A multi-modal urine volume measurement method based on a space-time graph neural network

Abstract

The application provides a kind of multi-modal urine volume measurement method based on space-time graph neural network, it is related to biomedical monitoring technical field, method includes: sensor array is mapped to dynamic heterogeneous graph and constructs time series graph sequence;Through three-level feature pyramid, micro gradient, mesoscopic morphology and macroscopic topological dominant mode are extracted;A space-time graph neural network model driven by physical information is constructed, Richards equation is embedded into the loss function as a physical constraint, and a gradient reversal layer is introduced for adversarial learning to eliminate the field bias caused by individual differences;Small dose calibration data is used to realize rapid adaptation and continuous updating of the model.The application significantly improves the accuracy, robustness and cross-individual generalization ability of urine volume monitoring through the deep fusion of physical logic and space-time topology perception.

Description

Technical Field

[0001] This invention relates to the field of biomedical monitoring technology, and in particular to a multimodal urine volume measurement method based on spatiotemporal graph neural networks. Background Technology

[0002] Urinary incontinence, especially stress urinary incontinence, is a common health problem that significantly impacts quality of life. Accurate assessment of this condition is crucial for clinical diagnosis, treatment planning, and efficacy evaluation. Traditional assessment methods, such as voiding diaries and pad weight gain tests, rely on subjective patient recordings or single, static measurements. These methods suffer from limitations such as low efficiency, discontinuous data, and susceptibility to human error, making it difficult to objectively and dynamically reflect the true pattern and magnitude of urine leakage.

[0003] In recent years, with the development of flexible electronics and sensing technology, smart incontinence pads / liners integrating sensors such as capacitance and pressure have emerged, providing a new technical path for objective and continuous monitoring of urinary incontinence. These solutions typically treat multiple sensors embedded in the pad as independent discrete detection points, estimating the occurrence of leakage events and approximate urine volume through threshold judgment or simple weighted aggregation algorithms. While this method achieves a degree of automation and objectivity, the diffusion of urine in a superabsorbent polymer (SAP) core is a continuous seepage process influenced by the coupling of multiple physical fields, including material properties, liquid properties, and external pressure. Simplifying the sensor array into independent points and only performing post-processing aggregation on its discrete signals essentially approximates the continuous physical field with a zero-order approximation, losing crucial dynamic information such as the velocity, direction, and morphological evolution of the diffusion front, resulting in insufficient ability to identify complex leakage modes. Because urine diffusion exhibits spatial neighborhood correlation and continuous temporal evolution, most existing methods do not explicitly model the spatial topological relationships between sensor nodes or the causal relationships of signals over time. This makes it difficult for the system to effectively distinguish between pressure redistribution caused by changes in body position and signal changes induced by actual urine diffusion, and also makes it difficult to accurately track diffusion paths and reconstruct wetting areas. Furthermore, purely data-driven models are prone to overfitting data distributions specific to particular patients or experimental conditions, and their performance may significantly degrade when faced with material variations due to different physiological characteristics, behavioral patterns, or product batches.

[0004] To address the aforementioned issues, a novel analytical method is urgently needed that can deeply integrate spatiotemporal information from sensors, prior knowledge of urine diffusion physics, and individual patient differences. Therefore, this invention introduces graph structure modeling, multi-scale physical feature extraction, and a hybrid neural network driven by physical information to construct a more accurate, robust, and generalizable intelligent quantitative assessment system for urinary incontinence that better reflects the intrinsic mechanism of urine diffusion. Summary of the Invention

[0005] The main objective of this invention is to provide a multimodal urine volume measurement method based on spatiotemporal graph neural networks, which solves the problem of inaccurate identification of urine leakage events caused by neglecting the spatiotemporal continuity and physical dynamics of urine diffusion in existing technologies. This method enables accurate identification of urine leakage events, accurate regression of urine volume, and improves the cross-individual generalization ability of the model among patients with different physiological characteristics.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a multimodal urine volume measurement method based on spatiotemporal graph neural network, which includes the following steps: S1. Acquire the sensor array data set inside the pad and construct a time sequence diagram. S2. Construct a three-level feature pyramid to extract and identify multi-stage urine leakage events; S3. Construct a hybrid neural network model driven by physical information to output predicted urine volume and measurement confidence level; S4. Use small-dose calibration data for rapid adaptation and continuously update the model through incremental learning.

[0007] Preferably, step S1 includes: A graph structure is built using capacitive and pressure sensors as nodes. The node attributes include real-time capacitance changes, location coordinates, pressure values, and activation timestamps. The weights of the edges are constructed based on spatial adjacency, Euclidean distance, and material permeability coefficient; The monitoring period is discretized into time steps with preset intervals to form a time-varying graph structure that includes time-series evolution characteristics.

[0008] Preferably, in step S2, the three-level feature pyramid includes: At the microscopic level, the capacitance gradient features of the local infiltration front are extracted, the impact response spectrum reflecting the abdominal pressure-urine leakage correlation is captured, and the slope of the capacitance rise edge is analyzed to distinguish between transient urinary leakage and slow infiltration. In the meso-level, Hu moment invariants are calculated based on the set of activated nodes to identify diffusion patterns, and the infiltration depth is inverted and pressure is compensated using the capacitance-pressure coupling equation. At the macroscopic level, the graph Laplace spectrum is computed to quantify diffusion connectivity, and the dominant modes of urine diffusion are extracted through dynamic mode decomposition.

[0009] Preferably, the inversion of the mesoscopic layer in this way is specifically achieved by eliminating the signal error caused by the compression deformation of the pad due to changes in body position through the capacitance-pressure coupling equation.

[0010] Preferably, in step S3, the physical information-driven hybrid neural network model includes: The spatiotemporal encoder uses graph convolutional neural networks to capture spatial dependencies and bidirectional gated recurrent units to model temporal dynamics. The PINN constraint layer embeds the Richards equation for urine in the absorbent material as a partial differential equation residual constraint into the loss function to penalize predictions that do not conform to the physical flow laws.

[0011] Preferably, the GCN layer specifically employs a Chebyshev polynomial approximation of the spectral graph convolution, which matches the permeation characteristic length of the absorbing material and reduces computational complexity to achieve real-time solution.

[0012] Preferably, the expression for the Richards equation constraint is as follows (1): (1); in, Water content by volume Permeability coefficient, As the matrix potential, It is a spatial gradient operator that calculates the spatial gradient through automatic differentiation to estimate the infiltration volume.

[0013] Preferably, in operation S3, the gradient reversal layer specifically performs the following adversarial learning: With urine volume regression as the core task, mean squared error is used as the optimization objective. Patient identification is used as an adversarial task, and cross-entropy loss is used as the optimization objective. During backpropagation, gradient inversion forces the model to extract universal urine diffusion features that are independent of BMI, urethral location, and gender.

[0014] Preferably, a multi-task decoding process is provided, which includes three branches: The left branch outputs the instantaneous amount of urine leakage; The middle branch uses a pressure-capacitance coupling model to estimate the material compression return flow caused by changes in body position. The right branch uses temperature sensor data to physically compensate for evaporation loss.

[0015] Preferably, in step S4, rapid adaptation specifically involves: obtaining initialization parameters using model-independent learning, and performing gradient updates for new patients using only known low-dose urine calibration data to achieve personalized adaptation.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Spatial topology sensing: By representing the sensor array through dynamic heterogeneous graphs, discrete sensor signals are transformed into topological structures that include spatial adjacency relationships and material permeability coefficient weights, which can effectively capture the real physical diffusion path of urine in absorbent materials. 2. Temporal dynamic fusion: By utilizing bidirectional GRU and graph convolutional layers in collaboration, the model can not only identify the current infiltration state, but also capture the dynamic evolution characteristics of urine diffusion over time (such as the slope of the rising edge), thereby accurately distinguishing between instantaneous leakage and slow infiltration events. 3. By introducing hard constraints based on physical laws and embedding the Richards equation as a soft constraint into the loss function of the physical information neural network, this invention forces the model prediction results to conform to the physical laws of porous media seepage, effectively overcoming the prediction distortion or logical violation problems that may occur when pure data-driven models face extreme out-of-sample data. 4. Multi-dimensional multi-physics compensation: Through a three-level feature pyramid, the infiltration depth is inverted using a capacitance-pressure coupling model, effectively eliminating deformation errors caused by material compression due to changes in body position. Through a multi-task decoder, the model can simultaneously output instantaneous urine leakage, material reabsorption, and evaporation compensation values. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of a multimodal urine volume measurement method based on a spatiotemporal graph neural network according to the present invention; Figure 2 This is a structural diagram of the dynamic graph construction method of the present invention; Figure 3 This is a flowchart of the spatiotemporal graph neural network of the present invention. Detailed Implementation

[0018] Example 1 like Figure 1 As shown, a multimodal urine volume measurement method based on a spatiotemporal graph neural network consists of steps S110-S140, including: In step S110, the graph structure representation of the sensor array maps the sensor data inside the pad into a dynamic heterogeneous graph and constructs a time-series graph sequence. In step S120, a three-level feature pyramid is designed to target the multi-scale characteristics of urine diffusion, extract the dominant modes of urine diffusion, and identify multi-stage urine leakage events. In step S130, a hybrid neural network model driven by physical information is constructed, with the Richards equation as the soft-constraint embedding loss function, and a gradient inversion layer is introduced to address the differences among different patients. In step S140, the calibration data is used to perform rapid adaptation by injecting small doses of urine. Data is continuously collected during the experiment, and the model is further updated using expert experience and historical data.

[0019] According to an embodiment of the present invention, in step S110, the eight capacitive sensors and two pressure sensors inside the pad are mapped into a dynamic heterogeneous diagram, as shown in the following equation (2): (2); Among them, nodes For each sensor unit, the attributes include real-time capacitance change, position coordinates, and activation timing, and the edges... Based on spatial adjacency relationships and prior knowledge of urine diffusion, the weights are jointly determined by Euclidean distance and material permeability coefficient, with the time dimension also considered. The 1-hour monitoring cycle was discretized into 600 time steps with an interval of 6 seconds, and a time series graph sequence was constructed, providing a complete data foundation including spatial topology, physical characteristics and temporal dynamics for subsequent multi-scale feature extraction.

[0020] According to an embodiment of the present invention, in step S120, a three-layer feature pyramid is designed to address the multi-scale characteristics of urine diffusion. In the microscopic layer, for the capacitive and pressure sensors, capacitance gradient features are obtained by capturing the local infiltration front. Impact response spectra from events such as coughing and walking are extracted to establish an abdominal pressure-leakage correlation model. The rising slope of capacitance changes is analyzed to distinguish between instantaneous leakage and slow infiltration. In the mesoscopic layer, Hu moment invariants are calculated based on the set of activated nodes to identify circular, elliptical, or irregular urine diffusion patterns. The infiltration depth is then inverted using the capacitance-pressure coupling equation, and pressure compensation eliminates compression deformation errors caused by body position. In the macroscopic layer, diffusion connectivity is quantified by calculating the Laplacian spectrum and clustering coefficients of the graph. Multi-stage leakage events are identified by extracting the dominant modes of urine diffusion, providing multi-level, highly discriminative input features for the neural network and improving the model's ability to identify leakage patterns.

[0021] According to an embodiment of the present invention, in step S130, a spatiotemporal graph neural network model based on physical information constraints is constructed. In the spatiotemporal encoder, a GCN layer captures spatial dependencies, and a bidirectional GRN models temporal dynamics. In the PINN constraint layer, the Richards equations for urine in SAP materials are used as soft constraints embedded in the loss function. A penalty term ensures that the model prediction conforms to the physical laws of permeation in porous media, solving the extrapolation failure problem of purely data-driven models. Addressing physiological differences among patients, including BMI and domain shift caused by urethral location, this invention introduces a gradient inversion layer. With urine volume regression as the core task, mean squared error is used as the loss function optimization objective. Patient identification is used as an adversarial task with cross-entropy loss as the optimization objective. Through this dual-task game, the model is forced to learn universal urine diffusion characteristics independent of individual patient characteristics, improving cross-individual generalization ability. Physical constraints ensure model accuracy, while adversarial learning ensures model universality. The combination of these two approaches achieves accurate and robust urine monitoring.

[0022] According to an embodiment of the present invention, in step S140, a prospective controlled trial of 30 cases is conducted in cooperation with a hospital to obtain clinical data. Addressing the data scarcity issue in medical settings, model-independent learning is employed, and the model is trained on 1000 virtual patient data generated in a simulator to obtain well-initialized parameters. For new patients, only the first 5 minutes of calibration data are used for rapid adaptation through gradient updates. Data is continuously collected during the trial, and the model is updated through incremental learning to improve accuracy.

[0023] like Figure 2 As shown, in order to model and learn the spatiotemporal continuity of urine diffusion, the diffusion process of the urine pad experiment is modeled as a spatiotemporal graph evolution problem, thereby converting the sensor array signal into a spatiotemporal graph structure, which serves as the input data for subsequent training of the spatiotemporal graph neural network.

[0024] In step S101, the spatiotemporal graph neural network model based on physical information constraints is input with sensor data, and the sensor unit inside the pad is... Extracting the original feature vector at each time step As shown in equation (3): (3); in, For sensor units, This represents the change in capacitance. This is the pressure value. For sensor plane coordinates, Activate the timestamp for the node.

[0025] In step S102, the sensor array is characterized as a time-varying graph structure, with nodes... For sensor units, edge The weights are dynamically adjusted based on spatial distance and diffusion state, as shown in equation (4) below: (4); in, Let i be the Euclidean distance between sensors i and j. Control the spatial perception range. This is the similarity weighting coefficient. For the maximum connection distance, The similarity of capacitance changes is shown in equation (5) below: (5); In step S103, the present invention identifies the infiltration area of ​​urine leakage based on the GCN (Graph Cut Network) algorithm. , The effective wetting radius is further calculated based on the sensor activation threshold, as shown in equation (6) below: (6); in, Let be the edge weight between node i and the center node. The sensor plane coordinates are used. The resulting spatiotemporal heterogeneous data stream... like Figure 3 As shown, when receiving spatiotemporally heterogeneous data streams from 10 sensor nodes, including 8 capacitive sensors and 2 pressure sensors, spatial diffusion dependence is captured in the graph convolutional layer of the GCN. A third-order expansion is performed using the spectral graph convolution approximated by Chebyshev polynomials, as shown in Equation (7) below: (7); in, For the first The input feature matrix, For the first The output deep features after spatial aggregation of layers It is a nonlinear transformation activation function. For the k-th order Chebyshev polynomial, Standardized graph Laplace matrix, For the first The trainable weight matrix of the k-th order polynomial in the layer has a polynomial order of k=3 in the physical prior, which matches the permeation characteristic length of the SAP material. The Chebyshev polynomial approximation pulls the graph convolution from the frequency domain back to the spatial domain, avoiding the computation of the eigenvalue decomposition of the graph Laplacian matrix. For 10 sensors, direct computation of the eigenvalue decomposition is not feasible, while the Chebyshev approximation can reduce computational complexity, thus meeting the real-time requirements.

[0026] When the urethral opening is near the center of the sensor, leakage occurs: When t=0.5s, directly opposite the capacitor Sudden rise, ,activation ; When t=2s, the urine diffuses to the nearby capacitor. , and , Aggregate neighbor information; When t=5s, the urine diffuses to a capacitor at a relatively far distance. and , Capture second-order transits; When t=10s, the capacitance of urine spreading to the edge of the incontinence pad and , Modeling long-range effects.

[0027] in, By integrating information from different diffusion radii, the model can accurately invert the location, volume, and diffusion rate of urine leakage.

[0028] The temporal dynamics are modeled using a gated recurrent unit (GRU). The bidirectional GRU shares parameters in both directions to capture the temporal causal chain of urinary incontinence events, including the increase in abdominal pressure caused by coughing, which further leads to the spatiotemporal movement of the urethral sphincter, and finally to the occurrence of urinary incontinence events. The final output is a spatiotemporal fusion feature vector.

[0029] The Physical Information Constraint Layer (PINN) embeds partial differential equation residual constraints in the network, estimates the infiltration volume based on the GCN output, and performs calculations using automatic differentiation, as shown in equation (8) below: (8); in, Water content by volume Permeability coefficient, As the matrix potential, The formula for calculating the spatial gradient operator is shown in equation (9) below: (9); This formula represents the urine diffusion coordinates. Differential calculations are performed on the distribution.

[0030] To show the change in moisture content of a certain area of ​​the sanitary pad over time, This represents the diffusion outflow of urine per unit time. When the loss function approaches 0, the diffusion rate matches the current wettability and material properties.

[0031] Dynamic feature weighting is achieved through a multi-head attention mechanism, which automatically identifies the central sensor of the leakage area and suppresses noise from malfunctioning nodes. Urine volume regression is performed in the pre-adaptive adversarial layer to minimize MSE, while adversarial tasks are used to identify patients. The feature extractor is multiplied by -0.1 during model backpropagation, thereby forcing the feature extractor to learn patient-independent general diffusion features. This solves the domain offset problem for patients with different BMIs, urethral opening locations, and genders, and increases the model's adaptability.

[0032] In the multi-task decoder, the left branch uses the ReLU activation function to output the instantaneous urine leakage. The middle branch detects fluid backflow caused by SAP material compression during a change in body position from lying to sitting using a pressure sensor, and estimates the leakage using a pressure-capacitance coupling model. The right branch measures ambient temperature using a temperature sensor to compensate for evaporating urine. The urine volume is decomposed into a multiphysics model of physiological leakage and physical compensation, ultimately outputting predicted urine volume and measurement confidence to aid clinical decision-making in urine leakage measurement.

[0033] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be defined as the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A multimodal urine volume measurement method based on spatiotemporal graph neural networks, characterized in that, Includes the following steps: S1. Acquire the sensor array data set inside the pad and construct a time sequence diagram. S2. Construct a three-level feature pyramid to extract and identify multi-stage urine leakage events; S3. Construct a hybrid neural network model driven by physical information to output predicted urine volume and measurement confidence level; S4. Use small-dose calibration data for rapid adaptation and continuously update the model through incremental learning.

2. The multimodal urine volume measurement method based on spatiotemporal graph neural network according to claim 1, characterized in that, Step S1 includes: A graph structure is built using capacitive and pressure sensors as nodes. The node attributes include real-time capacitance changes, location coordinates, pressure values, and activation timestamps. The weights of the edges are constructed based on spatial adjacency, Euclidean distance, and material permeability coefficient; The monitoring period is discretized into time steps with preset intervals to form a time-varying graph structure that includes time-series evolution characteristics.

3. The multimodal urine volume measurement method based on spatiotemporal graph neural network according to claim 1, characterized in that, In step S2, the three-level feature pyramid includes: At the microscopic level, the capacitance gradient features of the local infiltration front are extracted, the impact response spectrum reflecting the abdominal pressure-urine leakage correlation is captured, and the slope of the capacitance rise edge is analyzed to distinguish between transient urinary leakage and slow infiltration. In the meso-level, Hu moment invariants are calculated based on the set of activated nodes to identify diffusion patterns, and the infiltration depth is inverted and pressure is compensated using the capacitance-pressure coupling equation. At the macroscopic level, the graph Laplace spectrum is computed to quantify diffusion connectivity, and the dominant modes of urine diffusion are extracted through dynamic mode decomposition.

4. The multimodal urine volume measurement method based on spatiotemporal graph neural network according to claim 3, characterized in that, in The inversion of the immersion depth of the observation layer is specifically achieved by eliminating the signal error caused by the compression deformation of the pad due to changes in body position through the capacitance-pressure coupling equation.

5. The multimodal urine volume measurement method based on spatiotemporal graph neural network according to claim 1, characterized in that, In step S3, the physical information-driven hybrid neural network model includes: The spatiotemporal encoder uses graph convolutional neural networks to capture spatial dependencies and bidirectional gated recurrent units to model temporal dynamics. The PINN constraint layer embeds the Richards equation for urine in the absorbent material as a partial differential equation residual constraint into the loss function to penalize predictions that do not conform to the physical flow laws.

6. The multimodal urine volume measurement method based on spatiotemporal graph neural network according to claim 5, characterized in that: The GCN layer specifically employs Chebyshev polynomial approximation of spectral graph convolution, which matches the permeation characteristic length of the absorbing material and reduces computational complexity to achieve real-time solution.

7. The multimodal urine volume measurement method based on spatiotemporal graph neural network according to claim 5, characterized in that, the... The expressions for the Chaz equation constraints are as follows: (1)； in, Water content by volume Permeability coefficient, As the matrix potential, It is a spatial gradient operator that calculates the spatial gradient through automatic differentiation to estimate the infiltration volume.

8. The multimodal urine volume measurement method based on spatiotemporal graph neural network according to claim 1, characterized in that, In operation S3, the gradient reversal layer specifically performs the following adversarial learning: With urine volume regression as the core task, mean squared error is used as the optimization objective. Patient identification is used as an adversarial task, and cross-entropy loss is used as the optimization objective. During backpropagation, gradient inversion forces the model to extract universal urine diffusion features that are independent of BMI, urethral location, and gender.

9. The multimodal urine volume measurement method based on spatiotemporal graph neural network according to claim 1, characterized in that, It also includes a multi-task decoding process, which has three branches: The left branch outputs the instantaneous amount of urine leakage; The middle branch uses a pressure-capacitance coupling model to estimate the material compression return flow caused by changes in body position. The right branch uses temperature sensor data to physically compensate for evaporation loss.

10. The multimodal urine volume measurement method based on spatiotemporal graph neural network according to claim 1, characterized in that, In step S4, the rapid adaptation specifically involves: using model-independent learning to obtain initialization parameters, and performing gradient updates on the known low-dose urine calibration data used only for new patients to achieve personalized adaptation.

Images