Metabolic disease digital management system based on traditional Chinese and western medicine collaborative multi-source data fusion

Abstract

This application provides a digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine (TCM) and relates to the field of medical and health information processing technology. It first preprocesses the multi-source heterogeneous raw data according to sampling frequency characteristics, and then maps both to a unified feature space through a dual-stream coding network. In the fusion stage, TCM macroscopic constitution characteristics are used as query conditions to apply cross-modal attention modulation to Western medicine microscopic fluctuation characteristics, achieving deep fusion at the temporal and semantic levels. The fused features are simultaneously output as metabolic trend predictions and syndrome evolution inferences via a dual-path decoding network. Furthermore, a dynamic modulation mechanism for syndrome sensitivity is introduced in the intervention decision-making stage, using the syndrome probability distribution to nonlinearly modulate the risk penalty parameter in real time. This solves the problems of static threshold models failing to reflect the concept of treating different diseases with the same method and the break in the TCM-Western medicine synergy link at the decision-making level, realizing end-to-end TCM-Western medicine synergistic management from data collection to closed-loop intervention.

Description

Technical Field

[0001] This application relates to the field of medical and health information processing technology, and more specifically, to a digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine. Background Technology

[0002] With the continued rise in the prevalence of metabolic diseases, especially type 2 diabetes, the use of wearable sensors and smart terminals for continuous and digital health management of patients has become an important trend in the field of chronic disease prevention and control. Meanwhile, the holistic approach and syndrome differentiation-based treatment system of Traditional Chinese Medicine (TCM) has unique advantages in the long-term conditioning of the constitution for metabolic diseases, and a comprehensive management model integrating TCM and Western medicine is increasingly being adopted in clinical practice.

[0003] In existing technologies, digital management solutions on the Western medicine side mainly rely on high-frequency time-series data such as continuous blood glucose monitoring, using recurrent neural networks to predict short-term metabolic fluctuations. On the Traditional Chinese Medicine (TCM) side, intelligent auxiliary systems classify static constitutions and provide conditioning suggestions through tongue image recognition and body composition scale analysis. These two systems operate independently. Even those few solutions that attempt to combine Western and TCM data only employ simple feature vector concatenation, failing to establish a deep semantic connection between the two. The fundamental reason is that Western medicine data consists of precise quantitative signals sampled at the second to minute level, while TCM syndrome characteristics belong to macroscopic qualitative abstract patterns with daily or weekly cycles. There is a fundamental mismatch between the two in terms of sampling rate and semantic space, and traditional data-level alignment or feature-level concatenation methods cannot effectively bridge this gap. More importantly, TCM constitution essentially constitutes the macroscopic metabolic environment of the human body and should regulate and constrain the evolutionary trends of microscopic indicators, rather than simply being parallel to them. Existing solutions ignore this physiological logic, making it difficult for models to capture the nonlinear amplification mechanism of macroscopic constitution on microscopic fluctuations in complex clinical scenarios. Furthermore, in the final risk assessment and intervention decision-making stage, existing solutions generally use global static thresholds for discrimination, failing to establish a dynamic modulation relationship between TCM syndrome status and risk sensitivity parameters. This results in patients with different constitutions obtaining indiscriminate risk assessment results under the same blood glucose deviation, which violates the core concept of TCM of treating different diseases with the same method and causes a break in the TCM-Western medicine collaboration link at the decision-making level.

[0004] Therefore, there is a need for an optimized digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, comprising: The multi-source data preprocessing and digitization module is used to perform time-series cleaning, alignment, and visual feature digitization mapping on the collected multi-source heterogeneous raw data, including continuous glucose concentration sequences, photoplethysmography pulse wave sequences, traditional Chinese medicine body weight scale questionnaire data, and tongue images, to obtain high-frequency data sequences of Western medicine and low-frequency feature sets of traditional Chinese medicine. The multidimensional feature encoding and extraction module is used to encode and extract the temporal fluctuation features of high-frequency data sequences of Western medicine and to encode and extract the macroscopic constitution features of low-frequency feature sets of traditional Chinese medicine to obtain the microscopic feature vector of Western medicine and the macroscopic feature vector of traditional Chinese medicine. The cross-modal attention fusion module is used to perform time-semantic cross-modal attention fusion on macro-feature vectors of traditional Chinese medicine and micro-feature vectors of Western medicine to obtain a collaborative fusion mapping vector. The Health Trend and Syndrome Evolution Inference Module is used to predict the metabolic trend time series extension of the collaborative fusion mapping vector and to infer the syndrome state transition probability of the collaborative fusion mapping vector to obtain the metabolic prediction inference vector and the syndrome evolution inference vector. The digital closed-loop intervention strategy generation module is used to generate digital closed-loop intervention strategies based on the preset Western medicine drug rule base and the traditional Chinese medicine food homology and compatibility knowledge base to obtain a synergistic intervention plan of traditional Chinese and Western medicine.

[0006] Compared with existing technologies, the digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine provided in this application firstly preprocesses and encodes the time-series data of Western medicine, such as continuous glucose concentration sequences and pulse wave sequences, and the discrete data of traditional Chinese medicine, such as tongue images and body mass scales, independently to form a dimension-aligned dual-stream feature representation. Then, it abandons the traditional feature splicing paradigm and applies cross-attention modulation to the micro-fluctuation features of Western medicine as query conditions using the macro-constitutional characteristics of traditional Chinese medicine. This allows the macro-constitutional patterns to dynamically guide the feature extraction and weight allocation of micro-physiological fluctuations, achieving deep fusion at the temporal and semantic levels. On this basis, it synchronously outputs metabolic trend prediction and syndrome evolution inference results through a dual-branch decoding network, and introduces a syndrome sensitivity dynamic modulation mechanism in the final risk assessment stage. This enables patients with different constitutions to receive differentiated risk assessments and graded intervention responses, thereby solving the problems in existing solutions where static threshold discrimination cannot reflect the concept of treating different diseases with the same method and the link between traditional Chinese and Western medicine is broken at the decision-making level. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a system block diagram of a digital management system for metabolic diseases based on multi-source data fusion of traditional Chinese and Western medicine, according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in a digital management system for metabolic diseases based on multi-source data fusion of traditional Chinese and Western medicine, according to an embodiment of this application. Figure 3 This is a block diagram of the cross-modal attention fusion module in a digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, according to an embodiment of this application. Figure 4 This is a block diagram of the health trend and syndrome evolution inference module in the digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, according to an embodiment of this application. Figure 5 This is a block diagram of the digital closed-loop intervention strategy generation module in the digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, according to an embodiment of this application. Figure 6 This is a block diagram of the Chinese and Western medicine intervention instruction generation unit in the digital management system for metabolic diseases based on the fusion of multi-source data from integrated Chinese and Western medicine, according to an embodiment of this application. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] This application uses system block diagrams and data flow diagrams to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] In the field of digital management of metabolic diseases, there is a fundamental sampling rate mismatch and semantic space gap between the high-frequency quantitative time-series signals generated by continuous monitoring in Western medicine and the low-frequency discrete qualitative features relied upon by traditional Chinese medicine (TCM) constitution identification. Existing technologies only combine the two types of data through simple feature splicing, failing to establish a deep modulation correlation between the macroscopic constitution of TCM and the microscopic fluctuations of Western medicine. Furthermore, the risk assessment and intervention decision-making processes generally employ global static threshold discrimination, failing to incorporate the evolution of syndromes into the dynamic adjustment of risk parameters, resulting in a break in the TCM-Western medicine collaboration link at the decision-making level. Therefore, this application proposes a digital management system for metabolic diseases based on the fusion of multi-source data from TCM and Western medicine. Specifically, the system first preprocesses multi-source heterogeneous data, including continuous glucose concentration sequences, photoplethysmography pulse wave sequences, TCM constitution questionnaire data, and tongue images, according to their sampling frequency characteristics, forming high-frequency data sequences from Western medicine and low-frequency feature sets from TCM. These are then mapped to a unified feature space via a dual-stream coding network. The fusion phase abandons the traditional splicing paradigm, using macroscopic constitution characteristics of Traditional Chinese Medicine (TCM) as query conditions to apply cross-modal attention modulation to microscopic fluctuation characteristics of Western medicine, achieving deep fusion at the temporal and semantic levels. The fused features are simultaneously output via a dual-path decoding network, providing metabolic trend predictions and syndrome evolution deductions. A dynamic modulation mechanism for syndrome sensitivity is introduced in the intervention decision-making process, using real-time nonlinear modulation of risk penalty weights and warning thresholds based on syndrome probability distribution. Furthermore, a superlinear amplification function for syndrome perception simulates the differentiated pathological responses of patients with different constitutions to transgressive fluctuations. Finally, a tiered, step-by-step instruction mapping and pharmacological mutual exclusion safety verification generate a synergistic plan encompassing short-term Western medicine interventions and long-term TCM conditioning, achieving a complete closed-loop integration of TCM and Western medicine.

[0014] Figure 1 This is a system block diagram of a digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, according to an embodiment of this application. Figure 2 This is a schematic diagram illustrating the data flow of a digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, according to an embodiment of this application. Figure 1 and Figure 2As shown, the metabolic disease digital management system 100 based on multi-source data fusion of traditional Chinese and Western medicine according to an embodiment of this application includes: a multi-source data preprocessing and digitization module 110, used to perform temporal cleaning, alignment, and visual feature digitization mapping on the collected multi-source heterogeneous raw data, including continuous glucose concentration sequences, photoplethysmography pulse wave sequences, traditional Chinese medicine body constitution questionnaire data, and tongue images, to obtain a high-frequency data sequence of Western medicine and a low-frequency feature set of traditional Chinese medicine; and a multi-dimensional feature encoding and extraction module 120, used to perform temporal fluctuation feature encoding and extraction on the high-frequency data sequence of Western medicine and macroscopic body constitution feature encoding and extraction on the low-frequency feature set of traditional Chinese medicine to obtain a microscopic feature vector of Western medicine and a macroscopic feature vector of traditional Chinese medicine. The cross-modal attention fusion module 130 is used to perform time-semantic cross-modal attention fusion on the macro-feature vector of traditional Chinese medicine and the micro-feature vector of Western medicine to obtain a collaborative fusion mapping vector; the health trend and syndrome evolution inference module 140 is used to predict the metabolic trend time series extension of the collaborative fusion mapping vector and to infer the syndrome state transition probability of the collaborative fusion mapping vector to obtain the metabolic prediction inference vector and the syndrome evolution inference vector; the digital closed-loop intervention strategy generation module 150 is used to generate a digital closed-loop intervention strategy based on the preset Western medicine drug rule base and the traditional Chinese medicine food homology and compatibility knowledge base to obtain a collaborative intervention plan of traditional Chinese and Western medicine.

[0015] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the multi-source data preprocessing and digitization module 110 is used to perform temporal cleaning, alignment, and visual feature digitization mapping on the acquired heterogeneous raw data, including continuous glucose concentration sequences, photoplethysmography (PPG) sequences, TCM body weight questionnaire data, and tongue images, to obtain high-frequency data sequences from Western medicine and low-frequency feature sets from traditional Chinese medicine. It should be noted that in the digital management scenario of metabolic diseases, continuous glucose concentration sequences and PPG sequences are high-frequency time-series data sampled at the second to minute level, while TCM body weight questionnaire data and tongue images are low-frequency discrete data collected on a daily or weekly basis. These two types of data differ fundamentally in sampling rate, data format, and numerical dimensions. If they are directly mixed and calculated without preprocessing, the high-frequency signal will be distorted by the low-frequency data, or the low-frequency features will be submerged by high-frequency noise. Based on this, the technical solution of this application first performs memory buffering and routing, time-series data sliding denoising and interpolation stitching, tongue surface region of interest mask extraction and color and texture feature digital mapping, and dimensionless standardization calibration and stitching aggregation on the collected multi-source heterogeneous raw data. Through the above processing, the multi-source raw data with different sampling rates and data forms can be transformed into well-formatted and time-aligned high-frequency data sequences of Western medicine and low-frequency feature sets of traditional Chinese medicine, providing standardized input with controllable quality for subsequent cross-modal feature encoding and fusion.

[0016] More specifically, in a concrete example of this application, for the implementation of the data diversion routing unit, after the sensor array and mobile terminal synchronously upload multimodal raw data to the preprocessing pipeline, a memory buffer diversion routing operation is performed based on the physical polarity of the sampling period of each data source. The continuous glucose concentration sequence sampled every 5 minutes and the photoplethysmography pulse wave sequence sampled every second are identified as high-frequency time series types and packaged and written into the same hardware channel buffer to generate a Western medicine sequence group to be processed; the tongue images and TCM body composition questionnaire data submitted by users weekly are identified as low-frequency discrete types and imported into the asynchronous analysis channel buffer to generate a TCM feature group to be processed. This diversion routing operation ensures that subsequent processing units only receive homogeneous data streams that match their algorithm characteristics, avoiding processing logic conflicts caused by heterogeneous data mixing.

[0017] For the implementation of the Western medicine data processing unit, for the original waveforms of each channel after unpacking in the Western medicine sequence group to be processed, a fixed-length backward sliding observation window is first established on the time axis. The arithmetic mean of the sampling points within the window is taken to filter out motion artifacts and white noise from tissue fluid osmotic pressure measurement, obtaining a smoothed and denoised signal estimate. Subsequently, since the sampling interval of the continuous glucose concentration sequence is 5 minutes while the sampling interval of the photoplethysmography pulse wave sequence is 1 second, there is a sampling rate difference of hundreds of times between the two. Cubic spline interpolation technology is used to resample the dual-channel sequences to the same standardized timestamp reference, and physical-level splicing is completed in the feature channel dimension, thereby obtaining a Western medicine high-frequency data sequence with absolute time alignment.

[0018] For the implementation of the tongue image feature extraction unit, the first step is to perform a file format separation operation on the TCM feature group to be processed. The structured questionnaire data is temporarily stored as unnormalized questionnaire data and written into the secondary buffer pool. The separated tongue image is then fed into the semantic segmentation network. This segmentation network performs pixel-level annotation on the lips, teeth, and irrelevant background areas in the tongue image and generates a binary mask, based on which the pure tongue surface region of interest is extracted. On this region of interest, the pixel values ​​are converted from the RGB color space to the HSV color space and low-order color moment features are extracted. At the same time, a local binary pattern texture descriptor is calculated to quantify the tooth mark morphology and tongue coating roughness information. The above color and texture features are then flattened in one dimension to generate a tongue image visual feature vector.

[0019] For the implementation of the questionnaire data processing and feature aggregation unit, the questionnaire data to be normalized is read from the secondary buffer pool. Since the original scale scores of each questionnaire item differ significantly in numerical dimensions from the pixel statistics in the tongue image visual feature vector, the score of each valid item is subtracted from the arithmetic mean of the entire batch of records, and then divided by the corresponding sample standard deviation to complete the dimensionless standardization calibration, resulting in the questionnaire standard score tensor. Subsequently, this questionnaire standard score tensor and the tongue image visual feature vector output by the preceding tongue image feature extraction unit are concatenated and aggregated at the fully connected layer node level, solidifying into a low-frequency TCM feature set that jointly expresses the patient's static constitution characteristics. At this point, the entire processing flow of the multi-source data preprocessing and digitization module is completed. The output Western medicine high-frequency data sequence and the TCM low-frequency feature set are then fed as two parallel data streams into the subsequent multi-dimensional feature encoding and extraction modules.

[0020] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the multidimensional feature encoding and extraction module 120 is used to extract temporal fluctuation features from the high-frequency data sequence of Western medicine and to extract macroscopic constitution features from the low-frequency feature set of traditional Chinese medicine to obtain microscopic feature vectors of Western medicine and macroscopic feature vectors of traditional Chinese medicine. It should be noted that the high-frequency data sequence of Western medicine obtained after preprocessing is a multi-channel temporal continuous waveform containing blood glucose fluctuations and pulse changes, while the low-frequency feature set of traditional Chinese medicine is a static numerical vector composed of tongue visual features and questionnaire standard scores. The two differ in data form, information density, and semantic level, and cannot be directly input into the same fusion network for joint computation. Based on this, the technical solution of this application further performs multi-layer local temporal fluctuation feature perception and global max-pooling dimensionality reduction encoding on the high-frequency data sequence of Western medicine, performs affine transformation and high-order correlation capture mapping on the low-frequency feature set of traditional Chinese medicine, and performs covariance offset elimination and equal-dimensional orthogonal alignment projection on the encoded dual-path tensors. Through the above processing, two types of features with vastly different sampling rates and semantic levels can be compressed into Western medicine micro-feature vectors and traditional Chinese medicine macro-feature vectors with unified dimensions and consistent distribution characteristics, respectively, providing standardized inputs that can directly participate in matrix operations for subsequent cross-modal attention fusion.

[0021] More specifically, in a concrete example of this application, for the implementation of the Western medicine feature encoding unit, a high-frequency data sequence of Western medicine is input into a temporal convolutional encoder. This encoder adopts a dilated causal convolution structure, which expands the historical receptive field exponentially without introducing information leakage from future moments by progressively increasing the dilation factor layer by layer. In each convolutional layer, a local perceptual window is established along the time axis, and a weighted summation operation is performed on the sampling points within the window coverage area and the corresponding filter weights to capture the peak changes and periodic trend features of the blood glucose waveform at different time scales. After progressive abstraction through multiple convolutional layers, a global max pooling operation is performed along the time dimension on the temporal feature map output by the final convolutional layer to extract the time information with the strongest response on each feature channel and compress the time dimension to obtain the Western medicine microscopic latent space tensor.

[0022] For the implementation of the TCM feature encoding unit, the low-frequency feature set of TCM is input into a multi-layer feedforward network. Within each hidden layer, an affine transformation operation is performed on the input vector, which involves multiplying the input with the connection weight matrix of that layer and then adding a bias term. Subsequently, a nonlinear mapping capability is introduced through a leaky linear rectified activation function, breaking the linear dependency constraint between dimensions. Through alternating processing of multi-layered stacked affine transformations and nonlinear activations, the network gradually captures the implicit high-order correlation between tongue color and texture features and questionnaire constitution scores, upscaling the original low-dimensional static values ​​to a continuous latent space with rich constitution semantic expression capabilities, thus obtaining the TCM macroscopic latent space tensor.

[0023] For the implementation of the feature alignment projection unit, since the Western medicine microscopic latent space tensor is encoded by a convolutional network and the traditional Chinese medicine macroscopic latent space tensor is encoded by a feedforward network, the internal data distributions of the two exhibit covariance shifts, and their embedding dimensions may be inconsistent. Directly feeding them into the cross-attention network would cause the dot product operation between the query matrix and the key matrix to lose its physical meaning. Therefore, layer normalization operations are applied to both the Western medicine microscopic latent space tensor and the traditional Chinese medicine macroscopic latent space tensor to correct the numerical distributions on their respective feature dimensions to a zero-mean, unit-variance state, eliminating the covariance shift introduced by the difference in encoding paths. Subsequently, a shared global dimension projection operator is used to project the two normalized tensors onto the same target embedding dimension space, achieving equal-dimensional orthogonal alignment and obtaining the Western medicine microscopic feature vector and the traditional Chinese medicine macroscopic feature vector, respectively. At this point, the entire processing flow of the multi-dimensional feature encoding and extraction module is completed, and the two output equal-dimensional feature vectors serve as the query source and the queried source, respectively, and enter the subsequent cross-modal attention fusion module.

[0024] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from traditional Chinese and Western medicine, the cross-modal attention fusion module 130 is used to perform time-semantic cross-modal attention fusion on the macro-feature vectors of traditional Chinese medicine and the micro-feature vectors of Western medicine to obtain a collaborative fusion mapping vector. It should be noted that, given that the Western medicine micro-feature vector obtained after dual-stream encoding carries local temporal fluctuation information at the second to minute level, such as sudden changes in blood glucose peaks and periodic changes in pulse, while the TCM macro-feature vector carries overall constitution information at the daily or weekly scale, as reflected by tongue color and texture and body mass index scores, although the two have been aligned to a unified embedding dimension space, they still belong to the micro-dynamic domain and the macro-static domain in terms of temporal granularity and semantic level, respectively. If the traditional vector splicing method is used for fusion, it is essentially just mechanically juxtaposing the two types of features in the channel dimension, which cannot establish the modulation and constraint relationship between TCM constitution and Western medicine fluctuations. In the real metabolic disease management scenario, TCM constitution actually constitutes the macro-metabolic environment of the human body. For example, patients in the state of phlegm and dampness accumulation have abnormal spleen and stomach function and are less able to buffer postprandial blood glucose fluctuations than patients with qi deficiency. This means that macro-constitution features should dynamically adjust the attention weight of signals of different time periods and different amplitudes in micro-fluctuation features, rather than simply juxtaposing them with micro-features. Based on this, the technical solution of this application further performs time-semantic cross-modal attention fusion on the macroscopic feature vectors of Traditional Chinese Medicine (TCM) and the microscopic feature vectors of Western medicine. The macroscopic feature vectors of TCM are mapped to a query matrix through linear transformation, while the microscopic feature vectors of Western medicine are mapped to a key matrix and a value matrix, respectively. The scaling dot product operation of the query matrix and the key matrix is ​​used to evaluate the cross-modal correlation between macroscopic constitution patterns and high-frequency physiological fluctuations at different times. After generating an adaptive attention probability distribution through a normalized activation function, the value matrix is ​​weighted and aggregated. The aggregation result is then superimposed with the macroscopic feature vectors of TCM using residuals and subjected to layer normalization. Through the above processing, the holistic view of TCM can dynamically guide the selective extraction and amplification of signals from different frequency bands and different times in the microscopic fluctuations of Western medicine through attention weight allocation. This integrates the two heterogeneous features into a unified collaborative fusion mapping vector at the temporal and semantic levels, providing a joint representation basis that simultaneously contains details of microscopic fluctuations and macroscopic constitution constraints for subsequent metabolic trend prediction and syndrome evolution deduction.

[0025] Figure 3 This is a block diagram of the cross-modal attention fusion module in a digital management system for metabolic diseases based on multi-source data fusion of traditional Chinese and Western medicine, according to an embodiment of this application. Figure 3As shown, the cross-modal attention fusion module 130 includes: a feature projection transformation unit 131, used to project and transform the macro-feature vector of traditional Chinese medicine into a macro-constitution query matrix based on three independent learnable linear parameter matrices, and to project and transform the micro-feature vector of Western medicine into a micro-fluctuation key matrix and a micro-fluctuation value matrix respectively; a cross-modal modulation aggregation unit 132, used to perform cross-modal correlation evaluation on the macro-constitution query matrix and the micro-fluctuation key matrix to generate an adaptive attention probability distribution, and then use the adaptive attention probability distribution to perform weighted aggregation on the micro-fluctuation value matrix to obtain a cross-modal feature matrix; and a fusion stabilization unit 133, used to perform residual superposition of the cross-modal feature matrix and the macro-feature vector of traditional Chinese medicine, and then perform data distribution stabilization processing through a layer normalization function to obtain a collaborative fusion mapping vector.

[0026] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the feature projection transformation unit 131 is used to project and transform the macroscopic feature vectors of traditional Chinese medicine into a macroscopic constitution query matrix, and the microscopic feature vectors of Western medicine into a microscopic fluctuation key matrix and a microscopic fluctuation value matrix, respectively, based on three independent and learnable linear parameter matrices. It should be noted that although the microscopic feature vectors of Western medicine and the macroscopic feature vectors of traditional Chinese medicine are in the same embedding dimension space after feature alignment projection, their encoded semantic roles are fundamentally different. The macroscopic feature vectors of traditional Chinese medicine represent the patient's long-term stable constitution and syndrome tendencies, and should play a driving role in active retrieval and modulation in cross-modal fusion. The microscopic feature vectors of Western medicine represent the high-frequency dynamic fluctuations of blood glucose and pulse on the time axis, and should play a responding role in being retrieved and filtered. If both are directly input into attention calculation without role-based spatial projection, the dot product operation of the two types of vectors in the same linear space cannot distinguish the functional boundaries of the modulator and modulated party, resulting in the loss of cross-modal semantic orientation in attention weight allocation. Based on this, the technical solution of this application further transforms the macroscopic feature vector of Traditional Chinese Medicine (TCM) into a macroscopic constitution query matrix based on three independent and learnable linear parameter matrices, and transforms the microscopic feature vector of Western Medicine into a microscopic fluctuation key matrix and a microscopic fluctuation value matrix, respectively. Through the above processing, the two types of heterogeneous features can be projected from the shared alignment space to the query subspace, key subspace, and value subspace, which each undertake independent semantic functions. This allows the subsequent scaling dot product operation to be performed in the functionally decoupled projection domain, thereby providing a matrix input with a clear role and semantic separability for cross-modal attention fusion.

[0027] More specifically, in a concrete example of this application, three independent linear parameter matrices are pre-initialized, corresponding to the query projection, key projection, and value projection transformation paths, respectively. The weights of these three paths are independently updated during the training phase through backpropagation gradients. For the query projection path, a matrix multiplication operation is performed between the TCM macroscopic feature vector and the query projection weight matrix to encode the constitution state and syndrome tendency information into a macroscopic constitution query matrix for cross-modal retrieval. Each row of this matrix corresponds to a retrieval instruction in the constitution semantic dimension, used in subsequent steps to evaluate the correlation between the Western medicine microscopic fluctuation signals at each time step and the current constitution state. For the key projection path, a matrix multiplication operation is performed between the Western medicine microscopic feature vector and the key projection weight matrix to encode temporal fluctuation information such as sudden changes in blood glucose peaks and periodic changes in pulse into a microscopic fluctuation key matrix for relevance matching in the macroscopic constitution query. Each row of this matrix corresponds to a fluctuation feature signature at a time step, used to accept dot product addressing from the query matrix. For the value projection path, a matrix multiplication operation is performed between the same Western medicine microscopic feature vector and the value projection weight matrix. This re-encodes the original microscopic fluctuation information into a microscopic fluctuation value matrix that carries the actual physiological content. This matrix retains the effective information payload of blood glucose concentration changes and pulse morphology details at each time step, which are to be extracted with weights. The weight matrices of the three projection paths are independent of each other, so that the query space focuses on expressing the constitution retrieval intent, the key space focuses on providing temporal addressing identifiers, and the value space focuses on storing the physiological content to be aggregated. The three are decoupled at the functional level, laying a structured input foundation for the scaling dot product operation of the subsequent cross-modulation aggregation unit.

[0028] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the cross-modal aggregation unit 132 is used to perform cross-modal correlation assessment on the macro-constitution query matrix and the micro-fluctuation key matrix to generate an adaptive attention probability distribution. This adaptive attention probability distribution is then used to weight and aggregate the micro-fluctuation value matrix to obtain a cross-modal feature matrix. It should be noted that since the macro-constitution query matrix obtained after projection transformation encodes the retrieval instructions for the patient's constitution in each semantic dimension, while the micro-fluctuation key matrix encodes the feature signatures of blood glucose and pulse fluctuations at each time step, and these two belong to different semantic subspaces, quantitative correlation assessment is needed to determine which time steps of micro-fluctuations under the current constitution have higher reference value for metabolic management decisions. For example, patients in a state of yin deficiency and heat excess constitution are more sensitive to periods of rapid postprandial blood glucose rise than to periods of stable fluctuation. This constitution-dependent temporal attention preference cannot be achieved through fixed weight allocation and must be adaptively generated by dynamic dot product operations between the query matrix and the key matrix. Based on this, the technical solution of this application further evaluates the cross-modal correlation between the macroscopic physical condition query matrix and the microscopic fluctuation key matrix to generate an adaptive attention probability distribution. Then, this adaptive attention probability distribution is used to weighted aggregate the microscopic fluctuation value matrix to obtain a cross-modulation feature matrix. Through the above processing, macroscopic physical condition semantics can apply differentiated attention intensity to the microscopic fluctuation signals at each moment in the form of an attention probability distribution. This selectively amplifies fluctuation features highly correlated with the current physical condition while suppressing noise interference in low-correlation periods, while preserving all temporal information, thus generating a cross-modulation feature matrix that combines physical semantic directionality with temporal fluctuation details.

[0029] More specifically, in a concrete example of this application, matrix multiplication is first performed on the transpose of the macro-physicality query matrix and the micro-fluctuation key matrix to obtain a two-dimensional cross-modal relevance scoring matrix. Each element in this matrix represents the original dot product relevance score between a physiological semantic dimension and a micro-fluctuation feature at a time step. Since the original dot product value expands with increasing embedding dimension, causing the subsequent normalization function to enter the gradient saturation region, each element in the relevance scoring matrix is ​​divided by the arithmetic square root of the embedding dimension for scaling suppression, causing the numerical distribution to fall back to the effective gradient response range of the normalization function. Subsequently, a normalized exponential function is applied to the scaled relevance scoring matrix along the time step dimension of the key matrix, converting the original score of each row into a probability weight distribution whose sum is always equal to 1. This probability distribution is the adaptive attention probability distribution, where positions with higher probability values ​​correspond to time steps with stronger metabolic management reference significance under the current physiological state. Finally, matrix multiplication is performed on the adaptive attention probability distribution and the micro-fluctuation value matrix. The actual physiological content carried by each time step in the value matrix is ​​weighted and aggregated according to the above probability weights. This makes the blood glucose change amplitude and pulse morphology information at time steps that are highly correlated with the current physical condition receive a greater aggregate contribution, while the information contribution of stable fluctuation periods that are less correlated with the current physical condition is compressed accordingly. The final output is a cross-modulated feature matrix after being filtered and modulated by the macro-physical condition semantics of traditional Chinese medicine. This feature matrix is ​​used as the input of the subsequent fusion stabilization unit to participate in residual superposition and layer normalization processing.

[0030] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the fusion stabilization unit 133 is used to perform residual superposition of the cross-modulation feature matrix and the macro-feature vector of traditional Chinese medicine, and then perform data distribution stabilization processing through a layer normalization function to obtain a collaborative fusion mapping vector. It should be noted that since the cross-modulation feature matrix is ​​a purified information stream generated after attention-weighted aggregation, it only retains micro-fluctuation features highly correlated with the current physical condition. During the attention screening process, the baseline physical condition information at time steps assigned lower probability weights is compressed or even lost. If the cross-modulation feature matrix is ​​directly output as the fusion result, the complete physical condition information originally carried in the macro-feature vector of traditional Chinese medicine will not be able to play a global constraint role in subsequent metabolic trend prediction and syndrome evolution deduction. At the same time, the numerical cumulative offset generated by multi-layer attention operations will also cause the mean and variance of the output tensor to deviate from the stable range, affecting the gradient propagation efficiency of the downstream network layers. Based on this, the technical solution of this application further performs residual superposition of the cross-modulation feature matrix and the macro-feature vector of traditional Chinese medicine, and then performs data distribution stabilization processing through a layer normalization function to obtain a collaborative fusion mapping vector. Through the above processing, while preserving the constitution-sensitive fluctuation features extracted by cross-modal attention screening, the original TCM constitution baseline information can be seamlessly reinjected into the fusion result through residual path, and the superimposed numerical distribution can be corrected to a stable state with zero mean and unit variance through layer normalization. This provides a collaborative fusion mapping vector for the subsequent dual-path decoding network with controllable numerical distribution and containing both micro-fluctuation details and macro-constitutional overview.

[0031] More specifically, in a concrete example of this application, the cross-modulation feature matrix and the TCM macroscopic feature vector are first added element-wise at corresponding element positions to complete the residual superposition operation. Essentially, this operation establishes a skip connection path from the input to the output of the cross-modal attention network, allowing the tongue color and texture features encoded in the TCM macroscopic feature vector and the long-term constitution background information reflected in the body mass index score to bypass the information bottleneck of the attention operation and be directly superimposed into the fusion result. This avoids the gradual attenuation of the constitution baseline information in multi-layer operations due to the sparsity tendency of the attention probability distribution. After the residual superposition is completed, the arithmetic mean and sample variance of all values ​​along the feature dimension of the superimposed tensor are calculated to complete the layer normalization process. After this normalization operation, the numerical distribution of each feature dimension of the output tensor is corrected to a standardized state with zero mean and unit variance, eliminating the numerical scale drift introduced by the residual superposition. This ensures that the final output co-fusion mapping vector meets the input requirements of the subsequent metabolic trend prediction decoder and syndrome evolution assessment network in terms of both numerical stability and information integrity. At this point, the entire processing flow of the cross-modal attention fusion module is complete, and the output collaborative fusion mapping vector enters the subsequent health trend and syndrome evolution inference module.

[0032] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the health trend and syndrome evolution inference module 140 is used to predict the time-series extension of metabolic trends on the synergistic fusion mapping vector, and to infer the probability of syndrome state transition on the synergistic fusion mapping vector to obtain metabolic prediction inference vector and syndrome evolution inference vector. It should be noted that, given that the collaborative fusion mapping vector obtained after cross-modal attention fusion simultaneously contains details of micro-fluctuations in Western medicine and macro-constraints in traditional Chinese medicine, this vector only represents a snapshot of the patient's current combined Western and traditional Chinese medicine health status. It has not yet been extended and extrapolated along the time axis to the future, and cannot be directly used to predict whether acute hyperglycemia or hypoglycemia risk events will occur in the next 24 to 72 hours. Nor can it assess whether the continuous impact of high-frequency metabolic fluctuations on the patient's traditional Chinese medicine constitution will lead to the deterioration and migration of syndrome states. For example, the possibility that a patient with Qi deficiency constitution may transform into Qi and Yin deficiency after experiencing repeated and severe blood sugar fluctuations. These two types of prospective extrapolation results are precisely the decision-making basis necessary for the subsequent generation of digital closed-loop intervention strategies. Without metabolic trend prediction, short-term Western medicine intervention instructions cannot be triggered, and without syndrome evolution extrapolation, long-term conditioning strategies cannot be matched from the knowledge base of traditional Chinese medicine and food homology. Based on this, the technical solution of this application further performs metabolic trend time-series extension prediction on the collaborative fusion mapping vector, and performs syndrome state transition probability extrapolation on the collaborative fusion mapping vector to obtain metabolic prediction extrapolation vector and syndrome evolution extrapolation vector. Through the above processing, the predicted future glucose concentration sequence for the micro-temporal dimension and the probability distribution of syndrome transition for the macro-physical dimension can be output simultaneously based on the same collaborative fusion mapping vector. This allows the subsequent intervention strategy generation stage to trigger graded Western medicine intervention responses based on the risk information of cross-boundary in the metabolic prediction vector, and to retrieve and match TCM conditioning programs based on the transition probability information in the syndrome evolution inference vector. This provides a complete forward-looking inference basis for the collaborative generation of digital closed-loop intervention strategies through both TCM and Western medicine channels.

[0033] Figure 4 This is a block diagram of the health trend and syndrome evolution inference module in a digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, according to an embodiment of this application. Figure 4 As shown, the health trend and syndrome evolution deduction module 140 includes: a metabolic trend prediction unit 141, which is used to perform micro-time series autoregressive accurate trend prediction on the collaborative fusion mapping vector after initializing it to the decoder hidden state to obtain the metabolic prediction deduction vector; and a syndrome evolution evaluation unit 142, which is used to perform macro-physical syndrome transition probability nonlinear evaluation on the metabolic prediction deduction vector to obtain the syndrome evolution deduction vector.

[0034] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the metabolic trend prediction unit 141 is used to perform micro-temporal autoregressive precise trend prediction on the synergistic fusion mapping vector after initializing it to the decoder's hidden state to obtain a metabolic prediction inference vector. It should be noted that although the synergistic fusion mapping vector simultaneously encodes the details of micro-fluctuations in Western medicine and the macro-constitutional constraints in traditional Chinese medicine, it only represents the patient's current combined health status and has not yet been extended into the future along the time axis. The core requirement of closed-loop management of metabolic diseases is to predict the glucose concentration trend and extreme fluctuation risk within the next 24 to 72 hours in advance, so as to trigger an intervention response before acute hyperglycemia or hypoglycemia events actually occur. This requires using the synergistic fusion mapping vector as the starting point for inference and gradually expanding it along the temporal dimension to generate concentration estimates for each future time step. Based on this, the technical solution of this application further performs micro-temporal autoregressive precise trend prediction on the synergistic fusion mapping vector after initializing it to the decoder's hidden state to obtain a metabolic prediction inference vector. Through the above processing, the combined characteristics of traditional Chinese and Western medicine carried by the collaborative fusion mapping vector can be used as the global environmental foundation for future fluctuation inference. Through an autoregressive recursive mechanism, the glucose concentration prediction sequence at each discrete time step in the future can be gradually generated, providing continuous and temporally causal metabolic trend prediction data for subsequent over-threshold risk detection and graded intervention command triggering.

[0035] More specifically, in a concrete example of this application, the collaborative fusion mapping vector is directly assigned as the initial hidden state of the gated recurrent unit decoder, making the patient's physical background and current micro-fluctuation state contained in this vector the global memory starting point for the entire future time-series extrapolation process. After the decoder starts, a preset starting marker is used as the input for the first time step. The hidden state output for the first time step is generated through the collaborative operation of the reset gate and update gate inside the gated recurrent unit. This hidden state is then mapped through a fully connected layer and constrained to a non-negative value by a modified linear unit activation function to obtain the estimated glucose concentration for the first future time step. The autoregressive rolling decoding stage then begins, where the predicted output value of the current time step and the current hidden state are used together as input for the next time step. The prediction results for subsequent time steps are generated progressively according to the following recursive relationship: At the τ-th future time step, the predicted output value of the previous time step and the hidden state of the previous time step are fed into the gate control loop unit for state update calculation. Then, the updated hidden state is multiplied by the weight matrix of the fully connected layer, and a bias term is added. After processing by the modified linear unit activation function, the glucose concentration prediction value for the current time step is output. This recursive process continues along all discrete time steps within the future time window until the preset look-ahead prediction interval is covered. The glucose concentration prediction values ​​generated at all time steps are assembled into a one-dimensional data sequence in chronological order. This sequence is the metabolic prediction extrapolation vector, which contains the concentration prediction values ​​for each future time moment and their implied fluctuation trend information, for use by the subsequent syndrome evolution assessment unit and the digital closed-loop intervention strategy generation module.

[0036] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the syndrome evolution assessment unit 142 is used to perform a nonlinear assessment of the probability of macroscopic constitution syndrome transitions on the metabolic prediction vector to obtain the syndrome evolution vector. It should be noted that since the metabolic prediction vector contains the predicted glucose concentration and its fluctuation trend information at each future time step, and in the scenario of integrated traditional Chinese and Western medicine management of metabolic diseases, continuous microscopic metabolic fluctuations can have a cumulative impact on the patient's macroscopic TCM constitution. For example, repeated postprandial hyperglycemia events can accelerate the deterioration and migration of patients with Qi deficiency to Qi and Yin deficiency. This macroscopic syndrome transition process driven by microscopic fluctuations has nonlinear characteristics; the transition probability between different syndrome types and the fluctuation amplitude are not a simple linear proportional relationship, requiring a nonlinear mapping network for accurate capture. Based on this, the technical solution of this application further performs a nonlinear assessment of the probability of macroscopic constitution syndrome transitions on the metabolic prediction vector to obtain the syndrome evolution vector. Through the above processing, the information on the intensity of micro-metabolic fluctuations within the future time window can be transformed into the transition probability distribution between various TCM syndrome types, providing a probabilistic basis for syndrome evolution in the subsequent digital closed-loop intervention strategy generation module for matching TCM conditioning strategies and dynamically modulating syndrome sensitivity.

[0037] More specifically, in a concrete example of this application, a global average pooling operation is first applied to the metabolic prediction vector along the time step dimension. This compresses the glucose concentration predictions at each discrete time step into a concentrated feature scalar reflecting the overall intensity of metabolic load oscillations. This pooling operation eliminates redundant information in the time step dimension, retaining only the overall intensity characteristics of future metabolic fluctuations. Subsequently, this concentrated feature scalar is concatenated with the collaborative fusion mapping vector transmitted from the preceding metabolic trend prediction unit along the feature channel dimension to form a joint input vector that simultaneously carries the intensity of future micro-fluctuations and the current combined TCM and Western medicine constitution. This joint input vector is fed into a nonlinear syndrome inference network consisting of two fully connected layers. In the first hidden layer, matrix multiplication is performed between the joint input vector and the first layer weight matrix, and a first layer bias term is superimposed. After processing by a nonlinear activation function, the high-order interaction features between metabolic fluctuation intensity and constitution are captured. Then, the activated hidden layer output is multiplied with the second layer weight matrix, and a second layer bias term is superimposed to obtain the original score vectors in each syndrome type dimension. Finally, a normalized exponential function is applied to the original score vector, converting the scores on each syndrome dimension into a probability distribution whose sum is always equal to 1. Each probability value represents the likelihood that the patient will transition to the corresponding syndrome type under future metabolic fluctuations. For example, a higher probability value on the Yin deficiency and heat excess dimension indicates that future blood glucose fluctuations will accelerate the patient's deterioration towards Yin deficiency and heat excess. This probability distribution is the syndrome evolution projection vector, which, together with the metabolic prediction projection vector, serves as the final output of the health trend and syndrome evolution projection module, and is then fed into the subsequent digital closed-loop intervention strategy generation module.

[0038] Specifically, the tongue image semantic segmentation network in the multi-source data preprocessing and digitization module, the temporal convolutional encoder and multi-layer feedforward network in the multi-dimensional feature encoding and extraction module, the three sets of linear projection parameter matrices in the cross-modal attention fusion module, and the gated recurrent unit decoder and nonlinear syndrome inference network in the health trend and syndrome evolution inference module are all neural network components containing learnable parameters. These components require end-to-end joint training for parameter optimization before deployment. During the training phase, multi-source heterogeneous raw data from historical patients are used as input samples, and the corresponding future actual glucose concentration observation sequences and syndrome classification labels annotated by clinical experts are used as supervision signals to construct a multi-task joint loss function. This joint loss function consists of a weighted sum of two parts: the first part is the metabolic trend prediction loss, which uses the mean squared error function to measure the deviation between the predicted value and the actual observed value at each time step in the metabolic prediction inference vector; the second part is the syndrome classification loss, which uses the cross-entropy function to measure the distribution distance between the predicted probability distribution of each dimension in the syndrome evolution inference vector and the actual syndrome labels annotated by experts. The gradient of the joint loss function is propagated back through each layer of the network using a backpropagation algorithm. An adaptive moment estimation optimizer synchronously updates all learnable parameters, enabling the temporal convolutional encoder to extract the temporal fluctuation features most relevant to metabolic trend prediction. A multi-layer feedforward network learns to capture the higher-order constitution correlation between tongue appearance and questionnaire data. The projection parameter matrix in the cross-modal attention fusion module maps the macroscopic constitution semantics of Traditional Chinese Medicine (TCM) and the microscopic fluctuation features of Western medicine to the most favorable interaction subspace for joint prediction. The gated recurrent unit decoder and the syndrome inference network generate accurate metabolic trend predictions and syndrome evolution inference results based on the fused features, respectively. After the training process iterates until the joint loss function converges on the validation set, all learnable parameters are solidified and deployed to the inference environment. The Western medicine medication rule base, the TCM food and medicine homology knowledge base, and the pharmacological mutual exclusion rule graph in the subsequent digital closed-loop intervention strategy generation module are all pre-built static knowledge resources and do not participate in the above training process.

[0039] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the digital closed-loop intervention strategy generation module 150 is used to generate digital closed-loop intervention strategies for metabolic prediction vectors and syndrome evolution prediction vectors based on a preset Western medicine drug rule base and a traditional Chinese medicine food homology and compatibility knowledge base, so as to obtain a synergistic intervention plan of integrated traditional Chinese and Western medicine. It should be noted that, given that the metabolic prediction vector output by the health trend and syndrome evolution deduction module carries the estimated value and fluctuation trend information of glucose concentration at each discrete time step within the next 24 to 72 hours, while the syndrome evolution deduction vector carries the probability distribution of patients transitioning to various TCM syndrome types under the impact of future metabolic fluctuations, both provide forward-looking deduction basis from the micro-temporal dimension and the macro-constitutional dimension, respectively, but have not yet been transformed into specific intervention instructions that can be implemented clinically. In the actual management scenario of metabolic diseases, Western medicine needs to trigger a step-by-step intervention response from dietary fine-tuning to rapid-acting insulin injection according to the urgency of the risk of blood glucose exceeding the limit, while TCM needs to match a long-term conditioning plan that is suitable for the current direction of constitution migration from the knowledge base of medicine and food homology according to the syndrome evolution trend. Moreover, there may be pharmacological conflicts between the two types of intervention strategies. For example, Western medicine instructions require the supplementation of carbohydrates to deal with the risk of hypoglycemia, while TCM prescriptions contain herbs with hypoglycemic effects. If they are directly combined without safety verification, contradictory intervention instructions will be generated. Based on this, the technical solution of this application further utilizes a pre-defined Western medicine medication rule base and a traditional Chinese medicine food homology compatibility knowledge base to generate a digital closed-loop intervention strategy for metabolic prediction and syndrome evolution prediction vectors, thereby obtaining a synergistic intervention plan combining traditional Chinese and Western medicine. Through the above processing, microscopic metabolic trend prediction and macroscopic syndrome evolution prediction can be simultaneously transformed into a synergistic plan encompassing short-term Western medicine intervention instructions and long-term traditional Chinese medicine conditioning strategies. Furthermore, pharmacological mutual exclusion safety verification ensures the compatibility of the two types of strategies at the pharmacological level, thus achieving a full-link digital management closed loop for synergistic traditional Chinese and Western medicine, from multi-source data acquisition, cross-modal feature fusion, joint prediction and deduction to closed-loop intervention output.

[0040] Figure 5 This is a block diagram of the digital closed-loop intervention strategy generation module in a digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, according to an embodiment of this application. Figure 5As shown, the digital closed-loop intervention strategy generation module 150 includes: a Western medicine intervention instruction generation unit 151, used to perform dynamic weighted risk assessment of the hyperglycemia and hypoglycemia out-of-bounds deviation of the metabolic prediction deduction vector to determine whether to generate a short-acting Western medicine intervention instruction; a Traditional Chinese Medicine conditioning strategy generation unit 152, used to use the syndrome evolution deduction vector as a high-dimensional probability retrieval vector and perform matching degree optimization calculation with all candidate strategy templates in the Traditional Chinese Medicine and Food Homologous Compatibility Knowledge Base, extracting the medicinal diet formula and acupoint meridian intervention scheme corresponding to the highest matching degree index to obtain a long-term Traditional Chinese Medicine conditioning strategy; and a collaborative scheme encapsulation unit 153, used to perform pharmacological mutual exclusion security verification on the short-acting Western medicine intervention instruction and the long-term Traditional Chinese Medicine conditioning strategy, and to structure and encapsulate the valid strategies that pass the verification through a serialization protocol to generate a collaborative intervention scheme of Traditional Chinese and Western medicine.

[0041] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the Western medicine intervention instruction generation unit 151 is used to perform dynamic weighted risk assessment of the deviations of hyperglycemia and hypoglycemia from the metabolic prediction and extrapolation vector to determine whether to generate a short-term Western medicine intervention instruction. It should be noted that, given that the metabolic prediction and extrapolation vector contains the estimated glucose concentration at each discrete time step within the future time window, and patients with different TCM syndrome states exhibit different physiological vulnerabilities and organ damage sensitivities to the same magnitude of blood glucose fluctuations, if a globally static penalty weight and a fixed warning threshold are used to perform an indiscriminate risk assessment on all patients, the syndrome evolution extrapolation vector will only be transmitted without participating in risk calculation, leading to a break in the TCM-Western medicine collaboration link at the decision-making stage. Based on this, the technical solution of this application further performs dynamic weighted risk assessment on the deviations of hyperglycemia and hypoglycemia from the metabolic prediction vector to determine whether to generate a short-acting Western medicine intervention instruction. This risk assessment process introduces a dynamic modulation mechanism of syndrome sensitivity. Through a triple coupling mechanism of expected weighting of syndrome sensitivity probability, adaptive contraction of threshold driven by syndrome entropy, and nonlinear pathological amplification of syndrome perception, the risk penalty weight and the boundary of the warning threshold are dynamically adjusted in real time according to the patient's current syndrome probability distribution. Through the above processing, patients with different syndrome states can obtain differentiated risk assessment results and graded step-by-step intervention responses, realizing closed-loop dynamic modulation of Western medicine risk penalty parameters by the probability distribution of traditional Chinese medicine syndromes.

[0042] Figure 6 This is a block diagram of the traditional Chinese and Western medicine intervention instruction generation unit in a digital management system for metabolic diseases based on multi-source data fusion of traditional Chinese and Western medicine, according to a preferred embodiment of this application. Figure 6As shown, the Western medicine intervention instruction generation unit 151 includes: a threshold offset generation subunit 1511, used to generate a syndrome adaptive threshold offset based on the syndrome evolution inference vector; a dynamic early warning threshold adjustment subunit 1512, used to adjust the preset static high blood sugar early warning threshold and static low blood sugar early warning threshold inward based on the syndrome adaptive threshold offset to obtain the dynamic early warning threshold boundary; a coupling risk assessment subunit 1513, used to perform syndrome coupling dynamic risk assessment on each discrete predicted value of the metabolic prediction inference vector in the future time window based on the dynamic early warning threshold boundary to obtain the coupling risk assessment coefficient; and an intervention instruction generation subunit 1514, used to perform a down-rounding hierarchical operation on the coupling risk assessment coefficient using a preset level division granularity constant to obtain the intervention instruction urgency level number, and to perform a step-by-step instruction matching and retrieval from the hierarchical Western medicine prescription drug use rule map based on the intervention instruction urgency level number to generate a short-acting Western medicine intervention instruction.

[0043] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the threshold offset generation subunit 1511 is used to generate an adaptive threshold offset based on the syndrome evolution inference vector. It should be noted that, given the fundamental differences in the pathological sensitivity of different TCM syndrome types to hyperglycemic and hypoglycemic events—Yin deficiency with excessive heat syndrome being extremely sensitive to hyperglycemia while phlegm-dampness syndrome is more vulnerable to hypoglycemia—adjusting the penalty weight solely through the syndrome sensitivity modulation factor is insufficient to fully reflect the constraint effect of the syndrome state on the risk trigger boundary itself. When a patient's syndrome probability distribution clearly concentrates in a specific dangerous syndrome direction, it indicates that their constitution is deteriorating towards that dangerous state. At this time, not only is it necessary to increase the penalty intensity, but also to intercept in advance with a stricter warning threshold. However, the original static hyperglycemia warning threshold and static hypoglycemia warning threshold are globally fixed constants and cannot adaptively shrink according to the clarity of the syndrome evolution trend. Based on this, the technical solution of this application further generates an adaptive threshold offset based on the syndrome evolution inference vector. Through the above processing, the warning threshold boundary can be dynamically narrowed according to the concentration of the syndrome probability distribution, so that patients with high-risk syndromes can be warned in advance before their blood sugar reaches the original static threshold, while patients with relatively uniform syndrome distribution can maintain the original loose threshold to avoid clinical interference caused by excessive warning.

[0044] More specifically, in a concrete example of this application, the information entropy of the transition probabilities of each syndrome dimension in the syndrome evolution inference vector is first measured to quantify the degree of uncertainty in the current syndrome probability distribution. The information entropy value of the TCM syndrome evolution inference vector is calculated as follows: in, The information entropy value is the vector used to infer the evolution of TCM syndromes, and is used to measure the degree of uncertainty in the current syndrome probability distribution. This represents the total number of dimensions for classifying syndromes in Traditional Chinese Medicine. The first in the vector for inferring the evolution of syndromes in Traditional Chinese Medicine The transition probability value of syndrome type, To prevent the logarithmic function from having a singular zero point and a minimal positive real constant, the information entropy value approaches 0 when the syndrome probability distribution is highly concentrated in a certain dimension, indicating that the patient is clearly deteriorating towards a specific syndrome. When the syndrome probability distribution is relatively uniform, the information entropy value is higher, indicating that the patient's physical condition has not yet shown a clear tendency to deteriorate. Subsequently, the syndrome adaptive threshold offset is calculated based on this information entropy value: in, For the adaptive threshold offset of the syndrome, This is the maximum allowable amplitude reference constant for threshold offset. To control the positive real hyperparameter of entropy decay sensitivity, This refers to the syndrome information entropy value calculated above. When the patient's syndrome probability distribution is highly concentrated on a certain dangerous syndrome, the information entropy... Approaching 0, exponentially decaying term Approaching 1, the adaptive threshold offset of the syndrome Approaching maximum offset Subsequently, the high-sugar warning threshold will be compressed downwards, and the low-sugar warning threshold will be raised upwards, forming a more stringent warning window; conversely, when the probability distribution of symptoms is relatively uniform, the information entropy... The value is relatively high, the exponential decay term approaches 0, and the adaptive threshold offset of the syndrome is relatively small. Approaching zero, the warning threshold remains in its original, lenient state. Taking a patient whose syndrome probability distribution is highly concentrated in Qi and Yin deficiency as an example, due to the coexistence of insufficient vital energy and depletion of Yin fluid in the Qi and Yin deficiency syndrome state, the patient exhibits amplified pathological response characteristics to fluctuations in blood glucose at both high and low ends. At this time, the information entropy value is close to 0, and the syndrome adaptive threshold offset is close to the maximum allowable range. This causes the original static high glucose warning threshold and low glucose warning threshold to be inwardly contracted respectively. The warning assessment process is triggered when the blood glucose offset is still within the safe range for ordinary patients, thereby providing a more sufficient intervention response time window for such high-risk patients.

[0045] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the dynamic early warning threshold adjustment subunit 1512 is used to adjust the preset static high-glucose early warning threshold and static low-glucose early warning threshold inward based on the syndrome adaptive threshold offset to obtain the dynamic early warning threshold boundary. It should be noted that, given that the preceding subunit has already calculated the syndrome adaptive threshold offset based on the information entropy of the syndrome evolution inference vector, this offset reflects the concentration and severity of the current patient's syndrome probability distribution, but has not yet been applied to the actual early warning discrimination boundary. Furthermore, the preset static high-glucose and static low-glucose early warning thresholds are fixed constants set for the entire patient population, failing to distinguish the differentiated needs of high-risk syndrome patients and ordinary patients in terms of early warning sensitivity. Therefore, it is necessary to use the syndrome adaptive threshold offset to inwardly shrink the static thresholds to form a personalized dynamic early warning window. Based on this, the technical solution of this application further adjusts the preset static high-glucose and static low-glucose early warning thresholds inward based on the syndrome adaptive threshold offset to obtain the dynamic early warning threshold boundary. Through the above processing, patients with a clear trend of worsening symptoms can obtain a narrower warning window, thus being intercepted in advance when the blood glucose deviation is small, while patients with uniform symptom distribution can maintain a warning window close to the original width to avoid excessive warnings interfering with the clinical workflow.

[0046] More specifically, in a specific example of this application, the preset static high blood sugar warning threshold is subtracted from the syndrome adaptive threshold offset to obtain the dynamic high blood sugar warning threshold after inward contraction: in, The dynamic high blood sugar warning threshold is set. The preset static high sugar warning threshold, The symptom-adaptive threshold offset is calculated for the preceding subunit. Simultaneously, the preset static low-glycemia warning threshold is added to the symptom-adaptive threshold offset to obtain the dynamic low-glycemia warning threshold after inward contraction: in, The dynamic low sugar warning threshold is set. The preset static low sugar warning threshold, This refers to the adaptive threshold offset for the syndrome. The two dynamic thresholds mentioned above together constitute the dynamic warning threshold boundary. The essence of this inward contraction operation is to compress the high-sugar threshold downward and raise the low-sugar threshold upward, causing the upper and lower boundaries of the safe interval to move closer to the center simultaneously, forming a more stringent warning window than the original static threshold. The contraction magnitude is entirely controlled by the adaptive threshold offset for the syndrome, which is determined by the information entropy of the syndrome probability distribution through exponential decay mapping. Therefore, a monotonically increasing correspondence is formed between the degree of threshold contraction and the clarity of syndrome deterioration. Taking a type 2 diabetic patient with a syndrome probability distribution highly concentrated in Yin deficiency and heat excess as an example, since the syndrome information entropy is close to 0, the syndrome adaptive threshold offset is close to the maximum allowable amplitude benchmark constant. Assuming the static high blood sugar warning threshold is 13.9 mmol / L, the static low blood sugar warning threshold is 3.9 mmol / L, and the maximum allowable amplitude benchmark constant is 1.5 mmol / L, then after inward contraction, the dynamic high blood sugar warning threshold drops to about 12.4 mmol / L, and the dynamic low blood sugar warning threshold rises to about 5.4 mmol / L. The warning window shrinks from the original 10.0 mmol / L width to about 7.0 mmol / L, so that this patient with Yin deficiency and heat excess can enter the high blood sugar risk assessment process as soon as the blood sugar exceeds 12.4 mmol / L, without having to wait until the global static threshold of 13.9 mmol / L is reached. This provides a personalized discrimination boundary that matches the patient's current physical vulnerability for the subsequent syndrome-coupled dynamic risk assessment.

[0047] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the coupled risk assessment subunit 1513 is used to perform syndrome coupling dynamic risk assessment on the discrete predicted values ​​of the metabolic prediction vector within a future time window based on the dynamic early warning threshold boundary, so as to obtain the coupled risk assessment coefficient. It should be noted that, given that the preceding subunit has generated the high-glucose syndrome sensitivity modulation factor, the low-glucose syndrome sensitivity modulation factor, and the dynamic early warning threshold boundary, and the three syndrome modulation parameters are ready, they have not yet been injected into the risk penalty formula to replace the original global static constant. The formula of the existing method uses a linear penalty mode for the out-of-bounds deviation, which cannot simulate the superlinear pathological amplification effect of the fragility of the TCM constitution on large out-of-bounds events. In clinical practice, the organ damage risk caused by blood glucose exceeding the threshold of 2 mmol / L in patients with yin deficiency and heat excess is far more than twice that of blood glucose exceeding 1 mmol / L, but shows a superlinear growth characteristic, which the linear penalty mode cannot capture. Based on this, the technical solution of this application further uses a dynamic early warning threshold boundary to perform a dynamic risk assessment of the discrete predicted values ​​of the metabolic prediction vector within a future time window to obtain a coupled risk assessment coefficient. Through the above processing, the triple coupling mechanism of the syndrome sensitivity modulation factor, the syndrome adaptive threshold offset, and the syndrome perception nonlinear amplification function can be simultaneously injected into the risk penalty formula, so that the risk assessment results can reflect the differentiated pathological response characteristics of patients with different constitutions to the same blood glucose fluctuation amplitude.

[0048] More specifically, in a concrete example of this application, the metabolic prediction extrapolation vector is iterated through each discrete predicted value within a future time window. The high and low glucose warning boundaries are dynamically shrunk using a syndrome adaptive threshold offset. Furthermore, the high glucose syndrome sensitivity modulation factor and the low glucose syndrome sensitivity modulation factor are used to nonlinearly amplify the out-of-bounds deviation in syndrome perception, and the coupling risk assessment coefficient is calculated. in, For coupling risk assessment coefficients, The set of discrete time steps covered by a pre-defined future time window. As a modulatory factor for sensitivity to hyperglycemia symptoms, As a modulatory factor for sensitivity to hypoglycemia syndrome, and These are the basic penalty weight constants for high-sugar and low-sugar events, respectively. For the future generation of the metabolic prediction vector Estimated glucose concentration at time t, and These are the original static high-sugar and low-sugar warning thresholds, respectively. For the adaptive threshold offset of the syndrome, This is the nonlinear amplification function for symptom perception. The truncation function in this formula... This ensures that a non-zero out-of-bounds deviation only occurs when the predicted value exceeds the dynamic warning boundary after the symptom-adaptive threshold offset has been contracted; normal fluctuations within the dynamic safety range do not contribute to the risk score. The symptom-perception nonlinear amplification function in this formula... A superlinear pathological amplification effect specifically designed to simulate the vulnerability of traditional Chinese medicine constitutions on transgressive deviations is defined as follows: in, The non-negative out-of-bounds deviation value after threshold truncation. This is a second-order nonlinear amplification factor, controlling the intensity of the superlinear penalty imposed by the vulnerability of the syndrome on large out-of-bounds events. When the out-of-bounds deviation... When smaller, the quadratic term The contribution is negligible, the risk assessment is approximately linear, and it is suitable for scenarios with slight fluctuations; however, when the deviation exceeds the limit... When the value is large, the quadratic term dominates the penalty intensity, causing the risk coefficient to increase superlinearly, simulating the clinicopathological characteristics of a sharp increase in organ damage risk in high-risk patients when blood glucose levels significantly exceed the limit. Specifically, when... and and At this point, the reconstructed formula completely degenerates into the original static threshold linear penalty formula, ensuring backward compatibility. Taking a patient whose syndrome probability distribution is highly concentrated in Yin deficiency and heat excess as an example, their hyperglycemia syndrome sensitivity modulation factor... Much greater than the hypoglycemic syndrome sensitivity modulator factor Simultaneously, the adaptive threshold offset is a symptom. Approaching the maximum permissible range compresses the dynamic hyperglycemia warning threshold downwards. Assuming the patient's predicted glucose concentration at a future time step is 14.5 mmol / L, and the hyperglycemia warning boundary after dynamic threshold contraction is 12.4 mmol / L, then the deviation from the boundary is 2.1 mmol / L. After processing with a nonlinear amplification function, the actual penalty contribution is... Compared to version 2.1 in linear mode, it has increased... The superlinear penalty increment, then modulated by the hyperglycemic syndrome sensitivity modulator. The amplification of the risk coefficient will result in a coupling risk assessment coefficient that is much higher than the risk coefficient calculated for ordinary patients with uniform syndrome distribution under the same blood glucose prediction. This provides a continuous risk quantification basis for subsequent hierarchical intervention instruction mapping that is precisely matched with the patient's syndrome vulnerability.

[0049] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the intervention instruction generation subunit 1514 is used to perform a floor function on the coupling risk assessment coefficient using a preset grading granularity constant to obtain the intervention instruction urgency level number. Based on the intervention instruction urgency level number, a step-by-step instruction matching and retrieval is performed from the graded Western medicine prescription and medication rule map to generate a short-acting Western medicine intervention instruction. It should be noted that, given that the preceding subunit has already calculated the coupling risk assessment coefficient, which is a continuous positive real value, its magnitude reflects the current patient's comprehensive acute metabolic risk level under syndrome coupling modulation. However, continuous values ​​cannot be directly mapped to clinically executable intervention instructions. Furthermore, there is a fundamental difference in the intervention intensity required between slight and severe blood glucose exceedances. The former may only require minor dietary adjustments, while the latter requires immediate injection of rapid-acting insulin. If only judging whether the coupling risk assessment coefficient is greater than 0 triggers a single-level intervention response, this binary triggering logic is too crude in clinical scenarios and cannot distinguish the differentiated intervention measures that should be taken under different risk severity levels. Based on this, the technical solution of this application further utilizes a preset grading granularity constant to perform a floor function on the coupled risk assessment coefficient to obtain an intervention instruction urgency level number. Based on this intervention instruction urgency level number, a tiered instruction matching and retrieval is performed from the tiered Western medicine prescription drug use rule map to generate a short-acting Western medicine intervention instruction. Through the above processing, continuous risk values ​​can be transformed into discrete, gradient-based urgency levels, creating a refined tiered correspondence between clinical intervention intensity and risk severity, avoiding over-intervention in mildly fluctuating scenarios and under-intervention in severely fluctuating scenarios.

[0050] More specifically, in a particular example of this application, the coupling risk assessment coefficient is divided by a preset level division granularity constant, then rounded down and incremented by 1 to obtain the intervention instruction urgency level number: in, This is a positive integer representing the urgency level of the intervention instruction; a higher number indicates a greater degree of urgency. The coupling risk assessment coefficients are calculated for the preceding subunits. This is a preset granularity constant for classifying risk levels, i.e., the width of the risk coefficient range corresponding to each level. This is a floor function. The grading operation evenly divides the continuous risk coefficient range into several equally wide intervals, each corresponding to an integer grade number. The larger the coupled risk assessment coefficient, the higher the interval it falls into, and the larger the corresponding intervention urgency grade number. Based on the obtained intervention instruction urgency grade number, an intervention plan matching the urgency grade is precisely retrieved from a pre-constructed grading Western medicine prescription medication rule graph. This rule graph can be constructed based on a graph database (such as Neo4j), where node entities include urgency grade, drug category, specific drug name, and contraindication attributes, and edge relationships include grade matching, pharmacological mutual exclusion, and hierarchical progression. Grade 1 corresponds to dietary adjustments, Grade 2 corresponds to oral hypoglycemic drug dosage adjustments, and Grade 3 corresponds to rapid-acting insulin injection instructions. The retrieved intervention plans are encapsulated into short-acting Western medicine intervention instructions using a serialization protocol. Taking a patient whose syndrome probability distribution is highly concentrated in Yin deficiency and heat excess as an example, because both the hyperglycemia syndrome sensitivity modulation factor and the syndrome adaptive threshold offset are at a high level, the coupling risk assessment coefficient obtained after syndrome coupling dynamic risk assessment is 5.8. Assuming the grading granularity constant is 2.0, the result of rounding down is... The intervention instruction urgency level is 3. The rapid-acting insulin injection instruction corresponding to level 3 is retrieved from the hierarchical Western medicine prescription and medication rule map and packaged into a short-acting Western medicine intervention instruction. For a typical patient with a relatively uniform syndrome distribution, under the same blood glucose prediction, because the syndrome modulation parameters are close to the baseline level, the coupling risk assessment coefficient is only 1.6, and the rounded-down result is... The intervention instruction urgency level was 1, and only the dietary fine-tuning suggestions corresponding to level 1 were retrieved. The two patients received differentiated intervention response intensities that matched their respective syndrome vulnerabilities under the same blood glucose prediction.

[0051] Specifically, the generated short-term Western medicine intervention instructions and integrated Chinese and Western medicine intervention plans are pushed to the responsible physician's workstation as auxiliary treatment suggestions. They can only be implemented as formal prescriptions after being reviewed and confirmed by a licensed physician.

[0052] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the TCM conditioning strategy generation unit 152 is used to perform matching degree optimization calculations on the syndrome evolution inference vector as a high-dimensional probability retrieval vector and all candidate strategy templates in the TCM food homology knowledge base, and extract the medicinal diet formula and acupoint meridian intervention scheme corresponding to the highest matching degree index to obtain a long-term TCM conditioning strategy. It should be noted that while the preceding Western medicine intervention instruction generation unit has already completed the hierarchical triggering and encapsulation of short-term Western medicine intervention instructions based on the coupling risk assessment coefficient, the integrated management of traditional Chinese and Western medicine for metabolic diseases requires not only immediate intervention against acute blood glucose exceeding limits but also long-term conditioning and correction of the patient's deteriorating physical condition. The syndrome evolution projection vector carries the probability distribution information of the patient's transition to various TCM syndrome types under future metabolic fluctuations. This probability distribution directly reflects the direction of the patient's physical condition's migration and deterioration tendency. It needs to be semantically matched with the conditioning strategy templates pre-stored in the TCM food-medicine homology knowledge base to extract the long-term conditioning plan most suitable for the current syndrome evolution trend. Based on this, the technical solution of this application further uses the syndrome evolution projection vector as a high-dimensional probability retrieval vector and performs matching degree optimization calculations with all candidate strategy templates in the TCM food-medicine homology knowledge base, extracting the medicinal diet formula and acupoint meridian intervention plan corresponding to the highest matching degree index to obtain a long-term TCM conditioning strategy. Through the above processing, the most suitable combination of medicine and food and the acupoint and meridian guidance strategy that best match the patient's current symptom evolution direction can be accurately retrieved from the knowledge base based on the probability distribution of the patient's current symptom evolution. This provides a basis for TCM conditioning for long-term constitution reshaping for the subsequent encapsulation and merging of TCM and Western medicine collaborative solutions.

[0053] More specifically, in a concrete example of this application, the syndrome evolution projection vector is used as a high-dimensional probability retrieval vector. A weighted inner product operation is performed on each of the pre-stored candidate strategy templates in the TCM food homology knowledge base to find the treatment plan with the highest matching degree to the current syndrome evolution trend. Each candidate strategy template in this knowledge base is pre-labeled with its treatment contribution weight for each syndrome type, jointly determined by clinical evidence-based medicine data and TCM pharmacology expert consensus. The matching degree optimization calculation process involves traversing all candidate strategy templates in the knowledge base, calculating the weighted inner product sum of the transition probabilities of each dimension in the syndrome evolution projection vector and the corresponding dimension treatment contribution weight of each template for each candidate template, and taking the candidate template index with the largest sum as the optimal matching result. in, This is the index number of the optimal candidate strategy template. The first in the Traditional Chinese Medicine and Food Homologous Compatibility Knowledge Base Traversal index of candidate strategy templates, This represents the total number of dimensions for classifying syndromes in Traditional Chinese Medicine. The first vector in the symptom evolution deduction vector The transition probability value of syndrome type, For the first Set of candidate strategy templates for the first The weighted inner product operation essentially uses the syndrome transition probability as the demand-side weight and the treatment contribution weight of each candidate template as the supply-side response. It calculates the comprehensive suitability of each candidate template to the current patient's syndrome evolution trend through probability expectation aggregation. Syndrome dimensions with higher transition probabilities receive greater contribution weights in the matching calculation, ensuring that the final selected treatment plan prioritizes corrective measures for the syndrome direction most likely to worsen. After obtaining the optimal matching index, the complete treatment plan entity corresponding to that index is extracted from the knowledge base, including the specific names, dosage ratios, and treatment cycles of Chinese medicinal herbs in the medicinal diet formula, as well as the acupoint locations, stimulation methods, and treatment courses in the acupoint and meridian intervention plan. This content is then encapsulated into a long-term TCM treatment strategy. Taking a patient whose syndrome evolution inference vector has a transition probability of 0.65 for the Yin deficiency and heat excess dimension, a transition probability of 0.20 for the Qi and Yin deficiency dimension, and low probabilities for other dimensions as an example, the high probability value of the Yin deficiency and heat excess dimension will dominate the value trend of the internal accumulation summation during the matching calculation process. This will result in the highest matching score for the medicinal diet formula with the main function of nourishing Yin and clearing heat, which is specifically designed for Yin deficiency and heat excess syndrome in the knowledge base, and the corresponding acupoint and meridian scheme. Ultimately, this formula will be extracted as the patient's long-term TCM conditioning strategy.

[0054] In the aforementioned digital management system 100 for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, the collaborative scheme encapsulation unit 153 is used to perform pharmacological mutual exclusion safety verification on short-acting Western medicine intervention instructions and long-term traditional Chinese medicine conditioning strategies. The verified legitimate strategies are then structured, encapsulated, and merged using a serialization protocol to generate a collaborative intervention scheme between traditional Chinese and Western medicine. It should be noted that since the two preceding parallel units have already generated short-acting Western medicine intervention instructions and long-term traditional Chinese medicine conditioning strategies, which respectively address immediate prevention of acute metabolic risks and long-term correction of chronic physical deterioration, in actual metabolic disease management scenarios, there may be pharmacological mutual exclusion conflicts between Western medicine medications and traditional Chinese medicine formulations. For example, a Western medicine intervention instruction may require carbohydrate supplementation to address the risk of hypoglycemia, while a traditional Chinese medicine conditioning strategy may include berberine-like herbs with hypoglycemic effects; or a Western medicine instruction may require the injection of rapid-acting insulin, while a traditional Chinese medicine formulation may contain medicinal and food ingredients that enhance insulin sensitivity, leading to a cumulative risk of hypoglycemia. If the two types of strategies are directly merged and issued without safety verification, contradictory intervention instructions will be generated, potentially causing medical harm to the patient. Based on this, the technical solution of this application further performs pharmacological mutual exclusion safety verification on short-acting Western medicine intervention instructions and long-term traditional Chinese medicine conditioning strategies. The verified legitimate strategies are then structured, encapsulated, and merged using a serialization protocol to generate a synergistic intervention plan combining traditional Chinese and Western medicine. Through this process, pharmacological contraindications and conflicts can be eliminated before merging the dual-channel strategies, ensuring that the final output synergistic intervention plan meets clinical safety requirements in terms of pharmacological compatibility. This completes the entire data flow closure from multi-source data acquisition, cross-modal feature fusion, joint prediction and inference to closed-loop intervention output.

[0055] More specifically, in a concrete example of this application, the active pharmaceutical ingredient identifiers in short-acting Western medicine intervention instructions and the herbal ingredient identifiers in long-term TCM conditioning strategies are first extracted and parsed into standardized component coding sequences. Then, these two component coding sequences are fed into a pre-constructed pharmacological mutual exclusion rule map for pairwise cross-comparison. This pharmacological mutual exclusion rule map stores all known incompatible relationships between Western medicine active pharmaceutical ingredients and herbal ingredients in the form of a two-dimensional Boolean matrix. Each element in the matrix marks whether there is a pharmacokinetic antagonistic reaction or toxicity superposition risk between the corresponding pair of Western medicine ingredients and herbal ingredients. All cross-combinations of the two component coding sequences are iterated. If any pair of ingredients triggers a contraindication flag in the pharmacological mutual exclusion rule map, the specific entry containing the conflicting ingredient in the long-term TCM conditioning strategy is marked as invalid and removed. Simultaneously, a suboptimal alternative solution is retrieved from the TCM food-medicine homology knowledge base for backfilling, until all cross-combinations pass the security check. After verification, the verified short-term Western medicine intervention instructions and legitimate long-term TCM conditioning strategies are structurally encapsulated and merged using a serialization protocol. The specific content of the two types of strategies is written into a unified data payload block according to a predefined field structure. The short-term Western medicine intervention instruction part includes fields such as intervention instruction urgency level number, drug name, dosage, and execution time window. The long-term TCM conditioning strategy part includes fields such as medicinal diet formula name, TCM herb dosage ratio, treatment cycle, acupoint location and treatment course arrangement. The encapsulated data payload block is the TCM-Western medicine collaborative intervention plan, which is synchronously distributed to the user terminal and the responsible physician's workstation through the underlying communication driver layer. At this point, the entire processing flow of the digital closed-loop intervention strategy generation module is completed. From the collection and preprocessing of multi-source heterogeneous raw data, dual-stream feature encoding and cross-modal attention fusion, metabolic trend prediction and syndrome evolution deduction, to syndrome coupling dynamic risk assessment and graded intervention instruction generation, TCM conditioning strategy matching and pharmacological safety verification and encapsulation, the entire data flow achieves complete closure.

[0056] In summary, the digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, according to the embodiments of this application, is explained. First, the multi-source heterogeneous raw data is preprocessed by splitting it according to its sampling frequency characteristics, forming high-frequency data sequences for Western medicine and low-frequency feature sets for traditional Chinese medicine. Then, an independent dual-stream coding network maps both to a unified feature space. Based on this, the traditional feature splicing paradigm is abandoned, and cross-modal attention modulation is applied to the micro-fluctuation features of Western medicine using macro-constitutional characteristics of traditional Chinese medicine as query conditions. This allows the holistic view of traditional Chinese medicine to dynamically guide the weight allocation of local features of Western medicine, achieving deep fusion at both the temporal and semantic levels. The fused synergistic features are synchronously output through a dual-path decoding network to predict metabolic trends and infer syndrome evolution. In the final intervention decision-making stage, a dynamic modulation mechanism for syndrome sensitivity is introduced to modulate the risk penalty parameter in real time using the syndrome probability distribution. This enables patients with different constitutions to receive differentiated risk assessments and graded, step-by-step intervention responses. This solves the problems in existing schemes where static threshold models cannot reflect the concept of treating different diseases with the same method and where the synergistic link between traditional Chinese and Western medicine is broken at the decision-making level. It realizes full-link synergistic management of traditional Chinese and Western medicine from data collection to closed-loop intervention.

Claims

1. A digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, characterized in that, include: The multi-source data preprocessing and digitization module is used to perform time-series cleaning, alignment, and visual feature digitization mapping on the collected multi-source heterogeneous raw data, including continuous glucose concentration sequences, photoplethysmography pulse wave sequences, traditional Chinese medicine body weight scale questionnaire data, and tongue images, to obtain high-frequency data sequences of Western medicine and low-frequency feature sets of traditional Chinese medicine. The multidimensional feature encoding and extraction module is used to encode and extract the temporal fluctuation features of high-frequency data sequences of Western medicine and to encode and extract the macroscopic constitution features of low-frequency feature sets of traditional Chinese medicine to obtain the microscopic feature vector of Western medicine and the macroscopic feature vector of traditional Chinese medicine. The cross-modal attention fusion module is used to perform time-semantic cross-modal attention fusion on macro-feature vectors of traditional Chinese medicine and micro-feature vectors of Western medicine to obtain a collaborative fusion mapping vector. The Health Trend and Syndrome Evolution Inference Module is used to predict the metabolic trend time series extension of the collaborative fusion mapping vector and to infer the syndrome state transition probability of the collaborative fusion mapping vector to obtain the metabolic prediction inference vector and the syndrome evolution inference vector. The digital closed-loop intervention strategy generation module is used to generate digital closed-loop intervention strategies based on the preset Western medicine drug rule base and the traditional Chinese medicine food homology and compatibility knowledge base to obtain a synergistic intervention plan of traditional Chinese and Western medicine.

2. The digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, as described in claim 1, is characterized in that... The multi-source data preprocessing and digitization module includes: The data routing unit is used to perform memory buffering and routing of the collected continuous glucose concentration sequence, photoplethysmography pulse wave sequence, tongue image and TCM body weight scale questionnaire data to obtain the Western medicine sequence group and the TCM feature group to be processed. The Western medicine data processing unit is used to perform time-series data sliding denoising and interpolation splicing on the Western medicine sequence group to be processed in order to obtain Western medicine high-frequency data sequences. The tongue image feature extraction unit is used to extract the tongue image image separated from the TCM feature group to be processed, extract the region of interest mask of the tongue surface and digitally map the color and texture features to obtain the visual feature vector of the tongue image. The questionnaire data processing and feature aggregation unit is used to temporarily store the separated structured questionnaire data as questionnaire data to be normalized; to perform dimensionless standardization calibration on the questionnaire data to be normalized; and then to splice and aggregate the calibrated questionnaire standard score tensor and tongue image visual feature vector at the fully connected layer node level to obtain the TCM low-frequency feature set.

3. The digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, as described in claim 2, is characterized in that... High-frequency time-series continuous glucose concentration sequences and photoplethysmography pulse wave sequences were packaged into a Western medicine sequence group to be processed, and low-frequency discrete tongue images and traditional Chinese medicine body weight scale questionnaire data were combined into a traditional Chinese medicine feature group to be processed.

4. The digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, as described in claim 1, is characterized in that... The multidimensional feature encoding and extraction module include: The Western Medicine Feature Encoding Unit is used to perform multi-layer local temporal fluctuation feature perception and global max pooling dimensionality reduction encoding on high-frequency Western Medicine data sequences to obtain the Western Medicine microscopic latent space tensor. The TCM feature encoding unit is used to perform affine transformation and high-order correlation capture mapping on the TCM low-frequency feature set to obtain the TCM macroscopic latent space tensor. The feature alignment projection unit is used to perform covariance offset elimination and equal-dimensional orthogonal alignment projection on the microscopic latent space tensor of Western medicine and the macroscopic latent space tensor of traditional Chinese medicine to obtain the microscopic feature vector of Western medicine and the macroscopic feature vector of traditional Chinese medicine, respectively.

5. The digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, as described in claim 1, is characterized in that... The cross-modal attention fusion module includes: The feature projection transformation unit is used to project and transform the macroscopic feature vector of traditional Chinese medicine into a macroscopic constitution query matrix based on three independent and learnable linear parameter matrices, and to project and transform the microscopic feature vector of Western medicine into a microscopic fluctuation key matrix and a microscopic fluctuation value matrix, respectively. The cross-modulation aggregation unit is used to evaluate the cross-modal correlation between the macro-physical query matrix and the micro-fluctuation key matrix to generate an adaptive attention probability distribution, and then use the adaptive attention probability distribution to perform weighted aggregation on the micro-fluctuation value matrix to obtain the cross-modulation feature matrix. The fusion stabilization unit is used to superimpose the cross-modulation feature matrix and the macroscopic feature vector of traditional Chinese medicine with residuals, and then perform data distribution stabilization processing through the layer normalization function to obtain the collaborative fusion mapping vector.

6. The digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, as described in claim 1, is characterized in that... The health trend and syndrome evolution prediction module includes: The metabolic trend prediction unit is used to perform micro-temporal autoregressive accurate trend prediction on the co-fusion mapping vector after initializing the decoder hidden state to obtain the metabolic prediction inference vector. The syndrome evolution assessment unit is used to perform a macroscopic nonlinear assessment of the probability transition of syndrome transitions in the metabolic prediction vector to obtain the syndrome evolution inference vector.

7. The digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, as described in claim 1, is characterized in that... The digital closed-loop intervention strategy generation module includes: The Western medicine intervention instruction generation unit is used to perform dynamic weighted risk assessment of the deviation of hyperglycemia and hypoglycemia from the metabolic prediction and extrapolation vector to determine whether to generate a short-acting Western medicine intervention instruction. The TCM conditioning strategy generation unit is used to calculate the matching degree optimization between the syndrome evolution inference vector as a high-dimensional probability retrieval vector and all candidate strategy templates in the TCM food homology and compatibility knowledge base, and extract the medicinal diet formula and acupoint meridian intervention scheme corresponding to the highest matching degree index to obtain a long-term TCM conditioning strategy. The collaborative solution encapsulation unit is used to perform pharmacological mutual exclusion security verification on short-acting Western medicine intervention instructions and long-term TCM conditioning strategies. The valid strategies that pass the verification are structured, encapsulated, and merged through a serialization protocol to generate a collaborative intervention solution of TCM and Western medicine.

8. The digital management system for metabolic diseases based on the fusion of multi-source data from integrated traditional Chinese and Western medicine, as described in claim 7, is characterized in that... The Western medicine intervention instruction generation unit includes: The threshold offset generation subunit is used to generate adaptive threshold offsets for syndromes based on the syndrome evolution inference vector. The dynamic warning threshold adjustment subunit is used to adjust the preset static high sugar warning threshold and static low sugar warning threshold inward based on the syndrome adaptive threshold offset to obtain the dynamic warning threshold boundary. The coupled risk assessment subunit is used to perform syndrome-coupled dynamic risk assessment on each discrete predicted value of the metabolic prediction vector within a future time window based on the dynamic early warning threshold boundary, so as to obtain the coupled risk assessment coefficient. The intervention instruction generation subunit is used to perform a floor function on the coupling risk assessment coefficient using a preset grading granularity constant to obtain the intervention instruction urgency level number. Based on the intervention instruction urgency level number, a step-by-step instruction matching and retrieval is performed from the graded Western medicine prescription drug use rule map to generate a short-acting Western medicine intervention instruction.

Images