foot pressure signal sig r The system captured pressure signals from both feet for C walking cycles. Each walking cycle c was divided into S walking stages, and the pressure signal from the left foot for each walking stage s within walking cycle c was obtained. Pressure signal from the right foot 3. The preliminary screening method for sarcopenia based on paired foot pressure data according to claim 2, characterized in that, The walking cycle includes the following stages: initial contact, load response, intermediate standing, final standing, pre-swing, initial swing, intermediate swing and end swing. The walking stages are divided by pressure extreme value detection.

4. The method for preliminary screening of sarcopenia based on paired foot pressure data according to claim 2, characterized in that, Feature extraction employed a 1D CNN network as the foot pressure feature extractor. For the walking phase s within the walking cycle c, the left foot pressure features were extracted. Pressure characteristics of the right foot Represented as: CNN stands for 1D CNN network.

5. The method for preliminary screening of sarcopenia based on paired foot pressure data according to claim 1, characterized in that, In step 2, the overall pressure characteristics are expressed as follows: Among them, P c,s This represents the overall pressure characteristics of the walking phase s within the walking cycle c. and W represents the pressure characteristics of the left foot and the right foot, respectively. l W r b is the learning matrix corresponding to the pressure features of the left and right feet. l ,b r This is the bias vector during the learning process.

6. The method for preliminary screening of sarcopenia based on paired foot pressure data according to claim 1, characterized in that, In step 3, the plantar pressure pattern characteristics of walking phase s within walking cycle c. Represented as: Among them, P c,s This represents the overall pressure characteristics of the walking phase s within the walking cycle c.

7. The preliminary screening method for sarcopenia based on paired foot pressure data according to claim 1, characterized in that, In step 4, an attention mechanism is used to obtain fused pressure pattern features based on the plantar pressure pattern characteristics of the first s walking phases within the current walking cycle c. Represented as: in, The plantar pressure pattern characteristics of the walking phase s in the walking cycle c, a s W represents the attentional characteristics of each walking phase within the walking cycle. a W is the attention learning matrix during the walking cycle. f With b f For the transformation weights and biases of the pressure pattern characteristics within the walking cycle, A s Att represents the attention score for each walking phase within the walking cycle, and Att represents the attention mechanism. The plantar pressure pattern characteristics of the last walking phase S of the current walking cycle c are taken as the single-cycle fusion pressure pattern characteristics F of walking cycle c. c : For the single-cycle fusion stress pattern features of the first c walking cycles, an attention mechanism is used to obtain the multi-cycle fusion stress pattern features. Represented as: Among them, g c Attentional characteristics for each walking cycle; W b For attention learning matrices with multiple walking cycles, W F With b F Transformation weights and biases for single-cycle fusion pressure mode characteristics, γ c The attention score for each walking cycle.

8. The preliminary screening method for sarcopenia based on paired foot pressure data according to claim 1, characterized in that, In step 5, the time series features are represented as follows: r c =σ(W rm m c-1 +W rF F c +b rr ) m c =tanh(W mF F c +W mm (r c ⊙m c-1 )+b mm ) Among them, z c The temporal characteristics of the walking period c; W zm To update the weights for the features, W zF Update the weights of features for single-cycle fusion stress pattern features. To update the feature weights of multi-period fusion stress mode features, b zz Update the bias for the features; W rm For the feature forgetting weight, W rF b represents the forgetting weights for single-cycle fusion pressure pattern features. rr For forgetting bias; W mF W is the memory update weight for single-cycle fusion pressure mode features. mm To update the weights of the previous memory layer, b mm This is for memory bias.

9. The preliminary screening method for sarcopenia based on paired foot pressure data according to claim 1, characterized in that, In step 6, the predicted probability of sarcopenia is expressed as: in, z is the predictive probability of sarcopenia. C The temporal characteristics of the last walking cycle. The following are the multi-cycle fusion pressure pattern features corresponding to each walking cycle, and MLP stands for Multilayer Perceptron.

10. The preliminary screening method for sarcopenia based on paired foot pressure data according to claim 1, characterized in that, In step 6, the multilayer perceptron (MLP) layer consists of a linear layer and an activation function.

Images