Intelligent prediction system and hierarchical management method for sepsis-related acute kidney injury

Abstract

The invention discloses a sepsis-related acute kidney injury intelligent prediction system and a hierarchical management method, and the system comprises a multi-modal data collection module which synchronously collects an ultrasound contrast video, a urine biomarker and hemodynamic parameters in real time; the spatio-temporal feature extraction module is used for extracting multi-modal spatio-temporal features; modeling a dynamic interaction effect of heart and kidney nodes through a dynamic organ interaction graph network; the self-supervised pre-training module is used for improving the generalization of the model through comparative learning; the dynamic risk prediction module outputs a risk probability curve in the next 24 hours; the hierarchical decision management module is used for generating personalized strategies of liquid management, vasoactive drug adjustment and kidney replacement treatment opportunity; and the interpretability module is used for analyzing the key early warning characteristics based on the SHAP value. According to the method, the problems of deficiency of organ interaction modeling, poor generalization of small samples, stiffness of static decision and the like are solved, and the early warning and precise treatment capabilities of the sepsis-related kidney injury are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical artificial intelligence, and specifically relates to an intelligent prediction system and hierarchical management method for sepsis-associated acute kidney injury. Background Art

[0002] Sepsis-associated acute kidney injury (SA-AKI) is a common complication in patients in the intensive care unit (ICU), and its early diagnosis and intervention are crucial for improving the prognosis. At present, clinical practice mainly relies on single indicators such as serum creatinine (Scr) and urine output. However, the serum creatinine concentration only significantly increases 48-72 hours after renal parenchymal injury, resulting in an early missed detection rate of over 40%. In the prior art, the analysis methods based on single-modal data (such as ultrasound images or biochemical indicators) have obvious limitations:

[0003] 1. Insufficient data dimension: Although traditional contrast-enhanced ultrasound technology can evaluate renal blood perfusion, it cannot capture molecular-level injury markers (such as the early increase of urinary NGAL and KIM-1); while the models relying only on biochemical indicators (such as Scr) are limited by the detection lag and it is difficult to timely warn of kidney injury.

[0004] 2. Lack of organ interaction mechanism: The dynamic blood flow coupling effect between the heart and the kidney (such as the decrease in cardiac output leading to reduced renal perfusion) has not been effectively quantified and modeled. Existing methods mostly assume independent organ actions and ignore key pathological mechanisms such as cardiorenal syndrome.

[0005] 3. Poor generalization ability for small samples: The prediction models based on supervised learning rely heavily on labeled data. However, the data labeling cost of ICU patients is high and the consistency is low, resulting in insufficient generalization ability of the models in real scenarios.

[0006] 4. Static decision-making rigidity: The fixed thresholds recommended by clinical guidelines (such as MAP≥65 mmHg) cannot adapt to the real-time risk changes of patients, and the existing decision-making systems lack the ability of dynamic adjustment, which is likely to lead to under-treatment or over-intervention.

[0007] Although artificial intelligence technology has been gradually applied to the medical field in recent years, the above problems have not been systematically solved. How to integrate multi-modal data, model the dynamic interaction of organs and achieve personalized dynamic decision-making has become the key technical bottleneck for improving the diagnosis and treatment effect of SA-AKI. Summary of the Invention

[0008] Technical Objective: Aiming at the deficiencies of the prior art, the present invention discloses an intelligent prediction system and hierarchical management method for sepsis-associated acute kidney injury, which integrates multi-modal physiological data, dynamic organ interaction modeling and reinforcement learning decision-making, can early warn of acute kidney injury and improve the accuracy.

[0009] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0010] An intelligent prediction system for sepsis-related acute kidney injury, comprising:

[0011] A multimodal data acquisition module for real-time acquisition of contrast-enhanced ultrasound videos, urine biomarker concentrations, and hemodynamic parameters of a target object;

[0012] A spatio-temporal feature extraction module, comprising:

[0013] A contrast-enhanced ultrasound feature extraction unit that uses a 3D convolutional neural network to extract spatio-temporal features of kidney perfusion;

[0014] A urine biomarker encoding unit that generates concentration gradient feature vectors based on a time series Transformer model;

[0015] A hemodynamic parameter processing unit that models the dynamic time series relationship between cardiac output and systemic vascular resistance index through a bidirectional LSTM network;

[0016] A dynamic organ interaction graph network, comprising:

[0017] Hidden state representations of the heart node and the kidney node;

[0018] A dynamic edge weight calculation unit that generates a cardiac-kidney interaction weight matrix based on a gated recurrent unit;

[0019] A graph attention aggregation layer that fuses cross-organ features through a multi-head attention mechanism;

[0020] A self-supervised pre-training module that trains a feature extraction network based on a contrastive learning framework using unlabeled contrast-enhanced ultrasound video data;

[0021] A dynamic risk prediction module that outputs a risk probability curve of a patient over time using a deep neural network based on survival analysis;

[0022] A hierarchical decision-making management module that integrates a reinforcement learning algorithm and a constraint optimization model to generate personalized strategies for fluid management, vasoactive drug use, and the timing of renal replacement therapy;

[0023] An interpretability module that performs feature attribution analysis on the risk prediction results based on SHAP values and visualizes the evolution trend of dynamic edge weights.

[0024] Preferably, the dynamic variable weight calculation unit of the dynamic organ interaction network satisfies the following formula:

[0025]

[0026] Wherein, is the dynamic interaction weight between the heart and the kidney at time step t, σ represents the Sigmoid activation function, and W e is a learnable weight matrix with a dimension of 256×256. GRU is a gated recurrent unit. and are the historical hidden states of the heart and kidney nodes respectively. ΔCO t is the change rate of cardiac output, and ΔRBF t is the change rate of renal blood flow. b e is a bias term with a dimension of 256.

[0027] Preferably, the self-supervised pre-training module adopts the following method:

[0028] Perform spatio-temporal data augmentation on the contrast-enhanced ultrasound video, including random occlusion, temporal segment rearrangement, and contrast perturbation;

[0029] Construct positive sample pairs as contrast-enhanced ultrasound video segments of the same target object at different time periods, and negative sample pairs as contrast-enhanced ultrasound video segments of different target objects;

[0030] Use the normalized temperature cross-entropy loss to optimize the feature extractor.

[0031] Preferably, the survival analysis model of the dynamic risk prediction module satisfies:

[0032]

[0033] where is the risk probability of the target object having acute kidney injury at future time t, α m (x) is the weight of the m-th sub-distribution output by the multi-layer perceptron, β m (t) is the m-th B-spline basis function, γ is the organ interaction feature weight coefficient, and DOIN(x) is the feature vector output by the dynamic organ interaction graph network.

[0034] Preferably, the reinforcement learning strategy of the hierarchical decision management module satisfies:

[0035] The state space includes the current risk score, cumulative fluid balance, vasoactive drug dose, and blood potassium concentration;

[0036] The action space includes furosemide dose adjustment, norepinephrine infusion rate, and timing of initiating continuous renal replacement therapy;

[0037] The reward function is defined as:

[0038] R(s,a) = ω1·ΔeGFR - ω2·CRRT_indicator - ω3·Fluid_overload + ω4·Hemodynamic_stability

[0039] Among them, ΔeGFR is the change in estimated glomerular filtration rate, CRRT_indicator is the initiation marker of continuous renal replacement therapy, Fluid_overload is the amount of fluid overload, Hemodynamic_stability is the hemodynamic stability score, and ω1, ω2, ω3, and ω4 represent weight coefficients.

[0040] The constraint conditions include that the daily negative balance ≤ 1000 ml and the mean arterial pressure ≥ 65 mmHg.

[0041] Preferably, the input parameters of the multimodal data acquisition module include:

[0042] Contrast-enhanced ultrasound video, with a resolution ≥ 1280×720, a frame rate ≥ 25 fps, and the acquisition area covering the entire renal cortex and medulla;

[0043] Urine biomarkers, the detection threshold of neutrophil gelatinase-associated lipocalin concentration is 0.5 - 150 ng / mL, and the detection threshold of kidney injury molecule-1 concentration is 0.1 - 50 ng / mL;

[0044] Hemodynamic parameters, cardiac output, systemic vascular resistance index.

[0045] Preferably, the calculation of SHAP values of the interpretability module includes the following steps:

[0046] Sample the feature vectors output by the dynamic organ interaction graph network to generate a set of perturbed samples;

[0047] Obtain the change in risk probability of the perturbed samples through model prediction;

[0048] Calculate the feature contribution degree based on the Shapley value allocation rule:

[0049]

[0050] Among them, φ j is the SHAP value of feature j, F is the set of all features, S is the feature subset, and h(t|x S ) is the model prediction risk when using the features of subset S.

[0051] Preferably, the SHAP analysis results are presented in the following ways:

[0052] Time series feature contribution curve, showing the contribution weights of systemic vascular resistance index changing over time;

[0053] Interaction effect matrix diagram, explaining the synergistic or antagonistic effects between the heart node and the kidney node.

[0054] The present invention also provides a method for hierarchical management of sepsis-related acute kidney injury, which is applied to a sepsis-related acute kidney injury intelligent prediction system as described in any one of the above, and specifically includes the following steps:

[0055] S1. Real-time collect multi-modal physiological data of the target object through the multi-modal data acquisition module, and perform standardized preprocessing. The multi-modal physiological data includes contrast-enhanced ultrasound video, urine biomarkers, and hemodynamic parameters;

[0056] S2. Model the interaction effects between the heart node and the kidney node through the dynamic organ interaction graph network, and extract organ synergistic feature vectors;

[0057] S3. Predict the risk probability of acute kidney injury per hour within the next 24 hours based on the survival analysis model, and generate a dynamic risk curve;

[0058] S4. When the risk probability exceeds the preset threshold, trigger the reinforcement learning decision engine to generate a hierarchical intervention strategy, including fluid management, vasoactive drug adjustment, and suggestions for the timing of renal replacement therapy;

[0059] S5. Dynamically update the model parameters according to the clinical execution results and the physiological response of the target object, and optimize the decision-making strategy. Beneficial effects: The sepsis-related acute kidney injury intelligent prediction system and the hierarchical management method provided by the present invention have the following beneficial effects:

[0060] 1. By integrating the spatio-temporal features of contrast-enhanced ultrasound video, the temporal trends of urine biomarkers, and the dynamic parameters of hemodynamics, the present invention breaks through the limitations of traditional single-modal detection, and significantly improves the sensitivity and specificity of early kidney injury. The dynamic organ interaction graph network (DOIN) adopts a learnable edge weight mechanism to accurately quantify the lag effect of cardiac function changes on renal perfusion, and for the first time realizes the dynamic modeling of cross-organ pathological coupling effects, providing earlier and more reliable warning signals for clinical practice.

[0061] 2. Based on the self-supervised pre-training framework of contrastive learning, the present invention effectively utilizes unlabeled ultrasound video data through spatio-temporal data augmentation (occlusion, temporal rearrangement), and significantly reduces the dependence of the model on labeled data. Combining the dynamic decision-making mechanism of survival analysis and reinforcement learning, it generates personalized strategies (such as fluid management dose adjustment, suggestions for the timing of CRRT initiation) that match the patient's risk evolution in real time, overcomes the rigidity of traditional static guidelines, and achieves a balance between treatment accuracy and safety.

[0062] 3. Through the SHAP value analysis of key warning features (such as the mutation of systemic vascular resistance index and abnormal cardio-renal interaction weight) and the visualization of dynamic edge weights, the present invention intuitively reveals the decision-making basis of the model, assists doctors in understanding complex pathological mechanisms, and improves the clinical acceptance and trust of AI-assisted decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for the description of the embodiments or the prior art.

[0064] Figure 1 It is a block diagram of the intelligent prediction system of the present invention;

[0065] Figure 2 It is a schematic diagram showing the changes in the interaction between the cardio-renal nodes and the dynamic edge weights in the visualization of the dynamic organ interaction diagram network of the present invention;

[0066] Figure 3 It is the overall flowchart of the hierarchical management method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0067] The following more clearly and completely describes the present invention by way of a preferred embodiment in combination with the drawings, but the present invention is not limited to the scope of the described embodiments.

[0068] As Figure 1 shown, a sepsis-related acute kidney injury intelligent prediction system includes:

[0069] A multimodal data acquisition module for real-time acquisition of contrast-enhanced ultrasound videos, urine biomarker concentrations, and hemodynamic parameters of the target object;

[0070] In one embodiment, the contrast-enhanced ultrasound video uses a Philips EPIQ CVx ultrasound device, configured with a C5-1 probe, the mechanical index is set to 0.08 - 0.12 to avoid microbubble rupture, a long-axis section video of the kidney is acquired, the resolution is ≥1280×720, the frame rate is ≥25fps, the acquisition area covers the renal cortex and medulla, and it is acquired once every 30 minutes. ROI segmentation is performed on each frame of the image, the renal parenchymal area is extracted, and the contrast is enhanced by histogram equalization; the urine biomarkers include the concentration of neutrophil gelatinase-associated lipocalin (NGAL) and the concentration of kidney injury molecule-1 (KIM-1), and A bedside detector measures the concentrations of NGAL and KIM-1 in urine samples every hour, with the detection ranges being 0.5 - 150 ng / mL and 0.1 - 50 ng / mL respectively. Then, logarithmic transformation is performed on the concentration values to eliminate the dimensional difference. Hemodynamic parameters include cardiac output (CO, unit: L / min) and systemic vascular resistance index (SVRI, unit: dyn·s·cm -5 ·m 2 ). The PICCO2 monitor is used to collect CO and SVRI in real time, and the sampling interval is ≤ 1 minute. Taking the R wave of the electrocardiogram signal as the time reference, the contrast-enhanced ultrasound frames, urine concentrations, and hemodynamic parameters are synchronized.

[0071] The spatio-temporal feature extraction module includes:

[0072] The contrast-enhanced ultrasound feature extraction unit uses a 3D convolutional neural network to extract the spatio-temporal features of kidney perfusion. The 3D convolutional neural network is a 3D ResNet-50 network. The input is 30 consecutive frames of ultrasound videos with a time span of 1 second, and the output is a spatio-temporal distribution matrix of vascular density of 256×256×32. It is initialized with pre-trained weights, optimized by cross-entropy loss, with a learning rate of 0.001 and a batch size of 16.

[0073] The urine biomarker encoding unit generates concentration gradient feature vectors based on the time series Transformer model. The specific operation is that the input is the concentration sequences of NGAL and KIM-1 every hour within 24 hours, and a 512-dimensional trend vector is generated through the Transformer encoder (number of heads = 8, hidden layer dimension = 512). Then, sine position encoding is added to retain the time series information.

[0074] The hemodynamic parameter processing unit models the dynamic time series relationship between cardiac output and systemic vascular resistance index through a bidirectional LSTM network. The specific operation is that the input is the 5-minute sliding window data of CO and SVRI, and a 128-dimensional time series feature is output through the bidirectional LSTM (number of layers = 2, number of hidden units = 128). After concatenating the CO and SVRI features, the dimension is reduced to 64 through a fully connected layer.

[0075] As Figure 2 shown, the dynamic organ interaction graph network includes:

[0076] The hidden state representations of the heart node and the kidney node. The features of the heart node are the change rate of cardiac output (ΔCO) and the variability of stroke volume (SVV), and the features of the kidney node are the change rate of renal blood flow (ΔRBF) and the cortical perfusion rate.

[0077] The dynamic edge weight calculation unit generates a heart-kidney interaction weight matrix based on the gated recurrent unit. The dynamic variable weight calculation unit of the dynamic organ interaction network satisfies the following formula:

[0078]

[0079] Among them, is the dynamic interaction weight between the heart and the kidney at time step t, which quantifies the coupling strength of cardiac and renal blood flow. The value range is [0, 1]. The larger the weight, the more significant the interaction effect. σ represents the Sigmoid activation function, which constrains the edge weight to the range of [0, 1] to ensure biological rationality. W e is a learnable weight matrix with a dimension of 256×256, which maps the hidden state output by the GRU to the interaction weight space. GRU is a gated recurrent unit that models the temporal dependencies of the heart node and the kidney node and captures the lag effect (such as the delayed effect of CO change on RBF). and are the historical hidden states of the heart and kidney nodes respectively. ΔCO t is the change rate of cardiac output, which reflects the real-time change of cardiac function. A negative value indicates deterioration of cardiac function. ΔRBF t is the change rate of renal blood flow, which characterizes the dynamic change of renal perfusion. A negative value indicates an increased risk of renal ischemia. b e is a bias term with a dimension of 256.

[0080] In the calculation of the dynamic edge weight, the change rate of cardiac output ΔCO t is obtained in real time through a bedside monitoring device (such as the PICCO2 system). Its calculation needs to meet the sampling interval ≤ 1 minute to ensure temporal accuracy. The historical hidden states and are updated through the GRU unit, and the memory window is set to 6 hours to balance short-term fluctuations and long-term trends.

[0081] The graph attention aggregation layer fuses cross-organ features through the multi-head attention mechanism;

[0082] The self-supervised pre-training module is based on the contrastive learning framework and uses unlabeled contrast-enhanced ultrasound video data to train the feature extraction network. The self-supervised pre-training module adopts the following method:

[0083] Perform spatio-temporal data augmentation on the contrast-enhanced ultrasound video, including random occlusion, temporal segment rearrangement, and contrast perturbation;

[0084] Construct positive sample pairs as contrast-enhanced ultrasound video segments of the same target object at different time periods, and negative sample pairs as contrast-enhanced ultrasound video segments of different target objects;

[0085] Use the normalized temperature cross-entropy loss to optimize the feature extractor.

[0086] In one embodiment, the specific execution method of the self-supervised pre-training module is:

[0087] Data augmentation is performed, including spatial data augmentation and temporal data augmentation. Among them, spatial data augmentation includes random occlusion (maximum area 15%), contrast perturbation (±20%), and Gaussian noise (σ = 0.05), and temporal data augmentation includes video segment rearrangement (maximum offset ±5 frames) and random frame dropping (probability 10%);

[0088] A contrast learning task is carried out: the positive sample pair is the contrast-enhanced ultrasound video segment of the same target object ≤ 1 hour (time correlation constraint), and the negative sample pair is the contrast-enhanced ultrasound video segments of different target objects. The Adam optimizer is used, the learning rate is set to 0.0001, the batch size is 32, and the training period is 100 for training.

[0089] The dynamic risk prediction module uses a deep neural network based on survival analysis to output the risk probability curve of the patient over time. In one embodiment, the dynamic risk prediction module includes a survival analysis module. The architecture of the survival analysis module is DeepHit++, which includes 3 layers of MLP (512→256→10) and B-spline basis functions (number of nodes = 10, order = 3). The survival analysis model of the dynamic risk prediction module satisfies:

[0090]

[0091] Among them, is the risk probability that the target object will develop acute kidney injury at future time t, and the value range is [0,1], α m (x) is the weight of the m-th sub-distribution output by the multi-layer perceptron (MLP) (64→10), learning the contribution degree of different risk patterns, reflecting the individual differences of patients, β m (t) is the m-th B-spline basis function, modeling the non-linear trend of risk changing over time, such as a sharp rise in the early stage, a plateau period, etc. γ is the organ interaction feature weight coefficient, adjusting the contribution intensity of the dynamic organ interaction feature to the total risk. DOIN(x) is the feature vector output by the dynamic organ interaction graph network, with a dimension of 512, encoding the cardio-renal synergy effect and capturing the cross-organ pathological mechanism, such as cardio-renal syndrome.

[0092] The B-spline basis function β m (t) has nodes evenly distributed within the prediction time window (for example, 24 hours is divided into 10 intervals), and the historical data distribution is fitted by the least squares method. The sub-distribution weight α m (x) is calculated by the MLP network, and the ReLU activation function is used in the input layer to prevent gradient explosion.

[0093] The hierarchical decision-making management module integrates the reinforcement learning algorithm and the constraint optimization model to generate personalized strategies for fluid management, the use of vasoactive drugs, and the timing of renal replacement therapy. The reinforcement learning strategy of the hierarchical decision-making management module satisfies:

[0094] The state space includes the current risk score, cumulative fluid balance (±3000 mL), vasoactive drug dose, and serum potassium concentration;

[0095] Action space includes furosemide dose adjustment (0-40 mg / h), norepinephrine infusion rate (0-1 μg / kg / min), and timing of continuous renal replacement therapy (CRRT) initiation (0 / 1);

[0096] The reward function is defined as:

[0097] R(s,a)=ω1·ΔeGFR-ω2·CRRT_indicator-ω3·Fluid_overload+ω4·Hemodynamic_stability

[0098] Among them, ΔeGFR is the estimated change in glomerular filtration rate, which measures the dynamic changes in renal function. A positive value indicates improvement in renal function, while a negative value indicates deterioration. The calculation method is: CRRT_indicator is the starting mark of continuous renal replacement therapy, which punishes unnecessary continuous renal replacement therapy intervention and reduces the risk of excessive medical treatment. Fluid_overload is the fluid overload, which controls fluid balance and prevents pulmonary edema or volume insufficiency. Hemodynamic_stability is the hemodynamic stability score, which ensures circulatory stability and avoids hypotension or organ hypoperfusion. It is a comprehensive score based on mean arterial pressure (MAP) and cardiac output variability (SVV), with a value of 0-1, 1 indicating complete stability. ω1, ω2, ω3 and ω4 represent weight coefficients, which adjust the priority of various clinical goals. They are optimized through reinforcement learning training, with initial values ​​of ω1=0.5, ω2=0.3, ω3=0.2 and ω4=0.4.

[0099] Constraints included daily negative balance ≤1000ml and mean arterial pressure ≥65mmHg.

[0100] The explainability module performs feature attribution analysis on the risk prediction results based on the SHAP value and visualizes the dynamic edge weight evolution trend. The SHAP value calculation of the explainability module includes the following steps:

[0101] Sampling the feature vector output by the dynamic organ interaction graph network to generate a perturbation sample set, for example, randomly extracting 1,000 examples from the training data;

[0102] The risk probability change of the disturbance sample is obtained through model prediction, using Monte Carlo sampling with 10,000 iterations;

[0103] Calculate the feature contribution based on the Shapley value allocation rule:

[0104]

[0105] Among them, φ j is the SHAP value of feature j, F is the set of all features, S is the feature subset, and h(t|x S ) is the model prediction risk when using the features of subset S.

[0106] In one embodiment, the SHAP analysis result is presented in the following manner:

[0107] Time series feature contribution curve, showing the contribution weight of the systemic vascular resistance index changing with time;

[0108] Interaction effect matrix diagram, explaining the synergistic or antagonistic effect between the heart node and the kidney node.

[0109] As Figure 3 shown, the present invention also provides a hierarchical management method for sepsis-related acute kidney injury, which is applied to a sepsis-related acute kidney injury intelligent prediction system as described in any one of the above, and specifically includes the following steps:

[0110] S1. Real-time collect the multimodal physiological data of the target object through the multimodal data collection module, and perform standardized preprocessing. The multimodal physiological data includes contrast-enhanced ultrasound video, urine biomarkers, and hemodynamic parameters. The specific implementation process is as follows:

[0111] S11. Use a Philips EPIQ CVx ultrasound device, configure a C5-1 probe, set the mechanical index (MI) to 0.08 - 0.12, the frame rate to 30 fps, and the resolution to 1280×720.

[0112] S12. Collect the kidney long-axis section video once every 30 minutes, with a duration of ≥5 seconds, covering the cortical and medullary regions.

[0113] S13. Automatically segment the kidney parenchymal region (ROI) through a U-Net network to remove the interference of surrounding tissues.

[0114] S14. Perform contrast enhancement of limited contrast adaptive histogram equalization (CLAHE) on each frame of the image to enhance the microbubble perfusion signal.

[0115] S15. Use a bedside detector to detect the concentrations of NGAL (0.5 - 150 ng / mL) and KIM-1 (0.1 - 50 ng / mL) in the urine sample every hour.

[0116] S16. Perform Z-score standardization on the concentrations of NGAL and KIM-1;

[0117] S17. Real-time collect the cardiac output (CO) and systemic vascular resistance index (SVRI) through a PICCO2 monitor, with a sampling interval ≤ 30 seconds.

[0118] S18. Using the R wave of the electrocardiogram signal as a reference, synchronize the ultrasound video frames, urine concentration, and hemodynamic parameters, with a time error ≤ 0.5 seconds.

[0119] S19. Use a Hampel filter to remove outliers (window size = 5 minutes, threshold = 3σ).

[0120] S2. Model the interaction effect between the cardiac node and the renal node through a dynamic organ interaction graph network, and extract the organ cooperation feature vector. The specific implementation process is as follows:

[0121] S21. Define the cardiac node features including the change rate of cardiac output (ΔCO) and the variability of stroke volume (SVV), and define the renal node features including the change rate of renal blood flow (ΔRBF) and the cortical perfusion rate.

[0122] S22. Input the historical hidden state and real-time ΔCO, ΔRBF into the GRU unit (hidden unit = 128).

[0123] S23. Calculate the cardiac-renal interaction weight.

[0124] S24. Aggregate node features through a 4-head GATv2 network.

[0125] S25. Output a 512-dimensional organ cooperation feature vector, including the interaction intensity and perfusion imbalance index.

[0126] S3. Predict the acute kidney injury risk probability per hour within the next 24 hours based on a survival analysis model, and generate a dynamic risk curve. The specific implementation process is as follows:

[0127] S31. Input the organ cooperation features (512-dimensional), urine concentration trend vector (512-dimensional), and hemodynamic time series features (64-dimensional).

[0128] S32. Generate sub-distribution weights through a 3-layer MLP (512→256→10).

[0129] S33. Construct a risk function based on the B-spline basis function (number of nodes = 10, order = 3).

[0130] S34. Output the risk probability per hour within the next 24 hours (value range [0,1]).

[0131] S35. When the risk probability > 0.65 for 3 consecutive hours, an alarm is triggered and pushed at different levels. Among them, a high risk (> 0.8) is a red alarm and is pushed to the attending physician, and a medium risk (0.65 - 0.8) is a yellow alarm and is recorded in the nursing system.

[0132] S4. When the risk probability exceeds the preset threshold, the reinforcement learning decision engine is triggered to generate a hierarchical intervention strategy, including suggestions for fluid management, adjustment of vasoactive drugs, and timing of renal replacement therapy. The specific implementation process is as follows:

[0133] S41. Construct the state space, including real-time risk score (0 - 1), cumulative fluid balance (±3000 mL), norepinephrine dose (0 - 1 μg / kg / min), and blood potassium concentration (3.0 - 5.5 mmol / L);

[0134] S42. Define the action space, including adjustment of furosemide dose (0 - 40 mg / h, step size ≤ 10 mg / h), adjustment of norepinephrine infusion rate (±0.1 μg / kg / min), and CRRT start flag (0 / 1);

[0135] S43. Calculate the reward function;

[0136] S44. Adopt the PPO algorithm with Monte Carlo tree search (MCTS) to optimize the strategy, define the discount factor γ = 0.99, the constraint violation penalty coefficient λ = 10, with a daily negative balance ≤ 1000 mL, MAP ≥ 65 mmHg, and blood potassium ≥ 3.5 mmol / L.

[0137] S5. According to the clinical implementation results and the physiological responses of the target objects, dynamically update the model parameters and optimize the decision-making strategy. The specific implementation process is as follows:

[0138] S51. Record the actual intervention measures (such as drug dosage, CRRT time) and the physiological responses of the patients (urine output, changes in Scr);

[0139] S52. New data is added to the training set with a sliding window (window size = 7 days, step size = 1 day);

[0140] S53. Update the reinforcement learning policy network;

[0141] S54. Dynamically adjust the self-supervised pre-training parameters;

[0142] S55. Adjust the safety threshold according to the clinical feedback (such as the blood potassium threshold changing from 3.5 → 3.3 mmol / L).

[0143] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An intelligent prediction system for sepsis-related acute kidney injury, characterized in that, Including: A multimodal data acquisition module for real-time acquisition of contrast-enhanced ultrasound videos, urine biomarker concentrations, and hemodynamic parameters of a target object; A spatio-temporal feature extraction module, including: A contrast-enhanced ultrasound feature extraction unit that uses a 3D convolutional neural network to extract spatio-temporal features of kidney perfusion; A urine biomarker encoding unit that generates concentration gradient feature vectors based on a time series Transformer model; A hemodynamic parameter processing unit that models the dynamic time series relationship of cardiac output and systemic vascular resistance index through a bidirectional LSTM network; A dynamic organ interaction graph network, including: Hidden state representations of the heart node and the kidney node; A dynamic edge weight calculation unit that generates a heart-kidney interaction weight matrix based on a gated recurrent unit; A graph attention aggregation layer that fuses cross-organ features through a multi-head attention mechanism; A self-supervised pre-training module that trains a feature extraction network based on a contrastive learning framework using unlabeled contrast-enhanced ultrasound video data; A dynamic risk prediction module that uses a deep neural network based on survival analysis to output a risk probability curve of a patient over time; A hierarchical decision-making management module that integrates a reinforcement learning algorithm and a constrained optimization model to generate personalized strategies for fluid management, vasoactive drug use, and the timing of renal replacement therapy; An interpretability module that performs feature attribution analysis on the risk prediction results based on SHAP values and visualizes the evolution trend of dynamic edge weights.

2. The intelligent prediction system for sepsis-related acute kidney injury according to claim 1, wherein The dynamic variable weight calculation unit of the dynamic organ interaction network satisfies the following formula: Among them, is the dynamic interaction weight between the heart and the kidney at time step t, σ represents the Sigmoid activation function, and W e is a learnable weight matrix with a dimension of 256×256. GRU is a gated recurrent unit. and are the historical hidden states of the heart and kidney nodes respectively. ΔCO t is the change rate of cardiac output, and ΔRBF t is the change rate of renal blood flow. b e is a bias term with a dimension of 256.

3. The intelligent prediction system for sepsis-related acute kidney injury according to claim 1, characterized in that, The self-supervised pre-training module adopts the following method: Perform spatio-temporal data augmentation on the contrast-enhanced ultrasound video, including random occlusion, temporal segment rearrangement, and contrast perturbation; Construct positive sample pairs as contrast-enhanced ultrasound video segments of the same target object at different time periods, and negative sample pairs as contrast-enhanced ultrasound video segments of different target objects; Use normalized temperature cross-entropy loss to optimize the feature extractor.

4. The intelligent prediction system for sepsis-related acute kidney injury according to claim 1, wherein, The survival analysis model of the dynamic risk prediction module satisfies: wherein, is the risk probability of the target object having acute kidney injury at future time t, and α m (x) is the weight of the m-th sub-distribution output by the multi-layer perceptron, and β m (t) is the m-th B-spline basis function, γ is the weight coefficient of organ interaction features, and DOIN(x) is the feature vector output by the dynamic organ interaction graph network.

5. The intelligent prediction system for sepsis-related acute kidney injury according to claim 1, wherein The reinforcement learning strategy of the hierarchical decision-making management module satisfies: The state space includes the current risk score, cumulative fluid balance, vasoactive drug dose, and potassium concentration; The action space includes furosemide dose adjustment, norepinephrine infusion rate, and the timing of initiating continuous renal replacement therapy; The reward function is defined as: R(s,a)=ω1·ΔeGFR - ω2·CRRT_indicator - ω3·Fluid_overload + ω4·Hemodynamic_stability where ΔeGFR is the change in estimated glomerular filtration rate, CRRT_indicator is the continuous renal replacement therapy initiation flag, Fluid_overload is the fluid overload volume, Hemodynamic_stability is the hemodynamic stability score, and ω1, ω2, ω3, and ω4 represent weight coefficients. The constraint conditions include a daily negative balance ≤ 1000 ml and a mean arterial pressure ≥ 65 mmHg.

6. The intelligent prediction system for sepsis-related acute kidney injury according to claim 1, wherein The input parameters of the multimodal data acquisition module include: Contrast-enhanced ultrasound video, resolution ≥1280×720, frame rate ≥25fps, and the acquisition area covers the entire renal cortex and medulla; Urinary biomarkers, the detection threshold for the concentration of neutrophil gelatinase-associated lipocalin is 0.5 - 150 ng / mL, and the detection threshold for the concentration of kidney injury molecule-1 is 0.1 - 50 ng / mL; Hemodynamic parameters, cardiac output, systemic vascular resistance index.

7. An intelligent prediction system for sepsis-related acute kidney injury according to claim 1, characterized in that The SHAP value calculation of the interpretability module includes the following steps: Sample the feature vectors output by the dynamic organ interaction graph network to generate a perturbation sample set; Obtain the change in risk probability of the perturbation sample through model prediction; Calculate the feature contribution degree based on the Shapley value allocation rule: where φ j is the SHAP value of feature j, F is the set of all features, S is a subset of features, and h(t|x S ) is the model prediction risk when using the features in subset S.

8. An intelligent prediction system for sepsis-related acute kidney injury according to claim 7, characterized in that, The SHAP analysis results are presented in the following ways: Time series feature contribution degree curve, showing the contribution weight of the systemic vascular resistance index changing over time; Interaction effect matrix diagram, explaining the synergistic or antagonistic effect between the heart node and the kidney node.

9. A method for stratified management of sepsis-related acute kidney injury, characterized in that, Applied to a sepsis-related acute kidney injury intelligent prediction system according to any one of claims 1 - 8, specifically including the following steps: S1. Real-time collect the multimodal physiological data of the target object through the multimodal data acquisition module and perform standardized preprocessing. The multimodal physiological data includes contrast-enhanced ultrasound video, urinary biomarkers, and hemodynamic parameters; S2. Model the interaction effect between the heart node and the kidney node through the dynamic organ interaction graph network and extract the organ collaborative feature vectors; S3. Predict the acute kidney injury risk probability per hour within the next 24 hours based on the survival analysis model to generate a dynamic risk curve; S4. When the risk probability exceeds the preset threshold, trigger the reinforcement learning decision engine to generate a hierarchical intervention strategy, including fluid management, vasoactive drug adjustment, and suggestions for the timing of renal replacement therapy; S5. Dynamically update the model parameters according to the clinical execution results and the physiological response of the target object to optimize the decision-making strategy.

Images