Atrial fibrillation patient acute kidney injury early warning system and method based on deep learning

Abstract

The invention discloses an atrial fibrillation patient acute kidney injury early warning system based on deep learning. The system comprises a clinical data integration unit; a missing data reconstruction unit; a feature matrix construction unit; a deep neural network adaptive training unit; and a multi-dimensional model evaluation and decision interpretation unit. According to the method, the acute kidney injury complicated by the atrial fibrillation patient can be quickly predicted for 24 hours in advance, early discovery, diagnosis and treatment of the acute kidney injury patient are realized, the survival rate after the atrial fibrillation is improved, the prognosis of the atrial fibrillation patient is improved, and the occurrence rate of the acute kidney injury is reduced.

Description

Technical Field

[0001] The present invention relates to the field of predictive model construction, and in particular to a deep learning-based early warning system and method for acute kidney injury in patients with atrial fibrillation. Background Art

[0002] Atrial fibrillation (AF), the most common clinical arrhythmia, and its multi-organ interactions with acute kidney injury (AKI) have become a hot topic in critical care medicine research. According to the 2023 KDIGO guideline update, the incidence of AKI in patients with AF is 2.3 times higher than in those without AF (95% CI 1.7-3.1) and is significantly positively correlated with all-cause mortality (HR = 4.12, P < 0.001). Its pathological mechanisms involve multiple factors, including hemodynamic disturbances, the thrombotic-inflammatory cascade, and iatrogenic injury. These include fluctuations in renal perfusion pressure (a decrease ≥ 15 mmHg) caused by loss of atrial pump function and hemorrhagic tubular necrosis associated with anticoagulation.

[0003] Current clinical practice faces core challenges such as insufficient prediction timeliness, lack of model specificity, and inefficient data utilization. Traditional serum creatinine monitoring has a 6-12 hour delay, often missing the optimal intervention window. Existing AKI prediction tools (such as the NELA score) fail to fully account for the variability of hemodynamic parameters unique to atrial fibrillation. Furthermore, 90% of dynamic monitoring indicators (such as arterial blood pressure waveforms and electrolyte trends) in electronic medical records are not effectively characterized. Summary of the Invention

[0004] To address the above-mentioned shortcomings, the present invention innovatively proposes an AKI early warning framework based on the hemodynamic characteristics of atrial fibrillation. By integrating multi-dimensional time series statistics with a deep learning architecture, it breaks through the sensitivity limitation of single biomarker detection (AUC improvement ≥ 0.15), identifies the early compensatory characteristics of atrial fibrillation-related AKI (the warning time window is shifted forward by 8±2.6 hours), and achieves accurate stratification and early warning of AKI risk in patients with atrial fibrillation, providing a new paradigm for precision medicine.

[0005] According to one aspect of the present invention, a deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation is provided, comprising: The clinical data integration unit is used to extract clinical data based on the clinical data of hospitalized patients after atrial fibrillation, according to the inclusion and exclusion criteria, and perform clinical data purification, multi-source data integration and data verification; Missing data reconstruction unit, used to build a multi-dimensional feature space adaptive interpolation model based on the random forest algorithm to achieve nonlinear enhancement of clinical data, using a dynamic priority interpolation mechanism and recursive filling strategy to preserve the original data index structure; A feature matrix construction unit is used to screen out characteristic variables closely related to atrial fibrillation complicated by AKI based on LASSO regression, implement parallel loading and tensor storage of structured high-dimensional clinical matrices, and perform standardized conversion and structured adaptation of medical data through a multi-stage processing pipeline; Deep neural network adaptive training unit, which is used to integrate configurable convolutional neural network architecture and achieve parameter adaptive training through dynamic learning rate adjustment and loss function optimization; The multidimensional model evaluation and decision interpretation unit is used to output the 24-hour AKI risk probability value and simultaneously generate a three-dimensional evaluation matrix that includes confidence assessment, SHAP value-driven feature attribution analysis, and clinical decision recommendations.

[0006] In some technical solutions, it also includes: an external verification unit, which is used to execute the deep learning-based dynamic early warning system for acute kidney injury in patients with atrial fibrillation proposed by the present invention on another completely different data set to verify the universality of the method of the present invention.

[0007] As a further technical solution, the clinical data integration unit is further used to: Clinical data purification was performed to exclude cases with comorbidities such as renal cancer based on ICD codes, and data range verification was performed, such as age > 18 years; Multi-source data integration: Load structured clinical data from medical databases such as MIMIC-IV, integrating the following multidimensional data: patient basic characteristics, demographic parameters such as age, gender, and BMI; time-series physiological indicators, laboratory indicators such as creatinine, hemoglobin, and electrolytes dynamically monitored in the ICU ward; treatment intervention records, treatment characteristics such as NSAID use and mechanical ventilation; complication characteristics, quantitative parameters of the Elixhauser comorbidity index such as sepsis and malignant tumors; organ function scores, standardized critical care scoring systems such as SOFA and SAPSII; Data validation, automatically filtering abnormal records through range validation rules (such as pH∈[6.8,7.8]).

[0008] As a further technical solution, the missing data reconstruction unit is further used to: By integrating the feature importance analysis module of the random forest algorithm, an adaptive interpolation model for multidimensional feature space is constructed. Using subspace sampling technology under the bagging framework, feature interpolation points that conform to the original data distribution are dynamically generated to achieve nonlinear enhancement of medical time series data. Dynamic priority interpolation mechanism, builds a feature missing rate sorting queue (in ascending order), and prioritizes columns with lower missing rates; A recursive filling strategy is adopted, and the filled columns automatically participate in the reconstruction of subsequent features; The original data index structure is retained to ensure the temporal consistency of patients.

[0009] As a further technical solution, the feature matrix construction unit is further used to: Based on LASSO regression, characteristic variables closely related to atrial fibrillation complicated with AKI were screened out; Parallel loading: read and load the imputed clinical high-dimensional data from CSV files; Feature standardization, integrating the Z-score standardization algorithm to eliminate feature dimension differences; Tensor storage converts input clinical data into a deep learning-friendly tensor structure through dimension expansion and channel permutation operations, building a configurable batch data loading pipeline.

[0010] When the characteristic matrix construction unit screens characteristic variables based on LASSO regression, the final characteristic variables are determined through the regression path and cross-validation curve of LASSO regression, including 34 types, including demographic information, disease scores, laboratory indicators, clinical parameters, treatment interventions, etc.

[0011] As a further technical solution, the deep neural network adaptive training unit is further used to: Build an extensible modular neural network architecture library, including data loaders, hyperparameter setting units, multi-round training units, and verification units. Perform multiple rounds of training on the convolutional neural network model using the configured batchable training data, and visualize the training results through the verification unit to adjust the hyperparameter configuration.

[0012] The convolutional neural network architecture of the deep neural network adaptive training unit consists of three layers of convolution modules, each of which includes a convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer, and a random dropout layer; the size of all convolution kernels is 5, the step size is 1, the padding value is 2, the number of convolution kernels in the first layer is 16, the number of convolution kernels in the second layer is 32, and the number of convolution kernels in the third layer is 64; the pooling kernel size of the maximum pooling layer after each convolution layer is 2, and the step size is 2; a random dropout layer is used after each layer of pooling, and the random dropout probability is 0.3; after three layers of convolution and pooling operations, an adaptive average pooling layer is used to pool the feature map into a feature vector with a fixed length of 1.

[0013] As a further technical solution, the multidimensional model evaluation and decision interpretation unit is further used to: Output the confusion matrix and area under the receiver operating characteristic curve (AUC) of the model warning results; Calculate the model's accuracy, sensitivity, specificity, recall, and F1 score based on the confusion matrix; Based on the SHAP method, a feature importance visualization diagram, a single-patient AKI early warning process diagram, and a feature importance ranking diagram are output; Calibration curves were generated using 1000 bootstrap resamplings to assess the calibration accuracy of the model; The clinical utility of the prediction model was evaluated by decision curve analysis; Clinical impact curve analysis was performed to determine the optimal probability threshold for clinical decision making.

[0014] As a further technical solution, the calibration curve has the model-predicted probability value on the horizontal axis and the observed frequency of AKI in AF patients on the vertical axis. The closeness of the model calibration curve to the ideal calibration line indicates high model prediction accuracy. The clinical impact curve has the high-risk threshold and the corresponding cost-effectiveness ratio on the horizontal axis, and the number of high-risk patients on the vertical axis, demonstrating the impact of different risk threshold settings on the number of high-risk patients and the number of patients who actually experience AKI events. The decision curve analysis has the high-risk threshold on the horizontal axis and the net benefit on the vertical axis. By comparing the net benefit with the assumption that all patients do not develop AKI, the risk threshold is selected to balance the warning coverage and actual benefits.

[0015] According to one aspect of the present invention, a method for early warning of acute kidney injury in patients with atrial fibrillation based on deep learning is provided, comprising: The clinical data of hospitalized patients after atrial fibrillation is obtained through the clinical data integration unit, and the clinical data are extracted and processed according to the inclusion and exclusion criteria; A multi-dimensional feature space adaptive interpolation model is constructed based on the random forest algorithm using the missing data reconstruction unit to process the missing data; With the help of the feature matrix construction unit, feature variables are screened based on LASSO regression to perform relevant processing of the structured high-dimensional clinical matrix; Use the deep neural network adaptive training unit to integrate the convolutional neural network architecture, and perform parameter adaptive training through dynamic learning rate adjustment and loss function optimization; The multidimensional model evaluation and decision interpretation unit outputs the 24-hour AKI risk probability value, generating a three-dimensional evaluation matrix that includes confidence assessment, SHAP value-driven feature attribution analysis, and clinical decision recommendations.

[0016] According to one aspect of the present invention, a deep learning-based early warning device for acute kidney injury in patients with atrial fibrillation is provided, comprising a memory and a processor, wherein the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the deep learning-based early warning method for acute kidney injury in patients with atrial fibrillation.

[0017] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the deep learning-based early warning method for acute kidney injury in patients with atrial fibrillation.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The present invention can quickly predict acute kidney injury in patients with atrial fibrillation 24 hours in advance, achieve early detection, diagnosis and treatment of patients with acute kidney injury, increase the survival rate after atrial fibrillation and improve the prognosis of patients with atrial fibrillation, and reduce the incidence of acute kidney injury. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction will be given below to the drawings used in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 Schematic diagram of a deep learning-based dynamic early warning system for acute kidney injury in patients with atrial fibrillation in the present invention; Figure 2 Schematic diagram of the inclusion and exclusion criteria for clinical data in the present invention; Figure 3 Schematic diagram of the regression path of LASSO regression analysis for screening effective features in the present invention; Figure 4 Schematic diagram of the cross-validation curve of LASSO regression analysis for screening effective features in the present invention; Figure 5 This is a structural diagram of the deep neural network in the present invention; Figure 6 A schematic diagram of the confusion matrix of the output results of the early warning system model in the present invention; Figure 7 Schematic diagram of the ROC curve of the early warning system model validation set in the present invention; Figure 8 This is a visualization diagram of the feature importance of the early warning system model based on the SHAP method in the present invention; Figure 9 Schematic diagram of AKI warning for a single patient based on the SHAP method of the early warning system model of the present invention; Figure 10 This is the feature influence heat map of the early warning system model based on the SHAP method in the present invention; Figure 11 Schematic diagram of the calibration curve for warning of AKI in AF patients in the present invention; Figure 12 Schematic diagram of the clinical impact curve for warning of AKI in AF patients according to the present invention; Figure 13 Schematic diagram of the decision curve analysis for warning AF patients of AKI in the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0022] The terms "including" and "having" and any variations thereof in the description and claims of the present invention and the accompanying drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements, but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0023] In the description of the embodiments of the present invention, the words "exemplary," "for example," or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for example" in the embodiments of the present invention should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary," "for example," or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] In the description of the embodiments of the present invention, unless otherwise specified, “ / ” means or, for example, A / B can mean A or B; “and / or” in the text is merely a description of the association relationship of associated objects, indicating that three relationships may exist, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present invention, “multiple” refers to two or more than two.

[0025] In some processes described in the embodiments of the present invention, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order in which they appear in the embodiments of the present invention, or may be executed in parallel. The sequence numbers of the operations are only used to distinguish different operations and do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations or steps may be executed in sequence or in parallel, and these operations or steps may be combined.

[0026] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0027] First, refer to Figure 1 , Figure 1 Schematic diagram of a deep learning-based dynamic early warning system for acute kidney injury in patients with atrial fibrillation in the present invention. Figure 1 Figure 2 shows a schematic diagram of a deep learning-based dynamic early warning system for acute kidney injury in patients with atrial fibrillation, including: Clinical data integration unit, which is used to collect clinical information of hospitalized patients after atrial fibrillation occurs and extract clinical data according to inclusion and exclusion criteria; Missing data reconstruction unit, which is used to build a multi-dimensional feature space adaptive interpolation model based on the random forest algorithm to achieve nonlinear enhancement of clinical data; A feature matrix construction unit, which is used to screen out feature variables closely related to atrial fibrillation complicated by AKI based on LASSO regression, preprocess relevant feature variables, implement parallel loading and tensor storage of structured high-dimensional clinical matrices, and achieve standardized conversion and structured adaptation of medical data through a multi-level processing pipeline; A deep neural network adaptive training unit, which is used to integrate a configurable convolutional neural network architecture and achieve parameter adaptive training through dynamic learning rate adjustment and loss function optimization; A multidimensional model evaluation and decision interpretation unit is used to output a 24-hour AKI risk probability value and simultaneously generate a three-dimensional evaluation matrix that includes confidence assessment, SHAP value-driven feature attribution analysis, and clinical decision recommendations.

[0028] In this embodiment, the clinical data integration unit is Figure 2 , Figure 2 This is a schematic diagram of the inclusion and exclusion criteria for the clinical data in the present invention. There are 94,458 ICU records in the MIMIC-IV database, 9,216 non-initial admission records, 2,347 records with urinary tract complications, and 17,523 records with an ICU duration of less than 24 hours. On this basis, there are 8,897 records of patients with atrial fibrillation.

[0029] Among the above 8,897 records, 3,095 records had clinical data missing greater than 20%, and 22 records were missing two key physiological data: serum creatinine or urine volume. After exclusion, 5,780 valid data were obtained for the training and construction of a deep learning-based dynamic early warning system for acute kidney injury in patients with atrial fibrillation proposed in this invention.

[0030] It is worth noting that among the 5780 valid data, the number of concurrent AKI records was 3884, and the number of non-concurrent AKI records was 1896.

[0031] In this example, missing data is a common phenomenon in clinical medical records. The 5,780 valid data items screened according to the inclusion and exclusion criteria all contain some degree of missing data. When missing values are present in the data, directly addressing them can affect the statistical efficacy and skew the data distribution, leading to significant errors in the research results. Therefore, missing values must be addressed.

[0032] In this embodiment, the missing data reconstruction unit constructs an adaptive interpolation model for multidimensional feature space by integrating the feature importance analysis module of the random forest algorithm. Using subspace sampling technology under the bagging framework, it dynamically generates feature interpolation points that conform to the original data distribution, achieving nonlinear enhancement of medical time series data. Dynamic priority interpolation mechanism, builds a feature missing rate sorting queue (in ascending order), and prioritizes columns with lower missing rates; A recursive filling strategy is adopted, and the filled columns automatically participate in the reconstruction of subsequent features; The original data index structure is retained to ensure the temporal consistency of patients; In this embodiment, the feature matrix construction unit screens out feature variables closely related to atrial fibrillation complicated by AKI based on LASSO regression, preprocesses the relevant feature variables, implements parallel loading and tensor storage of structured high-dimensional clinical matrices, and achieves standardized conversion and structured adaptation of medical data through a multi-level processing pipeline. See also Figure 3 , Figure 3 Schematic diagram of the regression path of the LASSO regression analysis for screening effective features in the present invention. The LASSO regression, namely the Least Absolute Shrinkage and Selection Operator (LSS) regression analysis, is a variable selection and regularization method for linear regression models. LASSO regression adds a penalty term to ordinary linear regression to constrain and adjust the regression coefficients.

[0033] Figure 3The horizontal axis variable is the regularization parameter λ of the LASSO regression model, which is used to control the intensity of the penalty, and the vertical axis variable is the regression coefficient. In the LASSO regression model, the larger the λ, the greater the penalty for the regression coefficient, which will cause more coefficients to approach zero; the smaller the λ, the smaller the penalty for the regression coefficient, which will make the model closer to ordinary linear regression. LASSO regression can automatically screen out features that have an important impact on the response variable; Among them, see Figure 4 , Figure 4 This is a schematic diagram of the cross-validation curve of the LASSO regression analysis for screening effective features in the present invention, wherein the horizontal axis variable is λ, the vertical axis variable is the binomial deviation, and the top is the number of features corresponding to the corresponding λ. The binomial deviation is an indicator used to evaluate the goodness of fit of binomial distribution data. The smaller the value, the better the model fits the data, that is, the closer the model prediction value is to the actual observed value. In disease diagnosis and prediction, if the deviation between the model prediction of disease / non-disease and the actual situation is small, the binomial deviation value is small. In this embodiment, Figure 4 The vertical line on the left indicates that the model has the least binomial deviation when 48 features are included, but the complexity is relatively high. The vertical line on the right indicates the lambda value that includes the least number of features in the model within a standard error range, which corresponds to 34 features. In this case, the model can have a good fitting effect and achieve a certain degree of feature simplification.

[0034] In this embodiment, 34 characteristic variables are finally determined as follows: Demographic information: weight, age, length of hospital stay; Disease scores: Elixhauser comorbidity score, SAPS II score, SOFA score, APS III score; Laboratory parameters: hemoglobin, albumin, blood urea nitrogen (BUN), bilirubin, glucose, bicarbonate, base excess, lactate, chloride, potassium; Clinical parameters: pH, platelet count, calcium, phosphorus, troponin T, mean arterial blood pressure, heart rate, respiratory rate, systolic blood pressure, diastolic blood pressure, body temperature, blood oxygen saturation (SpO2), urine volume; Therapeutic interventions: Antibiotics, diuretics, vasopressors, mechanical ventilation. In Table 1 of this Example, the physiological indicators between the AKI-developing and non-AKI-developing groups are expressed as medians (quartiles) [M(Q1, Q3)]. M represents the median, Q1 represents the first quartile, and Q3 represents the third quartile. The t-test was used for comparison between the two groups. Count indicators for medical interventions were expressed as the sum of the number of cases, and the chi-square test was used for comparison between the two groups. In this Example, there were 3,884 patients in the AKI-developing group and 1,896 patients in the non-AKI group. There were significant differences in physiological indicators and medical interventions between the two groups, all less than 0.05 (p < 0.05), indicating that there were significant differences between the two groups in these indicators and that the two data sets had high intergroup discrimination. The selected features were closely associated with the occurrence of AKI, and the data had strong predictive power.

[0035] Table 1: Baseline characteristics of the above 34 characteristic variables in this example

[0036] In this embodiment, the deep neural network adaptive training unit is used to integrate a configurable convolutional neural network architecture and implement parameter adaptive training through dynamic learning rate adjustment and loss function optimization; See also Figure 5 , Figure 5 This is a structural diagram of the deep neural network of the present invention.

[0037] It is worth noting that the deep neural network described is a convolutional neural network (CNN) architecture, constructed as a three-layer convolutional module. Each convolutional module consists of a convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer, and a random dropout layer. All convolution kernels have a size of 5, a stride of 1, and a padding of 2. The first layer has 16 convolution kernels, the second layer has 32 convolution kernels, and the third layer has 64 convolution kernels. The maximum pooling layer after each convolution layer is the same, with a pooling kernel size of 2 and a stride of 2. The maximum pooling operation can reduce the length of the feature map and reduce computational complexity while preserving important feature information. A random dropout layer is used after each pooling layer. Random dropout is a regularization technique that randomly drops some neurons with a probability of 0.3 to prevent overfitting and improve model generalization. After the three convolution and pooling layers, an adaptive average pooling layer is used. This layer adaptively pools the input feature map into a feature vector with a fixed length of 1, so that regardless of the length of the input data, it can output fixed-length features, which is convenient for subsequent fully connected layer processing.

[0038] In this embodiment, the forward propagation process is as follows: the input one-dimensional data first passes through the first convolution layer for feature extraction, then passes through the maximum pooling layer for downsampling, and then passes through the random dropout layer for regularization. The processed feature map passes through the second and third convolution layers in sequence, and each layer of convolution is followed by maximum pooling and random dropout operations. After three layers of convolution and pooling, the feature map is pooled into a feature vector of fixed length using an adaptive average pooling layer. The feature vector is flattened and input into the first fully connected layer, and is nonlinearly transformed through the LeakyReLU activation function. After passing through the random dropout layer, it is input into the second fully connected layer, and finally a two-dimensional vector is output as the classification result.

[0039] In this example, stratified random partitioning was used to generate the training and validation groups, with partitioning ratios of 0.3 and 0.7. During model training, the training data batch size was 256, meaning that 256 samples were taken from the training dataset at a time to form a batch for model training. The learning rate was 0.0005, controlling the step size for parameter updates during model training. The number of training rounds was set to 200, meaning that the model training would traverse the entire training dataset 200 times.

[0040] In this embodiment, the validation data set is loaded after model training, and the batch size is also 256. 256 samples are taken out each time to form a validation batch. The validation results generate a confusion matrix to calculate the accuracy, recall, precision and specificity, and the model performance is evaluated by the area under the receiver operating characteristic curve (AUC).

[0041] It's worth noting that the confusion matrix is an important tool for evaluating the performance of classification models. It presents the model's predictions on a test dataset in matrix form, showing the correspondence between true and predicted categories through the intersection of rows and columns.

[0042] See also Figure 6 , Figure 6 This is a diagram of the confusion matrix output from the early warning system model used in this invention. In this confusion matrix, rows represent true labels and columns represent predicted labels. Each element in the matrix represents the proportion of samples that were predicted to be of the corresponding true class.

[0043] According to the confusion matrix, the accuracy, precision, recall and specificity can be calculated. It is worth noting that the accuracy measures the proportion of samples correctly predicted by the model in the total number of samples, reflecting the overall prediction accuracy of the model; the precision refers to the proportion of samples that are actually positive among all samples predicted to be positive, focusing on the accuracy of the model when predicting positive examples; the recall rate refers to the proportion of samples that are correctly predicted to be positive among all samples that are actually positive, reflecting the model's ability to recognize positive examples; the specificity represents the proportion of samples that are correctly predicted to be negative among all samples that are actually negative, reflecting the model's ability to recognize negative examples.

[0044] In addition, the F1 score is a metric that combines precision and recall. Because precision and recall often trade off against each other, a single metric cannot fully reflect model performance. The F1 score combines these two metrics by using a harmonic mean. The F1 score ranges from 0 to 1; the closer it is to 1, the better the model's performance, indicating a balance between precision and recall.

[0045] In this embodiment, the deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation has an accuracy of 0.884, a precision of 0.893, a recall of 0.872, a specificity of 0.896, and an F1 score of 0.882 in the performance of the validation data set, indicating that the early warning system proposed in the present invention has good predictive performance.

[0046] See also Figure 7 , Figure 7 The figure below shows a receiver operating characteristic (ROC) curve for the validation set of the early warning system model used in the present invention. The ROC curve, or receiver operating characteristic (ROC), depicts the relationship between the true positive rate and false positive rate of a model at different classification thresholds. Ideally, a model should have a high true positive rate and a low false positive rate. Specifically, the ROC curve should be as close as possible to the upper left corner (0, 1), indicating that the model can perfectly distinguish between positive and negative examples. The area under the ROC curve (AUC) is used as a metric to measure model performance. AUC values range from 0.5 to 1, with larger values indicating better model performance. Figure 7 It can be seen that the AUC value of the convolutional neural network model proposed in the present invention is 0.877, and the quality of model prediction is good.

[0047] In this embodiment, the multidimensional model evaluation and decision interpretation unit is used to output the 24-hour AKI risk probability value and simultaneously generate a three-dimensional evaluation matrix including model prediction accuracy evaluation, SHAP value-driven feature attribution analysis, and clinical decision recommendations.

[0048] In the above three-dimensional evaluation matrix, the first aspect, the model prediction accuracy evaluation has been described in the specific implementation of the deep neural network adaptive training unit, including accuracy, precision, recall, specificity, F1 score and AUC value.

[0049] The second aspect involves SHAP-value-driven feature attribution analysis. Specifically, the SHAP method is an effective tool in machine learning for explaining model predictions. It builds on the concept of Shapley values in game theory and decomposes model predictions into the contribution of each input feature to the prediction. Its core principle is to treat all features as "participants" in the model's predictions and, by calculating their marginal contributions, determine the degree of influence each feature has on the final prediction.

[0050] See also Figure 8 , Figure 8 This is a visualization diagram of the feature importance of the early warning system model based on the SHAP method in this invention, which is used to show the impact of different features on the model output. Specifically, Figure 8 The horizontal axis represents the SHAP value, a positive value indicates that the feature increases the model output, and a negative value indicates that the model output decreases; the vertical axis lists the names of the features; the color bar represents the high and low feature values, yellow indicates a high feature value, and purple indicates a low feature value.

[0051] Figure 8The feature names are annotated as follows: Sofa_Score: Sequential Organ Failure Assessment Score, Length_Stay: Length of Hospital Stay, Spo2: Blood Oxygen Saturation, Arterial Blood Pressure Systolic: Systolic Arterial Pressure, Weight: Weight, Ventilator: Ventilator Usage, Platelet_Count: Platelet Count, Respiratory_Rate: Respiratory Rate, Phosphorus: Blood Phosphorus Level, Arterial Blood Pressure Mean: Mean Arterial Pressure, Antibiotic: Antibiotic Usage, Urine_Output: Urine Output, Hemoglobin: Hemoglobin, Albumin: Albumin, PH: pH, Diuretics: Diuretic Usage, Bilirubin: Bilirubin, Apsiii_Score: Disease Severity Score, Lactic_Acid: Lactic Acid, Arterial Blood Pressure Diastolic: diastolic blood pressure, Troponin_T: Troponin T, Bun: Blood urea nitrogen, Heartrate: Heart rate, Chloride: Blood chloride level, Potassium: Blood potassium level, Sapsii_Score: Simplified Acute Physiology Score II, Age: Age, Bicarbonate: Bicarbonate, Base_Excess: Base excess, Vasopressor: Vasopressor use, Temperature: Body temperature, Glucose: Blood glucose, Elixhauser_Score: Elixhauser score (assessing disease severity and prognosis). Figure 9 and Figure 10 The feature names in have the same meaning.

[0052] Further explanation: Figure 8 The higher the feature is in the middle, the more dispersed its point distribution is, the larger the range of SHAP values it spans, and the greater its impact on the model output is. For example, Sofa_Score has a wide distribution of points, which means that its impact on the model output varies greatly in different samples and is an important feature. Regarding the relationship between eigenvalues and the impact on model output, if the yellow points (high eigenvalues) are mostly distributed in the positive area of the horizontal axis, it means that when the eigenvalue is high, the model output tends to increase; if the purple points (low eigenvalues) are mostly in the negative area, it means that when the eigenvalue is low, the model output decreases. For example, in this embodiment Figure 8 For the features in the middle, high eigenvalues (yellow dots) correspond to positive SHAP values, which have a positive effect on the model output.

[0053] See also Figure 9 , Figure 9The diagram of the single-patient AKI warning based on the SHAP method of the early warning system model in the present invention is used to show the direction and degree of influence of different features on the output results of a specific patient sample model. The red arrow area indicates that the feature makes the model output higher than the baseline value. The higher the feature value, the greater the positive impact on the model output; the blue arrow area indicates that the feature makes the model output lower than the baseline value. The higher the feature value, the greater the negative impact on the model output. In this embodiment Figure 9 In the figure, taking "Sofa_Score=7" as an example, it is in the red area, indicating that this eigenvalue makes the model output higher than the baseline value, which has a positive effect on the model output; while "Bun=59" is in the blue area, which means that this eigenvalue makes the model output lower than the baseline value, which has a negative effect on the model output.

[0054] See also Figure 10 , Figure 10 This is a heat map of the feature impact of the early warning system model in the present invention based on the SHAP method. Specifically, the horizontal axis represents the sample instance, ranging from 0 to 400, indicating different sample points, used to observe the impact of the features in each sample; the vertical axis lists different feature names; the color bar represents the degree of influence of the SHAP value on the model output, with red representing a positive impact and the larger the value, the stronger the impact; blue representing a negative impact and the smaller the value, the weaker the impact; white indicates an impact close to 0; f(x) represents the output value of the early warning system model of the present invention, and the middle dotted line is the classification threshold. When the f(x) value is above the threshold, the output prediction is AKI, and when it is below the threshold, the output prediction is Normal.

[0055] Specifically, in Figure 10 Taking Sofa_Score as an example, it appears red in some samples, indicating that Sofa_Score has a positive impact on the model output in these samples, that is, the higher the Sofa_Score value, the more likely it is to push the model output towards AKI; while it appears blue in other samples, indicating a negative impact.

[0056] The third aspect involves clinical decision recommendations. The decision interpretation unit in this embodiment of the present invention outputs three analytical curves for clinical decision recommendations: a calibration curve for warning of AKI in AF patients, a clinical impact curve for warning of AKI in AF patients, and a decision curve analysis for warning of AKI in AF patients.

[0057] Specifically, see Figure 11 , Figure 11This is a schematic diagram of the calibration curve for warning AF patients of AKI in the present invention. The horizontal axis is the probability value predicted by the model, reflecting the possibility of the model warning AF patients of AKI; the vertical axis is the frequency of AKI in AF patients actually observed. The red solid line in the figure represents the calibration of this model, and the dotted line is the ideal calibration line. If the model prediction is accurate, the red solid line should coincide with the dotted line. As can be seen, Figure 11 The overall trend of the calibration curve closely matches the ideal calibration line. Within most predicted probability intervals, the curve closely follows the ideal calibration line, indicating that the model-predicted probabilities are highly consistent with the observed frequency of AKI in AF patients. This demonstrates that the early warning model has high accuracy and reliability in assessing the risk of AKI in AF patients, accurately reflecting actual conditions and providing a solid and effective reference for clinical decision-making.

[0058] See also Figure 12 , Figure 12 This is a schematic diagram of the clinical impact curve for early warning of AKI in AF patients, as described in the present invention. The horizontal axis represents the high-risk threshold and corresponds to different cost-effectiveness ratios, reflecting the effects under different risk thresholds and resource input-output conditions. The vertical axis represents the number of high-risk patients, expressed in 2000 patients.

[0059] The solid red line in the figure represents the number of high-risk patients. As can be seen, as the high-risk threshold increases (i.e., the criteria for determining high risk become stricter), the number of high-risk patients gradually decreases. This suggests that as we raise the threshold for determining the risk of AKI in AF patients, the number of patients identified as high-risk decreases accordingly.

[0060] The blue dashed line represents the number of patients at high risk of developing an event. Its trend is similar to the solid red line, decreasing as the high-risk threshold increases. Notably, the blue dashed line remains below the solid red line, indicating that, at the same high-risk threshold, the actual number of patients at high risk of developing an AKI event is less than the number of patients identified as high risk.

[0061] This figure intuitively shows the impact of different risk threshold settings on the number of high-risk patients and the number of patients who actually experience AKI events. It helps to balance the scope of risk identification and resource investment in clinical decision-making, and provides a visual reference for optimizing early warning strategies for AKI in AF patients.

[0062] See also Figure 13 , Figure 13This figure shows a decision curve analysis for early warning of AKI in AF patients according to the present invention. The horizontal axis represents the high-risk threshold, reflecting the standard scale for determining a patient as high-risk. The value ranges from 0.0 to 1.0, representing a range from loose to strict risk assessment criteria. The vertical axis represents the net benefit, reflecting the net benefit obtained from implementing the corresponding early warning strategy at different risk threshold settings.

[0063] The gray line in the figure represents the net benefit assuming no patients develop AKI (i.e., no intervention). This line serves as a baseline. The red curve illustrates the net benefit of this early warning model at different high-risk thresholds. As can be seen, the red curve lies above the gray line across most high-risk thresholds, indicating that using this early warning model can result in a higher net benefit compared to not taking any intervention.

[0064] As the high-risk threshold changes from low to high, the red curve shows a gradually downward trend, which means that when the criteria for determining high risk become more and more stringent, the net benefit brought by the model is gradually decreasing. In practical applications, decision curve analysis can assist in selecting risk thresholds to balance the coverage of early warnings and actual benefits. Overall, Figure 13 The decision curve analysis diagram intuitively demonstrates the advantages and limitations of this early warning model at different risk thresholds, providing an important visual basis for how to set risk thresholds in clinical decision-making.

[0065] Based on the same inventive concept as the aforementioned system embodiment, an embodiment of the present invention further provides a deep learning-based early warning method for acute kidney injury in patients with atrial fibrillation, comprising the following steps: The clinical data of hospitalized patients after atrial fibrillation was collected through the clinical data integration unit, and the clinical data were extracted and processed according to the inclusion and exclusion criteria; A multi-dimensional feature space adaptive interpolation model is constructed based on the random forest algorithm using the missing data reconstruction unit to process the missing data; With the help of the feature matrix construction unit, feature variables are screened based on LASSO regression to perform relevant processing of the structured high-dimensional clinical matrix; Use the deep neural network adaptive training unit to integrate the convolutional neural network architecture, and perform parameter adaptive training through dynamic learning rate adjustment and loss function optimization; The multidimensional model evaluation and decision interpretation unit outputs the 24-hour AKI risk probability value, generating a three-dimensional evaluation matrix that includes confidence assessment, SHAP value-driven feature attribution analysis, and clinical decision recommendations.

[0066] In an embodiment of the method of the present invention, in the clinical data integration step, the specific operations of clinical data purification, multi-source data integration and data verification are consistent with the corresponding operations of the clinical data integration unit in the system embodiment; in the missing data reconstruction step, the specific operations of constructing a multidimensional feature space adaptive interpolation model, adopting a dynamic priority interpolation mechanism and a recursive filling strategy are consistent with the corresponding operations of the missing data reconstruction unit in the system embodiment; in the feature matrix construction step, the specific operations of screening feature variables based on LASSO regression, performing feature standardization and tensor storage are consistent with the corresponding operations of the feature matrix construction unit in the system embodiment; in the deep neural network adaptive training step, the composition of the convolutional neural network architecture and the parameter settings of each layer are consistent with the corresponding contents of the deep neural network adaptive training unit in the system embodiment; in the multidimensional model evaluation and decision interpretation step, the output evaluation indicators, analysis methods and clinical decision recommendation analysis curves are consistent with the corresponding contents of the multidimensional model evaluation and decision interpretation unit in the system embodiment.

[0067] In summary, the present invention provides a deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation, comprising: a clinical data integration unit; a missing data reconstruction unit; a feature matrix construction unit; a deep neural network adaptive training unit; and a multidimensional model evaluation and decision interpretation unit. This system can quickly predict acute kidney injury in patients with atrial fibrillation 24 hours in advance, enabling early detection, diagnosis, and treatment of patients with acute kidney injury, improving survival rates after atrial fibrillation and the prognosis of patients with atrial fibrillation, and reducing the incidence of acute kidney injury.

[0068] It should be noted that the above-mentioned order of the embodiments of the present invention is only for description and does not represent the advantages or disadvantages of the embodiments.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation, characterized by: include: The clinical data integration unit is used to extract clinical data based on the clinical data of hospitalized patients after atrial fibrillation, according to the inclusion and exclusion criteria, and perform clinical data purification, multi-source data integration and data verification; Missing data reconstruction unit, used to build a multi-dimensional feature space adaptive interpolation model based on the random forest algorithm to achieve nonlinear enhancement of clinical data, using a dynamic priority interpolation mechanism and recursive filling strategy to preserve the original data index structure; A feature matrix construction unit is used to screen out characteristic variables closely related to atrial fibrillation complicated by AKI based on LASSO regression, implement parallel loading and tensor storage of structured high-dimensional clinical matrices, and perform standardized conversion and structured adaptation of medical data through a multi-stage processing pipeline; Deep neural network adaptive training unit, which is used to integrate configurable convolutional neural network architecture and achieve parameter adaptive training through dynamic learning rate adjustment and loss function optimization; The multidimensional model evaluation and decision interpretation unit is used to output the 24-hour AKI risk probability value and simultaneously generate a three-dimensional evaluation matrix that includes confidence assessment, SHAP value-driven feature attribution analysis, and clinical decision recommendations.

2. A deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation according to claim 1, characterized in that: The clinical data integration unit is further used to: When performing clinical data purification, cases with comorbidities were excluded based on ICD codes, and data range verification was performed; When integrating multi-source data, structured clinical data is loaded from the MIMIC-IV medical database to integrate multidimensional patient data; When performing data validation, abnormal records are automatically filtered out through range verification rules.

3. The deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation according to claim 1, characterized in that: When the missing data reconstruction unit constructs the multi-dimensional feature space adaptive interpolation model, it is also used to: The subspace sampling technology under the Bagging framework is used to dynamically generate feature interpolation points that conform to the original data distribution. The dynamic priority interpolation mechanism constructs a feature missing rate sorting queue, giving priority to processing columns with lower missing rates. The recursive filling strategy enables the filled columns to automatically participate in the subsequent feature reconstruction.

4. The deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation according to claim 1, characterized in that: When the feature matrix construction unit screens feature variables based on LASSO regression, it is also used to: The final feature variables are determined through the regression path and cross-validation curve of LASSO regression; when performing feature standardization, the difference in feature dimensions is eliminated; when performing tensor storage, the input clinical data is converted into a deep learning-friendly tensor structure through dimension expansion and channel permutation operations, and a configurable batch data loading pipeline is constructed.

5. The deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation according to claim 1, characterized in that: The configurable convolutional neural network architecture integrated in the deep neural network adaptive training unit includes a three-layer convolution module, each of which contains a convolution layer, a batch normalization layer, an activation function layer, a maximum pooling layer, and a random dropout layer; after the three-layer convolution and pooling operations, an adaptive average pooling layer is used to pool the feature map into a fixed-length feature vector.

6. The deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation according to claim 1, characterized in that: The model prediction accuracy evaluation indicators output by the multidimensional model evaluation and decision interpretation unit include accuracy, precision, recall, specificity, F1 score, and area under the receiver operating characteristic curve; feature attribution analysis is performed based on the SHAP method, and the impact of different features on model output is demonstrated through feature importance visualization, single-patient AKI warning diagram, and feature impact heat map; The output clinical decision recommendation analysis curves include a calibration curve for warning AF patients of AKI, a clinical impact curve for warning AF patients of AKI, and a decision curve analysis for warning AF patients of AKI.

7. The deep learning-based early warning system for acute kidney injury in patients with atrial fibrillation according to claim 6, characterized in that: The calibration curve has the model-predicted probability value on the abscissa and the observed frequency of AKI in AF patients on the ordinate. A close trend between the model calibration curve and the ideal calibration line indicates high model prediction accuracy. The clinical impact curve has the high-risk threshold and the corresponding cost-effectiveness ratio on the abscissa, and the high-risk number on the ordinate, demonstrating the impact of different risk threshold settings on the number of high-risk patients and the number of patients who actually experience AKI. The decision curve analysis shows that the horizontal axis represents the high-risk threshold, and the vertical axis represents the net benefit. By comparing the net benefit with the assumption that all patients do not develop AKI, the risk threshold is selected to balance the warning coverage and actual benefits.

8. A deep learning-based early warning method for acute kidney injury in patients with atrial fibrillation, characterized in that: include: The clinical data of hospitalized patients after atrial fibrillation is obtained through the clinical data integration unit, and the clinical data are extracted and processed according to the inclusion and exclusion criteria; A multi-dimensional feature space adaptive interpolation model is constructed based on the random forest algorithm using the missing data reconstruction unit to process the missing data; With the help of the feature matrix construction unit, feature variables are screened based on LASSO regression to perform relevant processing of the structured high-dimensional clinical matrix; Use the deep neural network adaptive training unit to integrate the convolutional neural network architecture, and perform parameter adaptive training through dynamic learning rate adjustment and loss function optimization; The multidimensional model evaluation and decision interpretation unit outputs the 24-hour AKI risk probability value, generating a three-dimensional evaluation matrix that includes confidence assessment, SHAP value-driven feature attribution analysis, and clinical decision recommendations.

9. A deep learning-based early warning device for acute kidney injury in patients with atrial fibrillation, characterized by: It includes a memory and a processor, the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the deep learning-based acute kidney injury early warning method for patients with atrial fibrillation as described in claim 8.

10. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the deep learning-based early warning method for acute kidney injury in patients with atrial fibrillation as described in claim 8.

Images