Estimation method and estimation device for glomerular filtration rate of patient with asymmetric double renal functions

Abstract

The invention provides a glomerular filtration rate estimation method and device for a patient with asymmetric double renal functions. The method comprises the following steps: acquiring basic information, clinical data, laboratory detection data and iconography data of a patient, wherein the iconography data comprises kidney volume, blood flow volume, cortex and medullary thickness and kidney function expression of the right kidney and the left kidney; by analyzing iconography data, a functional difference measurement coefficient for quantifying the asymmetry of double kidney functions is calculated. Thirdly, extracting a global feature vector of the patient by using a deep learning model, and performing fusion processing on the iconography data, the clinical data and the laboratory detection data to generate a comprehensive feature vector; and on the basis, the functional weights of the right kidney and the left kidney are adjusted according to the functional difference measurement coefficient, finally, the adjusted functional weights and the global feature vectors are input into a deep regression model, and a corrected GFR estimated value is obtained through calculation. And more accurate renal function evaluation can be provided according to the condition that double renal functions are asymmetric.

Description

Technical Field

[0001] This invention relates to the field of glomerular filtration rate (GFR) calculation technology, specifically to a method and apparatus for estimating GFR in patients with asymmetrical renal function. Background Technology

[0002] Chronic kidney disease (CKD) has become a common health problem worldwide, especially with the increasing prevalence of chronic diseases such as diabetes and hypertension, the incidence of CKD is showing a year-on-year upward trend. Glomerular filtration rate (GFR) is the most important indicator for assessing kidney function, effectively reflecting the kidney's filtration capacity, and is crucial for the diagnosis, treatment planning, and disease monitoring of CKD.

[0003] Traditionally, GFR measurement has relied on the clearance rate of exogenous markers (such as inulin and iohexol). While these methods are accurate, they are complex and costly, making them unsuitable for large-scale clinical applications. A common alternative is estimation using formulas derived from endogenous metabolites such as serum creatinine and blood urea nitrogen. Common formulas include the Cockcroft-Gault formula and the MDRD formula. These methods are convenient for clinical use, but the estimation results can be biased in certain patient groups (such as those who have undergone unilateral nephrectomy or whose renal function is asymmetrical due to kidney disease).

[0004] Existing GFR estimation methods typically assume symmetrical renal function, but in clinical practice, many patients experience asymmetry in renal function due to disease, surgery, or other factors. Traditional estimation methods fail to effectively address the impact of renal function asymmetry, thus failing to provide accurate renal function assessments. To improve this, developing a deep learning-based dynamic correction method capable of considering differences in renal function has become a pressing technical challenge. Summary of the Invention

[0005] In view of the technical problems mentioned in the background, the present invention provides a method, device, electronic device and computer-readable medium for estimating glomerular filtration rate based on comprehensive data analysis.

[0006] A first aspect of the present invention provides a method for estimating glomerular filtration rate based on comprehensive data analysis, the method comprising the following steps:

[0007] S1, acquire the patient's basic information, clinical data, laboratory test data and imaging data; wherein, the imaging data includes at least the kidney volume, blood flow, thickness of the renal cortex and medulla, and renal function of the right and left kidneys;

[0008] S2, Based on the imaging data, a functional difference measurement coefficient for quantifying the asymmetry of renal function is calculated by comparing the differences in renal volume, blood flow, and thickness of the renal cortex and medulla between the right and left kidneys;

[0009] S3 extracts the patient's global feature vector through a deep learning model, including the imaging feature vector of each kidney extracted from imaging data, the physiological feature vector extracted from clinical data, and the laboratory test feature vector extracted from laboratory test data, and fuses the above feature vectors to generate a comprehensive feature vector.

[0010] S4, adjust the functional weights of the right and left kidneys in GFR estimation according to the functional difference measurement coefficient;

[0011] S5. Input the adjusted functional weights and the global feature vector into the deep regression model, calculate the corrected GFR estimate and output it.

[0012] A second aspect of the present invention provides a glomerular filtration rate estimation device based on comprehensive data analysis, the device comprising:

[0013] The data acquisition module is used to acquire the patient's basic information, clinical data, laboratory test data, and imaging data; wherein, the imaging data includes at least the kidney volume, blood flow, thickness of the renal cortex and medulla, and renal function of the right and left kidneys;

[0014] The functional difference quantification module is used to calculate a functional difference measurement coefficient for quantifying the asymmetry of bilateral renal function based on the imaging data by comparing the differences in renal volume, blood flow, and thickness of the renal cortex and medulla between the right and left kidneys.

[0015] The global feature extraction and fusion module is used to extract the patient's global feature vector through a deep learning model, including the imaging feature vector of each kidney extracted from imaging data, the physiological feature vector extracted from clinical data, and the laboratory test feature vector extracted from laboratory test data, and to fuse the above feature vectors to generate a comprehensive feature vector.

[0016] The functional weight dynamic adjustment module is used to adjust the functional weights of the right and left kidneys in GFR estimation based on the functional difference measurement coefficient.

[0017] The corrected estimation output module is used to input the adjusted function weights and the global feature vector into the deep regression model, calculate the corrected GFR estimate, and output it.

[0018] A third aspect of this application provides an electronic device, comprising: a processor; and a memory storing computer program instructions that, when executed by the processor, cause the processor to perform the method described above.

[0019] A fourth aspect of this application provides a computer-readable medium having computer program instructions stored thereon, which, when executed by a processor, cause the processor to perform the method described above.

[0020] The glomerular filtration rate (GFR) estimation method provided by this invention, based on comprehensive data analysis, can accurately reflect the impact of bilateral renal function asymmetry by combining imaging data, clinical data, laboratory test data, and renal function characteristics. Compared with traditional methods, this invention significantly improves the accuracy of GFR estimation by dynamically adjusting the weights of bilateral renal function and weightedly fusing multi-source data, making it particularly suitable for patients with bilateral renal function asymmetry. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for estimating glomerular filtration rate based on comprehensive data analysis disclosed in an embodiment of this application;

[0022] Figure 2 This is a schematic diagram of the structure of the deep learning model disclosed in the embodiments of this application;

[0023] Figure 3 This is a schematic diagram of the structure of the dynamic weight generation network disclosed in the embodiments of this application;

[0024] Figure 4 This is a schematic diagram of the structure of the deep regression model disclosed in the embodiments of this application;

[0025] Figure 5 This is a schematic diagram of a glomerular filtration rate estimation device based on comprehensive data analysis disclosed in an embodiment of this application;

[0026] Figure 6 This is a schematic diagram of the functional difference quantification module disclosed in an embodiment of this application;

[0027] Figure 7 This is a schematic diagram of the structure of the global feature extraction and fusion module disclosed in the embodiments of this application;

[0028] Figure 8 This is a schematic diagram of the structure of the corrected estimation output module disclosed in the embodiments of this application. Detailed Implementation

[0029] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0030] Please see Figure 1 This invention provides a method for estimating glomerular filtration rate (GFR) based on comprehensive data analysis. This method is particularly suitable for patients with asymmetrical renal function and can significantly improve the accuracy and clinical applicability of GFR estimation through multi-source data fusion and dynamic weight adjustment. The method includes the following steps:

[0031] S1, acquire the patient's basic information, clinical data, laboratory test data and imaging data; wherein, the imaging data includes at least the kidney volume, blood flow, thickness of the renal cortex and medulla, and renal function of the right and left kidneys;

[0032] The basic information includes, but is not limited to, the patient's age, gender, and other demographic characteristics; clinical data includes physiological indicators such as blood pressure, urine output, serum creatinine, and urine protein; and laboratory test data includes various biomarkers detected from blood and urine, such as cystatin C and β2-microglobulin.

[0033] Imaging data should include at least the following information for the right and left kidneys: kidney volume, which can be obtained through three-dimensional reconstruction of MRI or CT images; blood flow, for example, measured by Doppler ultrasound or dynamic contrast-enhanced MRI; thickness of the renal cortex and medulla, for example, measured from high-resolution cross-sectional images; and renal function data, such as the percentage of function of each kidney, renal imaging curve characteristics, urine test results, trends in renal function changes, and history of renal disease.

[0034] Understandably, the aforementioned data can be obtained through the integration of hospital information systems, image archiving and communication systems, and laboratory information systems.

[0035] S2, Based on the imaging data, a functional difference measurement coefficient for quantifying the asymmetry of renal function is calculated by comparing the differences in renal volume, blood flow, and thickness of the renal cortex and medulla between the right and left kidneys;

[0036] In clinical scenarios involving asymmetric renal function, such as unilateral renal atrophy, renal artery stenosis, and sequelae of pyelonephritis, traditional GFR estimation methods often overlook the functional differences between the left and right kidneys, leading to biased overall assessment results. To address this issue, this invention extracts key indicators reflecting renal structure and perfusion status from imaging data and utilizes a designed quantitative calculation method to objectively and comprehensively characterize the degree of functional asymmetry between the two kidneys.

[0037] As an example, a functional difference measure coefficient for quantifying the asymmetry of renal function is calculated by comparing the differences in renal volume, blood flow, and cortical-medullary thickness between the right and left kidneys. This includes:

[0038] S21, calculate the difference in kidney volume between the right and left kidneys to obtain the volume difference coefficient;

[0039] Measure the volume of the right kidney separately and left kidney volume Kidney volume can be obtained through three-dimensional segmentation and voxel accumulation based on medical images (such as CT or MRI). Kidney volume is an important structural indicator reflecting the total amount of renal parenchyma and potential functional reserve. Significant differences in volume between the two kidneys often indicate unilateral renal hypoplasia, atrophy, or space-occupying lesions.

[0040] Volume difference coefficient The calculation formula is as follows:

[0041]

[0042] This coefficient, based on the average volume of both kidneys, normalizes the absolute volume difference, making it a dimensionless measure of relative difference. When A value close to 0 indicates that the kidneys are symmetrical in volume, while a larger value indicates a greater degree of asymmetry.

[0043] S22, calculate the difference in blood flow between the right and left kidneys to generate a blood flow difference coefficient;

[0044] Renal blood flow is a key kinetic parameter for assessing renal perfusion status and filtration function, and can be obtained through Doppler ultrasound, dynamic contrast-enhanced MRI, or CT perfusion imaging. Right renal blood flow was extracted separately. With left kidney blood flow (Units are mL / min or relative perfusion values).

[0045] Blood flow difference coefficient The calculation method is as follows:

[0046]

[0047] This blood flow difference coefficient reflects the imbalance in blood supply between the two kidneys. Blood flow differences may not only originate from vascular lesions (such as renal artery stenosis), but may also be related to microcirculatory disturbances caused by renal parenchymal lesions, and are an important manifestation of functional asymmetry.

[0048] S23, calculate the difference in thickness between the renal cortex and medulla of the right and left kidneys to generate a thickness difference coefficient;

[0049] The thickness of the renal cortex and medulla is a micromorphological indicator for assessing the structural integrity and differentiation of the renal parenchyma. Cortical thickness is closely related to the number of glomeruli and filtration function, while medullary thickness affects urine concentrating ability. The thickness of the right renal cortex was measured separately from high-resolution cross-sectional images. medullary thickness and the corresponding value of the left kidney , .

[0050] To comprehensively reflect the overall differences between the cortex and medulla, a thickness difference coefficient was designed. as follows:

[0051]

[0052] The above formula also incorporates the differences between the cortex and medulla, and uses the average thickness of both sides as the standard benchmark to ensure that the calculation results can fully reflect the asymmetry of the renal parenchyma structure.

[0053] S24, by weighting and fusing the volume difference coefficient, the blood flow difference coefficient, and the thickness difference coefficient, a functional difference measurement coefficient is obtained.

[0054] To integrate the differences across the three dimensions into a single metric, a weighted fusion approach is used to generate the final functional difference metric coefficient. : .in, The weights for volume, blood flow, and thickness difference coefficients are respectively, and satisfy the following conditions: .

[0055] The aforementioned weights can be set based on prior clinical knowledge. For example, in chronic ischemic kidney disease, the weight of blood flow differences... The weighting of volume differences can be appropriately increased in cases of renal hypoplasia or atrophy. It can be set to a higher value. Alternatively, an end-to-end deep learning framework can be used to train the weight parameters along with the model, optimizing them through backpropagation. It is better able to predict the actual impact of renal function asymmetry on GFR.

[0056] Understandably, the functional difference measurement coefficient It is a continuous value between 0 and 1 (theoretically it can be greater than 1 but it is extremely rare). The larger the value, the higher the degree of asymmetry in the function of the two kidneys.

[0057] S3 extracts the patient's global feature vector through a deep learning model, including the imaging feature vector of each kidney extracted from imaging data, the physiological feature vector extracted from clinical data, and the laboratory test feature vector extracted from laboratory test data, and fuses the above feature vectors to generate a comprehensive feature vector.

[0058] Traditional GFR estimation models often rely on single-type data (such as serum creatinine) or simple linear combinations, making it difficult to capture the complex nonlinear relationships of kidney function and the interactions between multiple factors. This is especially true for patients with asymmetrical renal function, whose pathophysiological state is often reflected in changes in imaging structure, abnormal laboratory biomarkers, and evolving clinical history. This step employs a multi-branch deep learning network to process different types of data separately and uses an attention mechanism to achieve information complementarity and highlight key information, thereby constructing a unified feature representation that can both cover the overall picture and focus on crucial information.

[0059] As an example, the deep learning model includes multiple convolutional neural network layers, fully connected neural network layers, deep neural network layers, and attention mechanism layers; then, the deep learning model extracts the patient's global feature vector, and fuses these feature vectors to generate a comprehensive feature vector, including:

[0060] Please see Figure 2 The deep learning model used in this invention is a multi-branch fusion network, which mainly includes the following components:

[0061] Convolutional Neural Network (CNN) branch: used to process imaging data and extract spatial structural features.

[0062] Fully connected neural network (FCNN) branch: used to process structured data (basic information, clinical data) and learn higher-order relationships between features.

[0063] Deep Neural Networks (DNNs) branch: used to process sequential or high-dimensional laboratory and functional performance data, capturing deep nonlinear patterns.

[0064] Attention fusion layer: used to dynamically weigh the importance of features from each branch and fuse them.

[0065] S31, through a convolutional neural network layer, extracts the kidney volume, blood flow, and cortical and medullary thickness features of the right and left kidneys from imaging data to generate imaging feature vectors;

[0066] Imaging data (such as CT, MRI, and ultrasound) contains rich spatial and functional perfusion information of the kidneys. This step utilizes a CNN subnetwork containing multiple convolutional layers, pooling layers, and activation layers to process the preprocessed right and left kidney images separately. The specific processing steps are as follows: Input is registered and normalized two-dimensional slices or three-dimensional volumetric data of both kidneys; convolutional layers extract local features (such as edges and textures) through sliding windows; deeper convolutional layers progressively combine these local features to form higher-level semantic features (such as the boundary between the renal cortex and medulla, the overall morphology of the kidney, and areas of uneven perfusion); finally, a global average pooling layer maps the extracted spatial features into two fixed-length feature vectors: a right kidney imaging feature vector. and left kidney imaging feature vector .

[0067] Understandably, the two vectors above encode the volume, shape, perfusion intensity, and depth features of the internal structure (cortex / medulla) of their respective kidneys.

[0068] S32, The patient's basic information is processed through a fully connected neural network layer to generate a basic information feature vector; wherein, the basic information includes age and gender;

[0069] Patient demographic characteristics are an important baseline for assessing renal function. This branch consists of several fully connected layers. The specific processing is as follows: input is a numerical vector of basic information (e.g., age, gender encoding); the fully connected layers learn the complex relationships between these basic variables and renal function potential through nonlinear transformations; for example, deep learning models may learn the implicit correlation between "advanced age" and "glomerular sclerosis"; output is a feature vector of basic information. .

[0070] S33, The physiological features in the clinical data are processed through a fully connected neural network layer to generate a clinical feature vector; wherein, the physiological features include blood pressure, urine output, serum creatinine, and urine protein;

[0071] Real-time clinical indicators directly reflect a patient's physiological state. The specific processing involves: inputting a clinical data vector, including blood pressure (systolic and diastolic), 24-hour urine output, serum creatinine (Scr), and urine protein / creatinine ratio (UPCR); another fully connected network submodule processes these indicators, for example, learning the non-linear relationships between "hypertension" and "elevated glomerular pressure," and between "decreased urine output" and "decreased renal concentrating ability"; and outputting a clinical feature vector. .

[0072] S34, extracting laboratory test feature vectors from laboratory test data through a deep neural network layer; wherein, the laboratory test data includes biomarkers in blood and urine;

[0073] Specific biomarkers in blood and urine (such as cystatin C, NGAL, KIM-1, and β2-microglobulin) can more sensitively and specifically reflect renal tubular or glomerular damage. The specific processing involves: inputting a multi-dimensional vector of laboratory test indicators; a deep DNN (usually containing more hidden layers) processes this high-dimensional biomarker data, which may exhibit complex collinearity. This network can uncover the interactions between biomarkers and identify the most discriminative combination patterns for the current renal function status of the patient; outputting a vector of laboratory test features. .

[0074] S35, the renal function performance data is processed through a deep neural network layer to generate a renal function performance feature vector; wherein, the renal function performance data includes urine test results, renal function change trends, and history of renal disease;

[0075] The patient's historical and dynamic functional data provide contextual information about the disease progression. The specific processing involves: inputting serialized or summarized functional performance data, such as the results of previous urine tests (red blood cells, casts), eGFR trend sequences, and encoding of specific kidney disease histories (diabetic nephropathy, hypertensive nephropathy, etc.); this DNN branch is designed to process temporal or categorical information, capturing dynamic and contextual features such as "progressive decline in kidney function" or "a specific history of kidney disease," features that cannot be provided by single-point-of-time data; and outputting a vector of kidney function performance features. In particular, this vector implicitly encodes the dynamic pattern of changes in the patient's renal function over time, which can serve as a historical basis for subsequent dynamic weight adjustments.

[0076] S36, the attention mechanism layer uses a weighted fusion strategy to fuse imaging feature vectors, basic information feature vectors, clinical feature vectors, laboratory test feature vectors, and renal function performance feature vectors to generate a comprehensive feature vector.

[0077] After obtaining the above five feature vectors, an attention mechanism is further introduced to achieve adaptive fusion. The specific processing procedure is as follows:

[0078] Input five feature vectors (This is) (aggregation) , , , .

[0079] The attention mechanism processes the process as follows: Each feature vector is transformed into a query, key, and value vector through an independent linear projection. The similarity between the query of a feature vector and the keys of all other feature vectors is calculated (e.g., using a dot product). These similarities are then normalized using methods such as Softmax to obtain a set of attention weights. Using the attention weights as coefficients, a weighted sum is applied to the values ​​of the other feature vectors to obtain... Context-enhanced representation.

[0080] Repeat the above process (i.e., multi-head or cross-attention) for each feature vector so that each vector can obtain relevant information from other vectors.

[0081] Finally, all context-enhanced feature vectors are concatenated and then dimensionality-reduced and integrated through a fully connected layer to generate the final comprehensive feature vector. This vector not only contains the original multi-source information, but also enriches the interaction relationships between these information, and highlights key information based on the characteristics of the current data, providing a powerful and robust feature representation for downstream GFR regression prediction.

[0082] S4, adjust the functional weights of the right and left kidneys in GFR estimation according to the functional difference measurement coefficient;

[0083] Based on the functional difference quantification coefficient obtained in the previous steps With integrated feature vector A dynamic, adaptive functional weight (w) is calculated for the patient's right and left kidneys respectively. R and w L These two weights will directly determine the contribution ratio of left and right kidney-related features in the subsequent deep regression model, so that the final GFR estimate can truly reflect the overall kidney function under the condition of bilateral kidney asymmetry.

[0084] As an example, adjusting the functional weights of the right and left kidneys in GFR estimation based on the functional difference metric coefficient includes:

[0085] The functional difference metric coefficient and one or more key sub-features from the comprehensive feature vector are input into a pre-constructed dynamic weight generation network;

[0086] Please see Figure 3The dynamic weight generation network includes a gated recurrent unit network and a multilayer perceptron; wherein, the gated recurrent unit network is used to analyze the historical functional change trend implied by the renal function performance feature vector of the patient to capture the dynamic pattern of renal function compensation or decompensation; the multilayer perceptron is used to simultaneously process the functional difference measurement coefficient and the imaging sub-feature vector related to renal structure and perfusion.

[0087] Among them, Gated Recurrent Unit (GRU) network extraction Mid-renal Functional Feature Vector The implicit timing information. For example, This data encodes eGFR value sequences and trends in urinary protein changes over the past year. Gated Circulatory Unit (GRU) networks are well-suited for processing this type of sequence data, effectively capturing long-term dependencies in historical evolution and identifying dynamic patterns such as "a sustained, slow decline in left kidney function while right kidney function remains stable or even shows compensatory enhancement." The final hidden state output vector H of the GRU pathway... gru This represents an understanding of the dynamic compensatory pattern of renal function in patients.

[0088] Multilayer perceptron (MLP) uses the functional difference metric coefficient D, and Imaging feature vectors closely related to kidney structure and immediate perfusion (such as specific numerical features of left and right kidney volume, blood flow, and cortical thickness) are input into a multilayer perceptron (MLP). The MLP learns, through multilayer nonlinear transformations, how the combination of structural, perfusion, and morphological differences between the left and right kidneys at the current moment affects their functional weight allocation. Its output vector H... mlp This represents the assessment of functional asymmetry at the current moment.

[0089] The output of the gated recurrent unit network and the output of the multilayer perceptron interact in the fusion layer, and a set of intermediate weight vectors are calculated through a cross-attention mechanism.

[0090] The high-level representation of the output will be H. gru and H mlp A cross-attention fusion layer is used to achieve deep interaction between historical trends and the current state. The specific processing procedure is as follows:

[0091] With H gru As a query, using H mlpAttention is calculated as keys and values. Essentially, this process makes the system think: "Based on historical trends in kidney function (Query), what aspects of the current structural and functional differences (Key-Value) are particularly noteworthy?" For example, if historical trends show that the left kidney is continuously decompensated, then any subtle structural deterioration (Key) of the left kidney in the current state will be amplified (given high weight) in the attention.

[0092] Similarly, H can also be used. mlp For Query, H gru Perform reverse calculations for the key-value pair, considering: "Based on the current asymmetry, look for support or explanation in historical trends."

[0093] After cross-attention interaction, a more discriminative intermediate weight vector W is generated, which integrates temporal and real-time information. mid Understandably, this vector encodes the underlying rationale for the decision on how to assign weights to the left and right kidneys.

[0094] The intermediate weight vector is input to a differentiable soft-assignment layer, which outputs the first functional weight of the right kidney and the second functional weight of the left kidney. The temperature parameter of the soft-assignment layer can be learned during model training to control the confidence and sharpness of the weight assignment.

[0095] Among them, the intermediate weight vector W mid Input a differentiable soft-assignment layer (such as Gumbel-Softmax or a Softmax layer with a temperature parameter) to generate the final right kidney functional weights. and left kidney function weight w L .

[0096] Understandably, the design of this differentiable soft-assignment layer ensures that the entire weight generation process is completely differentiable, allowing gradients to propagate back from the final GFR loss function to all parameters of the DWGN, thus achieving end-to-end optimization. Through learning, the network can automatically find the weight assignment strategy that best improves the overall GFR estimation accuracy. Additionally, the soft-assignment layer includes a temperature parameter τ. In the early stages of training, a higher value for τ makes w... R and w L The output tends to be smoother (e.g., 0.55 / 0.45), encouraging the network to explore more extensively. As training progresses, the τ value can be learned or gradually decreased, making the output sharper (e.g., 0.85 / 0.15), indicating that the network has greater confidence in its weight allocation decisions.

[0097] The final output satisfies w R +w L =1 and wR ,w L ∈(0,1). For patients with symmetrical renal function, the ideal output will approach (0.5,0.5); for patients with unilateral non-functional kidney, the ideal output will approach (1.0,0.0) or (0.0,1.0); for patients with varying degrees of asymmetry, the corresponding intermediate value will be output.

[0098] S5. Input the adjusted functional weights and the global feature vector into the deep regression model, calculate the corrected GFR estimate and output it.

[0099] Traditional regression models often struggle to balance the holistic nature of global features with the specificity of local features when dealing with complex prediction tasks involving multiple sources, heterogeneity, and the need to consider local weight adjustments. To address this technical problem, this invention designs a dual-pathway gated fusion deep regression model architecture (see [link to relevant documentation]). Figure 4 The weighted feature pathway focuses on applying dynamic weights to emphasize the specific features of the left and right kidneys, generating a context vector that reflects the functional contribution of both kidneys after weighting. The original feature pathway is dedicated to mining deep and abstract semantic information related to GFR from the global comprehensive features. The gated fusion module adaptively determines the extent to which the weighted local kidney information and the global semantic information should be relied upon in the final decision.

[0100] As an example, the deep regression model includes a weighted feature path and a raw feature path; then, the adjusted function weights and the global feature vector are input into the deep regression model to calculate the corrected GFR estimate, including:

[0101] S51, the weighted feature pathway performs a Hadamard product operation on the first functional weight and the imaging and functional performance sub-feature vectors related to the right kidney to generate right kidney weighted features; similarly, it performs a Hadamard product operation on the corresponding sub-feature vectors related to the second functional weight and the left kidney to generate left kidney weighted features; the right kidney weighted features and the left kidney weighted features are aggregated to generate kidney-specific weighted context vectors.

[0102] The weighted feature pathway is specifically designed in this invention to address renal function asymmetry; its input is dynamic weights. and comprehensive feature vectors Imaging and functional features directly related to the left and right kidneys.

[0103] First, from Extract the sub-feature vector directly associated with the right kidney. (For example, imaging features of the right kidney, functional indicators of the right kidney's sub-renal segment, etc.) and sub-feature vectors directly related to the left kidney. .

[0104] Weight of right kidney function With right kidney feature vector Perform the Hadamard product (i.e., element-wise multiplication):

[0105]

[0106] This calculation will assign a scalar weight representing the contribution of the right kidney to function. This information is broadcast to all relevant feature dimensions, thereby uniformly gaining or attenuating each piece of information about the right kidney feature.

[0107] Similarly, calculate the weighted characteristics of the left kidney: .

[0108] Context vector generation: weighted right kidney features Features of the left kidney Aggregation is performed. Aggregation can be done by concatenation or element-wise addition to form a new feature vector. This vector encodes kidney-specific information adjusted for renal functional asymmetry. This vector is the kidney-specific weighted context vector. .

[0109] S52, the original feature path performs high-order nonlinear transformation and abstraction on the comprehensive feature vector through a deep network composed of multiple residual blocks to extract global deep semantic features related to GFR.

[0110] Based on this original feature pathway, it is possible to go beyond the local perspective of the kidney and uncover complex, higher-order patterns related to GFR from the patient's global information.

[0111] The original feature path uses a deep neural network (DNN) composed of multiple stacked residual blocks. The residual blocks effectively alleviate the vanishing gradient problem in deep networks through shortcut connections, allowing the network to be very deep and thus capable of learning extremely complex nonlinear relationships. The specific processing is as follows:

[0112] Combined feature vectors Input this deep residual network. Through layers of nonlinear transformations, the network gradually abstracts and combines information from all dimensions, including basic information, clinical indicators, laboratory data, and functional history. It can learn complex correlation patterns between factors such as advanced age, history of hypertension, specific combinations of biomarkers, and the rate of glomerular sclerosis—patterns that cannot be directly obtained from kidney imaging or local features alone.

[0113] The output of the original feature path is global deep semantic features. This represents a highly abstract and comprehensive understanding of the patient's overall health status, particularly the systemic factors associated with renal function decline.

[0114] S53, the kidney-specific weighted context vector is concatenated with the global deep semantic features and input into the gated fusion module; the gated fusion module learns to generate a set of dynamic fusion weights to adaptively control the contribution ratio of weighted context information and global semantic information in the final decision.

[0115] Focusing on the adjusted contribution of both kidneys and Focusing on the patient's overall state, directly concatenating or simply adding features may introduce noise or dilute information. To address this, a gating mechanism is introduced: a gating fusion module is designed, the core of which is a small neural network (typically consisting of several fully connected layers), and the features obtained from the previous concatenation step are used for fusion. For input.

[0116] The gating fusion module learns and outputs a pair of dynamic fusion weights. and (Typically constrained between 0 and 1 using the Sigmoid function, and normalizable). The generation process of these two weights is adaptive, depending on the current input features themselves.

[0117] For example, in a young patient with severe functional asymmetry due to unilateral renal artery stenosis but otherwise good health, the model can learn to confer... A relatively high value (close to 1). Lower values ​​indicate that decisions should primarily rely on weighted kidney-specific information. Conversely, for an elderly patient with mild asymmetry in renal function but complicated by severe diabetes, heart failure, and other systemic diseases, the model can assign... A higher weighting suggests that overall health status may have a greater impact on GFR than the asymmetry between the two kidneys itself.

[0118] The generated dynamic weights are used to weight and fuse the features of the two pathways to obtain the final feature representation used for prediction. This mechanism enables the model to have context-aware decision-making capabilities.

[0119] S54, the output of the gated fusion module is processed by a regression prediction head containing a dropout layer, and finally outputs a corrected GFR estimate.

[0120] The fused high-level feature representation The final regression prediction head is input. The regression prediction head consists of 1-3 fully connected layers, with the last layer being a linear output layer that outputs a scalar value. A Dropout layer is introduced in the key fully connected layers, randomly dropping a portion of neurons during training. This is an effective regularization technique that prevents the model from overfitting to the training data and enhances its generalization ability on new patient data. The regression prediction head ultimately outputs the corrected GFR estimate. The unit is mL / min / 1.73m².

[0121] It is understandable that the entire model (including all parts of feature extraction, weight generation, and regression prediction) works by minimizing the predicted value. End-to-end joint training is performed between the loss function (such as mean squared error MSE or smoothed L1 loss) and the true GFR value measured by the gold standard method (such as inulin clearance rate).

[0122] Please see Figure 5 This invention also provides a glomerular filtration rate estimation device 100 based on comprehensive data analysis, the device comprising:

[0123] The data acquisition module 101 is used to acquire the patient's basic information, clinical data, laboratory test data and imaging data; wherein, the imaging data includes at least the kidney volume, blood flow, thickness of the renal cortex and medulla, and renal function of the right and left kidneys;

[0124] The functional difference quantification module 102 is used to calculate a functional difference measurement coefficient for quantifying the asymmetry of bilateral renal function based on the imaging data by comparing the differences in renal volume, blood flow, and thickness of the renal cortex and medulla between the right and left kidneys.

[0125] The global feature extraction and fusion module 103 is used to extract the patient's global feature vector through a deep learning model, including the imaging feature vector of each kidney extracted from imaging data, the physiological feature vector extracted from clinical data, and the laboratory test feature vector extracted from laboratory test data, and to fuse the above feature vectors to generate a comprehensive feature vector.

[0126] The functional weight dynamic adjustment module 104 is used to adjust the functional weights of the right and left kidneys in GFR estimation according to the functional difference measurement coefficient.

[0127] The corrected estimation output module 105 is used to input the adjusted functional weights and the global feature vector into the deep regression model, calculate the corrected GFR estimate, and output it.

[0128] As an example, please refer to Figure 6The functional difference quantification module 102 includes:

[0129] The volume difference calculation unit 1021 is used to calculate the difference in kidney volume between the right and left kidneys to obtain the volume difference coefficient.

[0130] Blood flow difference calculation unit 1022 is used to calculate the difference in blood flow between the right kidney and the left kidney to generate a blood flow difference coefficient;

[0131] Thickness difference calculation unit 1023 is used to calculate the difference in thickness between the renal cortex and medulla of the right and left kidneys to generate a thickness difference coefficient.

[0132] The weighted fusion unit 1024 is used to obtain the functional difference measurement coefficient by weighted fusion of the volume difference coefficient, the blood flow difference coefficient and the thickness difference coefficient.

[0133] As an example, please refer to Figure 7 The deep learning model includes multiple convolutional neural network layers, fully connected neural network layers, deep neural network layers, and attention mechanism layers; the global feature extraction and fusion module 103 includes:

[0134] The imaging feature extraction unit 1031 is used to extract the kidney volume, blood flow, and cortical and medullary thickness features of the right and left kidneys from imaging data through a convolutional neural network layer, and generate imaging feature vectors.

[0135] The basic information feature extraction unit 1032 is used to process the patient's basic information through a fully connected neural network layer to generate a basic information feature vector; wherein, the basic information includes age and gender;

[0136] The clinical feature extraction unit 1033 is used to process the physiological features in clinical data through a fully connected neural network layer to generate a clinical feature vector; wherein, the physiological features include blood pressure, urine output, serum creatinine, and urine protein;

[0137] The laboratory testing feature extraction unit 1034 is used to extract laboratory testing feature vectors from laboratory testing data through a deep neural network layer; wherein, the laboratory testing data includes biomarkers in blood and urine;

[0138] The functional performance feature extraction unit 1035 is used to process renal function performance data through a deep neural network layer to generate a renal function performance feature vector; wherein, the renal function performance data includes urine test results, renal function change trends, and history of renal disease.

[0139] The attention fusion unit 1036 is used in the attention mechanism layer to fuse imaging feature vectors, basic information feature vectors, clinical feature vectors, laboratory test feature vectors and renal function performance feature vectors using a weighted fusion strategy to generate a comprehensive feature vector.

[0140] As an example, the function weight dynamic adjustment module 104 is configured as follows:

[0141] The functional difference metric coefficient and one or more key sub-features from the comprehensive feature vector are input into a pre-constructed dynamic weight generation network;

[0142] The dynamic weight generation network includes a gated recurrent unit network and a multilayer perceptron; wherein, the gated recurrent unit network is used to analyze the historical functional change trend implied by the renal function performance feature vector of the patient to capture the dynamic pattern of renal function compensation or decompensation; the multilayer perceptron is used to simultaneously process the functional difference measurement coefficient and the imaging sub-feature vector related to renal structure and perfusion.

[0143] The output of the gated recurrent unit network and the output of the multilayer perceptron interact in the fusion layer, and a set of intermediate weight vectors are calculated through a cross-attention mechanism.

[0144] The intermediate weight vector is input to a differentiable soft-assignment layer, which outputs the first functional weight of the right kidney and the second functional weight of the left kidney. The temperature parameter of the soft-assignment layer can be learned during model training to control the confidence and sharpness of the weight assignment.

[0145] As an example, please refer to Figure 8 The deep regression model includes a weighted feature pathway and a raw feature pathway; therefore, the corrected estimation output module 105 includes:

[0146] A weighted context generation unit 1051, deployed in the weighted feature pathway, is used to perform a Hadamard product operation on the imaging and functional sub-feature vectors related to the right kidney with the first functional weight to generate a right kidney weighted feature; similarly, it performs a Hadamard product operation on the corresponding sub-feature vectors related to the left kidney with the second functional weight to generate a left kidney weighted feature; and it aggregates the right kidney weighted feature and the left kidney weighted feature to generate a kidney-specific weighted context vector.

[0147] The global semantic feature extraction unit 1052 is deployed in the original feature path and is used to perform high-order nonlinear transformation and abstraction on the comprehensive feature vector through a deep network composed of multiple residual blocks to extract global deep semantic features related to GFR.

[0148] The gated fusion unit 1053 is used to concatenate the kidney-specific weighted context vector with the global deep semantic features and input them into the gated fusion module; the gated fusion module learns to generate a set of dynamic fusion weights to adaptively control the contribution ratio of weighted context information and global semantic information in the final decision.

[0149] The regression prediction output unit 1054 is configured to process the output of the gated fusion module via a regression prediction head containing a dropout layer, and finally output the corrected GFR estimate.

[0150] This invention also provides an electronic device, including: a processor; and a memory storing computer program instructions, which, when executed by the processor, cause the processor to perform the method described above.

[0151] This invention also provides a computer-readable medium storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the method described above.

[0152] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for estimating glomerular filtration rate based on comprehensive data analysis, characterized in that, The method includes the following steps: S1, acquire the patient's basic information, clinical data, laboratory test data and imaging data; wherein, the imaging data includes at least the kidney volume, blood flow, thickness of the renal cortex and medulla, and renal function of the right and left kidneys; S2, Based on the imaging data, a functional difference measurement coefficient for quantifying the asymmetry of renal function is calculated by comparing the differences in renal volume, blood flow, and thickness of the renal cortex and medulla between the right and left kidneys; S3 extracts the patient's global feature vector through a deep learning model, including the imaging feature vector of each kidney extracted from imaging data, the physiological feature vector extracted from clinical data, and the laboratory test feature vector extracted from laboratory test data, and fuses the above feature vectors to generate a comprehensive feature vector. S4, adjust the functional weights of the right and left kidneys in GFR estimation according to the functional difference measurement coefficient; S5. Input the adjusted functional weights and the global feature vector into the deep regression model, calculate the corrected GFR estimate and output it.

2. The method for estimating glomerular filtration rate based on comprehensive data analysis according to claim 1, characterized in that: Functional difference measures, used to quantify the asymmetry of renal function, are calculated by comparing the differences in kidney volume, blood flow, and cortical and medullary thickness between the right and left kidneys. These measures include: S21, calculate the difference in kidney volume between the right and left kidneys to obtain the volume difference coefficient; S22, calculate the difference in blood flow between the right and left kidneys to generate a blood flow difference coefficient; S23, calculate the difference in thickness between the renal cortex and medulla of the right and left kidneys to generate a thickness difference coefficient; S24, by weighting and fusing the volume difference coefficient, the blood flow difference coefficient, and the thickness difference coefficient, a functional difference measurement coefficient is obtained.

3. The method for estimating glomerular filtration rate based on comprehensive data analysis according to claim 1, characterized in that: The deep learning model includes multiple convolutional neural network layers, fully connected neural network layers, deep neural network layers, and attention mechanism layers. It extracts the patient's global feature vector using the deep learning model and fuses these feature vectors to generate a comprehensive feature vector, including: S31, through a convolutional neural network layer, extracts the kidney volume, blood flow, and cortical and medullary thickness features of the right and left kidneys from imaging data to generate imaging feature vectors; S32, The patient's basic information is processed through a fully connected neural network layer to generate a basic information feature vector; wherein, the basic information includes age and gender; S33, The physiological features in the clinical data are processed through a fully connected neural network layer to generate a clinical feature vector; wherein, the physiological features include blood pressure, urine output, serum creatinine, and urine protein; S34, extracting laboratory test feature vectors from laboratory test data through a deep neural network layer; wherein, the laboratory test data includes biomarkers in blood and urine; S35, the renal function performance data is processed through a deep neural network layer to generate a renal function performance feature vector; wherein, the renal function performance data includes urine test results, renal function change trends, and history of renal disease; S36, the attention mechanism layer uses a weighted fusion strategy to fuse imaging feature vectors, basic information feature vectors, clinical feature vectors, laboratory test feature vectors, and renal function performance feature vectors to generate a comprehensive feature vector.

4. The method for estimating glomerular filtration rate based on comprehensive data analysis according to claim 3, characterized in that: Adjusting the functional weights of the right and left kidneys in GFR estimation based on the aforementioned functional difference metric coefficient includes: S41, input the functional difference measurement coefficient and one or more key sub-features in the comprehensive feature vector into a pre-constructed dynamic weight generation network; The dynamic weight generation network includes a gated recurrent unit network and a multilayer perceptron; wherein, the gated recurrent unit network is used to analyze the historical functional change trend implied by the renal function performance feature vector of the patient to capture the dynamic pattern of renal function compensation or decompensation; the multilayer perceptron is used to simultaneously process the functional difference measurement coefficient and the imaging sub-feature vector related to renal structure and perfusion. The output of the gated recurrent unit network and the output of the multilayer perceptron interact in the fusion layer, and a set of intermediate weight vectors are calculated through a cross-attention mechanism. The intermediate weight vector is input to a differentiable soft-assignment layer, which outputs the first functional weight of the right kidney and the second functional weight of the left kidney. The temperature parameter of the soft-assignment layer can be learned during model training to control the confidence and sharpness of the weight assignment.

5. The method for estimating glomerular filtration rate based on comprehensive data analysis according to claim 4, characterized in that: The deep regression model includes a weighted feature pathway and a raw feature pathway; The adjusted function weights and the global feature vector are then input into the deep regression model to calculate the corrected GFR estimate, including: S51, the weighted feature pathway performs a Hadamard product operation on the first functional weight and the imaging and functional performance sub-feature vectors related to the right kidney to generate right kidney weighted features; similarly, it performs a Hadamard product operation on the corresponding sub-feature vectors related to the second functional weight and the left kidney to generate left kidney weighted features; the right kidney weighted features and the left kidney weighted features are aggregated to generate kidney-specific weighted context vectors. S52, the original feature path performs high-order nonlinear transformation and abstraction on the comprehensive feature vector through a deep network composed of multiple residual blocks to extract global deep semantic features related to GFR. S53, the kidney-specific weighted context vector is concatenated with the global deep semantic features and input into the gated fusion module; the gated fusion module learns to generate a set of dynamic fusion weights to adaptively control the contribution ratio of weighted context information and global semantic information in the final decision. S54, the output of the gated fusion module is processed by a regression prediction head containing a dropout layer, and finally outputs a corrected GFR estimate.

6. A glomerular filtration rate estimation device based on comprehensive data analysis, characterized in that, The device includes: The data acquisition module is used to acquire the patient's basic information, clinical data, laboratory test data, and imaging data; wherein, the imaging data includes at least the kidney volume, blood flow, thickness of the renal cortex and medulla, and renal function of the right and left kidneys; The functional difference quantification module is used to calculate a functional difference measurement coefficient for quantifying the asymmetry of bilateral renal function based on the imaging data by comparing the differences in renal volume, blood flow, and thickness of the renal cortex and medulla between the right and left kidneys. The global feature extraction and fusion module is used to extract the patient's global feature vector through a deep learning model, including the imaging feature vector of each kidney extracted from imaging data, the physiological feature vector extracted from clinical data, and the laboratory test feature vector extracted from laboratory test data, and to fuse the above feature vectors to generate a comprehensive feature vector. The functional weight dynamic adjustment module is used to adjust the functional weights of the right and left kidneys in GFR estimation based on the functional difference measurement coefficient. The corrected estimation output module is used to input the adjusted function weights and the global feature vector into the deep regression model, calculate the corrected GFR estimate, and output it.

7. The glomerular filtration rate estimation device based on comprehensive data analysis according to claim 6, characterized in that: The functional difference quantification module includes: The volume difference calculation unit is used to calculate the difference in kidney volume between the right and left kidneys to obtain the volume difference coefficient. The blood flow difference calculation unit is used to calculate the difference in blood flow between the right and left kidneys to generate a blood flow difference coefficient. The thickness difference calculation unit is used to calculate the difference in thickness between the renal cortex and medulla of the right and left kidneys to generate a thickness difference coefficient. The weighted fusion unit is used to obtain the functional difference measurement coefficient by weighted fusion of the volume difference coefficient, the blood flow difference coefficient and the thickness difference coefficient.

8. A glomerular filtration rate estimation device based on comprehensive data analysis according to claim 6, characterized in that: The deep learning model includes multiple convolutional neural network layers, fully connected neural network layers, deep neural network layers, and attention mechanism layers; the global feature extraction and fusion module includes: The imaging feature extraction unit is used to extract the kidney volume, blood flow, and cortical and medullary thickness features of the right and left kidneys from imaging data through a convolutional neural network layer, and generate imaging feature vectors. The basic information feature extraction unit is used to process the patient's basic information through a fully connected neural network layer to generate a basic information feature vector; wherein, the basic information includes age and gender; The clinical feature extraction unit is used to process the physiological features in clinical data through a fully connected neural network layer to generate a clinical feature vector; wherein the physiological features include blood pressure, urine output, serum creatinine, and urine protein; A laboratory testing feature extraction unit is used to extract laboratory testing feature vectors from laboratory testing data through a deep neural network layer; wherein, the laboratory testing data includes biomarkers in blood and urine; The functional performance feature extraction unit is used to process renal function performance data through a deep neural network layer to generate a renal function performance feature vector; wherein, the renal function performance data includes urine test results, renal function change trends, and history of renal disease; The attention fusion unit is used in the attention mechanism layer to fuse imaging feature vectors, basic information feature vectors, clinical feature vectors, laboratory test feature vectors, and renal function performance feature vectors using a weighted fusion strategy to generate a comprehensive feature vector.

9. An electronic device, comprising: processor; And a memory, wherein computer program instructions are stored in the memory, characterized in that: when the computer program instructions are executed by the processor, the processor causes the processor to perform the method as described in any one of claims 1-5.

10. A computer-readable medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by the processor, the processor causes the processor to perform the method as described in any one of claims 1-5.

Images