Artificial intelligence-based renal disease dialysis data processing method
By employing multimodal dynamic coupling graphs and adaptive causal reasoning techniques, the problems of single-feature prediction and black-box intervention decisions in dialysis have been solved, enabling accurate prediction and personalized intervention of hypotension during dialysis, thus improving the safety and efficiency of the dialysis process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 903 HOSPITAL OF THE JOINT LOGISTICS SUPPORT FORCE OF THE PEOPLES LIBERATION ARMY OF CHINA
- Filing Date
- 2026-03-02
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies for predicting hypotension during dialysis suffer from problems such as reliance on single features, static prediction mechanisms, and a black box approach to intervention decisions, leading to high false alarm rates, delayed intervention measures, and low physician trust.
The method employs multimodal dynamic coupled graph feature extraction, stage-adaptive temporal risk prediction, and counterfactual causal intervention scheme generation. By collecting high-frequency numerical signals from dialysis machines and low-frequency text events entered by nurses in real time, a dynamic coupled graph is constructed. Adaptive risk prediction and intervention decisions are then made using dilated causal convolutional networks and generative adversarial networks.
It enables accurate prediction and proactive prevention of hypotension during dialysis, improves prediction accuracy and the feasibility of intervention measures, shortens doctors' response time, and provides personalized treatment plans.
Smart Images
Figure CN122157970A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical data processing technology, and more specifically, to a method for processing kidney dialysis data based on artificial intelligence. Background Technology
[0002] Hemodialysis is one of the main treatments for end-stage renal disease patients. During dialysis, due to factors such as excessively rapid ultrafiltration rates and insufficient cardiovascular compensation, patients are highly susceptible to hypotension, leading to insufficient organ perfusion, arteriovenous fistula thrombosis, and even death. Currently, clinical monitoring of hypotension during dialysis mainly relies on blood pressure threshold alarms, which are triggered when systolic blood pressure drops below 90 mmHg or decreases by more than 20 mmHg from baseline. However, this reactive alarm mode means that the patient is already in a hypotensive state when it occurs, resulting in delayed intervention and a low success rate of resuscitation.
[0003] In recent years, with the development of medical informatization, some studies have attempted to use artificial intelligence technology to predict hypotension during dialysis. However, existing technologies still have the following significant drawbacks: First, the use of features is too simplistic, neglecting cross-modal coupling relationships. Existing solutions typically only collect numerical signals from a few dimensions such as blood pressure and heart rate for time-series modeling, failing to effectively integrate equipment parameters reported by the dialysis machine, such as ultrafiltration volume, dialysate sodium concentration, and blood flow, and even failing to incorporate textual information of subjective symptoms recorded by nurses, such as patient complaints, muscle spasms, and nausea. There are complex nonlinear coupling relationships between machine parameters and human responses (such as high sodium dialysis leading to thirst and thus increased implicit water intake, and excessively rapid ultrafiltration causing a sudden increase in hematocrit). Traditional methods analyze these modalities separately, failing to construct a feature representation that reflects the entire chain of "equipment intervention-physiological response," resulting in a one-sided perception dimension and a high false negative rate in the predictive model.
[0004] Secondly, the prediction mechanism is static and lacks adaptive capability during dialysis. Existing risk prediction models typically use fixed feature weights or a single warning threshold, failing to consider the highly non-stationary nature of the dialysis process. Patients' tolerance to volume changes varies significantly at different stages of dialysis: the initial 30 minutes of dialysis are a rapid adjustment period, with minor fluctuations in blood pressure mostly due to physiological adaptation; however, the two hours after dialysis mark the vulnerable period, where cardiovascular reserve decreases, making hypotension more likely to be induced by the same ultrafiltration volume. Current technology treats all time periods equally, leading to frequent false alarms during the rapid adjustment period and insufficient sensitivity during the vulnerable period, resulting in a severe "crying wolf" effect in clinical practice and a decline in physician trust.
[0005] Third, intervention decisions are opaque, lacking actionable counterfactual reasoning capabilities. Existing AI early warning systems only output abstract probability values such as "87% risk of low blood pressure," failing to explain the source of the risk or provide specific intervention suggestions such as "how much should the ultrafiltration rate be reduced, whether saline infusion is necessary, and how should the sodium concentration of the dialysate be adjusted." After receiving an early warning, doctors still need to rely on experience to blindly try various intervention measures, delaying the optimal intervention window. Furthermore, existing generative models, when simulating virtual intervention trajectories, do not incorporate medical physical constraints, often generating physiologically infeasible pseudo-trajectories such as blood pressure below the lethal threshold or negative ultrafiltration rates, making them unusable for real clinical decision-making.
[0006] In summary, there is an urgent need in this field for an intelligent data processing method that can deeply integrate multimodal dialysis data, adapt to the dynamic characteristics of dialysis stages, and output executable causal intervention plans. This method would overcome the limitations of existing early warning technologies, such as limited perception, time-domain mismatch, and diagnosis without treatment, and achieve accurate prediction and proactive prevention of hypotension during dialysis. Summary of the Invention
[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an artificial intelligence-based method for processing kidney dialysis data.
[0008] To achieve the above objectives, the present invention provides the following technical solution: The artificial intelligence-based method for processing kidney dialysis data includes the following steps: S1. Multimodal Dynamic Coupling Graph Feature Extraction: Real-time acquisition of cross-modal data streams during dialysis, including high-frequency time-series numerical signals from the dialysis machine and low-frequency clinical text event signals entered by nurses; construction of a dynamic coupling graph, using high-frequency values as dynamic node features and low-frequency events as event node features, calculating the dynamic mutual information coefficients between different modal nodes and establishing cross-modal coupling edges based on adaptive thresholds, generating a weighted dynamic graph reflecting the coupling relationship between machine parameters and human responses; neighborhood aggregation of the weighted dynamic graph through a graph attention network, outputting a deep temporal feature vector containing cross-modal coupling relationships; S2, Stage Adaptive Temporal Risk Prediction: The deep temporal feature vector is input into the dilated causal convolutional network, and non-equidistant intervals are divided according to the dialysis process. The historical features of different dialysis intervals are adaptively weighted by stage sensitivity through learnable interval embedding vectors and attention mechanism, and the probability of hypotension onset and its changing trend in the future preset time window are output. S3. Counterfactual Causal Intervention Program Generation: When a high-risk trend is identified, a structural causal model is constructed to identify the interventionable feature that contributes the most to the risk. At least two intervention operations of different magnitudes are performed. The virtual time-series trajectory under each intervention operation is simulated by generating adversarial networks. Based on medical safety constraints and dose-response monotonicity constraints, the intervention program with the greatest risk reduction is selected, and a structured medical order containing specific intervention parameters and expected effects is output.
[0009] Specifically, the calculation of the dynamic mutual information coefficient in S1 includes: Using the current time as the endpoint, take a fixed-duration time window and calculate the mutual information value between the feature sequences of high-frequency numerical nodes and the feature sequences of low-frequency event nodes within the window; The adaptive threshold is dynamically generated using a specified quantile of the average mutual information matrix of the patient's historical dialysis records; When the mutual information value is higher than the dynamic threshold, a cross-modal coupling edge is established, and the weight of the edge is the normalized mutual information value.
[0010] Specifically, when establishing cross-modal coupling edges in S1, a time decay factor is introduced to dynamically correct the edge weights: The normalized mutual information value calculated at the current time step is fused with the weight of the coupling edge at the previous time step using momentum, and an exponential decay related to the time interval is applied to the weight at the previous time step. If the coupling edge is not established at the current time, the weight of the previous time is used as the edge weight at the current time after exponential decay. When the decayed weight is lower than a preset threshold, the edge is completely deleted.
[0011] Specifically, the S2 dilated causal convolutional network employs a multi-scale feature fusion strategy: The feature maps output by convolutional layers with different dilation rates are concatenated along the channel dimension. The contribution of each scale feature is adaptively recalibrated through a learnable channel attention module. The fused feature map is then reduced to the original number of channels. The channel attention module includes global average pooling, a fully connected layer, and a Sigmoid activation function.
[0012] Specifically, in S2: The non-equidistant intervals are divided into the rapid adjustment period, the stable period, and the vulnerable period based on the pathophysiological characteristics of dialysis duration; The interval embedding vector is assigned a trainable vector to each interval, and is concatenated with the temporal feature vector of the current interval to form a query matrix. Historical features are weighted through a multi-head scaling dot product attention mechanism.
[0013] Specifically, in S2: The trend of the probability of hypotension is characterized by the first derivative. If the first derivative exceeds the rising threshold for multiple consecutive sampling points, it is judged as a steep rise in risk; otherwise, it is judged as a stable low risk.
[0014] Specifically, in S3: The generative adversarial network is a conditional temporal convolutional generator and a bidirectional long short-term memory discriminator, and medical physics constraint loss is introduced during the training process. The medical physical constraint loss includes penalties for predicted blood pressure below the safe lower limit and penalties for predicted ultrafiltration rate exceeding the safe range.
[0015] Specifically, the dose-response monotonicity constraint in S3 is as follows: The greater the reduction in intervention characteristics, the smaller the predicted decrease in blood pressure should be, or the trend should be positively correlated. By generating multiple virtual trajectories with varying downward adjustment magnitudes in parallel for the same input sample, the rank correlation coefficient between the blood pressure decrease rate and the downward adjustment magnitude is calculated. If the coefficient is lower than a preset threshold, it is determined to violate monotonicity and the generated result is rejected.
[0016] The technical effects and advantages of this invention are as follows: This invention achieves intelligent end-to-end processing of hypotension during dialysis—from feature perception to trend prediction to intervention decision-making—through the collaborative design of a multimodal dynamic coupled graph network, a stage-adaptive temporal convolutional network, and a counterfactual causal reasoning generator. Compared to existing technologies, this invention first breaks through the limitations of single numerical signals at the feature extraction level. By constructing a cross-modal dynamic coupled graph and introducing a time decay factor to dynamically correct edge weights, it can deeply explore the nonlinear coupling relationship between machine parameters and patient symptoms, significantly improving the ability to capture latent risk features. Secondly, at the risk prediction level, through multi-scale dilated convolution and adaptive weighting of learnable interval embedding vectors, the model can dynamically adjust its sensitivity according to the dialysis process, effectively suppressing false alarms during rapid adjustment periods and timely identifying steeply rising risks during vulnerable periods, substantially improving prediction accuracy and clinical acceptability.
[0017] This invention, for the first time at the decision support level, introduces counterfactual causal reasoning into the intervention scenario of dialysis-induced hypotension. It identifies key interventionable features through a structural causal model and utilizes a conditional generative adversarial network embedded with medical physical constraints and dose-response monotonicity constraints to simulate virtual physiological trajectories under various interventional procedures. This results in structured medical orders with clear parameter adjustment suggestions and quantified expected effects. This mechanism not only upgrades traditional AI warnings from abstract probabilistic outputs to executable and verifiable clinical prescriptions, effectively solving the "black box" dilemma of diagnosis without treatment, but also significantly shortens the doctor's reaction time from receiving the information to making a decision through a one-click confirmation human-computer interaction design. It provides an innovative and practical technical solution for the proactive prevention and personalized treatment of hypotension during dialysis. Attached Figure Description
[0018] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] like Figure 1 As shown, the modules of the artificial intelligence-based kidney dialysis data processing method are as follows: Step S1: Multimodal dynamic coupling graph feature extraction. High-frequency numerical signals from the dialysis machine and low-frequency clinical events entered by nurses are acquired in real time. Cross-modal dynamic mutual information is calculated using a sliding window to construct a weighted graph structure reflecting the coupling relationship between "machine parameters and human response." A time decay factor is introduced to dynamically correct the edge weights, and a graph attention network is used for neighborhood aggregation, outputting a temporal feature vector containing deep coupling relationships. Specific execution: First, outlier detection and missing value imputation were performed on the collected high-frequency time-series data. The 3-sigma principle was adopted: the mean μ and standard deviation σ of each parameter (systolic blood pressure, diastolic blood pressure, ultrafiltration rate, dialysate sodium ion concentration, and blood flow) over the most recent 10 minutes were calculated. Data points exceeding the μ±3σ range were identified as transient outliers caused by postural changes or sensor malfunctions and were removed. The resulting gaps were filled using linear interpolation, with the average of the two valid sampling points before and after the outlier. Low-frequency text event labels were standardized and mapped: colloquial complaints entered by nurses, such as "dizziness," "blurred vision," "palpitations," and "cold sweats," were uniformly categorized as "central nervous system symptoms"; "cramps" and "muscle tightness" were categorized as "muscle spasm symptoms"; and "nausea" and "vomiting" were categorized as "gastrointestinal symptoms." The mapped labels participated in the subsequent coupling graph construction with a unique encoding format.
[0021] The server collects cross-modal dialysis data streams in real time. The first modality is the high-frequency time-series numerical signal of the dialysis machine, including systolic blood pressure per minute, diastolic blood pressure per minute, ultrafiltration rate, dialysate sodium ion concentration, and blood flow. The second modality is the low-frequency clinical text event signal, including nurses' recorded complaints about ringing bells, muscle spasms, and nausea symptoms.
[0022] Constructing a dynamic coupling graph: For each sampling time T, the five high-frequency values at the current time are used as dynamic node features, and the cumulative low-frequency events within the previous 15 minutes are used as event node features. The dynamic mutual information coefficient between any two modal nodes is calculated, and cross-modal coupling edges are established for node pairs with mutual information values higher than a threshold, thereby generating a weighted dynamic graph reflecting the relationship between "machine parameters and human response". The dynamic mutual information coefficient is calculated in the following way: For each sampling time T, let the high-frequency numerical node feature sequence be... The characteristic sequence of low-frequency event nodes is Using the current time T as the endpoint, take a 20-minute time window (corresponding to approximately 40 sampling points on the dialysis machine, with a sampling frequency of 0.5Hz), and calculate the mutual information value between X and Y within this window: ; in, Mutual information is the value between random variables X and Y, which measures the degree of interdependence between the two variables. These are specific feature variables in sets X and Y, respectively. Let x be the marginal probability density function. Let y be the marginal probability density function. The natural logarithm is used to calculate information content. To estimate the joint probability density, a nonparametric kernel density estimation method is used. A Gaussian kernel is chosen as the kernel function, and the bandwidth is automatically determined using the Silverman rule. The kernel density estimation formula is as follows: ; in, Let be the probability density estimate of the random variable x, N be the total number of samples (the number of sampling points within the time window), and h be the bandwidth, which controls the smoothness of the kernel density estimate and is automatically determined by the Silverman rule. For the kernel function, a Gaussian kernel is selected. The eigenvalues are the current probability density values to be estimated. Let be the feature value of the i-th sample, and i be the sample index. For standardized variables, It is a natural constant; The threshold is not set as a fixed value, but rather adopts an adaptive threshold strategy: the average mutual information matrix of the four dialysis sessions preceding the current time (i.e., the patient's four most recent dialysis records) is calculated, and its 75th percentile is taken as the dynamic threshold for the current time. If the mutual information value at the current moment If the condition is met, a cross-modal coupling edge is established, and the weight of the edge is the normalized mutual information value; otherwise, no connection is established.
[0023] When establishing cross-modal coupling edges, a time decay factor is introduced to dynamically adjust the edge weights: Let the current time be T, and the weight of the coupling edge at the previous time (T−1) be... Then the final weight at the current moment ,in The final weight of the coupling edge at time T. This represents the weight of the coupling edge at the previous time step (T−1). The normalized mutual information value calculated at the current moment. The time interval (in minutes) since the last time the same coupling edge was established. The momentum update coefficient is set to 0.7. The time decay constant is set to 10. This mechanism allows strong coupling relationships that occur repeatedly in the short term to accumulate and strengthen, while occasional pseudo-coupling relationships decay rapidly, improving the temporal smoothness and physiological interpretability of the graph structure.
[0024] If the coupling edge has not been established at the current time (i.e.) Then the weight of the previous time step will be... Multiply by the attenuation factor The weight is then used as the edge weight at the current time (if the weight is less than 0.1 after decay, the edge is completely deleted) to maintain the sparsity of the graph structure.
[0025] Graph Convolutional Aggregation: A graph attention network is used to perform neighborhood aggregation on the dynamic graph, outputting a deep feature vector containing cross-modal coupling relationships. This step addresses the shortcomings of traditional methods that analyze blood pressure and ultrafiltration rate separately, failing to capture implicit coupling patterns such as "high sodium dialysis leading to thirst and thus increased hidden water intake".
[0026] Step S2: Stage-adaptive temporal risk prediction. The temporal feature vector is input into a multi-scale dilated causal convolutional network, and features from different receptive fields are adaptively fused using a channel attention mechanism. Non-equidistant intervals are divided according to the dialysis progress, and learnable interval embedding vectors and multi-head attention are used to calibrate the stage sensitivity of historical features. The probability of hypotension and its first derivative for the next 10 minutes and 20 minutes are output, achieving refined judgment of "sharply rising high risk" and "stable low risk." Specific execution: The temporal feature vector sequence generated in step S1 is input into the dilated causal convolutional network; the dilated causal convolutional network employs a multi-scale feature fusion strategy. The feature maps output from 4 layers of dilated convolution (dilation rates of 1, 2, 4, and 8) By stitching along the channel dimension, the fused features are obtained. Subsequently, the contributions of features at each scale are adaptively recalibrated through a learnable channel attention module. ; in, This is the fused feature map after channel attention recalibration. This is the fused feature obtained by concatenating the feature maps output by 4 layers of dilated convolution along the channel dimension. This is the weight matrix of the first fully connected layer. , This is the weight matrix for the second fully connected layer. ; It is a linear rectified activation function. For global average pooling, the time dimension is compressed to 1; d is the number of feature channels, and the reduction rate is... Take 16; The Sigmoid activation function outputs the attention weights for each channel. This is a channel-by-channel multiplication.
[0027] Ultimately The dimensionality is reduced to the original number of channels d using a 1×1 convolutional layer, which serves as the input to the subsequent interval adaptive weight calibration module. This strategy enables the network to dynamically integrate features with different inflation rates according to different dialysis stages—relying more on a small inflation rate (1,2) to capture instantaneous fluctuations during the rapid adjustment period, and relying more on a large inflation rate (4,8) to capture cumulative effects during the vulnerable period, thereby significantly improving the temporal adaptability of risk prediction.
[0028] Unlike conventional recurrent neural networks, this step divides the dialysis duration (4 hours) into 8 non-equidistant intervals (0-30min rapid adjustment period, 30-120min stable period, 120-240min vulnerable period, etc.).
[0029] Interval Adaptive Weight Calibration: A learnable interval embedding vector is set, and an attention mechanism is used to weight historical features of different dialysis intervals. For example, the sensitivity to total ultrafiltration volume and cardiovascular reserve features is significantly enhanced during vulnerable periods, while the weight of small fluctuations in blood pressure is weakened during rapid adjustment periods.
[0030] The interval embedding vector is generated as follows: After dividing the dialysis duration into 8 non-equidistant intervals, a trainable d-dimensional embedding vector is assigned to each interval. d is set to 64. This embedding vector is jointly optimized with the network backbone through backpropagation, and initialized using Xavier uniform initialization.
[0031] The attention mechanism is specifically a multi-head scaling dot product attention, calculated as follows: Let the time-series feature vector of the current dialysis interval be... Where t is the number of time steps within the interval. H is then compared with the interval embedding vector. By concatenating the matrices, we obtain the query matrix. Attention weights are calculated as follows: ; in To scale the output of the dot product attention mechanism, Q is the query matrix, which consists of the temporal feature vector H and the interval embedding vector. The resulting matrix is obtained by concatenating the two matrices; K is the key matrix, and V is the value matrix. This is the transpose of the key matrix. Scaling factor The attention score is converted into a probability distribution using a normalized exponential function, where H is the temporal feature vector of the current dialysis interval, and the dimension is... , Let be the learnable embedding vector for the i-th dialysis interval; An 8-head attention mechanism is used, where the outputs of each head are concatenated and then mapped back to the original dimension through a fully connected layer. The weights of the fully connected layer are L2 regularized with a regularization coefficient of 0.01.
[0032] This mechanism enables the model to explicitly perceive the current dialysis stage and dynamically adjust its sensitivity to features such as blood pressure fluctuations and ultrafiltration rate based on the pathophysiological characteristics of different stages. After training, the visualization of attention weights shows that during the vulnerable period, the model automatically increases the attention weight of the ultrafiltration accumulation feature to 3-5 times that during the rapid adjustment period; during the rapid adjustment period, the attention weight of small blood pressure fluctuations is significantly suppressed.
[0033] Generate a risk trend line: The network outputs the probability of hypotension attacks in the next 10 minutes and 20 minutes, and simultaneously generates the first derivative of the risk probability over time. If the first derivative is greater than 0.15 for three consecutive sampling points, it is judged as "sharply rising high risk"; if the derivative is flat and the absolute value is low, it is judged as "stable low risk". This step avoids the problem of frequent false alarms in traditional threshold alarms when blood pressure is close to the lower limit of normal by introducing the concept of time derivative.
[0034] Step S3: Counterfactual causal intervention plan generation. When a steeply rising high risk is identified, a structural causal model is constructed to lock the highest contributing feature (such as ultrafiltration rate); multi-level intervention operations are performed on the intervention variables, and the virtual time-series trajectory after intervention is simulated through conditional generative adversarial network; the optimal intervention plan is screened based on medical physical constraints and dose-response monotonicity, and an executable structured medical order is directly output. Specific execution: When step S2 determines the risk level to be "sharply escalating high risk," the intervention decision module is activated. This step constructs a structural causal model, with endogenous variables including: current ultrafiltration rate, cumulative water removal volume, serum sodium level, and hematocrit.
[0035] Counterfactual reasoning process: The model does not directly predict blood pressure, but answers the questions "If the current ultrafiltration rate is reduced from 500 ml / h to 300 ml / h, how much will the probability of blood pressure decrease change?" and "If the sodium concentration of the dialysate is adjusted from 140 to 138 and 100 ml of normal saline is infused, will the risk curve flatten?"
[0036] Specific execution logic: Based on the current patient characteristics, the model executes the following algorithm loop: Lock in the highest contributing feature identified in step S2 (e.g., feature attribution shows that "excessive ultrafiltration rate" is the main cause); Interventional operations were performed on the ultrafiltration rate characteristics, reducing their values by 5%, 10%, and 15% within the permissible safe range; A trained generative adversarial network is used to simulate the virtual temporal trajectory after intervention. Calculate the risk value of hypotension under the virtual trajectory and select the interventional procedure with the greatest risk reduction.
[0037] Outputting structured medical orders: The final terminal does not display an abstract "87% risk of low blood pressure," but instead displays structured intervention suggestions, such as: "Warning: IDH is expected in 15 minutes. It is recommended to immediately implement Plan A: reduce the ultrafiltration rate from 600 ml / h to 400 ml / h and maintain this rate for 20 minutes. This is expected to reduce the risk by 52%." An "Execution Confirmation" button is also provided for doctors to confirm with a single click. The simulation of the virtual temporal trajectory after intervention using a trained generative adversarial network includes the following sub-steps: Training data construction: Dialysis segments in historical dialysis records where hypotension events occurred are used as the positive sample set, while dialysis segments without hypotension events are randomly sampled at a 1:1 ratio as negative samples to balance the categories. Each sample includes: time-series data for the 30 minutes prior to the occurrence of hypotension (blood pressure, ultrafiltration rate, dialysate sodium ion concentration, hematocrit, etc.), and the corresponding actual intervention measures taken (ultrafiltration rate reduction, sodium concentration adjustment, fluid replacement volume, etc.). The intervention measures are encoded as a conditional vector c, with the time-series data for the 30 seconds prior to the intervention serving as the input Xgen, and the time-series trajectory for the 15 minutes following the intervention serving as the target Ygen.
[0038] Network Structure: The generator uses a conditional temporal convolutional network: the input Xgen is concatenated with the conditional vector c, and then passed through four layers of dilated causal convolutions (dilation rates of 1, 2, 4, and 8 respectively). Each convolutional layer has a kernel size of 3 and 64 channels, and the output dimension is the same as Ygen. The discriminator uses a bidirectional long short-term memory network: the input is either a real trajectory or a generated trajectory, which is passed through two layers of bidirectional LSTM (128 hidden units), and finally outputs a binary classification probability (true / false) through a fully connected layer.
[0039] Loss function: In addition to the usual adversarial loss In addition to L1 reconstruction loss, medical physical constraint loss is also introduced. : ; in, The medical physical constraint loss is used to penalize generated trajectories that do not meet clinical safety requirements. T represents the total number of time steps in the temporal trajectory, and t represents the time step index. The predicted systolic blood pressure value generated at time step t. ( () is the function that takes the maximum value. The predicted ultrafiltration rate is generated at time step t. The absolute value of the predicted ultrafiltration rate. The upper limit of ultrafiltration rate is the safe limit. The lower limit of systolic blood pressure is set at 90 mmHg. The first penalty is that the generated blood pressure is lower than the safe value. The second penalty is that the generated ultrafiltration rate is constrained to the range of 0~800 ml / h, and any excess is included in the loss. , These are the balance coefficients for the blood pressure constraint and the ultrafiltration rate constraint, respectively, taken as 0.6 and 0.4. Total loss . To counteract the losses generated by the game between the discriminator and the generator.
[0040] Counterfactual generation constraint: During the inference phase, when performing intervention operations on variables such as ultrafiltration rate, the generator must satisfy the dose-response monotonicity constraint: that is, the greater the reduction in ultrafiltration rate, the smaller the predicted decrease in blood pressure should be (or show a positive correlation). If the generated result violates this monotonicity constraint (for example, blood pressure is lower after a 5% reduction in ultrafiltration rate than after a 10% reduction), the generated result is rejected and resampling is performed. In the specific implementation, three virtual trajectories with different reduction magnitudes are generated in parallel for the same input sample. The Spearman rank correlation coefficient between the blood pressure decrease rate and the reduction magnitude is calculated. If the coefficient is less than 0.7, it is determined to be a violation of monotonicity, the output of that group is discarded, and regeneration is triggered.
[0041] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.
[0042] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.
[0043] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0044] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0045] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0046] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.
[0047] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0048] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0049] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for processing renal dialysis data based on artificial intelligence, characterized in that, Includes the following steps: S1. Multimodal Dynamic Coupling Graph Feature Extraction: Real-time acquisition of cross-modal data streams during dialysis, including high-frequency time-series numerical signals from the dialysis machine and low-frequency clinical text event signals entered by nurses; construction of a dynamic coupling graph, using high-frequency values as dynamic node features and low-frequency events as event node features, calculating the dynamic mutual information coefficients between different modal nodes and establishing cross-modal coupling edges based on adaptive thresholds, generating a weighted dynamic graph reflecting the coupling relationship between machine parameters and human responses; neighborhood aggregation of the weighted dynamic graph using a graph attention network, outputting a deep temporal feature vector containing cross-modal coupling relationships; S2, Stage Adaptive Temporal Risk Prediction: The deep temporal feature vector is input into the dilated causal convolutional network, and non-equidistant intervals are divided according to the dialysis process. The historical features of different dialysis intervals are adaptively weighted by stage sensitivity through learnable interval embedding vectors and attention mechanism, and the probability of hypotension onset and its changing trend in the future preset time window are output. S3. Counterfactual Causal Intervention Program Generation: When a high-risk trend is identified, a structural causal model is constructed to identify the interventionable feature that contributes the most to the risk. At least two intervention operations of different magnitudes are performed. The virtual time-series trajectory under each intervention operation is simulated by generating adversarial networks. Based on medical safety constraints and dose-response monotonicity constraints, the intervention program with the greatest risk reduction is selected, and a structured medical order containing specific intervention parameters and expected effects is output.
2. The method for processing renal dialysis data based on artificial intelligence according to claim 1, characterized in that, The calculation of the dynamic mutual information coefficient in S1 includes: Using the current time as the endpoint, take a fixed-duration time window and calculate the mutual information value between the feature sequences of high-frequency numerical nodes and the feature sequences of low-frequency event nodes within the window; The adaptive threshold is dynamically generated using a specified quantile of the average mutual information matrix of the patient's historical dialysis records; When the mutual information value is higher than the dynamic threshold, a cross-modal coupling edge is established, and the weight of the edge is the normalized mutual information value.
3. The method for processing renal dialysis data based on artificial intelligence according to claim 2, characterized in that, When establishing cross-modal coupling edges in S1, a time decay factor is introduced to dynamically correct the edge weights: The normalized mutual information value calculated at the current time step is fused with the weight of the coupling edge at the previous time step using momentum, and an exponential decay related to the time interval is applied to the weight at the previous time step. If the coupling edge is not established at the current time, the weight of the previous time is used as the edge weight at the current time after exponential decay. When the decayed weight is lower than a preset threshold, the edge is completely deleted.
4. The method for processing renal dialysis data based on artificial intelligence according to claim 1, characterized in that, The S2 dilated causal convolutional network employs a multi-scale feature fusion strategy: The feature maps output by convolutional layers with different dilation rates are concatenated along the channel dimension. The contribution of each scale feature is adaptively recalibrated through a learnable channel attention module. The fused feature map is then reduced to the original number of channels. The channel attention module includes global average pooling, a fully connected layer, and a Sigmoid activation function.
5. The method for processing renal dialysis data based on artificial intelligence according to claim 1, characterized in that, In S2: The non-equidistant intervals are divided into the rapid adjustment period, the stable period, and the vulnerable period based on the pathophysiological characteristics of dialysis duration; The interval embedding vector is assigned a trainable vector to each interval, and is concatenated with the temporal feature vector of the current interval to form a query matrix. Historical features are weighted through a multi-head scaling dot product attention mechanism.
6. The method for processing renal dialysis data based on artificial intelligence according to claim 1, characterized in that, In S2: The trend of the probability of hypotension is characterized by the first derivative. If the first derivative exceeds the rising threshold for multiple consecutive sampling points, it is judged as a steep rise in risk; otherwise, it is judged as a stable low risk.
7. The method for processing renal dialysis data based on artificial intelligence according to claim 1, characterized in that, In S3: The generative adversarial network is a conditional temporal convolutional generator and a bidirectional long short-term memory discriminator, and medical physics constraint loss is introduced during the training process. The medical physical constraint loss includes penalties for predicted blood pressure below the safe lower limit and penalties for predicted ultrafiltration rate exceeding the safe range.
8. The method for processing renal dialysis data based on artificial intelligence according to claim 1, characterized in that, The dose-response monotonicity constraint in S3 is: The greater the reduction in intervention characteristics, the smaller the predicted decrease in blood pressure should be, or the trend should be positively correlated. By generating multiple virtual trajectories with varying downward adjustment magnitudes in parallel for the same input sample, the rank correlation coefficient between the blood pressure decrease rate and the downward adjustment magnitude is calculated. If the coefficient is lower than a preset threshold, it is determined to violate monotonicity and the generated result is rejected.