Iodine metabolism abnormity early warning method, device and equipment based on multi-modal feature fusion

Abstract

The invention discloses an iodine metabolism abnormity early warning method, device and equipment based on multi-modal feature fusion, and the method comprises the steps: collecting a urine iodine concentration high-frequency oscillation signal and a blood metabolite concentration spectrum of a user, and constructing a time-aligned original metabolism data set; performing cross-modal feature extraction on the original metabolic data set by adopting an improved double-flow feature fusion network to generate a fusion feature matrix with physiological relevance; hierarchical analysis is conducted on the fusion feature matrix through a dynamic differentiable decision tree, a multi-dimensional risk probability distribution diagram is generated, and a metabolic feature interaction entropy optimization algorithm is adopted for a tree node splitting rule of the multi-dimensional risk probability distribution diagram; and constructing a dynamic early-warning strategy generator based on an improved meta-learning framework, and mapping the risk probability distribution diagram to an action space of an early-warning strategy so as to generate an iodine metabolism abnormity early-warning strategy of the user. According to the invention, accurate early warning can be carried out on the abnormal iodine metabolism of the user.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to an iodine metabolism abnormality early warning method, device and equipment based on multimodal feature fusion. Background Art

[0002] For abnormal iodine metabolism, the current conventional practice is to evaluate the user's iodine metabolism through a single detection method. For example, most rely only on measuring the user's urinary iodine concentration as a single biological feature, and simply judge whether the iodine nutrition status is normal based on the range of urinary iodine values, thereby indirectly inferring whether there may be abnormalities in iodine metabolism. Or simply test the levels of certain specific thyroid hormones in the blood, such as thyroid-stimulating hormone (TSH), free thyroxine (FT4), and free triiodothyronine (FT3), and evaluate iodine metabolism based on changes in these indicators. Some also investigate dietary iodine intake, and try to make a rough assessment of iodine metabolism by estimating the source and amount of iodine intake in the diet.

[0003] However, the existing technical solutions have significant technical problems, namely, a single detection method cannot fully and accurately reflect the complex iodine metabolism process of the human body, which is prone to missed diagnosis and misdiagnosis, and it is difficult to accurately warn users of abnormal iodine metabolism. Summary of the invention

[0004] The embodiments of the present invention provide a method, device and equipment for early warning of abnormal iodine metabolism based on multimodal feature fusion, which can accurately warn users of abnormal iodine metabolism.

[0005] An embodiment of the present invention provides an iodine metabolism abnormality early warning method based on multimodal feature fusion, comprising: Collect the user's urine iodine concentration high-frequency oscillation signal and blood metabolite concentration spectrum to construct a time-aligned raw metabolic data set; An improved dual-stream feature fusion network is used to perform cross-modal feature extraction on the original metabolic dataset to generate a fusion feature matrix with physiological relevance, wherein the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum is extracted using an improved graph attention network; The fusion feature matrix is ​​hierarchically analyzed by a dynamic differentiable decision tree to generate a multidimensional risk probability distribution map, wherein the tree node splitting rule of the multidimensional risk probability distribution map adopts a metabolic feature interactive entropy optimization algorithm; A dynamic early warning strategy generator is constructed based on an improved meta-learning framework, and the risk probability distribution map is mapped to the action space of the early warning strategy, so as to generate an iodine metabolism abnormality early warning strategy for the user.

[0006] As an improvement of the above scheme, the improved dual-stream feature fusion network is used to perform cross-modal feature extraction on the original metabolic data set to generate a fusion feature matrix with physiological relevance, wherein the high-frequency oscillation signal of urine iodine concentration is decomposed by an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum is extracted by using an improved graph attention network, including: Feature decomposition and correlation sub-step: performing multi-scale decomposition on the high-frequency oscillation signal of urine iodine concentration in the original metabolic data set, generating a time-frequency feature matrix based on dynamically optimized wavelet basis function parameters, wherein the scale factor of the wavelet basis function is adaptively adjusted through an error back propagation algorithm; Graph structure construction sub-step: constructing a dynamic association graph according to the blood metabolite concentration spectrum in the original metabolic data set, wherein the nodes of the dynamic association graph represent metabolite indicators, the edge weights of the dynamic association graph are jointly determined by the metabolic pathway correlation and the time correlation, and the key connection paths of the dynamic association graph are screened through a multi-head sparse attention mechanism; Cross-modal alignment sub-step: design a bidirectional feature interaction pipeline, use a gated recurrent unit to achieve temporal alignment of urine time-frequency features and blood map features, and eliminate the inter-modal dimensionality difference between the two through a learnable projection matrix; Fusion feature generation sub-step: The aligned dual-stream features are input into the feature compression network, and the channel attention mechanism is used to optimize the feature combination to generate a fusion feature matrix with unified dimension and physiological relevance.

[0007] As an improvement of the above scheme, the fusion feature matrix is ​​hierarchically parsed by a dynamic differentiable decision tree to generate a multidimensional risk probability distribution map, wherein the tree node splitting rule of the multidimensional risk probability distribution map adopts a metabolic feature interactive entropy optimization algorithm, including: The split optimization sub-step performs node split optimization on the fusion feature matrix, including: Based on the feature vectors in the fused feature matrix, a trainable splitting parameter matrix is ​​constructed to generate a dimensionally matched dynamic threshold vector set; Performing mutual information entropy calculation on the joint distribution of the dynamic threshold vector set and the clinical diagnosis label, and outputting a feature interaction entropy splitting indicator set; Calling the biological association rules in the thyroid metabolic pathway knowledge base, performing anatomical constraint correction on the split indicator set, and generating an optimized node splitting rule set; The structure optimization sub-step performs tree structure optimization on the node splitting rule set, including: Inputting the node splitting rule set into a reinforcement learning controller, generating an initial decision tree structure by defining a state space including node path purity and feature contribution and an action space including branch growth directions; Adopting a proximal strategy optimization algorithm to adjust the branch growth strategy of the initial decision tree structure, and outputting a dynamically optimized decision tree structure; Performing online pruning processing on the decision tree structure according to the product index of the node path purity and the feature contribution, and outputting a simplified decision tree; The sub-step of performing uncertainty quantification on the simplified decision tree includes: Inserting a random gating unit into each decision path of the simplified decision tree, dynamically adjusting the activation probability based on the feature vector L2 norm, and generating a probabilistic decision path; Performing Monte Carlo Dropout forward propagation on the probabilistic decision path to collect and generate a set of prediction results; Bootstrap resampling calculation is performed on the prediction result set, and a multidimensional risk probability distribution diagram with confidence intervals is output.

[0008] As an improvement of the above scheme, the dynamic early warning strategy generator is constructed based on the improved meta-learning framework, and the risk probability distribution map is mapped to the action space of the early warning strategy, so as to generate the user's iodine metabolism abnormality early warning strategy, including: The prototype construction sub-step is to build a prototype of the strategy execution strategy for historical diagnosis and treatment data, including: A variational autoencoder consisting of a bidirectional LSTM layer and an attention pooling layer is used to extract features from warning strategy records and generate a potential strategy vector set; Performing tensor splicing of the potential strategy vector set and the real-time environment parameter matrix to construct a multi-dimensional strategy feature space; Design a composite metric function that combines cosine similarity and clinical pathway matching to generate a standardized strategy prototype library; The matching optimization sub-step performs strategy matching optimization on the multi-dimensional risk probability distribution graph, including: Inputting the multi-dimensional risk probability distribution graph into a differentiable nearest neighbor algorithm, matching candidate strategies in a strategy prototype library, and generating an initial matching weight matrix; Performing feature interaction entropy analysis on the initial matching weight matrix, determining multi-head attention weight values, and generating a weighted fusion strategy set; Inputting the weighted fusion strategy set into a temporal convolutional network for time series modeling, and outputting a time-coherent early warning strategy sequence; The dynamic tuning sub-step performs dynamic tuning on the early warning strategy sequence, including: Construct a constrained reward function that includes the thyroid iodine metabolism kinetics equation; Performing an adjoint sensitivity analysis on the early warning strategy sequence to calculate a gradient response matrix of strategy parameters to metabolic equations; The parameters in the gradient response matrix are adjusted using a constrained optimization algorithm to output an iodine metabolism abnormality early warning strategy that has been verified by physiological rationality.

[0009] As an improvement of the above solution, after constructing a dynamic early warning strategy generator based on the improved meta-learning framework and mapping the risk probability distribution graph to the action space of the early warning strategy to generate the abnormal iodine metabolism early warning strategy for the user, the method further includes: Input the iodine metabolism abnormality warning strategy into a preset clinical diagnosis and treatment knowledge base, match the preset medication dosage template and dietary taboo rules, and generate text prompt content including medication time window and taboo foods; Performing display optimization processing on the text prompt content, and dynamically adjusting the font size, background color and information sorting weight according to the confidence interval; The optimized prompt content is pushed to the user's terminal device for display.

[0010] Another embodiment of the present invention provides an iodine metabolism abnormality early warning device based on multimodal feature fusion, comprising: The acquisition module is used to collect the user's urine iodine concentration high-frequency oscillation signal and blood metabolite concentration spectrum to construct a time-aligned original metabolic data set; A feature extraction module, for performing cross-modal feature extraction on the original metabolic data set using an improved dual-stream feature fusion network to generate a fusion feature matrix with physiological relevance, wherein the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum is extracted using an improved graph attention network; A parsing module, used for performing hierarchical parsing on the fusion feature matrix through a dynamic differentiable decision tree to generate a multidimensional risk probability distribution map, wherein the tree node splitting rule of the multidimensional risk probability distribution map adopts a metabolic feature interactive entropy optimization algorithm; The early warning generation module is used to build a dynamic early warning strategy generator based on an improved meta-learning framework, map the risk probability distribution map to the action space of the early warning strategy, and generate an iodine metabolism abnormality early warning strategy for the user.

[0011] As an improvement of the above solution, the feature extraction module is specifically used for: Performing multi-scale decomposition on the high-frequency oscillation signal of urine iodine concentration in the original metabolic data set, generating a time-frequency feature matrix based on dynamically optimized wavelet basis function parameters, wherein the scale factor of the wavelet basis function is adaptively adjusted through an error back propagation algorithm; Constructing a dynamic association map according to the blood metabolite concentration spectrum in the original metabolic data set, wherein the nodes of the dynamic association map represent metabolite indicators, the edge weights of the dynamic association map are jointly determined by the metabolic pathway correlation and the time correlation, and the key connection paths of the dynamic association map are screened by a multi-head sparse attention mechanism; Design a bidirectional feature interaction pipeline, use a gated recurrent unit to achieve temporal alignment between urine time-frequency features and blood map features, and eliminate the inter-modal dimensionality difference between the two through a learnable projection matrix; The aligned dual-stream features are input into the feature compression network, and the channel attention mechanism is used to optimize the feature combination to generate a fusion feature matrix with unified dimension and physiological relevance.

[0012] As an improvement of the above solution, the parsing module is specifically used for: Performing node splitting optimization on the fused feature matrix includes: Based on the feature vectors in the fused feature matrix, a trainable splitting parameter matrix is ​​constructed to generate a dimensionally matched dynamic threshold vector set; Performing mutual information entropy calculation on the joint distribution of the dynamic threshold vector set and the clinical diagnosis label, and outputting a feature interaction entropy splitting indicator set; Calling the biological association rules in the thyroid metabolic pathway knowledge base, performing anatomical constraint correction on the split indicator set, and generating an optimized node splitting rule set; Performing tree structure optimization on the node splitting rule set includes: Inputting the node splitting rule set into a reinforcement learning controller, generating an initial decision tree structure by defining a state space including node path purity and feature contribution and an action space including branch growth directions; Adopting a proximal strategy optimization algorithm to adjust the branch growth strategy of the initial decision tree structure, and outputting a dynamically optimized decision tree structure; Performing online pruning processing on the decision tree structure according to the product index of the node path purity and the feature contribution, and outputting a simplified decision tree; Performing uncertainty quantification on the streamlined decision tree includes: Inserting a random gating unit into each decision path of the simplified decision tree, dynamically adjusting the activation probability based on the feature vector L2 norm, and generating a probabilistic decision path; Performing Monte Carlo Dropout forward propagation on the probabilistic decision path to collect and generate a set of prediction results; Bootstrap resampling calculation is performed on the prediction result set, and a multidimensional risk probability distribution diagram with confidence intervals is output.

[0013] As an improvement of the above solution, the warning generation module is specifically used for: Prototype execution strategies for historical medical data, including: A variational autoencoder consisting of a bidirectional LSTM layer and an attention pooling layer is used to extract features from warning strategy records and generate a potential strategy vector set; Performing tensor splicing of the potential strategy vector set and the real-time environment parameter matrix to construct a multi-dimensional strategy feature space; Design a composite metric function that combines cosine similarity and clinical pathway matching to generate a standardized strategy prototype library; Performing strategy matching optimization on the multi-dimensional risk probability distribution graph includes: Inputting the multi-dimensional risk probability distribution graph into a differentiable nearest neighbor algorithm, matching candidate strategies in a strategy prototype library, and generating an initial matching weight matrix; Performing feature interaction entropy analysis on the initial matching weight matrix, determining multi-head attention weight values, and generating a weighted fusion strategy set; Inputting the weighted fusion strategy set into a temporal convolutional network for time series modeling, and outputting a time-coherent early warning strategy sequence; Dynamically tuning the early warning strategy sequence includes: Construct a constrained reward function that includes the thyroid iodine metabolism kinetics equation; Performing an adjoint sensitivity analysis on the early warning strategy sequence to calculate a gradient response matrix of strategy parameters to metabolic equations; The parameters in the gradient response matrix are adjusted using a constrained optimization algorithm to output an iodine metabolism abnormality early warning strategy that has been verified by physiological rationality.

[0014] Another embodiment of the present invention provides an iodine metabolism abnormality warning device based on multimodal feature fusion, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the iodine metabolism abnormality warning method based on multimodal feature fusion as described in the above-mentioned embodiment of the invention when executing the computer program.

[0015] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: By first collecting the high-frequency oscillation signal of the user's urine iodine concentration and the blood metabolite concentration spectrum, a time-aligned original metabolic data set is constructed, which can obtain information related to iodine metabolism from multiple dimensions, laying the foundation for subsequent precise analysis. Subsequently, with the help of an improved two-stream feature fusion network, the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum features are extracted using an improved graph attention network, cross-modal feature extraction is achieved, and a fusion feature matrix with physiological relevance is generated. This process deeply explores the intrinsic connection between different modal data, making feature analysis more targeted and accurate. Then, a dynamic differentiable decision tree is used in combination with a metabolic feature interactive entropy optimization algorithm to hierarchically analyze the fusion feature matrix and generate a multidimensional risk probability distribution map, thereby effectively processing complex features to accurately assess the risk of abnormal iodine metabolism. Finally, a dynamic early warning strategy generator is constructed based on an improved meta-learning framework, which maps the risk probability distribution map to the action space of the early warning strategy, generates the user's iodine metabolism abnormality early warning strategy, and enables the early warning strategy to be dynamically adjusted according to different situations. In summary, compared with the prior art that relies only on a single biometric feature and results in low warning accuracy, the embodiments of the present invention achieve a more comprehensive, accurate and dynamic warning of users' iodine metabolism abnormalities through multimodal biometric feature fusion and specific feature extraction and analysis methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of a method for early warning of abnormal iodine metabolism based on multimodal feature fusion provided by one embodiment of the present invention; Figure 2 It is a structural schematic diagram of an iodine metabolism abnormality early warning device based on multimodal feature fusion provided by one embodiment of the present invention; Figure 3 It is a structural schematic diagram of an iodine metabolism abnormality warning device based on multimodal feature fusion provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] See also Figure 1 , is a flow chart of an iodine metabolism abnormality early warning method based on multimodal feature fusion provided by an embodiment of the present invention. The iodine metabolism abnormality early warning method based on multimodal feature fusion comprises the following steps: S10, collects the user's urine iodine concentration high-frequency oscillation signal and blood metabolite concentration spectrum to construct a time-aligned original metabolic data set; S11, using an improved dual-stream feature fusion network to perform cross-modal feature extraction on the original metabolic data set to generate a fusion feature matrix with physiological relevance, wherein the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum is extracted using an improved graph attention network; S12, performing hierarchical analysis on the fusion feature matrix through a dynamic differentiable decision tree to generate a multidimensional risk probability distribution map, wherein the tree node splitting rule of the multidimensional risk probability distribution map adopts a metabolic feature interactive entropy optimization algorithm; S13, constructing a dynamic early warning strategy generator based on an improved meta-learning framework, mapping the risk probability distribution graph to the action space of the early warning strategy, thereby generating an iodine metabolism abnormality early warning strategy for the user.

[0019] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: By first collecting the high-frequency oscillation signal of the user's urine iodine concentration and the blood metabolite concentration spectrum, a time-aligned original metabolic data set is constructed, which can obtain information related to iodine metabolism from multiple dimensions, laying the foundation for subsequent precise analysis. Subsequently, with the help of an improved two-stream feature fusion network, the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum features are extracted using an improved graph attention network, cross-modal feature extraction is achieved, and a fusion feature matrix with physiological relevance is generated. This process deeply explores the intrinsic connection between different modal data, making feature analysis more targeted and accurate. Then, a dynamic differentiable decision tree is used in combination with a metabolic feature interactive entropy optimization algorithm to hierarchically analyze the fusion feature matrix and generate a multidimensional risk probability distribution map, thereby effectively processing complex features to accurately assess the risk of abnormal iodine metabolism. Finally, a dynamic early warning strategy generator is constructed based on an improved meta-learning framework, which maps the risk probability distribution map to the action space of the early warning strategy, generates the user's iodine metabolism abnormality early warning strategy, and enables the early warning strategy to be dynamically adjusted according to different situations. In summary, compared with the prior art that relies only on a single biometric feature and results in low warning accuracy, the embodiments of the present invention achieve a more comprehensive, accurate and dynamic warning of users' iodine metabolism abnormalities through multimodal biometric feature fusion and specific feature extraction and analysis methods.

[0020] As one example, in step S10, the specific working process is as follows: 1. Timestamp information added: High-frequency oscillation signal of urine iodine concentration: A high-precision clock chip is integrated in the data acquisition device (such as a high-precision electrochemical sensor). The chip uses an atomic clock or a high-precision crystal oscillator as the clock source to ensure that the time accuracy reaches the microsecond level. When the device starts to collect the high-frequency oscillation signal of urine iodine concentration, the clock chip automatically records the start time and synchronously records the corresponding timestamp at each sampling moment. These timestamp information is stored in the internal memory of the device in the form of binary coding, and the sampling frequency is set to 1000Hz, that is, data is collected once every millisecond, and each collection is associated with an accurate timestamp.

[0021] 2. Blood metabolite concentration spectrum: During the blood sample collection process, a sample collection tube with a time recording function is used. The collection tube has a built-in electronic tag, which records the time information accurate to the second at the moment of sample collection through wireless communication with the time synchronization equipment at the collection site. After the sample is sent to the laboratory, the laboratory's analytical instruments (such as liquid chromatography-mass spectrometry) will automatically record the time points of each key step (such as injection, separation, detection, etc.) when analyzing and processing the sample, and associate it with the time information in the electronic tag. This time information is ultimately stored in the laboratory information management system (LIMS) in a standardized time format (such as ISO 8601 format).

[0022] 3. Data preprocessing: normalize the acquired data.

[0023] 4. Use DTW algorithm for alignment: Since there may be some differences in the acquisition time of the high-frequency oscillation signal of urine iodine concentration and the blood metabolite concentration spectrum, time series alignment is required. The dynamic time warping (DTW) algorithm is used to find the optimal alignment path between the two time series without considering the difference in time series length, so that the data at the corresponding time points can be accurately matched. The aligned high-frequency oscillation signal of urine iodine concentration and the blood metabolite concentration spectrum data are fused to construct a unified original metabolic data set. The feature vector of the urine signal and the feature vector of the blood metabolite concentration spectrum can be spliced ​​together to form a high-dimensional feature matrix by feature splicing. It should be noted that human iodine metabolism is a dynamic process. The high-frequency oscillation signal of urine iodine concentration and the blood metabolite concentration spectrum will change with time, and it is difficult to collect the two at the same time in actual collection. There is a time offset. If they are not aligned, the temporal and spatial consistency of the data cannot be guaranteed, and it is difficult to accurately reflect the true correlation between the two data at the same physiological moment. At the same time, subsequent cross-modal feature extraction needs to accurately capture the intrinsic connection between the two data. Misalignment will lead to feature confusion and inaccurate extraction. In addition, time series misalignment will also affect the accuracy of the multidimensional risk probability distribution, thereby reducing the reliability of the iodine metabolism abnormality warning strategy generated based on this.

[0024] As one example, the improved dual-stream feature fusion network is used to perform cross-modal feature extraction on the original metabolic data set to generate a fusion feature matrix with physiological relevance, wherein the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum is extracted using an improved graph attention network, including: Feature decomposition and correlation sub-step: performing multi-scale decomposition on the high-frequency oscillation signal of urine iodine concentration in the original metabolic data set, generating a time-frequency feature matrix based on dynamically optimized wavelet basis function parameters, wherein the scale factor of the wavelet basis function is adaptively adjusted through an error back propagation algorithm; Graph structure construction sub-step: constructing a dynamic association graph according to the blood metabolite concentration spectrum in the original metabolic data set, wherein the nodes of the dynamic association graph represent metabolite indicators, the edge weights of the dynamic association graph are jointly determined by the metabolic pathway correlation and the time correlation, and the key connection paths of the dynamic association graph are screened through a multi-head sparse attention mechanism; Cross-modal alignment sub-step: design a bidirectional feature interaction pipeline, use a gated recurrent unit to achieve temporal alignment of urine time-frequency features and blood map features, and eliminate the inter-modal dimensionality difference between the two through a learnable projection matrix; Fusion feature generation sub-step: The aligned dual-stream features are input into the feature compression network, and the channel attention mechanism is used to optimize the feature combination to generate a fusion feature matrix with unified dimension and physiological relevance.

[0025] In this embodiment, an improved dual-stream feature fusion network is used to extract and fuse features for the two different modal data, namely, the high-frequency oscillation signal of urine iodine concentration and the blood metabolite concentration spectrum. By performing multi-scale decomposition on the urine signal, dynamically optimizing the wavelet basis function parameters, obtaining the time-frequency feature matrix, and mining the dynamic change information of urine iodine metabolism; constructing a dynamic correlation map based on the blood metabolite concentration spectrum, considering the correlation degree of metabolic pathways and time correlation to determine the edge weight, and using the multi-head sparse attention mechanism to screen the key connection path to reflect the complex relationship between blood metabolites. Then, by designing a bidirectional feature interaction pipeline, using the gated cyclic unit to achieve the time alignment of urine and blood features, combining the learnable projection matrix to eliminate the dimensional difference, and finally inputting the aligned features into the feature compression network, using the channel attention mechanism to optimize the combination, and generating a fusion feature matrix with physiological relevance. In summary, this embodiment comprehensively and deeply mines the intrinsic connection between different modal biometric data, effectively integrates multi-source information, and provides high-quality and highly correlated feature data for subsequent accurate analysis of the risk of abnormal iodine metabolism, thereby improving the accuracy and reliability of the assessment of abnormal iodine metabolism.

[0026] For ease of understanding, this embodiment is specifically described below: Eigendecomposition and association substeps: For urine iodine concentration high frequency oscillation signal , using a unique multi-scale adaptive wavelet decomposition method. Among them, the improved adaptive wavelet basis function for: ; in, is the scale factor. As in traditional wavelet analysis, its value determines the degree of expansion and contraction of the wavelet function. The value focuses on the low-frequency components of the signal. The value focuses on capturing high-frequency details. The adjustment of the signal can provide in-depth analysis of the signal at different scales. It is the translation factor, which controls the position of the wavelet function on the time axis and ensures that the wavelet can scan and analyze the signal at different times. The frequency modulation parameter gives the wavelet function a specific frequency characteristic, which helps to distinguish different frequency patterns in the signal. is a newly introduced parameter related to the local characteristics of the signal, which is expressed by This term makes the wavelet function around different time points t according to the translation position The frequency characteristics are dynamically fine-tuned according to the distance, so as to more accurately match the complex frequency changes of the high-frequency oscillation signal of urine iodine concentration. It is a wavelet function with good time-frequency localization characteristics. The adaptive adjustment of The minimization is achieved by: ; in, is the total number of signal sampling points, Indicates A sampling moment. is the number of decomposition levels, They are The corresponding translation factors, frequency modulation parameters and local characteristic parameters during layer decomposition. It is The sampling point is Wavelet coefficients under layer decomposition. is a trade-off factor that balances the signal approximation error and the smoothness of the scale factor change. , which represents the change in the scale factor when two adjacent layers are decomposed. By adding this item, the scale factor is prevented from drastically fluctuating during the adjustment process, ensuring the stability and rationality of the decomposition process. The scale factor is continuously updated through optimization algorithms such as gradient descent. , so that the error function Reach the minimum, and then dynamically optimize the wavelet basis function parameters to generate the time-frequency feature matrix .matrix Elements Represented in scale ,time , Frequency Modulation And local characteristics The wavelet coefficients under comprehensively and meticulously reflect the characteristics of the high-frequency oscillation signal of urine iodine concentration in the time-frequency domain.

[0027] Graph structure construction sub-steps: A dynamic correlation map is constructed based on the blood metabolite concentration spectrum. Assume that Blood metabolites, dynamic correlation map , node set , each node Represents a metabolite index. Edge weight It is determined by the metabolic pathway correlation and time correlation and is calculated by the following formula: ; in, It is a global balance parameter with a value range of [0, 1]. Its optimal value is determined by statistical analysis of a large amount of experimental data and model performance evaluation. It is used to balance the relative importance of metabolic pathway association, time correlation, and the newly introduced time change direction factor in the edge weight. Metabolites and The degree of association in metabolic pathways is obtained through complex semantic mining and network analysis of professional metabolic pathway databases. Reflecting metabolites and The correlation of concentration changes in time series is obtained by applying a complex calculation method based on mutual information to the concentration time series of the two. It is a correlation parameter that takes into account the rate of change of metabolite concentrations at different time points and is obtained by performing high-order statistical analysis on the time series of concentration change rates. is a parameter that controls the degree of influence of the time change direction. is the symbolic function, ,when hour, , indicating that time changes in a positive direction; when hour, , indicating the reverse change in time. Through this item, the time direction information of metabolite concentration change is included in the edge weight calculation, which more comprehensively reflects the dynamic relationship between blood metabolites. In order to screen the key connection paths of the dynamic association graph, an innovative multi-head sparse attention mechanism is adopted. Assume that the number of attention heads is , for the The attention of each head is calculated as follows: ; in, and They are the query matrix, key matrix and value matrix obtained by performing unique linear transformation on the node feature matrix, and their dimensions are . is a sparse mask matrix used to introduce sparsity, reduce computation and highlight key connections. is a newly introduced matrix related to node importance, whose elements The importance score of the metabolites represented by nodes m and n in the overall metabolic network is determined by comprehensively calculating multiple factors such as the topological structure of the metabolic network, metabolic flux analysis, and functional importance assessment related to iodine metabolism. , so that the attention mechanism pays more attention to the connection between important nodes during calculation, further improving the accuracy of key connection path screening. The Softmax function normalizes the attention score so that the sum of the attention distribution of each head is 1. The final attention output yes The weighted concatenation of individual attention results is used to screen out the key connection paths in the dynamic association graph, providing key information for subsequent analysis.

[0028] Cross-modal alignment sub-steps: A bidirectional feature interaction pipeline is designed, and an improved gated recurrent unit (GRU) is used to achieve the time sequence alignment of urine time-frequency features and blood graph features. Assume that the urine time-frequency feature sequence is , the blood map feature sequence is The improved GRU update process is as follows: ; in, Is the update gate, which determines the hidden state at the previous moment and the current candidate hidden state Hide state for the current moment contribution ratio. It is the reset gate, which controls the hidden state at the previous moment Hide the state of the current candidate degree of impact. is a candidate hidden state, taking into account the current input urine time-frequency characteristics and blood map features And the hidden state at the previous moment . is the hidden state at the current moment, through the update gate Hide the previous state and candidate hidden states Obtained by weighted combination.

[0029] They are all weight matrices, which control the connection weights between different inputs and hidden states. is the bias term that adjusts the activation threshold of the neuron. is the sigmoid activation function, which maps the input to the interval [0, 1], is the hyperbolic tangent activation function, which performs a nonlinear transformation on the candidate hidden states. is a newly introduced parameter. pass This item uses the difference information of urine and blood features at the previous moment to fine-tune the update gate, so that the model can better capture the dynamic alignment relationship between the two features. pass This item adjusts the reset gate according to the size relationship between urine and blood characteristics at the current moment, enhancing the model's sensitivity to the direction of feature changes. pass (here represents a bitwise XOR operation), and the binary representation information of urine and blood features is fused into the calculation of candidate hidden states to further enrich the feature representation.

[0030] At the same time, through the learnable projection matrix Except for the difference in the dimensions between the two modes. Assume that the urine time-frequency feature dimension is , the blood map feature dimension is , the projection matrix The dimension is . Projection transformation of blood map features , and the projection matrix is To update: ; in, is the learning rate parameter, which controls the step size of the projection matrix update. and They are urine time-frequency characteristic sequences and blood map feature sequence In this way, according to the difference information of the mean centering of urine and blood features, the projection matrix is ​​dynamically updated, so that the blood map features are better matched with the urine time-frequency features after projection, laying the foundation for subsequent feature fusion. Fusion feature generation sub-steps: The aligned dual-stream features and Input feature compression network, using innovative channel attention mechanism to optimize feature combination. Assume the input feature is First, perform global logarithmic pooling on the features to obtain channel feature description :in, and are the height and width of the feature map, respectively. is the feature map at position The elements at . Different from the traditional global average pooling, the global logarithmic pooling is used here, which can better highlight the small changes in the features and enhance the ability to capture feature details. Then, the channel attention weights are generated through two fully connected layers and activation functions. : ; in, and is the weight matrix of the fully connected layer, and is the bias term. It is an exponential linear unit activation function, which has a non-zero gradient when the input is negative, which can alleviate the gradient vanishing problem while retaining the computational efficiency advantage. is a newly introduced parameter. (here represents the cosine similarity calculation, is the characteristic description of all channels The mean vector of the current channel feature is used to integrate the similarity information between the current channel feature and the mean of the overall channel feature into the attention weight calculation, so that the model can dynamically adjust the attention allocation according to the relative relationship between channels. The Sigmoid function maps the output to the interval [0, 1] to obtain the channel attention weight. Finally, the channel attention weight is multiplied by the original feature to obtain the optimized fusion feature , generating a fusion feature matrix with unified dimensions and physiological relevance. This matrix fully integrates the characteristic information of the high-frequency oscillation signal of urine iodine concentration and the blood metabolite concentration spectrum, providing a high-quality data foundation for subsequent hierarchical analysis through a dynamic differentiable decision tree.

[0031] As one example, the fusion feature matrix is ​​hierarchically parsed by a dynamic differentiable decision tree to generate a multidimensional risk probability distribution map, wherein the tree node splitting rule of the multidimensional risk probability distribution map adopts a metabolic feature interactive entropy optimization algorithm, including: The split optimization sub-step performs node split optimization on the fusion feature matrix, including: Based on the feature vectors in the fused feature matrix, a trainable splitting parameter matrix is ​​constructed to generate a dimensionally matched dynamic threshold vector set; Performing mutual information entropy calculation on the joint distribution of the dynamic threshold vector set and the clinical diagnosis label, and outputting a feature interaction entropy splitting indicator set; Calling the biological association rules in the thyroid metabolic pathway knowledge base, performing anatomical constraint correction on the split indicator set, and generating an optimized node splitting rule set; The structure optimization sub-step performs tree structure optimization on the node splitting rule set, including: Inputting the node splitting rule set into a reinforcement learning controller, generating an initial decision tree structure by defining a state space including node path purity and feature contribution and an action space including branch growth directions; Adopting a proximal strategy optimization algorithm to adjust the branch growth strategy of the initial decision tree structure, and outputting a dynamically optimized decision tree structure; Performing online pruning processing on the decision tree structure according to the product index of the node path purity and the feature contribution, and outputting a simplified decision tree; The sub-step of performing uncertainty quantification on the simplified decision tree includes: Inserting a random gating unit into each decision path of the simplified decision tree, dynamically adjusting the activation probability based on the feature vector L2 norm, and generating a probabilistic decision path; Performing Monte Carlo Dropout forward propagation on the probabilistic decision path to collect and generate a set of prediction results; Bootstrap resampling calculation is performed on the prediction result set, and a multidimensional risk probability distribution diagram with confidence intervals is output.

[0032] In this embodiment, the fusion feature matrix is ​​hierarchically analyzed by a dynamic differentiable decision tree to generate an accurate and reliable multidimensional risk probability distribution map. First, in the split optimization sub-step, a trainable split parameter matrix is ​​constructed based on the fusion feature matrix to obtain a dynamic threshold vector set, and the feature interaction entropy split index set is output by mutual information entropy calculation, and the biological association rules of the thyroid metabolic pathway knowledge base are combined to correct it to ensure the scientificity and rationality of the split rules. Then in the structural optimization sub-step, the reinforcement learning controller is used to generate the initial decision tree structure in combination with a specific state and action space, and the branch growth strategy is adjusted by the proximal strategy optimization algorithm. Then, according to the product index of the node path purity and the feature contribution, the tree is pruned to obtain a streamlined decision tree, and the structural rationality and efficiency of the decision tree are improved. Finally, in the execution sub-step, a probabilistic decision path is generated by inserting a random gating unit in the decision path, and the Monte Carlo Dropout forward propagation and Bootstrap resampling are used to calculate and output a multidimensional risk probability distribution map with a confidence interval to quantify the uncertainty in the decision process. In summary, in this embodiment, on the one hand, the tree node splitting rules are optimized by the metabolic feature interactive entropy optimization algorithm, which fully considers the metabolic characteristics of biological features and clinical diagnosis information, so that the splitting of the decision tree is more targeted and accurate, and the potential patterns in the data can be better mined. On the other hand, the decision tree structure is optimized and pruned to avoid the overfitting problem and improve the generalization ability and computational efficiency of the model. At the same time, the quantification of uncertainty enables the multidimensional risk probability distribution map to not only give the risk probability, but also provide confidence intervals, providing doctors and patients with more comprehensive and reliable risk assessment information, so as to more accurately formulate iodine metabolism abnormality early warning strategies.

[0033] For ease of understanding, this embodiment is specifically described below: Split optimization substep 1. Construct a trainable splitting parameter matrix and a dynamic threshold vector set: For the fusion feature matrix , whose dimensions are ,in is the sample size, is the feature dimension. Construct a trainable split parameter matrix , the dimension is ,here is a hyperparameter that can be adjusted according to the complexity of the actual dataset and computing resources. By constructing a multi-layer perceptron (MLP), its input is the fusion feature matrix , after several hidden layers (the number of neurons in each layer is ), the output is the splitting parameter matrix, the output is the splitting parameter matrix During training, a stochastic gradient descent algorithm is used to minimize a custom loss function ,This loss function comprehensively considers the purity improvement of the node after splitting and its relevance to the clinical diagnosis label.

[0034] Based on the trained matrix , generate a dynamic threshold vector set For each feature dimension , threshold The calculation method is: ; in, It is a sigmoid activation function, which is used to map the output value to the [0,1] interval, making the threshold interpretable. is a matrix Middle Line Elements of a column. is a custom nonlinear transformation function, such as , which transforms the sum of all sample values ​​in the feature dimension to better capture the overall characteristics of the feature. is the fusion feature matrix Middle The sample in The value of a feature dimension.

[0035] 2. Calculate the feature interaction entropy split indicator set: Let the clinical diagnosis label be , the value is , is the number of label categories. For the dynamic threshold vector set Clinical diagnostic labels The joint distribution of For each threshold and tags , , in, is the indicator function, when the event It is 1 when it is true, otherwise it is 0. Then calculate the marginal probability distribution and Defining feature interaction entropy splitting index for: ; in, It is a weight parameter used to balance the influence of traditional mutual information entropy and correlation between features. It is the calculation feature and The corresponding eigenvector and (Here, the threshold on the feature dimension is expanded to a vector with the same dimension as the number of samples, by repeating the threshold The cosine similarity of the feature vector (value construction) measures the directional consistency between features. It is to calculate the Euclidean distance between two feature vectors, which is used to normalize the cosine similarity, making the measurement of the correlation between features more reasonable. In this way, the feature interaction entropy splitting index set is obtained: .

[0036] 3. Generate an optimized node splitting rule set: Calling the thyroid metabolic pathway knowledge base , the knowledge base contains the interrelationships between various substances in the thyroid metabolism process and the rules related to abnormal iodine metabolism. For each feature interaction entropy splitting index The anatomical constraints are modified according to the rules in the knowledge base. For example, if the knowledge base stipulates that two features have a strong correlation in the thyroid metabolic pathway, when the splitting indicators corresponding to these two features do not match the expectations in the knowledge base, they are adjusted. The specific adjustment method is: ; in, It is a subset of rules related to anatomical constraints in the knowledge base, which contains feature pairs with strong associations. . It is a weight set according to the importance of the rule, which can be determined through in-depth analysis and expert evaluation of the rules in the knowledge base. According to the rules, the features and The amount of adjustment made to the split indicator difference, for example, if the rule states that the feature and should have similar splitting indices, while the currently calculated and If the difference is large, is a value related to the degree of difference, which is used to narrow the gap between the two. After correction, the optimized node splitting rule set is generated .

[0037] Structural optimization sub-steps: 1. Generate the initial decision tree structure: Split the node rule set Enter the reinforcement learning controller. Define the state space , which is determined by the node path purity and feature contribution Composition. Node path purity It is measured by calculating the ratio of the number of samples of the category with the largest proportion in the samples in the node to the total number of samples in the node, that is: ; Among them, nodes is the sample set contained in the current node. Feature contribution The information gain of each feature in the node splitting process is weighted and summed. , its information gain According to the above calculation of interaction entropy split index And the calculation of sample distribution changes before and after node splitting, ,in It is a pre-set weight based on the importance of features, which can be determined through analysis of historical data and domain knowledge.

[0038] Action Space Including the direction of branch growth, such as left branch or right branch. Reinforcement learning controller is based on state space and action space By continuously trying different branch decisions, using the strategy network Generate initial decision tree structure ,in are the parameters of the policy network, which are updated by maximizing the cumulative reward during training. The cumulative reward function Not only the improvement of node path purity and feature contribution is considered, but also the degree of matching with the thyroid metabolic pathway knowledge base. For example, when the decision tree branch result is more matched with a typical thyroid metabolic abnormality pattern in the knowledge base, a higher reward is given.

[0039] 2. Adjust branch growth strategy: The improved proximal policy optimization algorithm (PPO) is used to optimize the initial decision tree structure. Adjust the branch growth strategy. In the traditional PPO algorithm, the cumulative reward function is , in the present invention, it is improved. Assume that the current state is , take action After reaching the new state ,award Defined as: ; in, is a weight parameter used to balance the impact of different reward factors. and They are the changes in node path purity and feature contribution before and after state transition. is the decision tree structure after taking action, is a typical pattern in the knowledge base. and When matching, , otherwise it is 0. By continuously adjusting the strategy parameters , so that the cumulative reward Maximize, thus obtaining a dynamically optimized decision tree structure .

[0040] 3. Perform online pruning processing: According to the node path purity and feature contribution The product index of , for the decision tree structure Perform online pruning. For product index Less than a dynamic threshold Nodes of , cut them off, and get a simplified decision tree . Dynamic Threshold According to the number of iterations in the decision tree construction process And the overall complexity of the current decision tree is dynamically adjusted, for example: ; in, is the initial threshold, is the preset maximum number of iterations. It is a weight parameter used to balance the impact of the number of iterations and the complexity of the decision tree on the threshold. Is a decision tree that measures The function of complexity can be determined, for example, by calculating comprehensive indicators such as the number of nodes and branches of the decision tree.

[0041] Execute substeps: 1. Generate probabilistic decision paths: In the simplified decision tree Insert random gating units into each decision path . Random Gating Unit The activation probability Based on feature vector The norm is dynamically adjusted. Let the eigenvector be , the activation probability calculation formula is: ; in, is a custom parameter used to adjust the relationship between activation probability and feature vector norm.

[0042] is the eigenvector of The square of the norm reflects the overall strength of the eigenvector. is the eigenvector of norm, taking into account the sparsity of the feature vector. In this way, the activation probability depends not only on the strength of the feature vector, but also on its sparsity, thereby making more comprehensive use of feature information. By dynamically adjusting the activation probability, a probabilistic decision path is generated .

[0043] 2. Collect the prediction result set: Probabilistic decision paths Perform Monte Carlo Dropout forward propagation. In each forward propagation, with probability Randomly discard some neurons, simulate different model structures, and obtain multiple prediction results. Forward propagation, collect and generate a set of prediction results ,in It is The prediction results are obtained by forward propagation. By simulating different model structures multiple times, the diversity of prediction results is increased and the robustness of the model is improved.

[0044] 3. Output a multidimensional risk probability distribution graph with confidence intervals: For the prediction result set Perform Bootstrap resampling calculation. From the original prediction result set Extraction with Replacement samples, construct A new sample set For each new sample set , calculate statistics such as the mean and variance Through multiple resampling, the distribution of statistics is obtained, thereby determining the confidence interval of the prediction results. The final output is a multidimensional risk probability distribution diagram with confidence intervals. , where each dimension represents a different risk factor. The values ​​in the figure represent the probability of abnormal iodine metabolism under different combinations of risk factors and the corresponding confidence intervals, providing more comprehensive and reliable risk assessment information for the subsequent early warning strategy generation.

[0045] As one example, the dynamic early warning strategy generator is constructed based on the improved meta-learning framework, and the risk probability distribution map is mapped to the action space of the early warning strategy, so as to generate the user's iodine metabolism abnormality early warning strategy, including: The prototype construction sub-step is to build a prototype of the strategy execution strategy for historical diagnosis and treatment data, including: A variational autoencoder consisting of a bidirectional LSTM layer and an attention pooling layer is used to extract features from warning strategy records and generate a potential strategy vector set; Performing tensor splicing of the potential strategy vector set and the real-time environment parameter matrix to construct a multi-dimensional strategy feature space; Design a composite metric function that combines cosine similarity and clinical pathway matching to generate a standardized strategy prototype library; The matching optimization sub-step performs strategy matching optimization on the multi-dimensional risk probability distribution graph, including: Inputting the multi-dimensional risk probability distribution graph into a differentiable nearest neighbor algorithm, matching candidate strategies in a strategy prototype library, and generating an initial matching weight matrix; Performing feature interaction entropy analysis on the initial matching weight matrix, determining multi-head attention weight values, and generating a weighted fusion strategy set; Inputting the weighted fusion strategy set into a temporal convolutional network for time series modeling, and outputting a time-coherent early warning strategy sequence; The dynamic tuning sub-step performs dynamic tuning on the early warning strategy sequence, including: Construct a constrained reward function that includes the thyroid iodine metabolism kinetics equation; Performing an adjoint sensitivity analysis on the early warning strategy sequence to calculate a gradient response matrix of strategy parameters to metabolic equations; The parameters in the gradient response matrix are adjusted using a constrained optimization algorithm to output an iodine metabolism abnormality early warning strategy that has been verified by physiological rationality.

[0046] In this embodiment, the purpose is to convert the multidimensional risk probability distribution map into an accurate and individual iodine metabolism abnormality warning strategy. In the prototype construction sub-step, the variational autoencoder composed of a bidirectional LSTM layer and an attention pooling layer is used to mine the potential features of the warning strategy in the historical diagnosis and treatment data. After splicing with the real-time environmental parameter tensor, a standardized strategy prototype library is constructed by integrating the composite metric function of cosine similarity and clinical pathway matching, providing rich prototypes for subsequent strategy matching. In the matching optimization sub-step, the differentiable nearest neighbor algorithm is used to match the candidate strategies that match the risk probability distribution map in the strategy prototype library, and the multi-head attention weights are determined by feature interaction entropy analysis to generate a weighted fusion strategy set, and then the time series modeling is performed through the time convolution network to output a warning strategy sequence that conforms to the time logic. The dynamic tuning sub-step constructs a constrained reward function containing the thyroid iodine metabolic kinetic equation, calculates the gradient response matrix of the strategy parameters to the metabolic equation through the accompanying sensitivity analysis, and uses the constrained optimization algorithm to adjust the parameters to ensure that the warning strategy conforms to physiological rationality. In summary, this embodiment uses multi-step fine processing to make full use of historical diagnosis and treatment data and real-time environmental information. The generated early warning strategy not only takes into account the similarity between strategies and the matching degree of clinical pathways, but also can be flexibly adjusted according to the risk probability distribution diagram, and meets the physiological laws of thyroid iodine metabolism. Compared with traditional methods, it can generate early warning strategies for individual iodine metabolism abnormalities more accurately and dynamically, provide strong support for the prevention and intervention of iodine metabolism abnormalities, improve the accuracy and reliability of early warnings, and provide more valuable references for clinical decision-making.

[0047] For ease of understanding, this embodiment is specifically described below: Prototype building sub-steps: 1. Feature extraction generates a potential strategy vector set: A variational autoencoder (VAE) consisting of a bidirectional long short-term memory network (LSTM) layer and an attention pooling layer is used to process the warning strategy records. Suppose the warning strategy record sequence is ,in Indicates The early warning strategy records of time steps, is the sequence length. The bidirectional LSTM layer processes the sequence from the forward and backward directions respectively, and the calculation formula of the forward LSTM unit is: ; in, They are input gate, forget gate and output gate respectively. It is a sigmoid activation function, which is used to map the input to the [0,1] interval and determine the degree of information transmission. enter To the weight matrix of each control gate and memory unit, is the hidden state at the previous moment To the weight matrix of each control gate and memory unit, is the bias term. is a candidate memory cell, is a memory unit, Represents element-by-element multiplication. The backward LSTM unit is calculated in a similar way, but the processing order is reversed.

[0048] Suppose the forward LSTM output is , the backward LSTM output is , concatenate the two to get Then, through the attention pooling layer, the attention weight is calculated : ; in, is the weight matrix, is a learnable attention vector. Perform weighted summation to obtain feature representation Finally, through the encoding process of the variational autoencoder, Mapping to latent space, generating a latent policy vector , the encoding process can be expressed as ,in and is obtained through the fully connected layer The calculated mean and standard deviation vectors, is a random noise vector that follows a standard normal distribution. After processing all warning strategy records, we get the potential strategy vector set , The number of records for the alert strategy.

[0049] 2. Construct a multi-dimensional strategy feature space: Assume that the real-time environment parameter matrix is , whose dimensions are ,in is the number of environmental parameters, is the time step (if the environment parameters have time series characteristics). and real-time environment parameter matrix For each potential policy vector , which is expanded to Dimensionally compatible tensor form, for example, if is a one-dimensional vector, repeating it Form a Then the expanded potential policy vector tensor is combined with Splicing on a specific dimension to obtain a multi-dimensional strategy feature space Assuming that in dimension If The dimension is , after expansion The dimension is ,but The dimension is Through this splicing method, the potential characteristics of historical early warning strategies are combined with real-time environmental parameters to provide more comprehensive information for subsequent strategy generation.

[0050] 3. Generate a standardized strategy prototype library: Design a composite metric function that combines cosine similarity and clinical pathway matching. Suppose two potential strategy vectors and , cosine similarity for: ; in, yes The transpose of yes of Clinical pathway matching is calculated through a complex function Calculation, this function is based on the clinical pathway knowledge base and considers the degree of match between the disease stage, treatment methods, etc. involved in the strategy and the clinical standard pathway. For example, suppose the clinical pathway knowledge base defines a set of standard strategy feature vectors for different disease stages: , for each potential policy vector , calculate its similarity with each standard strategy feature vector , and then the clinical pathway matching degree is obtained by weighted summation ,in The weights are determined based on clinical experience and expert evaluation. Composite metric function ,in is a weight parameter, and its optimal value is determined through experiments and clinical verification to balance the influence of cosine similarity and clinical pathway matching. Through this composite metric function, all potential strategy vectors are compared pairwise, similar strategies are clustered, and representative strategies are selected as strategy prototypes to build a standardized strategy prototype library. .

[0051] Matching optimization sub-steps: 1. Generate the initial matching weight matrix: Multidimensional risk probability distribution map Enter the differentiable nearest neighbor algorithm. Let The dimension is ,in For the risk dimension, It is the dimension of probability value or other related indicators. Each candidate strategy in , calculate its Distance . Here we use a custom distance metric function, for example: ; in, It is a weight matrix set according to the importance of risk dimension and indicator dimension, which can be determined through analysis of historical data and clinical feedback. is a weight parameter used to balance the impact of correlations between different risk dimensions on the distance. Is the dimension of calculating risk and The corresponding eigenvector and (Here the value on the risk dimension is expanded into a vector form) cosine similarity, is to calculate the Euclidean distance between two eigenvectors. By distance calculation, find The closest candidate strategies to generate the initial matching weight matrix ,in Indicates Candidate strategies and The matching weight can be normalized according to the inverse of the distance, for example .

[0052] 2. Determine the multi-head attention weight values ​​to generate a weighted fusion strategy set: For the initial matching weight matrix Perform feature interaction entropy analysis. Let feature interaction entropy be , for each candidate strategy and risk dimensions , calculate its interaction entropy with other candidate strategies on this risk dimension: ; in, Is a candidate strategy and In the risk dimension , indicator dimension The joint probability distribution on is obtained by statistics of corresponding strategies and risk indicators in historical data. and is the marginal probability distribution. According to the feature interaction entropy, the multi-head attention weight value is determined . Assume the number of attention heads is , for the The weight of each head is calculated as: ; in, is a weight parameter used to adjust the influence of feature interaction entropy on attention weight. Perform weighted fusion to obtain the weighted fusion strategy set ,For example .

[0053] 3. Output a time-coherent sequence of early warning strategies: The weighted fusion strategy set Input the temporal convolutional network (TCN) for time series modeling. Suppose the input of TCN is , whose dimensions are ,in is the number of candidate strategies, time step (assuming the strategy has a time series nature), is the strategy feature dimension. TCN processes the input through a series of convolutional layers. The convolution kernel of the convolution layer is , the bias is , the convolution operation can be expressed as ,in , ReLU is a linear rectifier function. Through multi-layer convolution, the time-dependent features in the strategy sequence are extracted and a time-coherent warning strategy sequence is output. , whose dimensions are , To output the feature dimension, after further processing (such as the fully connected layer), the final time-coherent warning strategy sequence is obtained for subsequent dynamic tuning.

[0054] Dynamic tuning sub-steps: 1. Construct a constrained reward function: Constructing a constrained reward function that includes the thyroid iodine metabolism kinetics equation Assume that the thyroid iodine metabolism kinetic equation is ,in is the iodine content in the thyroid gland. is the iodine uptake rate, is the kinetic parameter, and is thyroid hormone. Suppose the early warning strategy sequence is , the strategy parameters are , the reward function Defined as: ; in, is the thyroid iodine content at time The target value can be determined based on clinical criteria and individual patient conditions. is a weight parameter used to balance the impact of different reward items. The first item measures the deviation of thyroid iodine content from the target value, the second item measures the degree of satisfaction of the thyroid iodine metabolism kinetic equation, and the third item measures the degree of change of the strategy parameters to prevent over-adjustment of the strategy parameters.

[0055] 2. Calculate the gradient response matrix: Early warning strategy sequence Perform accompanying sensitivity analysis and calculate strategy parameters Gradient response matrix for metabolic equation . Let \(L = R(\mathbf{Y},\theta)\), for each policy parameter , and calculate its gradient by the chain rule : ; in, and By deriving the relationship between the thyroid iodine metabolism kinetics equation and the early warning strategy and thyroid iodine content, the gradient of all strategy parameters is formed into a gradient response matrix , whose dimensions are ,in is the time step, is the number of strategy parameters.

[0056] 3. Output an early warning strategy for abnormal iodine metabolism that has been verified by physiological rationality: Use constrained optimization algorithms (such as projected gradient descent) to adjust the gradient response matrix Parameters in . Let the constraints be ,For example The value range of the strategy parameters, the logical relationship between the strategy parameters, etc. In each iteration, the current parameters are calculated. The gradient at , and then satisfy the constraints In the case of : ; in, is a projection operator that projects the updated parameters into a space that satisfies the constraints. is the learning rate. Through multiple iterations, the reward function Maximize, while satisfying physiological rationality constraints, and ultimately output an iodine metabolism abnormality early warning strategy that has been verified by physiological rationality This strategy can provide accurate and reasonable early warning information based on the patient's specific situation and the physiological characteristics of thyroid iodine metabolism.

[0057] As one example, after constructing a dynamic early warning strategy generator based on the improved meta-learning framework and mapping the risk probability distribution graph to the action space of the early warning strategy to generate the abnormal iodine metabolism early warning strategy for the user, the method further includes: Input the iodine metabolism abnormality warning strategy into a preset clinical diagnosis and treatment knowledge base, match the preset medication dosage template and dietary taboo rules, and generate text prompt content including medication time window and taboo foods; Performing display optimization processing on the text prompt content, and dynamically adjusting the font size, background color and information sorting weight according to the confidence interval; The optimized prompt content is pushed to the user's terminal device for display.

[0058] In the present embodiment, after generating the iodine metabolism abnormal warning strategy, further with the help of the preset clinical diagnosis and treatment knowledge base, the warning strategy is matched with it, the information such as the dosage template and the dietary taboo rules are extracted, and the detailed text prompt content is generated, and then through the analysis of the confidence interval, the font size, background color and information sorting weight of the text prompt content are dynamically adjusted to achieve display optimization, and finally the optimized prompt content is accurately pushed to the user terminal device. This process is intended to combine professional medical knowledge with personalized early warning strategies, and improve the effectiveness of information transmission and the user's attention through display optimization. The technical effect of the present embodiment is that it can provide users with more targeted and operational iodine metabolism abnormal warning information, medication time window and clear prompts of taboo foods, which helps users to manage and prevent diseases more scientifically. Display optimization processing improves the readability and attractiveness of information, makes it easier for users to obtain key information, and enhances the practical application value of the early warning strategy. Pushing the optimized content to the terminal device realizes the timely transmission of information, which is convenient for users to view at any time, thereby improving the user's attention to iodine metabolism abnormalities and coping capabilities.

[0059] For ease of understanding, this embodiment is specifically described below: The first step is to match the clinical diagnosis and treatment knowledge base to generate text prompt content: 1. Knowledge base construction and indexing mechanism: The preset clinical diagnosis and treatment knowledge base is a carefully organized and structured knowledge collection. ( is the total number of drugs) as the primary index, under each drug, the severity of the disease is (Assume the value range is - Mild to - severe), age range ( The number of age intervals divided into ( is the number of weight ranges) as a secondary index. For example, for drugs , the severity of the disease is , age range is , weight range is Dosage , determined through a large amount of clinical experimental data and medical research.

[0060] Dietary taboo rules are based on the type of iodine metabolism abnormalities ( The total number of abnormal iodine metabolism types and the level of iodine in the body (divided into - Iodine deficiency -iodine excess) as the index. For example, in the case of iodine metabolism abnormality (Assuming hyperthyroidism with iodine excess) and the iodine level in the body is The corresponding set of taboo foods Contains high-iodine foods such as kelp, seaweed, etc.

[0061] 2. Matching algorithm and calculation: When the iodine metabolism abnormality warning strategy is input into the clinical diagnosis and treatment knowledge base, the patient's condition assessment results in the warning strategy are first analyzed to determine the severity of the condition. , query the patient's age range and weight range For recommended drugs (determined by the early warning strategy), the dosage Calculated by the following formula: ; in, Represents the number of different data sources or research results in the knowledge base for this drug, disease severity, age range, and weight range. It is the weight of each data source, which is determined through a comprehensive evaluation of factors such as data reliability and research sample size. It is The dosage of the drug under the corresponding conditions in each data source.

[0062] For dietary taboo rule matching, according to the type of iodine metabolism abnormality of the patient in the early warning strategy and iodine content in the body , determine the corresponding set of taboo foods At the same time, the determination of the medication time window is combined with the characteristics of the drug and the patient's physical condition to search in the knowledge base. For example, for a certain type of drug, its medication time window is before meals. Minutes to meal minute, and The value of is determined by factors such as the drug's absorption characteristics, metabolic rate, and the patient's gastrointestinal function, and is stored in the knowledge base with a similar indexing mechanism. This information is integrated to generate text prompts, such as "It is recommended that you take medication , the dosage is The medication time window is before meals Minutes to meal minutes. Avoid: ”.

[0063] Step 2: Optimize the display of text prompt content: 1. Relationship between confidence interval and font size: Assume that the confidence interval in the multidimensional risk probability distribution diagram is . Font size adjustment factor Calculated by the following formula: ; in, is a parameter that controls the sensitivity of font size to changes in confidence interval width, The larger it is, the more drastic the response of font size to changes in confidence interval width. is a threshold value, when the confidence interval width Greater than , the font size starts to change dramatically. Used to adjust the font size change. is a basic offset that ensures that the font size has a reasonable minimum value. For example, when the confidence interval is very narrow, that is, hour, , the font size is smaller; when the confidence interval is wide, that is, hour, , the font size is larger to highlight the uncertainty of the risk.

[0064] 2. Background color adjustment algorithm: The background color is adjusted based on the confidence interval and the preset risk threshold. , the high risk threshold is The RGB value of the background color is calculated as follows: ; (Assume that only the red and green channels are used to indicate the risk level, and the blue channel is set to 0); in, and Controls the sensitivity of the red and green channel values ​​to changes in the confidence interval relationship to the risk threshold, respectively. and is the corresponding threshold, and Adjust the range of red and green channel values. When the risk probability is low and the confidence interval is narrow, Smaller, Small value, Larger, When the value is large, the background color is green; when the risk probability is high and the confidence interval is wide, Larger, Big value, Smaller, The smaller the value, the redder the background color.

[0065] 3. Determination of information ranking weight: Assume that the text prompt content contains information collection ,in For medication time window information, Information on prohibited foods, etc. The sorting weight Calculated by the following formula: ; in, It is the number of factors related to abnormal iodine metabolism, such as the severity of the disease, the degree of deviation of iodine content, etc. It is the weight of each factor, which is determined through clinical expert evaluation and analysis of a large amount of user feedback data. Display information Factors The relevance score is, for example, calculated by Keywords and factors The semantic similarity of related words is used to determine the

[0066] Step 3, push the optimized prompt content to the user terminal device: send the encrypted content to the user terminal device through the network. After the user terminal device receives the encrypted content, it uses the pre-stored key and the corresponding decryption function to decrypt it. The decrypted prompt content is displayed to the user in a friendly interface. At the same time, a user feedback function is set, and the user can evaluate the accuracy and clarity of the prompt content. The user's evaluation results are transmitted back to the system server through the network. Based on user feedback, the system optimizes the data in the clinical diagnosis and treatment knowledge base, the parameters in the display optimization algorithm, and the early warning strategy generation process. For example, if a large number of users report that the dietary taboos under a certain type of iodine metabolism abnormality are inaccurate, the system will re-evaluate the relevant dietary taboo rules in the knowledge base, update the corresponding indexes and data, so as to improve the quality of subsequent early warning prompts.

[0067] See also Figure 2 , is a schematic diagram of a structure of an iodine metabolism abnormality early warning device based on multimodal feature fusion provided by an embodiment of the present invention. The iodine metabolism abnormality early warning device based on multimodal feature fusion comprises: The acquisition module 10 is used to collect the high-frequency oscillation signal of the user's urine iodine concentration and the blood metabolite concentration spectrum to construct a time-aligned original metabolic data set; A feature extraction module 11 is used to perform cross-modal feature extraction on the original metabolic data set using an improved dual-stream feature fusion network to generate a fusion feature matrix with physiological relevance, wherein the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum is extracted using an improved graph attention network; The parsing module 12 is used to perform hierarchical parsing on the fusion feature matrix through a dynamic differentiable decision tree to generate a multidimensional risk probability distribution map, wherein the tree node splitting rule of the multidimensional risk probability distribution map adopts a metabolic feature interactive entropy optimization algorithm; The warning generation module 13 is used to construct a dynamic warning strategy generator based on the improved meta-learning framework, map the risk probability distribution map to the action space of the warning strategy, and generate the iodine metabolism abnormality warning strategy for the user.

[0068] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: By first collecting the high-frequency oscillation signal of the user's urine iodine concentration and the blood metabolite concentration spectrum, a time-aligned original metabolic data set is constructed, which can obtain information related to iodine metabolism from multiple dimensions, laying the foundation for subsequent precise analysis. Subsequently, with the help of an improved two-stream feature fusion network, the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum features are extracted using an improved graph attention network, cross-modal feature extraction is achieved, and a fusion feature matrix with physiological relevance is generated. This process deeply explores the intrinsic connection between different modal data, making feature analysis more targeted and accurate. Then, a dynamic differentiable decision tree is used in combination with a metabolic feature interactive entropy optimization algorithm to hierarchically analyze the fusion feature matrix and generate a multidimensional risk probability distribution map, thereby effectively processing complex features to accurately assess the risk of abnormal iodine metabolism. Finally, a dynamic early warning strategy generator is constructed based on an improved meta-learning framework, which maps the risk probability distribution map to the action space of the early warning strategy, generates the user's iodine metabolism abnormality early warning strategy, and enables the early warning strategy to be dynamically adjusted according to different situations. In summary, compared with the prior art that relies only on a single biometric feature and results in low warning accuracy, the embodiments of the present invention achieve a more comprehensive, accurate and dynamic warning of users' iodine metabolism abnormalities through multimodal biometric feature fusion and specific feature extraction and analysis methods.

[0069] As one example, the feature extraction module is specifically used for: Performing multi-scale decomposition on the high-frequency oscillation signal of urine iodine concentration in the original metabolic data set, generating a time-frequency feature matrix based on dynamically optimized wavelet basis function parameters, wherein the scale factor of the wavelet basis function is adaptively adjusted through an error back propagation algorithm; Constructing a dynamic association map according to the blood metabolite concentration spectrum in the original metabolic data set, wherein the nodes of the dynamic association map represent metabolite indicators, the edge weights of the dynamic association map are jointly determined by the metabolic pathway correlation and the time correlation, and the key connection paths of the dynamic association map are screened by a multi-head sparse attention mechanism; Design a bidirectional feature interaction pipeline, use a gated recurrent unit to achieve temporal alignment between urine time-frequency features and blood map features, and eliminate the inter-modal dimensionality difference between the two through a learnable projection matrix; The aligned dual-stream features are input into the feature compression network, and the channel attention mechanism is used to optimize the feature combination to generate a fusion feature matrix with unified dimension and physiological relevance.

[0070] As one example, the parsing module is specifically used for: Performing node splitting optimization on the fused feature matrix includes: Based on the feature vectors in the fused feature matrix, a trainable splitting parameter matrix is ​​constructed to generate a dimensionally matched dynamic threshold vector set; Performing mutual information entropy calculation on the joint distribution of the dynamic threshold vector set and the clinical diagnosis label, and outputting a feature interaction entropy splitting indicator set; Calling the biological association rules in the thyroid metabolic pathway knowledge base, performing anatomical constraint correction on the split indicator set, and generating an optimized node splitting rule set; Performing tree structure optimization on the node splitting rule set includes: Inputting the node splitting rule set into a reinforcement learning controller, generating an initial decision tree structure by defining a state space including node path purity and feature contribution and an action space including branch growth directions; Adopting a proximal strategy optimization algorithm to adjust the branch growth strategy of the initial decision tree structure, and outputting a dynamically optimized decision tree structure; Performing online pruning processing on the decision tree structure according to the product index of the node path purity and the feature contribution, and outputting a simplified decision tree; Performing uncertainty quantification on the streamlined decision tree includes: Inserting a random gating unit into each decision path of the simplified decision tree, dynamically adjusting the activation probability based on the feature vector L2 norm, and generating a probabilistic decision path; Performing Monte Carlo Dropout forward propagation on the probabilistic decision path to collect and generate a set of prediction results; Bootstrap resampling calculation is performed on the prediction result set, and a multidimensional risk probability distribution diagram with confidence intervals is output.

[0071] As one example, the warning generation module is specifically used for: Prototype execution strategies for historical medical data, including: A variational autoencoder consisting of a bidirectional LSTM layer and an attention pooling layer is used to extract features from warning strategy records and generate a potential strategy vector set; Performing tensor splicing of the potential strategy vector set and the real-time environment parameter matrix to construct a multi-dimensional strategy feature space; Design a composite metric function that combines cosine similarity and clinical pathway matching to generate a standardized strategy prototype library; Performing strategy matching optimization on the multi-dimensional risk probability distribution graph includes: Inputting the multi-dimensional risk probability distribution graph into a differentiable nearest neighbor algorithm, matching candidate strategies in a strategy prototype library, and generating an initial matching weight matrix; Performing feature interaction entropy analysis on the initial matching weight matrix, determining multi-head attention weight values, and generating a weighted fusion strategy set; Inputting the weighted fusion strategy set into a temporal convolutional network for time series modeling, and outputting a time-coherent early warning strategy sequence; Dynamically tuning the early warning strategy sequence includes: Construct a constrained reward function that includes the thyroid iodine metabolism kinetics equation; Performing an adjoint sensitivity analysis on the early warning strategy sequence to calculate a gradient response matrix of strategy parameters to metabolic equations; The parameters in the gradient response matrix are adjusted using a constrained optimization algorithm to output an iodine metabolism abnormality early warning strategy that has been verified by physiological rationality.

[0072] See also Figure 3 , is a schematic diagram of an iodine metabolism abnormality warning device based on multimodal feature fusion provided by an embodiment of the present invention. The iodine metabolism abnormality warning device based on multimodal feature fusion of this embodiment includes: a processor 100, a memory 101, and a computer program stored in the memory 101 and executable on the processor 100, such as an iodine metabolism abnormality warning program based on multimodal feature fusion. When the processor 100 executes the computer program, the steps in the above-mentioned embodiments of the iodine metabolism abnormality warning method based on multimodal feature fusion are implemented. Alternatively, when the processor 100 executes the computer program, the functions of each module / unit in the above-mentioned device embodiments are implemented.

[0073] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program in the iodine metabolism abnormality early warning device based on multimodal feature fusion.

[0074] The iodine metabolism abnormality early warning device based on multimodal feature fusion can be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The iodine metabolism abnormality early warning device based on multimodal feature fusion may include, but is not limited to, a processor and a memory. Those skilled in the art will appreciate that the schematic diagram is merely an example of an iodine metabolism abnormality early warning device based on multimodal feature fusion, and does not constitute a limitation on the iodine metabolism abnormality early warning device based on multimodal feature fusion, and may include more or less components than shown in the figure, or combine certain components, or different components, for example, the iodine metabolism abnormality early warning device based on multimodal feature fusion may also include input and output devices, network access devices, buses, etc.

[0075] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the iodine metabolism abnormality early warning device based on multimodal feature fusion, and uses various interfaces and lines to connect the various parts of the iodine metabolism abnormality early warning device based on multimodal feature fusion.

[0076] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the iodine metabolism abnormality early warning device based on multimodal feature fusion by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0077] Wherein, if the module / unit integrated in the iodine metabolism abnormality early warning device based on multimodal feature fusion is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media does not include electrical carrier signals and telecommunication signals.

[0078] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.

[0079] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for early warning of abnormal iodine metabolism based on multimodal feature fusion, characterized in that: The following steps are involved: Collect the user's urine iodine concentration high-frequency oscillation signal and blood metabolite concentration spectrum to construct a time-aligned raw metabolic data set; An improved dual-stream feature fusion network is used to perform cross-modal feature extraction on the original metabolic dataset to generate a fusion feature matrix with physiological relevance, wherein the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum is extracted using an improved graph attention network; The fusion feature matrix is ​​hierarchically analyzed by a dynamic differentiable decision tree to generate a multidimensional risk probability distribution map, wherein the tree node splitting rule of the multidimensional risk probability distribution map adopts a metabolic feature interactive entropy optimization algorithm; A dynamic early warning strategy generator is constructed based on an improved meta-learning framework, and the risk probability distribution map is mapped to the action space of the early warning strategy, so as to generate an iodine metabolism abnormality early warning strategy for the user.

2. The iodine metabolism abnormality early warning method based on multimodal feature fusion according to claim 1, characterized in that: The improved dual-stream feature fusion network is used to perform cross-modal feature extraction on the original metabolic data set to generate a fusion feature matrix with physiological relevance, wherein the urine iodine concentration high-frequency oscillation signal is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum is extracted using an improved graph attention network, including: Feature decomposition and correlation sub-step: performing multi-scale decomposition on the high-frequency oscillation signal of urine iodine concentration in the original metabolic data set, generating a time-frequency feature matrix based on dynamically optimized wavelet basis function parameters, wherein the scale factor of the wavelet basis function is adaptively adjusted through an error back propagation algorithm; Graph structure construction sub-step: constructing a dynamic association graph according to the blood metabolite concentration spectrum in the original metabolic data set, wherein the nodes of the dynamic association graph represent metabolite indicators, the edge weights of the dynamic association graph are jointly determined by the metabolic pathway correlation and the time correlation, and the key connection paths of the dynamic association graph are screened through a multi-head sparse attention mechanism; Cross-modal alignment sub-step: design a bidirectional feature interaction pipeline, use a gated recurrent unit to achieve temporal alignment of urine time-frequency features and blood map features, and eliminate the inter-modal dimensionality difference between the two through a learnable projection matrix; Fusion feature generation sub-step: The aligned dual-stream features are input into the feature compression network, and the channel attention mechanism is used to optimize the feature combination to generate a fusion feature matrix with unified dimension and physiological relevance.

3. The iodine metabolism abnormality early warning method based on multimodal feature fusion according to claim 2, characterized in that: The fusion feature matrix is ​​hierarchically analyzed by a dynamic differentiable decision tree to generate a multidimensional risk probability distribution map, wherein the tree node splitting rule of the multidimensional risk probability distribution map adopts a metabolic feature interactive entropy optimization algorithm, including: The split optimization sub-step performs node split optimization on the fusion feature matrix, including: Based on the feature vectors in the fused feature matrix, a trainable splitting parameter matrix is ​​constructed to generate a dimensionally matched dynamic threshold vector set; Performing mutual information entropy calculation on the joint distribution of the dynamic threshold vector set and the clinical diagnosis label, and outputting a feature interaction entropy splitting indicator set; Invoking biological association rules in a thyroid metabolic pathway knowledge base, performing anatomical constraint correction on the split indicator set, and generating an optimized node splitting rule set; The structure optimization sub-step performs tree structure optimization on the node splitting rule set, including: Inputting the node splitting rule set into a reinforcement learning controller, generating an initial decision tree structure by defining a state space including node path purity and feature contribution and an action space including branch growth directions; Adopting a proximal strategy optimization algorithm to adjust the branch growth strategy of the initial decision tree structure, and outputting a dynamically optimized decision tree structure; Performing online pruning processing on the decision tree structure according to the product index of the node path purity and the feature contribution, and outputting a simplified decision tree; The sub-step of performing uncertainty quantification on the simplified decision tree includes: Inserting a random gating unit into each decision path of the simplified decision tree, dynamically adjusting the activation probability based on the feature vector L2 norm, and generating a probabilistic decision path; Performing Monte Carlo Dropout forward propagation on the probabilistic decision path to collect and generate a set of prediction results; Bootstrap resampling calculation is performed on the prediction result set, and a multidimensional risk probability distribution diagram with confidence intervals is output.

4. The iodine metabolism abnormality early warning method based on multimodal feature fusion according to claim 3, characterized in that: The dynamic early warning strategy generator is constructed based on the improved meta-learning framework, and the risk probability distribution map is mapped to the action space of the early warning strategy, so as to generate the user's iodine metabolism abnormality early warning strategy, including: The prototype construction sub-step is to build a prototype of the strategy execution strategy for historical diagnosis and treatment data, including: A variational autoencoder consisting of a bidirectional LSTM layer and an attention pooling layer is used to extract features from warning strategy records and generate a potential strategy vector set; Performing tensor splicing of the potential strategy vector set and the real-time environment parameter matrix to construct a multi-dimensional strategy feature space; Design a composite metric function that combines cosine similarity and clinical pathway matching to generate a standardized strategy prototype library; The matching optimization sub-step performs strategy matching optimization on the multi-dimensional risk probability distribution graph, including: Inputting the multi-dimensional risk probability distribution graph into a differentiable nearest neighbor algorithm, matching candidate strategies in a strategy prototype library, and generating an initial matching weight matrix; Performing feature interaction entropy analysis on the initial matching weight matrix, determining multi-head attention weight values, and generating a weighted fusion strategy set; Inputting the weighted fusion strategy set into a temporal convolutional network for time series modeling, and outputting a time-coherent early warning strategy sequence; The dynamic tuning sub-step performs dynamic tuning on the early warning strategy sequence, including: Construct a constrained reward function that includes the thyroid iodine metabolism kinetics equation; Performing an adjoint sensitivity analysis on the early warning strategy sequence to calculate a gradient response matrix of strategy parameters to metabolic equations; The parameters in the gradient response matrix are adjusted using a constrained optimization algorithm to output an iodine metabolism abnormality early warning strategy that has been verified by physiological rationality.

5. The iodine metabolism abnormality early warning method based on multimodal feature fusion as claimed in claim 4, characterized in that: After constructing a dynamic early warning strategy generator based on the improved meta-learning framework and mapping the risk probability distribution graph to the action space of the early warning strategy to generate the abnormal iodine metabolism early warning strategy for the user, the method further includes: Input the iodine metabolism abnormality warning strategy into a preset clinical diagnosis and treatment knowledge base, match the preset medication dosage template and dietary taboo rules, and generate text prompt content including medication time window and taboo foods; Performing display optimization processing on the text prompt content, and dynamically adjusting the font size, background color and information sorting weight according to the confidence interval; The optimized prompt content is pushed to the user's terminal device for display.

6. An iodine metabolism abnormality early warning device based on multimodal feature fusion, characterized in that: include: The acquisition module is used to collect the user's urine iodine concentration high-frequency oscillation signal and blood metabolite concentration spectrum to construct a time-aligned original metabolic data set; A feature extraction module, for performing cross-modal feature extraction on the original metabolic data set using an improved dual-stream feature fusion network to generate a fusion feature matrix with physiological relevance, wherein the high-frequency oscillation signal of urine iodine concentration is decomposed using an improved adaptive wavelet basis function, and the blood metabolite concentration spectrum is extracted using an improved graph attention network; A parsing module, used for performing hierarchical parsing on the fusion feature matrix through a dynamic differentiable decision tree to generate a multidimensional risk probability distribution map, wherein the tree node splitting rule of the multidimensional risk probability distribution map adopts a metabolic feature interactive entropy optimization algorithm; The early warning generation module is used to build a dynamic early warning strategy generator based on an improved meta-learning framework, map the risk probability distribution map to the action space of the early warning strategy, and generate an iodine metabolism abnormality early warning strategy for the user.

7. The iodine metabolism abnormality early warning device based on multimodal feature fusion according to claim 6, characterized in that: The feature extraction module is specifically used for: Performing multi-scale decomposition on the high-frequency oscillation signal of urine iodine concentration in the original metabolic data set, generating a time-frequency feature matrix based on dynamically optimized wavelet basis function parameters, wherein the scale factor of the wavelet basis function is adaptively adjusted through an error back propagation algorithm; Constructing a dynamic association map according to the blood metabolite concentration spectrum in the original metabolic data set, wherein the nodes of the dynamic association map represent metabolite indicators, the edge weights of the dynamic association map are jointly determined by the metabolic pathway correlation and the time correlation, and the key connection paths of the dynamic association map are screened by a multi-head sparse attention mechanism; Design a bidirectional feature interaction pipeline, use a gated recurrent unit to achieve temporal alignment between urine time-frequency features and blood map features, and eliminate the inter-modal dimensionality difference between the two through a learnable projection matrix; The aligned dual-stream features are input into the feature compression network, and the channel attention mechanism is used to optimize the feature combination to generate a fusion feature matrix with unified dimension and physiological relevance.

8. The iodine metabolism abnormality early warning device based on multimodal feature fusion according to claim 7, characterized in that: The analysis module is specifically used for: Performing node splitting optimization on the fused feature matrix includes: Based on the feature vectors in the fused feature matrix, a trainable splitting parameter matrix is ​​constructed to generate a dimensionally matched dynamic threshold vector set; Performing mutual information entropy calculation on the joint distribution of the dynamic threshold vector set and the clinical diagnosis label, and outputting a feature interaction entropy splitting indicator set; Invoking biological association rules in a thyroid metabolic pathway knowledge base, performing anatomical constraint correction on the split indicator set, and generating an optimized node splitting rule set; Performing tree structure optimization on the node splitting rule set includes: Inputting the node splitting rule set into a reinforcement learning controller, generating an initial decision tree structure by defining a state space including node path purity and feature contribution and an action space including branch growth directions; Adopting a proximal strategy optimization algorithm to adjust the branch growth strategy of the initial decision tree structure, and outputting a dynamically optimized decision tree structure; Performing online pruning processing on the decision tree structure according to the product index of the node path purity and the feature contribution, and outputting a simplified decision tree; Performing uncertainty quantification on the streamlined decision tree includes: Inserting a random gating unit into each decision path of the simplified decision tree, dynamically adjusting the activation probability based on the feature vector L2 norm, and generating a probabilistic decision path; Performing Monte Carlo Dropout forward propagation on the probabilistic decision path to collect and generate a set of prediction results; Bootstrap resampling calculation is performed on the prediction result set, and a multidimensional risk probability distribution diagram with confidence intervals is output.

9. The iodine metabolism abnormality early warning device based on multimodal feature fusion according to claim 8, characterized in that: The warning generation module is specifically used for: Prototype execution strategies for historical medical data, including: A variational autoencoder consisting of a bidirectional LSTM layer and an attention pooling layer is used to extract features from warning strategy records and generate a potential strategy vector set; Performing tensor splicing of the potential strategy vector set and the real-time environment parameter matrix to construct a multi-dimensional strategy feature space; Design a composite metric function that combines cosine similarity and clinical pathway matching to generate a standardized strategy prototype library; Performing strategy matching optimization on the multi-dimensional risk probability distribution graph includes: Inputting the multi-dimensional risk probability distribution graph into a differentiable nearest neighbor algorithm, matching candidate strategies in a strategy prototype library, and generating an initial matching weight matrix; Performing feature interaction entropy analysis on the initial matching weight matrix, determining multi-head attention weight values, and generating a weighted fusion strategy set; Inputting the weighted fusion strategy set into a temporal convolutional network for time series modeling, and outputting a time-coherent early warning strategy sequence; Dynamically tuning the early warning strategy sequence includes: Construct a constrained reward function that includes the thyroid iodine metabolism kinetics equation; Performing an adjoint sensitivity analysis on the early warning strategy sequence to calculate a gradient response matrix of strategy parameters to metabolic equations; The parameters in the gradient response matrix are adjusted using a constrained optimization algorithm to output an iodine metabolism abnormality early warning strategy that has been verified by physiological rationality.

10. An iodine metabolism abnormality warning device based on multimodal feature fusion, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and when the processor executes the computer program, it implements the iodine metabolism abnormality early warning method based on multimodal feature fusion as described in any one of claims 1 to 5.

Images