Uric acid level prediction method based on multi-source data fusion

Abstract

The invention relates to the technical field of medical artificial intelligence, in particular to a uric acid level prediction method based on multi-source data fusion, and the method comprises the steps: carrying out the synchronous signal collection of a finger tip of a target user through a near-infrared hyperspectral collection device and a flexible microfluidic impedance sensor; obtaining a subcutaneous microvascular hyperspectral image sequence and an epidermal sweat electrochemical impedance spectrum; and performing Euler video amplification processing on the subcutaneous microvascular hyperspectral image sequence to obtain a blood flow micro-dynamic enhancement sequence. According to the method, a calculation and compensation mechanism for physiological response time lag is introduced, nonlinear time sequence alignment is carried out by further utilizing a dynamic time warping algorithm and taking the time lag as a constraint window when the global time lag is obtained, and compared with a traditional simple timestamp alignment method, it is guaranteed that actually-related physiologically features can be accurately matched, and the accuracy of time sequence alignment is improved. High-quality and synchronous input data is provided for subsequent feature fusion, and the logic leakproofness and accuracy of the whole prediction method are improved.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to a method for predicting uric acid levels based on multi-source data fusion. Background Technology

[0002] Hyperuricemia is a chronic metabolic disease caused by disordered purine metabolism in the body. It is the main biochemical basis for gout attacks and is also closely related to kidney disease, cardiovascular disease, and diabetes. Therefore, monitoring serum uric acid levels is of significant clinical importance for the early warning of hyperuricemia. Traditional invasive blood tests suffer from drawbacks such as pain and discontinuity, while existing single-modal noninvasive tests suffer from insufficient accuracy due to low signal-to-noise ratio and susceptibility to interference. Therefore, noninvasive tests that integrate multiple physiological information sources are a direction that urgently needs to be explored in this field.

[0003] Multi-source data fusion is considered an effective way to improve the robustness of non-invasive detection, such as combining hemodynamics and metabolite excretion. However, existing technologies face two major challenges when fusing heterogeneous data. First, changes in uric acid in the blood are transmitted to epidermal sweat for an electrochemical response, which involves an inherent, non-linear physiological response lag. Traditional simple timestamp alignment methods ignore this lag, leading to mismatches between asynchronous blood flow features and metabolic features. Second, high-dimensional image sequences and low-dimensional feature sequences differ greatly in dimensionality and structure, making it difficult to effectively stitch or couple them to construct a unified spatiotemporal feature representation. Given these shortcomings, traditional multi-source fusion models cannot effectively capture the complex spatiotemporal dependencies and deep coupling relationships between blood flow microdynamics and sweat metabolites, resulting in limited generalization ability and accuracy of prediction models. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting uric acid levels based on multi-source data fusion, comprising: Synchronous signal acquisition of the target user's fingertip is performed using near-infrared hyperspectral acquisition equipment and flexible microfluidic impedance sensor to obtain subcutaneous microvascular hyperspectral image sequence and epidermal sweat electrochemical impedance spectrum; The hyperspectral image sequence of subcutaneous microvessels was subjected to Euler video amplification processing to obtain a blood flow micro-dynamic enhancement sequence; the electrochemical impedance spectroscopy of epidermal sweat was subjected to wavelet packet decomposition processing to obtain metabolic impedance characteristic components of different frequency bands. Calculate the physiological response time delay between the blood flow microdynamic enhancement sequence and the metabolic impedance characteristic component; based on the physiological response time delay, perform nonlinear temporal alignment on the blood flow microdynamic enhancement sequence and the metabolic impedance characteristic component to obtain the aligned blood flow characteristic matrix and the aligned impedance characteristic vector; The aligned impedance feature vectors are broadcast spatially and channel-cascaded with the aligned blood flow feature matrix to construct a vascular-metabolic coupled multidimensional feature tensor. The blood vessel-metabolism coupled multidimensional feature tensor is input into a pre-trained spatiotemporal attention map neural network model for feature encoding to obtain a high-dimensional latent space vector. The high-dimensional latent space vector is input into a preset nonlinear regression mapping function for solution to obtain the predicted value of serum uric acid concentration for the target user.

[0005] Preferably, the subcutaneous microvascular hyperspectral image sequence is subjected to Euler video amplification processing to obtain a blood flow microdynamic enhancement sequence, including: The subcutaneous microvascular hyperspectral image sequence was subjected to Laplacian pyramid decomposition to obtain multi-scale spatial frequency components; The multi-scale spatial frequency components are subjected to time-domain bandpass filtering to extract weak blood flow fluctuation signals consistent with the target user's heart rate range; The amplitude of the weak blood flow fluctuation signal is amplified and reconstructed by weighting it with the corresponding level of the Laplace pyramid to generate the blood flow micro-dynamic enhancement sequence.

[0006] Preferably, the electrochemical impedance spectroscopy of the epidermal sweat is subjected to wavelet packet decomposition to obtain metabolic impedance characteristic components in different frequency bands, including: Electrochemical impedance spectroscopy of the epidermal sweat was performed using a pre-defined Symlets wavelet basis. Full layer decomposition to obtain the first layer. Layer Wavelet packet coefficients; Based on a pre-set uric acid impedance response database, the characteristic frequency bands that are correlated with changes in uric acid concentration higher than a preset correlation threshold are matched and locked. Extracting the characteristic frequency band in the The wavelet packet coefficients corresponding to the layer total decomposition are used as the target node; The total energy and Shannon entropy of the target node within the acquisition time window are calculated, and the total energy and Shannon entropy are normalized to generate the metabolic impedance characteristic components of the different frequency bands.

[0007] Preferably, calculating the physiological response lag between the blood flow microdynamic enhancement sequence and the metabolic impedance characteristic components includes: Data within the same predetermined time window are extracted from the blood flow micro-dynamic enhancement sequence and the metabolic impedance characteristic component, respectively, and the extracted data are preprocessed to obtain blood flow time-series signal and impedance time-series signal; Set a maximum expected time delay value to define a time delay search window; Within the time delay search window, a step-wise cross-correlation operation is performed on the blood flow time-series signal and the impedance time-series signal to generate a cross-correlation coefficient curve, wherein the horizontal axis of the cross-correlation coefficient curve represents the time delay and the vertical axis represents the correlation coefficient value. Locate the peak point of the cross-correlation coefficient curve and extract the horizontal coordinate time delay corresponding to the peak point as the physiological response time delay.

[0008] Preferably, based on the physiological response time delay, nonlinear temporal alignment is performed on the blood flow microdynamic enhancement sequence and the metabolic impedance feature components to obtain an aligned blood flow feature matrix and an aligned impedance feature vector, including: The pulse wave peak point in the blood flow micro-dynamic enhancement sequence is extracted as the first time reference, and the impedance peak point in the metabolic impedance characteristic component is extracted as the second time reference. A search constraint window based on the physiological response time delay is constructed, and within the search constraint window, the minimum bend path between the first time series reference and the second time series reference is calculated using a dynamic time warping algorithm; Based on the synchronization mapping index generated by the minimum bend path, the blood flow micro-dynamic enhancement sequence and the metabolic impedance feature components are resampled to generate the aligned blood flow feature matrix and the aligned impedance feature vector.

[0009] Preferably, the aligned impedance eigenvectors are spatially broadcast and channel-cascaded with the aligned blood flow eigenvector matrix, including: Obtain the spatial dimension of the aligned blood flow feature matrix; The aligned impedance feature vector is copied as a feature layer with the same spatial dimension; The aligned blood flow feature matrix and the feature layer are concatenated along the channel dimension to construct the blood vessel-metabolism coupled multidimensional feature tensor.

[0010] Preferably, the vascular-metabolic coupled multidimensional feature tensor is input into a pre-trained spatiotemporal attention map neural network model for feature encoding to obtain a high-dimensional latent space vector, including: The pixel region of the spatial feature layer corresponding to each time step in the blood vessel-metabolism coupled multidimensional feature tensor is defined as a graph node, and the dynamic connection edge weights between the graph nodes are defined according to the metabolic impedance feature components to construct time series graph structure data. The temporal graph structure data is input into the spatiotemporal attention graph neural network model, and the graph convolutional layer is used to capture the hemodynamic correlation of the graph nodes in the spatial dimension to generate spatial fusion features; The spatial fusion features are input into the long short-term memory network layer in the spatiotemporal attention map neural network model to extract the evolution law of the spatial fusion features in the time dimension and generate time-dependent features; The key time steps in the temporal-dependent features are weighted by the attention mechanism layer in the spatiotemporal attention graph neural network model, and the weighted features are then globally pooled to output the high-dimensional latent space vector.

[0011] Preferably, the high-dimensional latent space vector is input into a preset nonlinear regression mapping function for solution to obtain the predicted serum uric acid concentration of the target user, including: The high-dimensional latent space vector is used as input and fed into the nonlinear regression mapping function; wherein, the nonlinear regression mapping function consists of at least three fully connected layers and an activation function; The high-dimensional latent space vector is subjected to dimensional transformation and nonlinear fitting through the fully connected layer to extract deep abstract features related to uric acid concentration layer by layer. The feature vector output by the last fully connected layer is passed through a linear output layer to map it to the target numerical range of uric acid concentration. The results of the linear output layer are output as the predicted serum uric acid concentration for the target user.

[0012] Preferably, the training process of the pre-trained spatiotemporal attention map neural network model includes: Multiple sets of the aforementioned vascular-metabolic coupling multidimensional feature tensors with known serum uric acid concentration labels were collected to construct a training dataset; The training dataset is input into the spatiotemporal attention map neural network model, and the loss between the model prediction value and the serum uric acid concentration label is calculated using the mean squared error loss function. An adaptive moment estimator optimizer is used to minimize the loss function, and the trainable parameters in the graph neural network model are iteratively adjusted. The parameters of the trained graph neural network model are frozen and used as a preset model for feature encoding.

[0013] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention introduces a calculation and compensation mechanism for physiological response time delay. After obtaining the global time delay, the dynamic time warping algorithm is further used and the time delay is used as a constraint window for nonlinear time alignment. Compared with the traditional simple timestamp alignment method, it ensures that physiologically relevant features can be accurately matched, eliminates the error caused by time mismatch, provides high-quality and synchronous input data for subsequent feature fusion, and improves the logical rigor and accuracy of the entire prediction method. (2) This invention successfully constructs a vascular-metabolic coupled multidimensional feature tensor by broadcasting the aligned impedance feature vector in the spatial dimension and cascading it with the aligned blood flow feature matrix. This tensor simultaneously represents the spatial hemodynamic information of the fingertip microvascular region and the local metabolite concentration information of epidermal sweat in the same data structure. This coupling method makes full use of the complementarity of the two types of data, forming a feature representation that is more physiologically meaningful and discriminative than a single modality or simple splicing, providing rich and strongly correlated inputs for deep learning models. (3) This invention encodes features by inputting the multidimensional feature tensor of blood vessel-metabolism coupling into a spatiotemporal attention graph neural network model. The model captures the hemodynamic correlation of features in the spatial dimension through graph convolutional layers, captures the evolution law of features in the temporal dimension through long short-term memory network, and weights key temporal information through attention mechanism. This enables efficient modeling of complex spatiotemporal dependencies and deep nonlinear couplings in high-dimensional features, enhances the model's ability to extract deep abstract features related to uric acid concentration, and thus ensures the high accuracy and high reliability of the final serum uric acid concentration prediction value. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart of the overall method in one embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 This invention provides a technical solution: a method for predicting uric acid levels based on multi-source data fusion, comprising: S1. Synchronously acquire signals from the fingertips of the target user using near-infrared hyperspectral acquisition equipment and flexible microfluidic impedance sensor to obtain subcutaneous microvascular hyperspectral image sequences and epidermal sweat electrochemical impedance spectra. S2. Euler video amplification processing was performed on the hyperspectral image sequence of subcutaneous microvessels to obtain the enhanced sequence of blood flow micro-dynamics; wavelet packet decomposition processing was performed on the electrochemical impedance spectrum of epidermal sweat to obtain the metabolic impedance characteristic components of different frequency bands. S3. Calculate the physiological response time delay between the blood flow micro-dynamic enhancement sequence and the metabolic impedance characteristic components; based on the physiological response time delay, perform nonlinear time alignment on the blood flow micro-dynamic enhancement sequence and the metabolic impedance characteristic components to obtain the aligned blood flow characteristic matrix and the aligned impedance characteristic vector. S4. Broadcast the aligned impedance feature vectors spatially and cascade them with the aligned blood flow feature matrix to construct the vascular-metabolic coupling multidimensional feature tensor. S5. Input the vascular-metabolic coupled multidimensional feature tensor into a pre-trained spatiotemporal attention map neural network model for feature encoding to obtain a high-dimensional latent space vector. S6. Input the high-dimensional latent space vector into a preset nonlinear regression mapping function for solution to obtain the predicted value of serum uric acid concentration for the target user.

[0017] It should be noted that the near-infrared hyperspectral acquisition device is a device capable of simultaneously acquiring spatial image information and spectral information. It utilizes near-infrared light to penetrate the skin and reach subcutaneous microvessels, capturing the reflection or absorption of light at different wavelengths to obtain data related to blood components and blood flow dynamics. The flexible microfluidic impedance sensor is a sensor that can be attached to the skin surface. It analyzes the concentration information of metabolites such as uric acid in sweat by applying a weak current to the epidermal sweat and measuring its resistance and capacitance parameters, i.e., electrochemical impedance spectroscopy. Synchronous signal acquisition refers to the coordinated acquisition of data from the same part of the user's fingertip by two devices within the same time period to ensure the temporal comparability of subsequent data. Physiological response lag refers to the time difference between the change in uric acid concentration in the blood and the final manifestation of this change in sweat. It is a key physiological parameter for accurate data alignment in this invention. Nonlinear temporal alignment refers to the use of algorithms such as dynamic time warping to stretch and compress two feature sequences on the time axis according to the physiological response lag to eliminate nonlinear time differences in data acquisition and physiological transmission, thereby achieving accurate matching.

[0018] In an optional embodiment, the subcutaneous microvascular hyperspectral image sequence is subjected to Euler video amplification processing to obtain a blood flow microdynamic enhancement sequence, including: Laplacian pyramid decomposition was performed on subcutaneous microvascular hyperspectral image sequences to obtain multi-scale spatial frequency components; Time-domain bandpass filtering is performed on the multi-scale spatial frequency components to extract weak blood flow fluctuation signals consistent with the target user's heart rate range; The amplitude of weak blood flow fluctuation signals is amplified and reconstructed by weighting with the corresponding level of the Laplace pyramid to generate a blood flow micro-dynamic enhancement sequence.

[0019] It's important to note that Euler video upscaling is a signal processing-based video processing technique. Instead of physically moving the camera, it extracts and amplifies the minute vibrations of pixels in a video image sequence over time, such as changes in skin color and brightness caused by a heartbeat, making these subtle movements easier for the human eye or algorithms to detect. Laplacian pyramid decomposition is an image decomposition technique that breaks down each frame of a hyperspectral image into a series of sub-bands with different spatial frequencies, i.e., multi-scale spatial frequency components. For example, low-frequency components represent smooth areas and macroscopic structures, while high-frequency components represent texture and details. Temporal bandpass filtering refers to allowing only signals within a specific frequency range to pass through in the time dimension. Here, it specifically refers to extracting blood flow fluctuation frequencies consistent with the user's heart rate, such as 60 to 120 times per minute, thereby effectively filtering out environmental noise and random motion.

[0020] Specifically, this embodiment employs the Laplacian pyramid decomposition algorithm to decompose each frame of the hyperspectral image into... Layers contain different spatial frequency components. The bottom layer (layer 0) contains high-frequency details of the image, such as skin texture, while the top layer contains... Each layer contains a low-frequency overview of the image, such as illumination distribution. For each level of the pyramid, pixel value changes at the same spatial location are monitored over time. An infinite impulse response bandpass filter is used to extract signals within a specific frequency range, which is set to match the normal and pathological heart rate ranges of the human body. This separates the weak color or brightness change signals caused by blood flow pulsation, i.e., weak blood flow fluctuation signals. The amplitude of the weak blood flow fluctuation signals is amplified and weighted for reconstruction with the corresponding level of the Laplacian pyramid. The extracted weak fluctuation signals are then multiplied by a preset amplification factor. The enhanced wave signal is obtained; then, the enhanced wave signal is superimposed back onto the original Laplace pyramid component; finally, the blood flow micro-dynamic enhancement sequence is reconstructed by inverse transformation of the Laplace pyramid, that is, by upsampling layer by layer from the top layer and superimposing it with the next layer.

[0021] As an application illustration of the above operations, regarding the number of decomposition levels in the Laplace pyramid decomposition... The pyramid is set to 4 to 6 layers; too few layers will not effectively separate noise, while too many layers will result in computational redundancy. When constructing the pyramid, a standard 5×5 Gaussian kernel is used for smoothing and downsampling operations. The Gaussian kernel parameters are... The frequency response is generally set between 1.0 and 1.4 to ensure a smooth transition of image edge information. For the infinite impulse response bandpass filter, a third- or fourth-order Butterworth bandpass filter is preferred. Compared to an ideal filter, it has a flat frequency response within the passband, effectively avoiding ringing effects. The cutoff frequency setting considers that the target user group may include healthy individuals and those with metabolic abnormalities, resulting in a wider range of heart rate fluctuations. Therefore, the low-frequency cutoff frequency of the bandpass filter is... Set to 0.6Hz, corresponding to a heart rate of 36 bpm, high-frequency cutoff frequency. Set to 2.5Hz, corresponding to a heart rate of 150 bpm; this setting accurately covers most physiological blood flow pulsation frequencies while filtering out extremely low-frequency baseline drift such as respiratory movements and high-frequency camera sensor thermal noise; amplification factor The value range is typically from 10 to 50. In this embodiment, to prevent signal oversaturation or excessive noise amplification, an adaptive amplification strategy is adopted, and the specific formula can be expressed as: ; in The original signal, This is the filtered fluctuation signal.

[0022] It should be further explained that, due to the inherent thermal noise of hyperspectral image sensors, and the fact that minute textures on the skin surface, such as pores and skin lines, correspond to the high spatial frequency layer of the Laplacian pyramid, directly amplifying this layer would lead to severe artifacts and increased noise, masking real blood flow changes. Therefore, this embodiment introduces a spatial cutoff frequency. To suppress noise, for spatial wavelengths smaller than 100 nm in the Laplace pyramid... The high-frequency stages reduce their corresponding amplification factor through a linear attenuation function. For example, setting For pixels smaller than this value, no magnification or significant reduction in magnification is applied. This processing can significantly improve the signal-to-noise ratio of blood flow micro-dynamic enhancement sequences, ensuring that the data input to the neural network mainly reflects the spectral absorption differences caused by changes in blood volume, rather than sensor noise.

[0023] In an optional embodiment, wavelet packet decomposition is performed on the electrochemical impedance spectroscopy of epidermal sweat to obtain metabolic impedance characteristic components in different frequency bands, including: Electrochemical impedance spectroscopy of epidermal sweat was performed using a pre-defined Symlets wavelet basis. Full layer decomposition to obtain the first layer. Layer Wavelet packet coefficients; Based on a pre-set uric acid impedance response database, the characteristic frequency bands that are correlated with changes in uric acid concentration higher than a preset correlation threshold are matched and locked. Extracting feature frequency bands in The wavelet packet coefficients corresponding to the layer total decomposition are used as the target nodes; The total energy and Shannon entropy of the target node within the acquisition time window are calculated, and the total energy and Shannon entropy are normalized to generate metabolic impedance characteristic components of different frequency bands.

[0024] It should be noted that wavelet packet decomposition is a refined time-frequency analysis method that decomposes the electrochemical impedance spectrum into multiple non-overlapping frequency bands, offering higher frequency resolution compared to wavelet decomposition; Symlets wavelet basis is a specific type of wavelet function chosen as the basis for signal decomposition. Layer decomposition refers to the process of decomposing the original signal into layers. The next iteration of decomposition eventually produces Wavelet packet coefficients reflecting details at different frequencies; the uric acid impedance response database is a pre-established medical / biochemical database that records the response characteristics of sensor impedance spectra at different frequencies and their correlation with uric acid concentration at different uric acid concentrations; Shannon entropy is a concept in information theory, used here to quantify the complexity or information richness of the characteristic frequency band represented by the wavelet packet coefficients, and together with the total energy, it is an important component of metabolic impedance components.

[0025] As an application illustration of the above operations, the Symlets wavelet basis is preferably Sym4 or Sym5 wavelets. These wavelets have good symmetry and relatively compact support, making them suitable for processing mutations and oscillations in bioelectrochemical signals; decomposition layer number The preferred setting is 4 layers, at which point it will generate The number of characteristic frequency bands, while ensuring frequency resolution, avoids feature redundancy that may result from excessive decomposition. Assuming the sampling frequency of the original electrochemical impedance spectroscopy is... Then the bandwidth of each characteristic frequency band is Next, based on the pre-established uric acid impedance response database, queries were performed. Within each wavelet packet frequency band, which bands exhibit a high correlation between impedance changes and actual uric acid concentration changes? By matching and identifying frequency bands with correlation coefficients higher than a preset correlation threshold, such as the Pearson correlation coefficient, the wavelet packet coefficients corresponding to these bands are defined as target nodes. Then, the preset correlation threshold is set to the Pearson correlation coefficient. This high threshold ensures that only those frequency bands that have a strong specific response to changes in uric acid concentration are selected, thereby effectively filtering out non-specific frequency bands related to sweat pH, temperature, or other interfering metabolites. It should be further explained that the formula for calculating the total energy is: ; in, The total number of coefficients within this node, The modulus of the coefficient; The formula for calculating Shannon entropy is: ; in, Represents the Shannon entropy value; Indicates the first The relative probability proportion of the energy of each wavelet packet coefficient in the total energy of the current frequency band, i.e. When agreed hour, ; The normalization calculation formula is: ; .

[0026] In an optional embodiment, calculating the physiological response lag between the blood flow microdynamic enhancement sequence and the metabolic impedance characteristic components includes: Data within the same predetermined time window were extracted from the blood flow microdynamic enhancement sequence and the metabolic impedance characteristic component, respectively, and the extracted data were preprocessed to obtain the blood flow time series signal and the impedance time series signal. Set a maximum expected time delay value to define a time delay search window; Within the time delay search window, a step-wise cross-correlation operation is performed on the blood flow time-series signal and the impedance time-series signal to generate a cross-correlation coefficient curve, where the horizontal axis of the cross-correlation coefficient curve represents the time delay and the vertical axis represents the correlation coefficient value. Locate the peak point of the cross-correlation coefficient curve and extract the corresponding horizontal axis time delay as the physiological response time delay.

[0027] It should be noted that physiological response lag is the objective time difference between the transmission of dynamic changes in blood flow to changes in sweat metabolites; step-wise cross-correlation is a signal processing method used to measure the similarity of two time-series signals at different time offsets; specifically, one signal is fixed, and the other signal is shifted in time by a preset minimum step size, i.e., the lag, and the correlation coefficient between the two signals is calculated for each shift; the cross-correlation curve is a curve with the lag as the horizontal axis and the correlation coefficient as the vertical axis, which shows the correlation strength of the two signals under all possible lags; the peak point is the point with the largest cross-correlation coefficient, and the corresponding lag is the delay time when the two signals reach the best match, which is considered to be the best estimate of physiological response lag.

[0028] As an application of the above operations, for blood flow micro-dynamic enhancement sequences, video frames within a predetermined time window are extracted from the blood flow micro-dynamic enhancement sequences. A region of interest is defined for each frame, and the average grayscale value or specific color channel value of that region is calculated. These values ​​are then arranged sequentially by frame to form a blood flow temporal signal reflecting the pulsation state of blood flow. For extracting metabolic impedance feature components from data within the same predetermined time window, since step S3 calculates the energy entropy feature based on wavelet packets, a sliding window sampling method is used here to arrange the continuously calculated normalized feature values ​​in chronological order, forming an impedance time-series signal reflecting the trend of metabolite concentration fluctuations. Subsequently, on and Z-score normalization was performed separately to eliminate the effects of differences in dimensions and amplitudes between the two; a maximum expected time delay value was set. Based on the principles of human physiology, an asymmetric time-delay search window is defined. This window excludes cases with negative time delay, i.e., the unreasonable situation where changes in sweat precede changes in blood, and limits the search range to reduce computational load; in the time delay search window Within, with a preset minimum step size Sliding blood flow timing signal And calculate its relationship with impedance timing signal. Normalized cross-correlation coefficients between them, generating cross-correlation coefficient curves. ; Traversing the generated cross-correlation coefficient curves The global maximum point of the location curve, i.e., the peak point, is then used to extract the horizontal coordinate time delay corresponding to that peak point. This was identified as a physiological response time delay.

[0029] In an optional embodiment, based on physiological response time delay, nonlinear temporal alignment is performed on the enhanced blood flow microdynamic sequence and metabolic impedance characteristic components to obtain an aligned blood flow characteristic matrix and an aligned impedance characteristic vector, including: The pulse wave peak point in the blood flow micro-dynamic enhancement sequence was extracted as the first time reference, and the impedance peak point in the metabolic impedance characteristic component was extracted as the second time reference. A search constraint window based on physiological response time delay is constructed, and within the search constraint window, the minimum bend path between the first time series reference and the second time series reference is calculated using a dynamic time warping algorithm; Based on the synchronization mapping index generated by the minimum bend path, the blood flow microdynamic enhancement sequence and metabolic impedance feature components are resampled to generate aligned blood flow feature matrices and aligned impedance feature vectors.

[0030] It should be noted that the pulse wave peak point and impedance peak point represent significant physiological events in blood flow microdynamics and metabolite response, respectively. They are selected as time series benchmarks to guide the nonlinear alignment process. The search constraint window is a time range set based on the physiological response delay calculated in the previous step. It is used to limit the search space of the dynamic time warping algorithm, ensuring that the alignment results conform to a physiologically reasonable range and preventing the algorithm from forcibly matching irrelevant points. The dynamic time warping algorithm is a method for calculating the similarity between two time series. Even if they have nonlinear changes or distortions on the time axis, it can find the optimal match. The minimum bend path is the optimal path calculated by the dynamic time warping algorithm to describe how the two time series benchmark points correspond to each other. The synchronization mapping index is the mapping relationship provided by this path. It tells the system how to stretch, compress, or repeat the original blood flow and impedance sequences on the time axis to obtain fully synchronized alignment features on the time axis.

[0031] As an application illustration of the above operations, the constructed blood flow timing signal... Perform peak detection and set the minimum peak spacing parameter. (e.g., 0.5) Identify all local maxima and construct a blood flow pulse wave peak sequence as the first time series reference. For the constructed impedance timing signal Peak detection was performed to identify local maxima reflecting dramatic fluctuations in metabolite concentrations, forming an impedance peak sequence as a second time-series reference. Then assume the blood flow sequence length is The impedance sequence length is Build a The mesh is then defined, and a constraint window is defined that only allows offsets on the diagonal. A path search is performed within the nearby area, and the specific constraints can be expressed as follows: ,in , These are the time indices of the two sequences. The constraint window, with its preset bandwidth, forces the algorithm to find the best match only within a reasonable physiological delay range, preventing the algorithm from incorrectly aligning irrelevant signal segments. It calculates the Euclidean distance between points in the two sequences, fills the grid points within the constraint window, and forms the cumulative cost matrix, starting from the grid endpoint. Begin by finding a path to the starting point that minimizes cumulative cost; this path consists of a series of index pairs: ,in Represents the first blood flow sequence Frame and impedance sequence Each sampling point is physiologically synchronously corresponding, and finally, they are arranged according to the index sequence. Image frames are extracted from the original enhanced blood flow micro-dynamic sequence. If the index is a duplicate value, the frame is copied; if the index span is large, inter-frame interpolation is performed. The resulting images are then stacked to form a sequence with dimensions of [dimension not specified]. The three-dimensional matrix, i.e., the aligned blood flow feature matrix; according to the index sequence The corresponding feature values ​​are extracted from the metabolic impedance feature components, and similarly replicated or interpolated to form a length of [length missing]. A one-dimensional vector, i.e., an aligned impedance eigenvector.

[0032] In an optional embodiment, spatial dimension broadcasting of the aligned impedance eigenvectors and channel concatenation with the aligned blood flow feature matrix includes: Obtain the spatial dimension of the aligned blood flow feature matrix; Copy the aligned impedance eigenvectors to a feature layer with the same spatial dimensions; The aligned blood flow feature matrix and feature layer are concatenated along the channel dimension to construct a vascular-metabolic coupled multidimensional feature tensor.

[0033] It should be noted that spatial dimension broadcasting is a data processing operation used to expand low-dimensional features to have the same spatial size as high-dimensional features. Specifically, impedance feature vectors themselves do not contain spatial information. Through a copying operation, they are transformed into a two-dimensional feature layer with the same value in all spatial locations, enabling them to be fused with the blood flow feature matrix, which contains spatial information. Channel cascading refers to stacking and splicing two or more feature layers with the same spatial dimension in the channel dimension, thereby forming a vascular-metabolic coupled multidimensional feature tensor containing dynamic spatial information of blood flow and global metabolic information, enabling neural networks to process these two heterogeneous data simultaneously.

[0034] As an application illustration of the above operations, the aligned blood flow feature matrix is ​​read. Assuming its dimension is ,in The time step (number of frames) after alignment. and For the height and width of the space, The number of channels in the hyperspectral image, while simultaneously reading the aligned impedance eigenvectors. Its dimensions are ,in This refers to the dimension of impedance characteristics; since impedance characteristics are global quantities reflecting systemic or local metabolic levels and lack spatial resolution, they need to be expanded to the same spatial scale as blood flow characteristics. Each time step in eigenvectors In spatial dimension and To perform tiling, that is, to copy and tile for each pixel position in the image. All are filled with the same impedance characteristic value. After the operation is completed, a dimension is generated. Impedance feature layer In the channel dimension, the aligned blood flow feature matrix will be... With impedance feature layer The concatenation is performed, and the dimensions of the concatenated tensor become... This tensor is the vascular-metabolic coupling multidimensional feature tensor. .

[0035] It should be further explained that the physical significance of this fusion method lies in the fact that each spatial node in the tensor contains both local microvascular hemodynamic information and is endowed with the current metabolic environment context information. This pixel-level embedding enables the subsequent neural network to learn what specific spatiotemporal response pattern microvascular blood flow should exhibit at a specific metabolic level.

[0036] In an optional embodiment, the vascular-metabolic coupled multidimensional feature tensor is input into a pre-trained spatiotemporal attention map neural network model for feature encoding to obtain a high-dimensional latent space vector, including: The pixel region of the spatial feature layer corresponding to each time step in the vascular-metabolic coupled multidimensional feature tensor is defined as a graph node, and the dynamic connection edge weights between graph nodes are defined according to the metabolic impedance feature components to construct time series graph structure data. The temporal graph structure data is input into the spatiotemporal attention graph neural network model, and the graph convolutional layer is used to capture the hemodynamic correlation of graph nodes in the spatial dimension to generate spatial fusion features. Spatial fusion features are input into the long short-term memory network layer in the spatiotemporal attention graph neural network model to extract the evolution law of spatial fusion features in the time dimension and generate time-dependent features; By using the attention mechanism layer in the spatiotemporal attention graph neural network model, the key time steps in the temporally dependent features are weighted, and the weighted features are then globally pooled to output a high-dimensional latent space vector.

[0037] It should be noted that the spatiotemporal attention graph neural network model is a deep learning architecture, particularly suitable for processing complex data with both spatial correlation and temporal dependence. Graph nodes correspond to each pixel or local region of the image in the coupled feature tensor, representing the vascular-metabolic state of that region. Graph convolutional layers are used for feature transfer and aggregation on the graph structure, capable of capturing the spatial interactions between different pixels within the fingertip microvascular region, i.e., hemodynamic correlations. Long Short-Term Memory (LSTM) layers are a special type of recurrent neural network specifically designed to capture dependencies and evolutionary patterns in long-term time series, used here to analyze feature changes over time. The attention mechanism layer automatically learns and determines which key time steps, i.e., which physiological changes at which moments are most important for the final uric acid prediction, and assigns them higher weights, thereby improving the model's ability to focus on key information.

[0038] As an application illustration of the above operations, the image in each frame... Each grid region is defined as Each graph node, i.e. Construct an adjacency matrix between nodes. Unlike traditional fixed-distance matrices, this embodiment introduces metabolic impedance characteristic components. As a dynamic adjustment factor, two nodes and In time edge weight The calculation formula is: ; in, It is a node With nodes The Euclidean distance between them The width of the Gaussian kernel. The learnable adjustment coefficient, The aligned impedance eigenvector obtained in step S4 at time step The value indicates that the strength of the synergistic relationship between microvascular network nodes will also dynamically adjust when the body fluid impedance changes. The constructed graph structure data is input into the graph convolutional layer, and Chebyshev graph convolution or a first-order approximate graph convolution operator is used, with the following formula: ; , For adjacency matrix with added self-loops The model contains two graph convolutional layers. The first layer has 64 output channels and the second layer has 128 output channels. This process captures the spatial cooperative fluctuation pattern of local microvessels and generates spatial fusion features. then After the spatially fused feature sequence output by the graph convolution is flattened, it is input into a bidirectional long short-term memory (LSTM) network. The bidirectional LSM network layer contains 256 hidden units, which processes the time series from both the forward and backward directions, capturing the evolution of hemodynamics with the cardiac cycle and the metabolic delay effect, and outputting time-dependent features. Then, a temporal attention mechanism is introduced to calculate the attention score at each time step. Using scores right We perform weighted summation to obtain feature vectors that can focus on the most diagnostically valuable moments, and finally output a fixed-length high-dimensional latent space vector through global average pooling. In an optional embodiment, a high-dimensional latent space vector is input to a preset nonlinear regression mapping function for solution to obtain a predicted value of serum uric acid concentration for the target user, including: The high-dimensional latent space vector is used as input and fed into a nonlinear regression mapping function; the nonlinear regression mapping function consists of at least three fully connected layers and an activation function. By performing dimensional transformation and nonlinear fitting on high-dimensional latent space vectors through fully connected layers, deep abstract features related to uric acid concentration are extracted layer by layer. The feature vector output from the last fully connected layer is passed through a linear output layer to map it to the target numerical range of uric acid concentration. The results of the linear output layer are used as the predicted serum uric acid concentration for the target user.

[0039] It should be noted that the high-dimensional latent space vector is a refined feature representation containing rich vascular-metabolic coupling information, obtained after complex encoding by a spatiotemporal attention map neural network model; the nonlinear regression mapping function is a model constructed by a multi-layer neural network, whose function is to map high-dimensional abstract features to a specific, continuous value, namely serum uric acid concentration; the fully connected layer is the most basic layer in the neural network, where each neuron is connected to all neurons in the previous layer, used to perform complex dimensional transformations and nonlinear fitting; the linear output layer is the last layer in the regression task, which does not use an activation function, directly mapping the abstract features of the previous layer to the actual uric acid concentration value, ensuring that the final output result is within a reasonable biological and medical numerical range.

[0040] In an optional embodiment, the training process of the pre-trained spatiotemporal attention map neural network model includes: Multiple sets of vascular-metabolic coupling multidimensional feature tensors with known serum uric acid concentration labels were collected to construct a training dataset. The training dataset is input into the spatiotemporal attention graph neural network model, and the mean squared error loss function is used to calculate the loss between the model's predicted value and the serum uric acid concentration label. An adaptive moment estimator optimizer is used to minimize the loss function, and the trainable parameters in the graph neural network model are iteratively adjusted. The parameters of the trained graph neural network model are frozen and used as a pre-set model for feature encoding.

[0041] It should be noted that serum uric acid concentration labels refer to the uric acid concentration values ​​obtained through traditional invasive blood tests and considered to be true values, used as the learning target of the model; the training dataset is a collection of a large number of feature tensors and their corresponding true labels used to train the model; the mean squared error loss function is a commonly used loss function for regression tasks, which calculates the average of the squared differences between the model's predicted values ​​and the true label values. The smaller this value, the more accurate the model's prediction; the adaptive moment estimator optimizer is an efficient neural network optimization algorithm that accelerates the model training process by dynamically adjusting the learning rate, with the aim of finding model parameters that minimize the loss function; trainable parameters refer to the weights and biases in the model that are continuously adjusted and optimized during training; frozen parameters refer to the weights that are fixed after the model training is completed and will not be changed again for use in actual predictions.

[0042] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for predicting uric acid levels based on multi-source data fusion, characterized in that, include: Synchronous signal acquisition of the target user's fingertip is performed using near-infrared hyperspectral acquisition equipment and flexible microfluidic impedance sensor to obtain subcutaneous microvascular hyperspectral image sequence and epidermal sweat electrochemical impedance spectrum; The hyperspectral image sequence of subcutaneous microvessels was subjected to Euler video amplification processing to obtain a blood flow micro-dynamic enhancement sequence; Wavelet packet decomposition was performed on the electrochemical impedance spectrum of the epidermal sweat to obtain metabolic impedance characteristic components in different frequency bands. Calculate the physiological response time lag between the blood flow microdynamic enhancement sequence and the metabolic impedance characteristic components; Based on the physiological response time delay, the blood flow microdynamic enhancement sequence and the metabolic impedance feature components are nonlinearly time-aligned to obtain the aligned blood flow feature matrix and the aligned impedance feature vector. The aligned impedance feature vectors are broadcast spatially and channel-cascaded with the aligned blood flow feature matrix to construct a vascular-metabolic coupled multidimensional feature tensor. The blood vessel-metabolism coupled multidimensional feature tensor is input into a pre-trained spatiotemporal attention map neural network model for feature encoding to obtain a high-dimensional latent space vector. The high-dimensional latent space vector is input into a preset nonlinear regression mapping function for solution to obtain the predicted value of serum uric acid concentration for the target user.

2. The method for predicting uric acid levels based on multi-source data fusion according to claim 1, characterized in that, The subcutaneous microvascular hyperspectral image sequence was subjected to Euler video amplification processing to obtain a blood flow micro-dynamic enhancement sequence, including: The subcutaneous microvascular hyperspectral image sequence was subjected to Laplacian pyramid decomposition to obtain multi-scale spatial frequency components; The multi-scale spatial frequency components are subjected to time-domain bandpass filtering to extract weak blood flow fluctuation signals consistent with the target user's heart rate range; The amplitude of the weak blood flow fluctuation signal is amplified and reconstructed by weighting it with the corresponding level of the Laplace pyramid to generate the blood flow micro-dynamic enhancement sequence.

3. The method for predicting uric acid levels based on multi-source data fusion according to claim 2, characterized in that, Wavelet packet decomposition was performed on the electrochemical impedance spectroscopy of the epidermal sweat to obtain metabolic impedance characteristic components in different frequency bands, including: Electrochemical impedance spectroscopy of the epidermal sweat was performed using a pre-defined Symlets wavelet basis. Full layer decomposition to obtain the first layer. Layer Wavelet packet coefficients; Based on a pre-set uric acid impedance response database, the characteristic frequency bands that are correlated with changes in uric acid concentration higher than a preset correlation threshold are matched and locked. Extracting the characteristic frequency band in the The wavelet packet coefficients corresponding to the layer total decomposition are used as the target node; The total energy and Shannon entropy of the target node within the acquisition time window are calculated, and the total energy and Shannon entropy are normalized to generate the metabolic impedance characteristic components of the different frequency bands.

4. The method for predicting uric acid levels based on multi-source data fusion according to claim 3, characterized in that, Calculating the physiological response lag between the blood flow microdynamic enhancement sequence and the metabolic impedance characteristic components includes: Data within the same predetermined time window are extracted from the blood flow micro-dynamic enhancement sequence and the metabolic impedance characteristic component, respectively, and the extracted data are preprocessed to obtain blood flow time-series signal and impedance time-series signal; Set a maximum expected time delay value to define a time delay search window; Within the time delay search window, a step-wise cross-correlation operation is performed on the blood flow time-series signal and the impedance time-series signal to generate a cross-correlation coefficient curve, wherein the horizontal axis of the cross-correlation coefficient curve represents the time delay and the vertical axis represents the correlation coefficient value. Locate the peak point of the cross-correlation coefficient curve and extract the horizontal coordinate time delay corresponding to the peak point as the physiological response time delay.

5. The method for predicting uric acid levels based on multi-source data fusion according to claim 4, characterized in that, Based on the physiological response time delay, the blood flow microdynamic enhancement sequence and the metabolic impedance characteristic components are nonlinearly time-aligned to obtain an aligned blood flow characteristic matrix and an aligned impedance characteristic vector, including: The pulse wave peak point in the blood flow micro-dynamic enhancement sequence is extracted as the first time reference, and the impedance peak point in the metabolic impedance characteristic component is extracted as the second time reference. A search constraint window based on the physiological response time delay is constructed, and within the search constraint window, the minimum bend path between the first time series reference and the second time series reference is calculated using a dynamic time warping algorithm; Based on the synchronization mapping index generated by the minimum bend path, the blood flow micro-dynamic enhancement sequence and the metabolic impedance feature components are resampled to generate the aligned blood flow feature matrix and the aligned impedance feature vector.

6. The method for predicting uric acid levels based on multi-source data fusion according to claim 5, characterized in that, The aligned impedance eigenvectors are spatially broadcast and channel-concatenated with the aligned blood flow feature matrix, including: Obtain the spatial dimension of the aligned blood flow feature matrix; The aligned impedance feature vector is copied as a feature layer with the same spatial dimension; The aligned blood flow feature matrix and the feature layer are concatenated along the channel dimension to construct the blood vessel-metabolism coupled multidimensional feature tensor.

7. The method for predicting uric acid levels based on multi-source data fusion according to claim 6, characterized in that, The blood vessel-metabolism coupled multidimensional feature tensor is input into a pre-trained spatiotemporal attention map neural network model for feature encoding to obtain a high-dimensional latent space vector, including: The pixel region of the spatial feature layer corresponding to each time step in the blood vessel-metabolism coupled multidimensional feature tensor is defined as a graph node, and the dynamic connection edge weights between the graph nodes are defined according to the metabolic impedance feature components to construct time series graph structure data. The temporal graph structure data is input into the spatiotemporal attention graph neural network model, and the graph convolutional layer is used to capture the hemodynamic correlation of the graph nodes in the spatial dimension to generate spatial fusion features; The spatial fusion features are input into the long short-term memory network layer in the spatiotemporal attention map neural network model to extract the evolution law of the spatial fusion features in the time dimension and generate time-dependent features; The key time steps in the temporal-dependent features are weighted by the attention mechanism layer in the spatiotemporal attention graph neural network model, and the weighted features are then globally pooled to output the high-dimensional latent space vector.

8. The method for predicting uric acid levels based on multi-source data fusion according to claim 7, characterized in that, The high-dimensional latent space vector is input into a preset nonlinear regression mapping function for solution to obtain the predicted serum uric acid concentration of the target user, including: The high-dimensional latent space vector is used as input and fed into the nonlinear regression mapping function; wherein, the nonlinear regression mapping function consists of at least three fully connected layers and an activation function; The high-dimensional latent space vector is subjected to dimensional transformation and nonlinear fitting through the fully connected layer to extract deep abstract features related to uric acid concentration layer by layer. The feature vector output by the last fully connected layer is passed through a linear output layer to map it to the target numerical range of uric acid concentration. The results of the linear output layer are output as the predicted serum uric acid concentration for the target user.

9. The method for predicting uric acid levels based on multi-source data fusion according to claim 8, characterized in that, The training process of the pre-trained spatiotemporal attention map neural network model includes: Multiple sets of the aforementioned vascular-metabolic coupling multidimensional feature tensors with known serum uric acid concentration labels were collected to construct a training dataset; The training dataset is input into the spatiotemporal attention map neural network model, and the loss between the model prediction value and the serum uric acid concentration label is calculated using the mean squared error loss function. An adaptive moment estimator optimizer is used to minimize the loss function, and the trainable parameters in the graph neural network model are iteratively adjusted. The parameters of the trained graph neural network model are frozen and used as a preset model for feature encoding.

Images