AI-based peripheral pressure stimulation quantification assessment method and system

Abstract

The application provides an AI-based peripheral pressure stimulation quantification evaluation method and system. It relates to the fields of digital medicine and artificial intelligence technology. The method comprises the following steps: collecting multi-dimensional pressure stimulation data of peripheral pressure stimulation applied by different body parts or different points of the same part of the user in real time, including pressure intensity, stimulation mode, cycle and frequency parameters; obtaining multi-modal physiological feedback data of the skin and superficial tissues under pressure stimulation through a bionic skin sensing module, including stress distribution changes, local blood flow changes and electromyographic signals; performing AI-driven physiological effect prediction based on the multi-dimensional pressure stimulation data and the multi-modal physiological feedback data to obtain a prediction result; dynamically adjusting pressure stimulation parameters through an adaptive optimization module based on the prediction result and individual characteristics; and generating a personalized stimulation scheme, real-time monitoring data and recorded curative effect through an intelligent control interface. The application improves the precision and intelligent level of peripheral pressure stimulation intervention.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital medicine and artificial intelligence, and particularly relates to an AI-based peripheral pressure stimulation quantification evaluation method and system. BACKGROUND

[0002] Peripheral pressure stimulation is widely used in chronic pain management, neurological rehabilitation, sports injury repair, and biomechanical research. Traditional pressure stimulation evaluation methods mainly rely on manual experience or simple physical parameter settings, lacking precise response to individual physiological differences and real-time biological feedback. At the same time, existing physiological effect modeling mostly relies on static statistical methods, which are difficult to dynamically adapt to complex and variable body response processes, and have obvious shortcomings in data noise processing, multi-granularity information fusion, and modeling accuracy. In addition, the traditional Daubechies wavelet has significant loss of high-frequency details when denoising, which can cause the characteristics of electromyographic signals to be excessively smoothed.

[0003] With the development of artificial intelligence (AI) technology, flexible sensing technology, and intelligent control theory, there is an urgent need for an intelligent peripheral pressure stimulation quantification evaluation system that can integrate multi-modal physiological signals and has self-adaptive learning ability, so as to realize high-precision prediction and intervention optimization of the physiological state of the body and improve the personalization and scientificity of treatment. SUMMARY

[0004] The present application aims to solve the above problems and proposes an AI-based peripheral pressure stimulation quantification evaluation method and system, which systematically improves the precision and intelligence level of peripheral pressure stimulation intervention.

[0005] To achieve the above-mentioned purpose, the following technical solutions are implemented:

[0006] According to an aspect of the present application, an AI-based peripheral pressure stimulation quantification evaluation method is provided, comprising the following steps:

[0007] S1: Real-time acquisition of multi-dimensional pressure stimulation data of peripheral pressure stimulation applied to different body parts or different points of the same part of the user, including pressure intensity, stimulation mode, cycle, and frequency parameters;

[0008] S2: Obtain multi-modal physiological feedback data of skin and superficial tissues under pressure stimulation through a bionic skin sensing module, including stress distribution changes, local blood flow changes, and electromyographic signals;

[0009] S3: AI-driven physiological effect prediction based on the multi-dimensional stress stimulation data and the multi-modal physiological feedback data, obtaining a prediction result, comprising: data preprocessing on the multi-dimensional stress stimulation data and the multi-modal physiological feedback data; inputting the data after data preprocessing into a CMBlock_Mamba model constructed based on a Mamba mechanism for deep feature extraction and physiological effect prediction, and outputting prediction values of a plurality of physiological effect indexes, the plurality of physiological effect indexes including blood flow, muscle response and pain perception;

[0010] S4: dynamically adjusting stress stimulation parameters by an adaptive optimization module according to the prediction result and individual characteristics;

[0011] S5: generating a personalized stimulation scheme, real-time monitoring data and recording curative effect through an intelligent control interface.

[0012] According to another aspect of the present application, an AI-based peripheral stress stimulation quantitative evaluation method system is also provided, comprising:

[0013] A stimulation data acquisition module: real-time acquisition of multi-dimensional stress stimulation data of peripheral stress stimulation applied to different body parts or different points of the same part of a user, including stress intensity, stimulation mode, cycle and frequency parameters;

[0014] A bionic skin sensing module: used for acquiring multi-modal physiological feedback data of skin and shallow tissue under stress stimulation, including stress distribution change, local blood flow change and electromyographic signal;

[0015] A physiological effect prediction module: AI-driven physiological effect prediction based on the multi-dimensional stress stimulation data and the multi-modal physiological feedback data, obtaining a prediction result, comprising: data preprocessing on the multi-dimensional stress stimulation data and the multi-modal physiological feedback data; inputting the data after data preprocessing into a CMBlock_Mamba model constructed based on a Mamba mechanism for deep feature extraction and physiological effect prediction, and outputting prediction values of a plurality of physiological effect indexes, the plurality of physiological effect indexes including blood flow, muscle response and pain perception;

[0016] An adaptive optimization module: dynamically adjusting stress stimulation parameters according to the prediction result and individual characteristics;

[0017] An intelligent control interface: used for generating a personalized stimulation scheme, real-time monitoring data and recording curative effect.

[0018] Compared with the prior art, the present application has the following beneficial effects:

[0019] 1、The application adopts Coiflets wavelet decomposition combined with a minimax threshold adaptive algorithm, combined with a soft threshold rule, to complete multi-scale feature extraction and efficient denoising, ensure that the input signal retains physiological intrinsic characteristics and effectively suppresses noise influence, and provide high-fidelity input data for subsequent modeling.

[0020] 2、The application designs a residual de-redundancy module for the redundant information existing in physiological data of different granularities, systematically eliminates the redundancy between multi-scale features through three-dimensional tensorization, unified embedding space processing and residual connection, improves the independence and effectiveness of feature expression, and enhances the convergence and generalization ability of modeling.

[0021] 3、The application constructs a CMBlock_Mamba deep neural network architecture based on the Mamba state modeling mechanism. The model combines normalization processing, dynamic feature extraction (Mamba module), deep perception (multilayer perceptron module) and feature fusion mechanism, can efficiently capture the time sequence dynamic change and cross-scale correlation of physiological signals, realize accurate prediction of physiological effects such as local blood flow change, muscle response intensity and pain perception. The CMBlock architecture based on the Mamba state space model reduces the calculation complexity through lightweight convolution and state recursion, reduces the modeling error of long-term dependence of blood flow, electromyography and other signals, and improves the real-time inference speed.

[0022] 4、The application introduces a weighted loss function of Jensen-Shannon (JS) divergence and coefficient of determination (R ). The JS divergence measures the similarity between the output distribution and the real distribution, and the R index evaluates the numerical fitting quality, forming a comprehensive optimization strategy, effectively improving the learning and generalization ability of the model to complex physiological response patterns, and improving the compatibility of the prediction model to distribution consistency and numerical fitting accuracy. Avoid the shortcomings of single task regression model (such as SVR) or mean square error (MSE) loss function that is difficult to balance distribution consistency and numerical accuracy.

[0023] In summary, the application realizes the leap of peripheral pressure stimulation evaluation from "experience-driven" to "data-intelligent-driven", and is significantly superior to the prior art in terms of denoising accuracy, model efficiency, personalized adaptation and clinical safety, providing a reliable intelligent solution for chronic pain management, sports rehabilitation and other fields.

[0024] It is to be understood that the description in the summary is not intended to identify key or essential features of embodiments of the application, nor is it intended to limit the scope of the application. Other features, aspects, and advantages of the application will become apparent from the following description, which is given for the purpose of illustration and is not intended to limit the scope of what the inventors regard as their application. BRIEF DESCRIPTION OF DRAWINGS

[0025] The above and other features, aspects, and advantages of embodiments of the present application will become more apparent from the following description when taken in conjunction with the accompanying drawings. The drawings are intended to be not limiting and are provided for purposes of illustration only. In the drawings:

[0026] Figure 1 is a specific step schematic diagram of an AI-based peripheral pressure stimulation quantification evaluation method of an embodiment of the present application;

[0027] Figure 2 is a flow schematic diagram of an AI-based peripheral pressure stimulation quantification evaluation method and system of an embodiment of the present application;

[0028] Figure 3 is a data processing and prediction model construction process of the present application;

[0029] Figure 4 is a CMBlock stack structure schematic diagram of the present application;

[0030] Figure 5 is a module schematic diagram of an AI-based peripheral pressure stimulation quantification evaluation system of an embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to make the objects, technical solutions and advantages of embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some, but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0032] In addition, the term "and / or" in this paper is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents an "or" relationship between the front and rear associated objects.

[0033] In the present application, peripheral pressure stimulation refers to the controllable mechanical pressure applied to the peripheral parts of the human body (such as skin, muscle, joint, etc.) by external equipment, the purpose of which is to induce specific physiological responses through physical stimulation for treatment, rehabilitation or physiological state evaluation.

[0034] Peripheral refers to the parts of the human body other than the central nervous system, such as the skin, muscles, blood vessels, and peripheral nerves in areas such as the limbs, back, and soles. Unlike interventions targeting the central nervous system (such as the brain and spinal cord), peripheral pressure stimulation directly acts on the body surface or superficial tissues, affecting physiological functions through local mechanical effects.

[0035] In this invention, the specific forms of pressure stimulation include: pressure intensity: the size of the applied force (in units of kPa or mmHg), such as light touch (low intensity) or deep pressing (high intensity). Stimulation mode: continuous mode (continuous pressure application), intermittent mode (periodic pressure application and release), pulse mode (high-frequency short-time impact). Frequency and period: the rhythm of pressure change (such as 1 or 5 times per second), affecting tissue adaptability and physiological response speed. Spatial distribution: through multi-channel sensor synchronous monitoring of force in different parts, optimizing pressure distribution (such as avoiding local high pressure leading to injury).

[0036] The present invention aims at the problem of serious noise interference in the collected physiological time series signals. It first introduces Coiflets wavelet decomposition and minimax threshold estimation method, combined with soft threshold rule, to complete multi-scale feature extraction and efficient denoising, ensuring that the input signal not only retains the physiological intrinsic characteristics, but also effectively suppresses the noise influence, providing pure and stable data basis for subsequent modeling. In view of the redundant information existing in physiological data of different granularity, the present invention designs a residual de-redundancy module, which systematically eliminates the redundancy between multi-scale features through three-dimensional tensorization, unified embedding space processing and residual connection, improves the independence and effectiveness of feature expression, and enhances the convergence and generalization ability of modeling. The present invention innovatively constructs CMBlock_Mamba deep neural network architecture based on Mamba state modeling mechanism. This model integrates normalization processing, dynamic feature extraction (Mamba module), deep perception (multi-layer perceptron module) and feature fusion mechanism, which can efficiently capture the time series dynamic changes and cross-scale correlation of physiological signals, and realize accurate prediction of physiological effects such as local blood flow changes, muscle response intensity, pain perception, etc. In order to improve the compatibility of the prediction model for distribution consistency and numerical fitting precision, the present invention introduces the weighted loss function of Jensen-Shannon (JS) divergence and coefficient of determination (R2). The JS divergence measures the similarity between the output distribution and the true distribution, and the R2 index evaluates the numerical fitting quality, forming a comprehensive optimization strategy, which effectively improves the learning and generalization ability of the model for complex physiological response patterns.

[0037] Figure 1 A specific step diagram of an AI-based peripheral pressure stimulation quantification evaluation method is shown, Figure 2 ​​A flowchart of an AI-based peripheral pressure stimulation quantification evaluation method and system is shown. As shown in Figure 1 and Figure 2 An AI-based peripheral pressure stimulation quantification evaluation method includes the following steps:

[0038] S1: Real-time acquisition of multi-dimensional pressure stimulation data of peripheral pressure stimulation applied by different body parts or different points of the same part of the user, including pressure intensity, stimulation mode, cycle and frequency parameters;

[0039] This step S1 is mainly responsible for high-precision and all-around real-time acquisition of various parameters during the application of peripheral pressure stimulation, ensuring the comprehensiveness and accuracy of the input data of the subsequent analysis model. Multiple points are allowed to be measured synchronously, that is, data is collected at different body parts or different points of the same part at the same time, and a complete pressure stimulation spatial distribution map is constructed, which is convenient for subsequent spatial feature analysis. Multi-dimensional pressure stimulation parameters are collected through a stimulation data collection module, which specifically includes the following units:

[0040] (1) Pressure intensity acquisition unit: the pressure acting on the skin surface is measured in real time by a high-sensitivity pressure sensor (such as a capacitive or piezoelectric sensor), usually in units of kPa or mmHg. The collection range is wide, which can cover different treatment needs from gentle touch to deep compression, and the accuracy range is ±0.05 kPa, the range is 0-200 kPa, and when collecting, the sensor is evenly distributed on the body part to be measured to ensure accurate measurement of the pressure intensity at different points. The system can alarm according to the set threshold value to avoid excessive stimulation causing discomfort or tissue damage.

[0041] (2) Stimulation mode recognition unit: by analyzing the time sequence characteristics of pressure stimulation, the applied stimulation mode is automatically recognized and recorded, including continuous mode, intermittent mode, pulse mode, while recording the stimulation rhythm (such as frequency, duty cycle) and stimulation duration, to ensure that the subsequent model can fully understand the influence of different modes on physiological response.

[0042] (3) Stimulation cycle and frequency acquisition unit: records the time points of stimulation application and removal, calculates the stimulation cycle (number of stimulations in a unit of time) and frequency (Hz), which is used to evaluate the response characteristics of the body under different rhythms. The system supports millisecond-level time resolution to ensure that dynamic stimulation modes (such as rapid frequency changes) can also be accurately captured.

[0043] S2: Obtain multi-modal physiological feedback data of skin and superficial tissue under pressure stimulation through a bionic skin sensing module, including stress distribution changes, local blood flow changes, and electromyographic signals;

[0044] The step S2 simulates the multi-layer structure and sensing characteristics of human skin through the bionic skin sensing module, combines various advanced sensing technologies, and collects biophysical and biochemical feedback signals in real time under the action of peripheral pressure stimulation. The entire bionic skin sensing module uses high flexibility and low modulus materials, and the thickness can be controlled to be in the order of millimeters or even sub-millimeters, ensuring adhesion and wearing comfort; adapting to the pressure sensing needs of complex surfaces of different parts (such as palms, soles, and backs), avoiding the influence of the sensor itself on the physiological state. The bionic skin sensing module includes the following units:

[0045] (1) Tissue stress distribution sensing unit: A large-area stress sensing array is made of flexible sensing materials (such as graphene, nanosilver wire, and flexible capacitive grid) to capture the stress distribution changes of the skin and shallow tissues under pressure stimulation in real time. Through two-dimensional or three-dimensional pressure imaging technology, the system can accurately identify local high-pressure points and shear force concentration areas, providing a basis for avoiding pressure injury and optimizing stimulation location. Alternatively, a Myoton Pro muscle tester can be used to measure muscle mechanics indicators.

[0046] (2) Local blood flow change monitoring unit: Integrating optical sensors (such as near-infrared spectroscopy NIRS and laser speckle imaging LSI) or miniature ultrasonic sensors, it can monitor the changes in blood flow and blood flow velocity in the pressure area in real time. NIRS detects changes in blood oxygen saturation, reflecting tissue ischemia or hyperemia, and LSI can dynamically observe microcirculation blood flow changes and capture short-term blood flow recovery or blockage phenomena. For example, using an fNIRS near-infrared spectroscopy imaging system, the probe is fixed on the frontal lobe position, and the signal is collected in a quiet state. The collection time is 5 minutes, the sampling frequency is 10 Hz, and during the collection process, the subject is kept in a quiet state to avoid head movement interference on the data, and the subject's blood flow change data is recorded.

[0047] (3) Electromyographic signal acquisition unit: Through a surface electromyography (sEMG) electrode array, the electrical physiological response of the local muscle group during pressure stimulation is recorded synchronously. It monitors the muscle activation degree, relaxation degree, and reflexive contraction response; analyzes the influence of different stimulation modes (such as continuous pressure and intermittent pressure) on muscle nerve control; and provides important indicators for nerve rehabilitation and muscle function evaluation. For example, using the American DELSYS Trigno surface electromyography testing system, electrode patches are pasted on the right triceps brachii, long head of biceps brachii, and short head of biceps brachii of the subject.

[0048] S3: AI-driven physiological effect prediction based on the multi-dimensional stress stimulation data and the multi-modal physiological feedback data, to obtain a prediction result, including: data preprocessing on the multi-dimensional stress stimulation data and the multi-modal physiological feedback data; inputting the data after data preprocessing into a CMBlock_Mamba model based on a Mamba mechanism for deep feature extraction and physiological effect prediction, to output prediction values of a plurality of physiological effect indexes, including blood flow, muscle response and pain perception;

[0049] This step S3 is based on a deep learning algorithm, and establishes a mapping relationship between stress stimulation and physiological response, which can predict physiological indexes such as local blood flow change, muscle response and pain perception under different stress stimulation conditions.

[0050] The input data is obtained from two parts of data collected by step S1 and step S2: one is the collected multi-dimensional stress stimulation parameters, such as the size, frequency and action time of the applied force; the other is the multi-modal physiological feedback data collected by the bionic skin sensing module, such as stress change, blood flow, electromyography (EMG) and other time series data. First, the multi-dimensional stress stimulation data and the multi-modal physiological feedback data of the above process are preprocessed to ensure strict alignment of various signals on the time axis, facilitating subsequent feature extraction and model training. For example, Figure 3 is the data processing and prediction model construction process of the present application, and this step S3 specifically includes the following sub-steps:

[0051] S31: Data preprocessing on the multi-dimensional stress stimulation data and the multi-modal physiological feedback data, including noise reduction processing and residual redundancy reduction processing;

[0052] S311: Coiflets wavelet denoising processing on the collected time series data, and the processing flow includes Coiflets wavelet decomposition, minimum maximum threshold estimation, soft threshold rule denoising and Coiflets wavelet reconstruction, specifically including:

[0053] S3111: Multi-scale decomposition of the original signal by using the Coiflets wavelet basis function;

[0054] Coiflets wavelet has compact support, orthogonality and approximate symmetry, and is suitable for physiological signal processing. Among them, the Coiflets wavelet basis function is defined as:

[0055] ;

[0056] Among them, represents the wavelet basis function of the layer and the th translation position; is the number of wavelet decomposition layers; is the translation factor is the time variable;

[0057] S3112: The original signal is decomposed into multi-scale approximation coefficients and detail coefficients by determining the optimal decomposition layer number and wavelet order and detail coefficients ;

[0058] wherein the determination of the optimal decomposition layer number comprises: (1) calculating the maximum feasible layer number: according to the signal length n (i.e. the number of signal sample points), according to the formula:

[0059] -YL, wherein YL is a safety margin, and YL=3 is optional to avoid excessive decomposition leading to too short low-frequency approximation coefficients. (2) Energy proportion analysis: for each layer of decomposition , the approximation coefficient energy proportion is calculated layer by layer: , wherein are the approximation coefficients and detail coefficients of the s-th layer, respectively; the minimum s satisfying ≥0.95 is selected as the optimal decomposition layer number CS, ensuring that more than 95% of the signal energy is retained.

[0060] (3) Termination condition: if the variance of the detail coefficients changes by less than 10% between two consecutive layers, it is considered that the noise has been sufficiently separated, and the decomposition is terminated.

[0061] wherein the wavelet order is determined by selecting a wavelet order that balances the frequency domain resolution and time domain compactness. Generally, the feasible range of the Coiflets wavelet order JS can be empirically limited to 1≤JS≤5. High-order wavelets (such as Coif5) have better frequency domain resolution, but the calculation cost increases. Therefore, based on the frequency band matching method, the main frequency range of the physiological signal (such as the high-frequency component of the EMG signal, 20-500Hz) can be further analyzed; the wavelet passband with the highest overlap degree with the signal main frequency is selected as the order JS, specifically: Coif1: suitable for low-frequency signals (such as blood flow changes, 0.1-2Hz); Coif3: suitable for medium-frequency signals (such as electromyographic signals, 20-150Hz); Coif5: suitable for high-frequency noise suppression (such as electrode motion artifacts).

[0062] S3113: The optimal threshold value of each layer of detail coefficients is adaptively calculated based on the min-max criterion;

[0063] The threshold value of each layer of detail coefficients (high-frequency components) after decomposition is processed to remove noise interference. The optimal threshold value λ is adaptively determined using the min-max criterion, and the calculation of the threshold value is based on the noise statistical characteristics, and the optimal threshold value λ is:

[0064] ;

[0065] where n is the number of signal sample points; λ is the optimal threshold value; is the noise standard deviation estimate value, and the noise estimate is usually obtained by using the median absolute deviation method, by taking the first layer of detail coefficients The estimation is realized as follows:

[0066] ;

[0067] where is the median absolute deviation operation; is the first layer of detail coefficients.

[0068] The minimax criterion can realize the minimization of the signal error under the condition of unknown noise distribution, and ensure the robustness and universality of denoising.

[0069] S3114: According to the optimal threshold value of each layer of detail coefficients, soft threshold processing is performed on each layer of detail coefficients to obtain the denoised detail coefficients;

[0070] The threshold processing adopts a soft threshold strategy, and specifically, soft threshold processing is performed on each layer of detail coefficients , and the formula is as follows:

[0071] ;

[0072] where is the original detail coefficient; is the detail coefficient after threshold processing, i.e., the denoised detail coefficient; λ is the optimal threshold value; sign is the sign function; max is the maximum value operation; is the absolute value operation.

[0073] Compared with the hard threshold, the soft threshold method can effectively avoid the oscillation phenomenon in the denoising process, can smoothly suppress small amplitude noise, can retain important features, and can ensure the continuity and smoothness of the signal after denoising.

[0074] S3115: The denoised detail coefficients are combined with the approximation coefficients to reconstruct the denoised signal :

[0075] ;

[0076] where is the denoised signal; is the coefficient after threshold processing; denotes the wavelet basis function of the th layer and the th translation position; is the approximation coefficient of the lowest decomposition level, is the corresponding scale function.

[0077] After data denoising, further residual de-redundancy is used to eliminate redundant information between multi-granularity data to ensure the purity of data input.

[0078] S312: residual de-redundancy processing is performed on the denoised signal data, specifically by tensor conversion, uniform embedding space mapping and residual connection to eliminate multi-granularity redundant information, including:

[0079] S3121: tensor dimension conversion:

[0080] The multi-scale features extracted from the same physiological signal data from different time / frequency resolution levels are divided into coarse-grained data and fine-grained data. Among them, the coarse-grained data refers to the feature data extracted from the time series with a longer time window or lower resolution, which provides the global change characteristics of physiological response, and is suitable as the "main feature" or "reference signal" of the model. Fine-grained data refers to feature data extracted from the same physiological signal with a shorter time window or higher resolution. Such data has high dimension and rapid change, but is more susceptible to noise interference or redundant repetition, and its role is to supplement detailed information and reflect short-term fluctuations and local responses. Specifically:

[0081] The multi-scale features of the same physiological signal after denoising are divided into coarse-grained data and fine-grained data ;

[0082] The coarse-grained data and fine-grained data are respectively converted as follows:

[0083] The coarse-grained data is converted as follows:

[0084] ;

[0085] Num is the number of samples; is the feature dimension; is the number of channels (C) );

[0086] The fine-grained data is converted as follows:

[0087] ;

[0088] is the feature dimension of the fine-grained data; is the number of channels, which is consistent with the coarse-grained data.

[0089] Suppose the original fine-grained data (100 samples, 256 features), converted to a 3D tensor (channel number 64, feature length 4).

[0090] The coarsest-grained data and the fine-grained data are converted in dimension by the above process, from 2D data to 3D tensors, and the 3D tensor structure (number of samples x number of channels x feature length) is more suitable for the input requirements of convolutional networks (CNN) or state space models (such as Mamba), improving computational efficiency. Then each granular data segment of the fine-grained data after dimension conversion is sampled into a subsequence;

[0091] After conversion, each granular data segment is sampled into the first subsequence , the expression is:

[0092] ; is the length of a single subsequence.

[0093] Segmented sampling divides the fine-grained data into sub-sequences of length c , capturing short-term fluctuations or local spatial patterns. For example, in a time series, if c=10, each sub-sequence contains data from 10 time points, and the model can learn local temporal correlations.

[0094] S3122: Uniform embedding space mapping:

[0095] Use a multi-layer perceptron to adjust the size of each fine-grained tensor to a uniform embedding space, and obtain the mapped feature tensor , the mapping formula is:

[0096] ;

[0097] is each fine-grained tensor converted by step S3121; is the mapped feature tensor; represents a fully connected layer; is an activation function; represents a flattening operation, since the fully connected layer (FC) usually accepts a two-dimensional input, before inputting the fully connected layer, the three-dimensional tensor needs to be flattened into a two-dimensional matrix, for example: , .

[0098] Joint modeling of channel and feature dimensions through the fully connected layer can enhance cross-channel information interaction. For example, each fine-grained tensor The dimensions are 100×64×4 (Num=100, Channel=64, L=4); after flattening, it is 100×256, which is mapped to a unified embedding space (e.g., 256-dimensional) through a fully connected layer, and the output is... .

[0099] Through the above process, the model can effectively extract local features of fine-grained signals and fuse them with coarse-grained features, ultimately improving the accuracy of physiological effect prediction.

[0100] S3123: Residual connection eliminates redundancy:

[0101] Construct residual connections according to the following rules to eliminate multi-granularity redundant information:

[0102] ;

[0103] in, Indicates the first The feature tensor after layer processing is passed to the CMBlock_Mamba model in step S32 after information embedding and processing. This indicates a redundant feature used for the next residual connection. This indicates that the number of channels is adjusted using a 1x1 convolution. The mapped feature tensor After eliminating redundancy through residual connections, the data is finally input into the CMBlock_Mamba model.

[0104] By employing dimensionality transformation and piecewise sampling, high-dimensional, fine-grained data is regularized into a parseable tensor form, enhancing the model's ability to perceive local features. Consistent channel count design ensures seamless fusion of multi-granularity data, strengthening the model's ability to jointly model global and local physiological responses. The 3D tensor structure is compatible with mainstream deep learning frameworks (such as PyTorch and TensorFlow), reducing implementation complexity. Residual connections dynamically subtract redundant features between adjacent layers. This approach preserves the differences across data granularities, avoiding feature duplication or noise interference. By suppressing redundant feature propagation, it reduces the complexity of network parameter training, accelerates model convergence, and improves generalization ability, ultimately enhancing the accuracy and robustness of physiological effect prediction.

[0105] S32: Input the preprocessed data into the CMBlock_Mamba model built on the Mamba mechanism for deep feature extraction and physiological effect prediction, and output the predicted values ​​of multiple physiological effect indicators, including blood flow, muscle response and pain perception.

[0106] After the noise reduction and redundancy removal preprocessing in step S31, the data is sent to the CMBlock_Mamba model constructed based on the Mamba mechanism for deep feature extraction and physiological effect prediction. The CMBlock_Mamba model is composed of multiple CMBlocks.

[0107] S321: Constructing CMBlock

[0108] Each CMBlock contains a Mamba module, a residual connection module, and a two-layer multilayer perceptron (MLP) module. The Mamba module is responsible for modeling dynamic features over time. The residual connection module preserves the original input features to prevent gradient vanishing. The multilayer perceptron (MLP) module is a two-layer fully connected network that enhances the feature representation capability of the current CMBlock. Specifically:

[0109] (1) Dynamic feature modeling based on Mamba module

[0110] The Mamba module is a state space modeling structure that can effectively capture long-term dependencies and dynamic changes in time series. The state update formula (expressed in discrete form) is as follows:

[0111] ;

[0112] : hidden state vector at time ; x : input feature vector, initialized as the output of step S3123 ; : state transition matrix and observation matrix, both are learnable parameters. Initialization method of : A and B are initialized with normal distribution, and C and D are set as learnable diagonal matrices; training process: Adam optimizer is used with learning rate 1e-4 and batch size 32.

[0113] The Mamba module combines convolution and state recursion, which has the characteristics of light weight, good parallelism, and strong dynamic perception ability.

[0114] (2) Residual connection module

[0115] After the dynamic feature modeling based on the Mamba module, a residual connection is added to alleviate the gradient vanishing problem and maintain the original information flow:

[0116] ; represents the features extracted by the Mamba module; is the output feature after the residual connection module.

[0117] (3) Constructing a multi-layer perceptron module

[0118] After each CMBlock, a two-layer multi-layer perceptron (MLP) module is set to further improve the feature expression capability:

[0119] (3.1) The first layer of full connection (with activation function ReLU):

[0120] ;

[0121] is the output feature after the residual connection module; is the activation function; is the weight matrix; is the bias term; is the output after the first layer of full connection.

[0122] (3.2) The second layer of full connection (without activation function):

[0123] ;

[0124] wherein, is the weight matrix; is the bias term; and the final output is the output result of the current CMBlock.

[0125] S322: Hierarchical stacking and feature fusion

[0126] By stacking multiple CMBlocks, a deep-level feature extraction structure is formed, and a CMBlock stacking structure as shown in Figure 4 is obtained, and the number of CMBlocks is NUM. The implementation of hierarchical stacking and feature fusion of the CMBlock_Mamba model is as follows: multiple CMBlocks are stacked to generate multi-level features; the multi-level features are weighted and fused to obtain fused features , so that the features output by different levels are integrated through the feature fusion operation. The fusion formula is as follows:

[0127] ;

[0128] wherein, denotes the output of the th CMBlock; is the total number of CMBlocks; is the learnable fusion weight; and the final fused feature is used as the input of the physiological response prediction.

[0129] The historical dataset containing 500 patients' pressure stimulation parameters and corresponding physiological feedback data is used for model training. The training parameters of the model are as follows: the optimizer is Adam, the learning rate is 10 −4 , the batch size is 32, and the training round is 200.

[0130] S34: Output prediction

[0131] Suppose there are M sub-tasks in the final prediction target, corresponding to different physiological effect indicators (such as local blood flow changes, muscle response intensity, pain perception, etc.), then the output is:

[0132] ;

[0133] wherein, is the fusion feature obtained through step S33; is the output layer weight matrix; is the output layer bias; is the predicted value of the physiological effect indicator output by the model; is the Tanh activation function.

[0134] This step builds a regressor for each physiological indicator separately, sharing a set of forward networks (sharing the bottom layer representation), supporting the model to output multiple continuous physiological responses simultaneously, which is a typical multi-task regression architecture with shared features and independent output, which is beneficial to improve modeling efficiency and generalization ability.

[0135] S35: Loss function optimization

[0136] In order to measure the deviation of the probability distribution output by the model and the numerical fitting effect at the same time, the CMBlock_Mamba model training adopts a weighted loss function of Jensen-Shannon divergence and coefficient of determination, and the total loss is expressed as:

[0137] ;

[0138] wherein, , are the weight coefficients of the divergence and loss terms, respectively; the experience value range is: α∈[0.3,0.7], β=1-α, used to balance the distribution and numerical loss, preferably, α=0.5, β=0.5.

[0139] represents the predicted distribution Jensen-Shannon divergence between the real distribution Y, Y is the real value of the physiological effect index, such as blood flow change, muscle response intensity, pain perception, etc. Based on the definition of Kullback-Leibler divergence, it is used to measure the similarity between two probability distributions, and the calculation formula is as follows:

[0140] ;

[0141] Wherein, is the real distribution is the normalized vector (such as the normalized vector of blood flow change, muscle response intensity, etc.) is the predicted distribution is the normalized vector is the intermediate distribution, which is the average of P and Q, used to balance the difference between and ; represents the Kullback-Leibler divergence, reflecting the deviation of the real distribution P from the intermediate distribution M, the smaller the value, the closer P and M, that is, the closer the model prediction distribution to the real distribution, defined as:

[0142] ;

[0143] Wherein, represents the probability value of the i-th element (i.e. sample) in the vector P, for example, the predicted blood flow change, then represents the probability of the real blood flow change value of the i-th element; represents the probability value of the i-th element (i.e. sample) in the vector Q, for example, the probability of the blood flow change value of the i-th element predicted by the model; , represents the i-th element of the intermediate distribution M; the multi-task prediction of the present application simultaneously predicts blood flow, muscle response and pain perception, etc., i corresponds to the joint distribution dimension of different physiological indicators.

[0144] JS divergence is symmetric (KL divergence is not symmetric), and the value range is [0, 1], which is more suitable for measuring the similarity of two distributions. When predicting multiple physiological indicators (such as blood flow, muscle response, pain perception), JS divergence can uniformly measure the global difference of multi-task distribution, and provide an intermediate reference benchmark for each task. By balancing the difference between P and Q, the modeling accuracy and generalization ability of the model for complex physiological response patterns are improved.

[0145] represents the coefficient of determination, which is used to measure the consistency of the model prediction value and the real value, and is defined as:

[0146] ;

[0147] in, It is the sample size; For true distribution The mean, i.e. .

[0148] This step S35 minimizes the total loss. The model simultaneously optimizes: distribution consistency (JS divergence): ensuring that the probability distribution of the predicted results matches the actual physiological response pattern; numerical fitting accuracy (R²). 2 Coefficient of determination: ensures that the absolute error of the predicted value is minimized.

[0149] S4: Based on the prediction results and individual characteristics, the stress stimulation parameters are dynamically adjusted through the adaptive optimization module;

[0150] In step S4, the adaptive optimization module is the core of the system's dynamic adjustment and personalization. It aims to continuously optimize peripheral pressure stimulation parameters based on individual subject characteristics and real-time physiological feedback, thereby improving intervention effectiveness. Its main functions and design include: dynamically adjusting pressure stimulation parameters through the adaptive optimization module based on prediction results and individual characteristics, including:

[0151] (1) Individual feature modeling:

[0152] Before use, the system inputs or measures the subject's basic physiological characteristics, including but not limited to:

[0153] Age, gender, weight, height;

[0154] Skin thickness, muscle condition (muscle mass, tension);

[0155] Indicators of nerve reflex sensitivity (such as nerve conduction velocity and reflex threshold);

[0156] Past medical history (such as diabetic neuropathy, history of chronic pain, records of previous treatment responses, etc.).

[0157] These characteristics are used to build a personalized basic model, providing a preliminary basis for adjusting stress stimulus parameters. Specifically:

[0158] The above features are encoded into feature vectors. This indicates that the feature vector, which encodes the basic physiological characteristics of the subject, is a d-dimensional real space. This vector is input into a pre-defined rule engine or a lightweight neural network to generate an initial set of stimulus parameters. ,in, including initial pressure intensity, stimulation mode (continuous / intermittent / pulsed), frequency range, action area, etc. Optionally, the lightweight neural network used is a three-layer fully connected neural network, the number of input layer neurons is the individual feature dimension d = 10, = [age, weight, height, skin thickness, muscle mass, nerve conduction velocity, blood pressure, blood glucose level, pain threshold, motor ability score], the number of hidden layer neurons is 32, and the number of output layer neurons is the dimension of the initial stimulation parameter. ReLU activation function and Adam optimizer are used for training, the learning rate is 0.001, and the batch size is 16.

[0159] Safety constraints: set the upper limit of pressure intensity according to skin thickness and medical history to avoid tissue damage; set the minimum stimulation interval according to the sensitivity of nerve reflex to prevent overstimulation.

[0160] (2) Real-time feedback analysis:

[0161] Compare the physiological feedback collected by the bionic skin sensing module, such as tissue stress distribution, local blood flow, and electromyographic signals, with the prediction results of the physiological effects driven by AI in step S4; use dynamic difference analysis algorithm to evaluate the stimulation effect; monitor short-term adaptability (such as immediate blood flow improvement) and long-term trend change (such as muscle tension relief).

[0162] Use dynamic difference analysis algorithm to evaluate the stimulation effect and real-time feedback analysis results, including: calculate the instantaneous deviation mean square sum of real physiological data and prediction results to get difference measure index D:

[0163] ;

[0164] Where T is the total number of time steps; is the real data collected at each time step t, that is, the tissue stress, blood flow, electromyographic signal (sEMG) and the like collected by the bionic skin sensor; is the prediction value of each time step t.

[0165] Difference type judgment: short-term adaptability (T is small): evaluate the immediate intervention effect (such as blood flow improvement speed); long-term trend analysis (T is large): monitor muscle tension relief, pain perception change and other chronic responses.

[0166] The smaller the difference D is, the more accurate the model prediction is, otherwise the stimulation parameters or the model need to be adjusted.

[0167] Multi-task prediction output: the model of the present application simultaneously predicts multiple physiological indicators (such as blood flow, electromyography, pain, etc.), is a vector, for example: Each subtask calculates the loss independently, and finally optimizes the weighted loss function.

[0168] (3) Dynamic adjustment of stimulation parameters:

[0169] According to the difference metric D and individual characteristics, the adaptive optimization module adjusts the pressure stimulation parameters through a policy gradient algorithm, including at least one of pressure intensity fine-tuning, stimulation mode switching, stimulation frequency / period optimization, and local offset of stimulation area. Specifically:

[0170] Pressure intensity fine-tuning (80-110 mmHg);

[0171] Stimulation mode switching (such as switching from continuous pressure to pulse pressure);

[0172] Stimulation frequency, period optimization (lengthen or shorten the stimulation interval, 20-70s);

[0173] Local offset of stimulation area (avoid long-term pressure on local tissues).

[0174] If the difference metric D exceeds the safety threshold or abnormal signals (such as local tissue ischemia) are detected, an emergency stop or parameter reset is triggered.

[0175] According to and deviation, the adaptive optimization module dynamically adjusts the pressure stimulation parameters through a policy gradient algorithm, aiming to minimize the difference D and improve the intervention effect.

[0176] A policy gradient-based method is adopted to iteratively optimize the stimulation scheme during continuous stimulation, ensuring that each application tends to be optimal. The goal is to maximize the cumulative reward or maximize the expected return (treatment effect score):

[0177] ;

[0178] where R(s,a) is the immediate reward (such as the amount of physiological state improvement) obtained by taking action ; is the action (such as increasing pressure , switching modes); is the current state (such as the degree of blood flow improvement, muscle electrical recovery level); is the policy network parameter.

[0179] The reward function R(s,a) is defined as follows: ;

[0180] where is the weight, For the index improvement amount, D is the instantaneous deviation.

[0181] The policy gradient update formula is used:

[0182] ;

[0183] Indicates the derivative of the parameter with respect to ; is the stimulation parameter adjustment strategy.

[0184] The formula indicates that the optimization strategy of adjusting the stimulation parameter is adjusted according to the actual reward feedback at each step.

[0185] Through the comparison and optimization of the predicted and measured dynamic feedback, the precision and individualization control of peripheral pressure stimulation is realized.

[0186] (4) Adaptive threshold update:

[0187] During the long-term intervention process, the system automatically adjusts the response threshold according to the physiological changes of the subjects (such as the improvement of tolerance and the decrease of sensitivity), avoids the invalidation of early setting, and improves the individualization adaptability of long-term intervention. The initial threshold is set as , and the tolerance score is defined as : based on historical data, the pressure intensity is calculated in combination with subjective feedback (such as pain score), etc. ;

[0188] , the current maximum tolerance pressure: by gradually increasing the pressure intensity, until the physiological abnormalities (such as blood flow obstruction, abnormal electromyographic signal) or the pressure value (unit: kPa) when the patient actively feedbacks discomfort is detected. Initial tolerance pressure: the maximum safe pressure value that the patient can withstand in the first treatment cycle (initially set by the doctor or the system). Pain score: visual analog scale (VAS, 0-10 points) is used, 0 is painless, and 10 is severe pain. Decrease rate: the decrease percentage of the current cycle pain score relative to the initial score, that is: . Divide by 10 to standardize the treatment, scale the decrease rate to 0-10 points, and make the pressure intensity dimension consistent.

[0189] According to the change of patient tolerance (tolerance score ), the threshold is dynamically updated:

[0190] ;

[0191] , where is the stimulation threshold (such as the maximum allowable pressure) of the current cycle ; is the learning rate for adjusting the amplitude; is the tolerance score of the current cycle , is the set target tolerance score (set according to individual rehabilitation goals), which is a pre-set ideal tolerance score (e.g. =1.2), indicating the expected tolerance improvement and pain relief effect.

[0192] If , it indicates that the tolerance is enhanced, and the threshold p is increased to enhance the stimulation intensity;

[0193] If , it indicates that the tolerance is decreased, and the threshold p is decreased to ensure the safety of personalized stimulation.

[0194] The updated threshold is written into the individual model as a constraint condition for the next cycle parameter adjustment.

[0195] S5: Generate personalized stimulation programs, real-time monitoring data, and record treatment effects through the intelligent control interface.

[0196] The intelligent control interface is an important window for the system to interact with users (operators / doctors / patients), and the design focuses on intuitiveness, personalization, and feedback to ensure user experience and data visualization. Its main functions include:

[0197] (1) Personalized stimulation program generation interface:

[0198] Automatically generate personalized pressure stimulation programs based on individual information of subjects and system recommendations; provide three options: default program, recommended program, and manual editing mode, support doctors or professionals to adjust according to specific needs; display expected physiological effects (such as blood flow improvement percentage, muscle tension reduction expected value).

[0199] (2) Real-time monitoring and feedback display interface:

[0200] Visualize pressure distribution map, local blood flow dynamic map, and muscle electrical signal change curve; provide real-time stimulation state monitoring (such as current stimulation intensity, mode, remaining time); set up abnormal alarm (such as detecting abnormal reflex, high pressure risk) to remind users in time.

[0201] (3) Interactive parameter adjustment function:

[0202] Users can adjust key parameters such as pressure intensity, stimulation rhythm, and stimulation mode through the interface in real time; support multiple interaction methods such as touch and voice commands to improve convenience, especially suitable for scenarios where patients with limited mobility use.

[0203] (4) Efficacy record and tracking:

[0204] Automatic record each treatment process of stimulation parameters, subject feedback (such as pain score), physiological effect change data; Provide time axis type efficacy tracking chart, facilitate doctor or user to evaluate long-term intervention effect, adjust treatment strategy; Support one key export treatment record, facilitate medical record management or scientific research data accumulation.

[0205] (5) Intelligent recommendation and self-learning function:

[0206] Based on the treatment data and efficacy feedback accumulated by the system, the recommendation algorithm is continuously optimized using federated learning or local training mechanism; For patients with similar characteristics, automatically recommend better stimulation programs to improve the overall intelligence level of the system and the success rate of individual intervention.

[0207] Suppose an athlete has a muscle strain and receives rehabilitation treatment, the application can monitor muscle tension and blood flow speed in real time by attaching a flexible sensor to the affected area; After the AI model analyzes the data, it automatically controls the pressure device to apply intermittent pulse pressure to stimulate local microcirculation; If insufficient blood flow is detected, the system will increase the pulse frequency while avoiding excessive pressure that may cause secondary damage. Traditional methods rely solely on therapists manually pressing the affected area and adjusting the force based on experience, which cannot quantify the effect and may result in slow recovery due to uneven force. Therefore, compared to traditional methods, the application has obvious technical advantages. The application converts traditional empirical operations into precise "data-driven" treatment through AI technology, which can improve the safety, individualization and scientificity of treatment effect.

[0208] Figure 5 is a module schematic diagram of an AI-based peripheral pressure stimulation quantification evaluation method system according to an embodiment of the application. As shown in Figure 5 An AI-based peripheral pressure stimulation quantification evaluation method system 200 includes:

[0209] Stimulation data acquisition module 210: Real-time acquisition of multi-dimensional pressure stimulation data of peripheral pressure stimulation applied to different body parts or different points of the same part of the user, including pressure intensity, stimulation mode, cycle and frequency parameters;

[0210] Bionic skin sensing module 220: used to obtain multi-modal physiological feedback data of skin and superficial tissue under pressure stimulation, including stress distribution change, local blood flow change and electromyographic signal;

[0211] The physiological effect prediction module 230: according to the multidimensional stress stimulation data and the multi-modal physiological feedback data, AI-driven physiological effect prediction is performed to obtain a prediction result, including: performing data preprocessing on the multidimensional stress stimulation data and the multi-modal physiological feedback data; inputting the data after data preprocessing into the CMBlock_Mamba model constructed based on the Mamba mechanism for deep feature extraction and physiological effect prediction, and outputting prediction values of a plurality of physiological effect indexes, the plurality of physiological effect indexes including blood flow, muscle response and pain perception;

[0212] The adaptive optimization module 240: dynamically adjusting the stress stimulation parameters according to the prediction result and the individual characteristics;

[0213] The intelligent control interface 250: generating a personalized stimulation scheme, monitoring data in real time and recording the therapeutic effect.

[0214] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described modules can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0215] It should be noted that the embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0216] It should also be noted that in the embodiments of the present application, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or equipment including the element.

[0217] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An AI-based peripheral pressure stimulus quantification evaluation method, characterized by, The method comprises the following steps: S1: collecting multi-dimensional pressure stimulation data of peripheral pressure stimulation applied by a user on different body parts or different points of the same part in real time, including pressure intensity, stimulation mode, cycle and frequency parameters; S2: obtaining multi-modal physiological feedback data of the skin and superficial tissue under pressure stimulation through a bionic skin sensing module, including stress distribution changes, local blood flow changes and muscle electrical signals; S3: performing AI-driven physiological effect prediction based on the multi-dimensional pressure stimulation data and the multi-modal physiological feedback data to obtain a prediction result, including: data preprocessing of the multi-dimensional pressure stimulation data and the multi-modal physiological feedback data; inputting the data after data preprocessing into a CMBlock_Mamba model constructed based on a Mamba mechanism for deep feature extraction and physiological effect prediction, and outputting prediction values of a plurality of physiological effect indicators, including blood flow, muscle response and pain perception; Wherein, the data preprocessing includes noise reduction processing and residual redundancy processing; wherein, the noise reduction processing: adopting Coiflets wavelet decomposition, minimum maximum threshold estimation and soft threshold rule for noise reduction and Coiflets wavelet reconstruction, specifically including: adopting Coiflets wavelet basis function for multi-scale decomposition of the original signal; determining the optimal decomposition layer number and wavelet order, and decomposing the original signal into multi-scale approximation coefficients and detail coefficients; based on the minimum maximum criterion, the optimal threshold of each layer of detail coefficients is adaptively calculated; according to the optimal threshold of each layer of detail coefficients, the soft threshold value of each layer of detail coefficients is processed to obtain the denoised detail coefficients; the denoised detail coefficients and the approximation coefficients are combined to reconstruct the denoised signal; Wherein, the CMBlock_Mamba model is composed of a plurality of CMBlock, each CMBlock internally includes a Mamba module, a residual connection module and a two-layer multilayer perceptron module; the CMBlock_Mamba model adopts a weighted loss function of Jensen-Shannon divergence and determination coefficient for training; S4: dynamically adjusting the pressure stimulation parameters through the self-adaptive optimization module according to the prediction result and individual characteristics; S5: generating a personalized stimulation scheme, real-time monitoring data and recording efficacy through an intelligent control interface. 2.The AI-based peripheral pressure stimulus quantification evaluation method of claim 1, wherein, Wherein, The residual redundancy processing: eliminating multi-granularity redundant information through tensor conversion, uniform embedding space mapping and residual connection for the denoised signal data, including: Divide the multi-scale features of the denoised same physiological signal into coarse-grained data and fine-grained data; Respectively convert the dimensions of the coarse-grained data and the fine-grained data; Segmentally sample each granular data in the dimensionally converted fine-grained data into a subsequence; Use a multilayer perceptron to adjust the size of each fine-grained tensor to a uniform embedding space to obtain a mapped feature tensor; Construct a residual connection to eliminate multi-granularity redundant information according to the following rules: ； wherein, represents the layer processed feature tensor; is the mapped feature tensor; represents redundant features for use in the next residual connection.

3. The AI-based peripheral pressure stimulation quantification evaluation method according to claim 2, characterized in that, The multi-layer perceptron module comprises: two full connection layers are arranged after each CMBlock, and each of the two full connection layers outputs respectively: ； wherein, , is a weight matrix, , is a bias term, is an activation function; is the output feature after the residual connection. 4.The AI-based peripheral pressure stimulus quantification evaluation method of claim 3, wherein, Wherein, The CMBlock_Mamba model further comprises: forming a deep feature extraction structure by stacking the plurality of CMBlocks; The implementation manner of the hierarchical stacking and feature fusion of the CMBlock_Mamba model is: A plurality of CMBlocks are stacked to generate multi-level features; weighting and fusing the multi-level features to obtain a fused feature . 5.The AI-based peripheral pressure stimulus quantification evaluation method of claim 4, wherein, Wherein, The data after data preprocessing is input into the CMBlock_Mamba model constructed based on the Mamba mechanism for deep feature extraction and physiological effect prediction, and the prediction values of a plurality of physiological effect indexes are output, the plurality of physiological effect indexes include blood flow, muscle response and pain perception, comprising: On fusion features Perform multi-task regression, output layer formula is: ； wherein, is the fused feature; is the output layer weight matrix; is the output layer bias; is the predicted value of the physiological effect indicator of the model output; is the Tanh activation function; Each physiological effect index corresponds to an independent regressor, shares the bottom network features, and the model simultaneously outputs a plurality of continuous physiological effect indexes. 6.The AI-based peripheral pressure stimulus quantification evaluation method of claim 5, wherein, Wherein, Weighted loss function is represented as: ； wherein, , are respectively divergence and weighting coefficients of loss terms; denotes the predictive distribution Jensen-Shannon divergence between the true distribution and the predictive distribution is calculated as follows: ； wherein, is the true distribution is the normalized vector, is the predicted distribution is the normalized vector; is the intermediate distribution, which is the mean of P and Q; denotes the Kullback-Leibler divergence, defined as: ； wherein, denotes the probability value of the i-th element in the vector P; denotes the probability value of the i-th element in the vector Q; denotes the i-th element of the intermediate distribution M; R2represents the coefficient of determination, which measures how well the model predicts the true values, defined as: ； wherein, is the number of samples; is the mean of the true distribution i.e. . 7.The AI-based peripheral pressure stimulus quantification evaluation method of claim 6, wherein, The S4: according to the prediction result and the individual characteristics, the adaptive optimization module dynamically adjusts the stress stimulation parameters, Including: Individual characteristic modeling: by inputting or measuring the age, gender, weight, height, skin thickness, muscle state, nerve reflex sensitivity index and past medical history of the subject, a personalized basic model is established; Adopting dynamic difference analysis algorithm to evaluate the stimulation effect and real-time feedback analysis result, including: calculating the instantaneous deviation mean square sum of the real physiological data and the prediction result to obtain the difference measure index D: ； wherein T is the total number of time steps; is the real data collected at time step t; is the predicted value of the physiological effect indicator output by the model at time step t; According to the difference measure index D and the individual characteristics, the adaptive optimization module adjusts the stress stimulation parameters by strategy gradient algorithm, including: at least one of stress intensity fine tuning, stimulation mode switching, stimulation frequency / period optimization, and local offset of stimulation area; The objective function is to maximize the cumulative reward expectation, and the stimulation threshold is dynamically adjusted according to the tolerance score; If the difference measure index D exceeds the safety threshold, emergency stop or parameter reset is triggered. 8.The AI-based peripheral pressure stimulus quantification evaluation method of claim 7, wherein, Wherein, The stimulation threshold is dynamically adjusted according to the tolerance score, including: ； wherein, is the stimulation threshold of the current cycle ; is the learning rate for adjusting the amplitude; is the tolerance score of the current cycle , is the set target tolerance score; If , it indicates that the tolerance is enhanced, and the threshold value p is increased to enhance the stimulation intensity; If , it indicates a decrease in tolerance, the threshold value p is reduced to ensure the safety of the individualized stimulation.

9. An AI-based peripheral pressure stimulus quantification evaluation method system, characterized by, Including: Stimulation data acquisition module: real-time acquisition of multi-dimensional pressure stimulation data of peripheral pressure stimulation applied to different body parts or different points of the same part of the user, including pressure intensity, stimulation mode, cycle and frequency parameters; Bionic skin sensing module: used for acquiring multi-modal physiological feedback data of skin and shallow tissue under pressure stimulation, including stress distribution change, local blood flow change and electromyographic signal; Physiological effect prediction module: AI-driven physiological effect prediction is performed according to the multi-dimensional pressure stimulation data and the multi-modal physiological feedback data to obtain a prediction result, including: data preprocessing is performed on the multi-dimensional pressure stimulation data and the multi-modal physiological feedback data; the data after data preprocessing is input into the CMBlock_Mamba model constructed based on the Mamba mechanism for deep feature extraction and physiological effect prediction, and the prediction values of a plurality of physiological effect indexes are output, the plurality of physiological effect indexes include blood flow, muscle response and pain perception; The data preprocessing comprises noise reduction processing and residual de-redundancy processing; the noise reduction processing comprises Coiflets wavelet decomposition, minimum maximum threshold estimation and soft threshold rule, and Coiflets wavelet reconstruction, and specifically comprises: performing multi-scale decomposition on an original signal by using a Coiflets wavelet basis function; determining an optimal decomposition layer number and a wavelet order, and decomposing the original signal into multi-scale approximation coefficients and detail coefficients; adaptively calculating optimal thresholds of the detail coefficients of each layer based on a minimum maximum rule; performing soft threshold processing on each layer of the detail coefficients according to the optimal thresholds of the detail coefficients, to obtain denoised detail coefficients; and combining the denoised detail coefficients with the approximation coefficients to reconstruct a denoised signal; The CMBlock_Mamba model is composed of multiple CMBlocks, each of which comprises a Mamba module, a residual connection module and a two-layer multi-layer perceptron module; and the CMBlock_Mamba model is trained by using a weighted loss function of Jensen-Shannon divergence and a determination coefficient; An adaptive optimization module is configured to dynamically adjust the stress stimulation parameters according to the prediction results and individual characteristics; An intelligent control interface is configured to generate a personalized stimulation scheme, monitor data in real time and record therapeutic effects.

Images