Intelligent pulse diagnosis method for type 2 diabetes mellitus based on Mamba and KAN architecture

Abstract

The invention discloses an intelligent pulse diagnosis method for type 2 diabetes mellitus based on Mamba and KAN architectures, and belongs to the field of diabetes mellitus diagnosis and traditional Chinese medicine pulse signal analysis. The method specifically comprises the following steps: collecting pulse condition data; preprocessing the pulse condition data and dividing a data set; a Mama and KAN hybrid neural network model is built; training a Mama and KAN hybrid network model by utilizing the pulse condition data in the training set, and storing network model parameters converged after training; and finally, performing type 2 diabetes diagnosis on the pulse condition data by utilizing the trained model.

Description

Technical Field

[0001] The present invention relates to the fields of diabetes diagnosis and traditional Chinese medicine pulse analysis, and specifically relates to a type 2 diabetes intelligent pulse diagnosis method based on the Mamba and KAN architectures. Background Art

[0002] Diabetes, as a common chronic metabolic disease, its core pathological mechanism is the disorder of blood glucose metabolism caused by absolute or relative deficiency of insulin secretion. Among the diabetes classifications, type 2 diabetes accounts for up to 90%, becoming the most main clinical type. The chronic hyperglycemic state can cause damage to various tissues, especially prone to cause chronic damage and dysfunction of the eyes, kidneys, heart, blood vessels and nerves. Therefore, the early and accurate diagnosis of type 2 diabetes has important clinical value.

[0003] Traditional Chinese medicine diagnostics, as a core component of traditional Chinese medicine, shows unique advantages in the screening of type 2 diabetes. Compared with Western medicine diagnosis, traditional Chinese medicine can achieve non-invasive and convenient early identification through observation, auscultation, interrogation and palpation, and does not need to rely on complex equipment, and has significant potential in the early intervention of diseases and the regulation of sub-healthy states. Among them, in the clinical diagnosis and treatment of traditional Chinese medicine, pulse diagnosis (pulse palpation) is an important diagnostic method, and the pulse of the human body can reflect the physical condition of the human body to a certain extent. Relevant research has found that the pulses of type 2 diabetes patients are different from those of normal people to a certain extent. For example, thready pulse is common in qi deficiency syndrome, while taut and rapid pulse is common in yin deficiency syndrome. However, the diagnosis of type 2 diabetes by traditional Chinese medicine still highly relies on the subjective experience of experts and lacks objective and quantitative standards, which hinders the development and spread of traditional Chinese medicine to a certain extent.

[0004] In recent years, artificial intelligence technology has made breakthrough progress in the field of medical diagnosis. By combining artificial intelligence technology with traditional Chinese medicine pulse diagnosis and constructing an intelligent diagnosis model for type 2 diabetes based on pulse information, it can not only make up for the deficiency of traditional Chinese medicine quantification, but also provide a new paradigm for the integrated diagnosis and treatment of traditional Chinese and Western medicine, with great research significance and application value. Summary of the Invention

[0005] Aiming at the problems described in the background art, the present invention uses pulse information to propose a type 2 diabetes intelligent pulse diagnosis method based on the Mamba and KAN architectures. By integrating deep learning and traditional Chinese medicine diagnosis experience, it realizes non-invasive and rapid screening of type 2 diabetes and promotes the digital upgrade of traditional Chinese medicine diagnosis technology.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] S1, Pulse data collection;

[0008] S2, Pulse data preprocessing and dataset division;

[0009] S3. Build a hybrid neural network model of Mamba and KAN;

[0010] S4. Use the pulse condition data in the training set to train the hybrid network model of Mamba and KAN, and save the network model parameters that converge after training;

[0011] S5. Use the trained model to diagnose type 2 diabetes for pulse condition data;

[0012] Furthermore, in S1, a pulse condition collector is used to collect the pulse conditions of type 2 diabetes patients and those of the healthy control group respectively; during the process of collecting pulse conditions, by adjusting the measurement pressure of the pulse condition collector, pulse diagrams under different pressure values are obtained, and the pulse data corresponding to the pressure are recorded. The time for each pulse condition collection is 10 s, the sampling frequency is 200 Hz, and there are a total of 2000 data points.

[0013] Furthermore, in S2, preprocessing operations are performed on the one-dimensional pulse condition signals obtained in S1. Variational Mode Decomposition (VMD) is used to denoise the pulse condition data collected by the pulse diagnosis instrument to suppress electromyogram interference and baseline drift; after the pulse condition data is decomposed by VMD, multiple Intrinsic Mode Functions (IMFs) are generated, and the baseline drift components are concentrated in several high-order IMFs; by passing them through a group of low-pass filters and then reconstructing the IMFs, the preprocessed pulse condition data can be obtained.

[0014] Subsequently, the preprocessed pulse condition data is divided into a training set and a test set: the training set data is used to train the model parameters, and the test set data is used to test the generalization and robustness of the model. Usually, the data in the training set and the test set need to be ensured to be independent of each other to avoid affecting the model evaluation results.

[0015] Furthermore, in S3, a hybrid model of Mamba and KAN is constructed to analyze and identify pulse condition data; the model is divided into three parts. First, a multi-scale convolution module is used as the feature extraction front end for pulse condition data to extract key features in the pulse condition data. The multi-scale convolution module consists of cascaded Downsample modules and Dilation modules. Among them, the Downsample module is used to gradually reduce the spatial dimension of the feature map, and the Dilation module uses dilated convolution operations to increase the receptive field; subsequently, the Mamba module is used to perform global temporal dependence modeling on the features to capture the long-term and short-term dependence relationships in the pulse signal; finally, the features are output to the KAN (Kolmogorov - Arnold Networks) network for generating the final type 2 diabetes diagnosis result (classification result).

[0016] Further, in S4, the model is trained using the pulse condition data in the training set, and the neural network parameters are continuously adjusted until the model output tends to be stable. Specifically, x represents the input pulse condition data, y represents its corresponding true label, and p represents the classification probability distribution obtained by the model for the input pulse condition data x. Then, the objective during the model training process is to minimize the loss function L loss :

[0017]

[0018] where N represents the number of samples; y i represents the true label of the i-th sample; p i represents the predicted output of the model for the sample. By minimizing this loss function, the model parameters are dynamically adjusted to achieve the detection of type 2 diabetes by the model.

[0019] Further, in S5, the trained diabetes intelligent pulse diagnosis model based on the Mamba and KAN architectures is deployed in the actual application scenario. The pulse condition information is collected using the pulse condition acquisition device, and the pulse condition data is uploaded. Subsequently, the model is used to analyze and process the pulse condition data, and finally, the rapid and accurate diagnosis of type 2 diabetes is achieved.

[0020] The advantages of the present invention are as follows: The present invention innovatively integrates the CNN, Mamba, and KAN architectures to construct a type 2 diabetes intelligent diagnosis model for pulse condition time series features. Based on the one-dimensional time series characteristics of the pulse signal, the long-range dependence modeling ability of the Mamba architecture is utilized, and combined with the local feature extraction advantages of the CNN and the non-linear mapping characteristics of the KAN, high-precision non-invasive screening of type 2 diabetes is achieved. This solution breaks through the subjective limitations of traditional Chinese medicine pulse diagnosis and provides a quantifiable technical path for the digitalization of traditional Chinese medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings, as a part of the present invention, mainly aim to facilitate the understanding of the present invention and are used for explanation, and should not constitute an improper limitation to the present invention.

[0022] Figure 1 is a flowchart of the method of the present invention.

[0023] Figure 2 is a type 2 diabetes intelligent pulse diagnosis model based on the Mamba and KAN architectures in the method of the present invention.

[0024] Figure 3 is the mapping relationship between the KAN layers in the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0026] As Figure 1 shown, the method for intelligent pulse diagnosis of type 2 diabetes based on the Mamba and KAN architectures of the present invention includes the following steps:

[0027] S1, pulse condition data acquisition;

[0028] Use a pulse condition acquisition instrument to collect the pulse conditions of type 2 diabetes patients and the pulse conditions of the healthy control group respectively; during the process of collecting the pulse condition, by adjusting the measurement pressure of the pulse condition acquisition instrument, obtain the pulse diagrams under different pressure values, and record the pulse data under the corresponding pressure. The time for each pulse condition collection is 10s, and the sampling frequency is 200Hz, with a total of 2000 data points.

[0029] The ZM300 type intelligent pulse condition instrument (Shanghai Shike Science and Education Equipment Co., Ltd.) is used for pulse condition data acquisition. Different pressures are output through a precision pressure sensor to simulate the light-taking and heavy-taking pulse-taking in traditional Chinese medicine clinical diagnosis, ensuring that the pulse condition signals meeting the traditional Chinese medicine diagnosis standards are obtained. The collected pulse condition signals can be transmitted to the computer in real time and digitally processed by relevant software.

[0030] S2, preprocessing of pulse condition data and dataset division;

[0031] Perform preprocessing operations on the one-dimensional pulse condition signals obtained in S1. Use Variational Mode Decomposition (VMD) to denoise the pulse condition data collected by the pulse diagnosis instrument to suppress electromyogram interference and baseline drift; after the pulse condition data is decomposed by VMD, multiple Intrinsic Mode Functions (IMFs) are generated, and the baseline drift components are concentrated in several high-order IMFs; make it pass through a group of low-pass filters, and then reconstruct the IMFs to obtain the preprocessed pulse condition data.

[0032] Subsequently, divide the preprocessed pulse condition data into a training set and a test set: the training set data is used to train the model parameters, and the test set data is used to test the generalization and robustness of the model. Usually, the data in the training set and the test set need to be ensured to be independent of each other to avoid affecting the model evaluation results.

[0033] S3, build a hybrid neural network model of Mamba and KAN;

[0034] Construct a hybrid model of Mamba and KAN to analyze and identify pulse data; as Figure 2 shown, the model is divided into three parts:

[0035] First, a multi-scale convolution module is used as the front-end for feature extraction of pulse data to extract key features in the pulse data. The multi-scale convolution module consists of cascaded Downsample modules and Dilation modules. Among them, the Downsample module is used to gradually reduce the spatial dimension of the feature map, and the Dilation module uses dilated convolution operations to increase the receptive field; the core of the convolutional layer is the convolution kernel, which has a powerful local feature extraction ability. The input data of the upper layer is convolved through the convolution kernel to extract the features of the data. The mathematical operation of the convolutional layer is as follows:

[0036]

[0037] Among them, M k is the connection relationship between the input and output feature maps in the convolutional layer, is the l-th output of the k-th layer of neurons, w ik is the weight parameter of the convolution kernel, b k is the bias coefficient of the k-th layer, · is the discrete convolution, and f is the activation function.

[0038] After completing the convolutional feature extraction, this study uses the Mamba module to perform global temporal dependence modeling on the extracted features to capture the long-term and short-term dependence relationships in the pulse signal; the State Space Models (SSM) is a commonly used method for constructing deep learning architectures to process time series signals.

[0039] Traditional SSMs can be formalized as linear time-invariant systems. Such systems map the input sequence x(t) ∈ R N to the output response y(t) ∈ R N through the hidden state y(h) ∈ R N . This mapping process can be represented by a linear ordinary differential equation:

[0040]

[0041] where t is the current input time, h′(t) represents the hidden state of the current input x(t), h(t) represents the hidden state of the previous time, A ∈ R N×N represents the state transition matrix, B ∈ R N×N is the input matrix, describing the influence of the external input on the system state together, and C ∈ R 1×N is the output matrix, used to control the influence of the state on the output.

[0042] Traditional SSM is constructed based on a continuous-time framework, which is essentially different from the discrete-time series processing methods commonly used in deep learning. Therefore, it cannot be directly applied to deep learning architectures. Thus, it is necessary to discretize SSM, and the zero-order hold method is often used to achieve discretization. Specifically, A and B are discretely transformed by combining the zero-order hold technique and the time-scale parameter Δ, and the transformation process is as follows:

[0043]

[0044] where I represents the identity matrix, and the discretized SSM equation can be expressed as:

[0045]

[0046] where h k is the hidden state at time step k; x k , y k are the input and output sequences at time step k, respectively.

[0047] However, due to the linear time-invariance of SSM, its ability to perform content-aware reasoning is limited, that is, it is difficult to dynamically adjust its behavior according to the content of the input data. Therefore, on this basis, Mamba introduces a selection mechanism, which allows the model to dynamically filter out relevant information according to the input data, thereby enhancing the adaptability and flexibility of the model. The architecture of Mamba is as Figure 2 shown.

[0048] Finally, the features are output to the KAN (Kolmogorov-Arnold Networks) network for generating the final type 2 diabetes diagnosis result (classification result).

[0049] The theoretical basis of KAN is the Kolmogorov-Arnold representation theorem. The theorem states that any multivariate continuous function can be represented as a combination of a finite number of univariate continuous functions. Compared with the limitations of traditional multi-layer perceptrons (MLPs) and their derivative architectures (such as Transformers) that rely on linear combinations and fixed activation functions, KAN realizes a dynamic learning mechanism through parameterized non-linear activation functions (such as spline functions). It not only breaks through the bottleneck of linear modeling but also significantly improves the model representation ability and parameter efficiency by autonomously optimizing the network structure. The model formula of KAN can be expressed as:

[0050]

[0051] where f(x1,x2,...,x n ) represents a multivariate continuous function, Φ q and Φ q,pRepresent some univariate continuous functions. The core idea of this theorem is to decompose a multivariate function into a combination and superposition of some univariate functions, and perform mutual superposition on Φ q,p and perform mutual superposition on Φ q and perform mutual superposition on Φ, so as to simplify the complexity. For each KAN with L layers, the mapping relationship between neurons from the l-th layer to the (l + 1)-th layer is as Figure 3 shown.

[0052] S4. Use the pulse condition data in the training set to train the Mamba and KAN hybrid network model, and save the network model parameters that converge after training;

[0053] Use the pulse condition data in the training set to train the model, and continuously adjust the neural network parameters until the model output tends to be stable. Specifically, x represents the input pulse condition data, y represents its corresponding true label, and p represents the classification probability distribution obtained by the model for the input pulse condition data x. Then the goal during the model training process is to minimize the loss function L loss :

[0054]

[0055] where N represents the number of samples; y i represents the true label of the i-th sample; p i represents the predicted output of the model for the sample. By minimizing this loss function, the model parameters are dynamically adjusted to achieve the detection of type 2 diabetes by the model.

[0056] S5. Use the trained model to diagnose type 2 diabetes for the pulse condition data;

[0057] Deploy the trained diabetes pulse diagnosis model based on the Mamba and KAN architectures in the actual application scenario. Use the pulse condition acquisition device to collect the pulse condition information and upload the pulse condition data. Then use this model to analyze and process the pulse condition data, and finally achieve the rapid and accurate diagnosis of type 2 diabetes.

[0058] The above has described in detail the preferred specific embodiments of the present invention. It should be understood that those of ordinary skill in the art can make many modifications and variations according to the concept of the present invention without creative labor. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A type 2 diabetes intelligent pulse diagnosis method based on the Mamba and KAN architectures, characterized in that: The method performs the following steps: S1, Pulse condition data acquisition; S2, Pulse condition data preprocessing and dataset division; S3, Construct a hybrid neural network model of Mamba and KAN; S4, Use the pulse condition data in the training set to train the hybrid network model of Mamba and KAN, and save the network model parameters that converge after training; S5, Use the trained model to diagnose type 2 diabetes for pulse condition data; 2. The intelligent pulse diagnosis method for type 2 diabetes based on the Mamba and KAN architectures according to claim 1, characterized in that: In step S1: Use a pulse condition acquisition instrument to collect the pulse conditions of type 2 diabetes patients and the pulse conditions of the healthy control group respectively; during the process of collecting the pulse condition, by adjusting the measurement pressure of the pulse condition acquisition instrument, obtain the pulse diagrams under different pressure values, and record the pulse data under the corresponding pressure. The time for each pulse condition collection is 10 s, the sampling frequency is 200 Hz, and a total of 2000 data points are obtained.

3. The intelligent pulse diagnosis method for type 2 diabetes based on the Mamba and KAN architectures according to claim 1, wherein: In step S2: Perform preprocessing operations on the one-dimensional pulse condition signals obtained in S1. Use variational mode decomposition (VMD) to denoise the pulse condition data collected by the pulse diagnosis instrument to suppress interference and baseline drift; after the pulse condition data is decomposed by VMD, multiple intrinsic mode functions (IMFs) are generated, and the baseline drift components are concentrated in several high-order IMFs; make it pass through a group of low-pass filters, and then reconstruct the IMFs to obtain the preprocessed pulse condition data. Subsequently, divide the preprocessed pulse condition data into a training set and a test set: the training set data is used to train the model parameters, and the test set data is used to test the generalization and robustness of the model. Usually, the data in the training set and the test set need to be ensured to be independent of each other to avoid affecting the model evaluation results.

4. The intelligent pulse diagnosis method for type 2 diabetes based on the Mamba and KAN architectures according to claim 1, characterized in that: Step S3 includes: Construct a hybrid model of Mamba and KAN to analyze and identify pulse condition data; the model is divided into three parts. First, use a multi-scale convolution module as the feature extraction front end for pulse condition data to extract key features in the pulse condition data. The multi-scale convolution module is composed of cascaded Downsample modules and Dilation modules. Among them, the Downsample module is used to gradually reduce the spatial dimension of the feature map, and the Dilation module uses dilated convolution operations to increase the receptive field; subsequently, use the Mamba module to perform global temporal dependence modeling on the features to capture the long-term and short-term dependence relationships in the pulse signal; finally, output the features to the KAN (Kolmogorov-Arnold Networks) network for generating the final type 2 diabetes diagnosis result (classification result).

5. The intelligent pulse diagnosis method for type 2 diabetes based on the Mamba and KAN architectures according to claim 1, characterized in that: In step S4: The model is trained using the pulse condition data in the training set, and the neural network parameters are continuously adjusted until the model output tends to be stable. Specifically, x represents the input pulse condition data, y represents its corresponding true label, and p represents the classification probability distribution obtained by the model for the input pulse condition data x. Then the goal during the model training process is to minimize the loss function L loss : Among them, N represents the number of samples; y i represents the true label of the i-th sample; p i represents the predicted output of the model for the sample. By minimizing this loss function, the model parameters are dynamically adjusted to achieve the detection of type 2 diabetes by the model.

6. The intelligent pulse diagnosis method for type 2 diabetes based on the Mamba and KAN architectures according to claim 1, characterized in that: In step S5: Deploy the trained diabetes diagnosis model based on the Mamba and KAN architectures in the actual application scenario, use the pulse condition acquisition device to collect pulse condition information and upload the pulse condition data, and then use this model to analyze and process the pulse condition data, and finally achieve a fast and accurate diagnosis of type 2 diabetes.

Images