A dialysis patient multi-risk early warning intervention method and system

Abstract

This invention discloses a multi-risk early warning intervention method and system for dialysis patients. The method includes: acquiring multi-source time-series monitoring data; performing continuous time state alignment based on neural differential equations; jointly inferring volume status, hypotension, and hypoglycemia risks using a causal graph neural network; generating structured intervention instruction sequences through deep reinforcement learning; achieving continuous learning using Bayesian online variational inference; and combining federated learning to achieve cross-center collaborative training. This invention can achieve second-level time-series alignment, reduce the misjudgment rate of hypotension and hypoglycemia, and provide hemodialysis patients with a precise, real-time, executable, and continuously optimized intelligent monitoring solution.

Description

Technical Field

[0001] This invention belongs to the field of medical information technology and intelligent monitoring technology, and in particular relates to a method and system for multi-risk early warning intervention for dialysis patients. Background Technology

[0002] Maintenance hemodialysis is the primary renal replacement therapy for patients with end-stage renal disease. During dialysis treatment, volume management, blood pressure stability, and blood glucose control are the three core safety indicators. However, current clinical monitoring technologies have the following limitations:

[0003] (1) Isolated data acquisition: Dialysis machines, vital sign monitors, and blood glucose monitoring devices are usually from different manufacturers and use their own internal clocks and data formats, resulting in timestamp deviations and sampling rate differences in multi-source data. Traditional manual recording methods cannot achieve time alignment with second-level accuracy, affecting the accuracy of subsequent fusion analysis.

[0004] (2) Delayed risk warning: Existing monitoring systems usually only trigger alarms when blood pressure is below 90 mmHg or obvious symptoms appear, at which point the patient has already entered the compensatory or decompensatory phase. Alarms based on fixed thresholds cannot use time-series trend information for early prediction, thus missing the optimal intervention window.

[0005] (3) Confusion of multiple risk signals and lack of combined discriminative ability: The clinical manifestations of hypotension during dialysis (IDH) and hypoglycemia during dialysis are highly overlapping, both presenting with symptoms such as dizziness, sweating, palpitations, and decreased blood pressure. Studies have shown that relying solely on symptoms can lead to a clinical misdiagnosis rate of over 30%. Incorrect differential diagnosis leads to inappropriate intervention—misdiagnosing hypoglycemia as hypotension and simply administering fluids can delay glucose supplementation and, in severe cases, cause irreversible brain damage.

[0006] (4) Unstructured decision output and experience-dependent execution: The existing system only outputs risk warning text and does not form an executable sequence of operation instructions. Nurses with different seniority have significantly different understandings of vague prompts such as "strengthen observation" and "timely treatment", resulting in insufficient standardization and timeliness of intervention measures.

[0007] (5) Existing models are mostly static and cannot adapt to individual differences: patients' physiological state changes over time, and different individuals respond significantly differently to dialysis prescriptions. Traditional static models experience performance degradation in long-term use and lack a continuous learning mechanism.

[0008] Therefore, there is an urgent need for an intelligent system that can integrate multi-source heterogeneous data, achieve time series prediction, have the ability to make joint decisions on multiple risks, output structured intervention instructions, and support continuous learning. Summary of the Invention

[0009] This invention aims to solve the technical problems in the prior art, such as isolated data acquisition, delayed early warning, high clinical misjudgment rate of hypotension and hypoglycemia, unstructured decision output, and inability to continuously optimize models.

[0010] In a first aspect, the present invention provides a multi-risk early warning intervention method for dialysis patients, comprising:

[0011] Acquire multi-source time-series monitoring data, which includes at least dialysis machine operating parameters, continuous blood pressure signals, continuous blood glucose signals, and patient medical history data from the hospital information system;

[0012] Based on neural differential equations, continuous time state modeling and spatiotemporal alignment are performed on the multi-source time-series monitoring data to obtain time-synchronized multimodal physiological feature sequences.

[0013] The time-synchronized multimodal physiological feature sequence is input into a causal graph neural network for multi-risk joint inference, and the joint distribution of volume state level, probability of hypotension and probability of hypoglycemia is output.

[0014] Based on the optimal intervention strategy model of deep reinforcement learning, a structured intervention instruction sequence containing intervention action type, execution timing and dose parameters is generated according to the current joint distribution and individual patient status.

[0015] The structured intervention instruction sequence is sent to the nurse station terminal for visualization and execution feedback.

[0016] Secondly, the present invention provides a multi-risk early warning and intervention system for dialysis patients, comprising:

[0017] The acquisition module is configured to acquire multi-source time-series monitoring data, which includes at least dialysis machine operating parameters, continuous blood pressure signals, continuous blood glucose signals, and patient medical history data from the hospital information system.

[0018] The modeling module is configured to perform continuous time state modeling and spatiotemporal alignment on the multi-source time-series monitoring data based on neural differential equations, so as to obtain a time-synchronized multimodal physiological feature sequence.

[0019] The output module is configured to input the time-synchronized multimodal physiological feature sequence into a causal graph neural network for multi-risk joint inference, and output the joint distribution of volume state level, probability of hypotension, and probability of hypoglycemia.

[0020] The generation module is configured as an optimal intervention strategy model based on deep reinforcement learning. Based on the current joint distribution and the individual patient status, it generates a structured sequence of intervention instructions containing the type of intervention action, execution timing, and dosage parameters.

[0021] The sending module is configured to send the structured intervention instruction sequence to the nurse station terminal for visualization and execution feedback.

[0022] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the dialysis patient multi-risk early warning intervention method according to any embodiment of the present invention.

[0023] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the steps of the multi-risk early warning intervention method for dialysis patients according to any embodiment of the present invention.

[0024] The multi-risk early warning intervention method and system for dialysis patients proposed in this application have the following beneficial effects:

[0025] Based on continuous-time state estimation using neural differential equations, this method solves the problems of inconsistent sampling frequencies, timestamp bias, and missing data from multiple sources. Specifically, it unifies heterogeneous data from dialysis machines, continuous blood pressure monitors, blood glucose meters, etc., into a latent variable evolution process driven by ordinary differential equations, and uses the adjoint method for efficient training.

[0026] The graph neural network that integrates causal structure learning enables the model to move from simply learning statistical correlations to identifying the causal direction between variables. It automatically constructs a causal graph between volume, blood pressure, blood sugar, heart rate, and ultrafiltration rate through PC algorithm, and then propagates the causal effect through graph convolutional layers.

[0027] Hypoglycemia prediction based on a denoised diffusion probability model overcomes the shortcomings of discriminative models such as XGBoost, which cannot model complex blood glucose distributions and are insensitive to rare events.

[0028] The deep reinforcement learning-driven dynamic intervention strategy models intervention decisions as a Markov decision process and uses a soft actor-critic algorithm to optimize long-term cumulative rewards. The state space includes capacity level, dual-risk probability, patient characteristics, and historical actions; the action space includes combined interventions such as glucose injection, fluid resuscitation, and suspension of ultrafiltration; the reward function simultaneously penalizes adverse events (hypotension, hypoglycemia) and ultrafiltration actions, and rewards capacity balance.

[0029] In summary, this invention constructs a complete technical loop encompassing data alignment, causal inference, probabilistic prediction, dynamic decision-making, continuous adaptation, and privacy collaboration. Each module is mutually supportive and indispensable: Neural ODE provides high-quality temporal features, causal GNN ensures the causal reliability of inference, diffusion models overcome the performance bottleneck of hypoglycemia prediction, deep reinforcement learning achieves long-term optimal intervention, Bayesian mechanisms endow the model with evolutionary capabilities, and federated learning breaks down multi-center barriers. The above effects have been verified by retrospective and prospective clinical studies, providing a precise, real-time, executable, and continuously optimized intelligent solution for volume management and multi-risk prevention in maintenance hemodialysis patients, demonstrating clear clinical promotion value and industrial application prospects. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A flowchart of a multi-risk early warning intervention method for dialysis patients provided in an embodiment of the present invention;

[0032] Figure 2 This is a structural block diagram of a multi-risk early warning and intervention system for dialysis patients provided in an embodiment of the present invention;

[0033] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0035] Please see Figure 1 The diagram shows a flowchart of a multi-risk early warning intervention method for dialysis patients according to this application.

[0036] like Figure 1 As shown, the multi-risk early warning intervention method for dialysis patients specifically includes the following steps:

[0037] Step S101: Obtain multi-source time-series monitoring data, which includes at least dialysis machine operating parameters, continuous blood pressure signals, continuous blood glucose signals, and patient medical history data from the hospital information system;

[0038] Step S102: Based on neural differential equations, the multi-source time-series monitoring data is modeled in continuous time and spatiotemporally aligned to obtain a time-synchronized multimodal physiological feature sequence.

[0039] In this step, the latent variable state is defined. The evolution over time is controlled by ordinary differential equations parameterized by a neural network:

[0040] ,

[0041] In the formula, For deep neural networks, For any observation of sensor data at time t;

[0042] State estimation at any missing sampling time is achieved through backpropagation training using the adjoint method. The training loss function is:

[0043] ,

[0044] In the formula, For the observation time The true hidden state;

[0045] An attention mechanism is introduced to dynamically weight the reliability of different sensors. The attention weights are calculated as follows:

[0046] ,

[0047] .

[0048] Step S103: Input the time-synchronized multimodal physiological feature sequence into the causal graph neural network for multi-risk joint inference, and output the joint distribution of volume state level, probability of hypotension and probability of hypoglycemia.

[0049] In this step, the temporal causal graph construction module uses volume, blood pressure, blood sugar, heart rate, and ultrafiltration rate as nodes. It learns the causal direction between nodes from the data through a causal structure learning algorithm to obtain the prior structure of the causal graph.

[0050] Graph convolutional layers are used to update node representations:

[0051] ,

[0052] In the formula, For nodes The causal parent node, For activation function, , The weight matrix is ​​a learnable weight matrix;

[0053] Temporal causal convolutional layers are used to capture causal effects across time steps. The output of the causal convolution is:

[0054] ,

[0055] In the formula, For causal convolution kernel, The expansion rate;

[0056] The output layer is used to output the joint distribution of capacity state level, probability of hypotension, and probability of hypoglycemia.

[0057] .

[0058] Using the blood glucose sequence, heart rate sequence, and conditional information from a preset time window as input, Gaussian noise is gradually added to the actual blood glucose trajectory through a forward noise addition process. The forward process is defined as follows:

[0059] ,

[0060] In the formula, The noise variance is pre-defined for scheduling;

[0061] Training the neural network through a reverse denoising process Predict noise, and optimize the objective function as follows:

[0062] ,

[0063] In the formula, For conditional information, ;

[0064] After generating multiple future blood glucose trajectories, the probability of hypoglycemia is calculated as follows:

[0065] ,

[0066] In the formula, Low blood glucose threshold, This represents the number of sampling trajectories.

[0067] Step S104: Based on the optimal intervention strategy model of deep reinforcement learning, a structured intervention instruction sequence containing intervention action type, execution timing and dose parameters is generated according to the current joint distribution and individual patient status.

[0068] In this step, the optimal intervention strategy model based on deep reinforcement learning is trained using the soft actor-critic algorithm or the proximal policy optimization algorithm;

[0069] The state space is defined as follows:

[0070] ,

[0071] The action space consists of discrete action combinations, including: no action, glucose injection, fluid resuscitation, pause ultrafiltration, calling the doctor, and combined actions;

[0072] The reward function is defined as:

[0073] ,

[0074] in, For preset weighting coefficients, This is the adjustment amount for the ultrafiltration rate;

[0075] The policy network uses a Transformer structure, and the action probability distribution output is as follows:

[0076] ,

[0077] In the formula, For temperature coefficient, This is the action value function.

[0078] Generating a structured sequence of intervention instructions that includes the type of intervention action, execution timing, and dosage parameters also includes:

[0079] An asymmetric decision tree with hypoglycemia as the root node was used for intervention grading.

[0080] If there is a high risk of hypoglycemia and a high risk of hypotension, a three-level intervention instruction sequence will be output, and the mandatory timing is: intravenous glucose injection, saline infusion, suspension of ultrafiltration, elevation of lower limbs and notification of doctor;

[0081] If there is a high risk of hypoglycemia and a moderate or low risk of hypotension, then the decision to implement a secondary or tertiary intervention is based on the absolute blood glucose level. The decision function is as follows:

[0082] ,

[0083] If the risk of hypoglycemia is moderate, then a primary or secondary intervention should be initiated based on the patient's volume status.

[0084] If the risk of hypoglycemia is low, then a level 3 or level 1 intervention should be initiated based on the risk of hypotension.

[0085] It should be noted that the continuous learning mechanism based on Bayesian online variational inference is as follows:

[0086] Treating the model parameters as random variables, and assuming that the prior distribution of the parameters is Gaussian:

[0087] ,

[0088] After each batch of new data arrives, the approximate posterior is updated using online variational Bayesian methods:

[0089] ,

[0090] In the formula, This is the learning rate decay coefficient;

[0091] Using the diagonal Gaussian approximation The parameter update rule is as follows:

[0092] ,

[0093] ,

[0094] In the formula, This is the step size hyperparameter.

[0095] Step S105: The structured intervention instruction sequence is sent to the nurse station terminal for visualization and execution feedback.

[0096] In summary, the method of this application addresses the problems of isolated data acquisition, delayed early warning, high misjudgment rates of hypotension and hypoglycemia, unstructured decision output, and the inability to continuously optimize models in existing technologies. This invention acquires multi-source time-series monitoring data, performs continuous-time state alignment based on neural differential equations, utilizes causal graph neural networks to jointly infer volume status, hypotension, and hypoglycemia risk, generates structured intervention instruction sequences through deep reinforcement learning, and employs Bayesian online variational inference for continuous learning, combined with federated learning for cross-center collaborative training. This invention achieves second-level time-series alignment, reduces the misjudgment rate of hypotension and hypoglycemia, and provides hemodialysis patients with a precise, real-time, executable, and continuously optimized intelligent monitoring solution.

[0097] Please see Figure 2 The diagram shows a structural block diagram of a multi-risk early warning and intervention system for dialysis patients according to this application.

[0098] like Figure 2 As shown, the dialysis patient multi-risk early warning intervention system 200 includes an acquisition module 210, a modeling module 220, an output module 230, a generation module 240, and a sending module 250.

[0099] The system includes: an acquisition module 210 configured to acquire multi-source time-series monitoring data, which includes at least dialysis machine operating parameters, continuous blood pressure signals, continuous blood glucose signals, and patient medical history data from the hospital information system; a modeling module 220 configured to perform continuous time-state modeling and spatiotemporal alignment of the multi-source time-series monitoring data based on neural differential equations to obtain a time-synchronized multimodal physiological feature sequence; an output module 230 configured to input the time-synchronized multimodal physiological feature sequence into a causal graph neural network for multi-risk joint inference, and output a joint distribution of volume state level, probability of hypotension, and probability of hypoglycemia; a generation module 240 configured to generate an optimal intervention strategy model based on deep reinforcement learning, which generates a structured intervention instruction sequence containing intervention action type, execution timing, and dose parameters according to the current joint distribution and the patient's individual state; and a sending module 250 configured to send the structured intervention instruction sequence to the nurse station terminal for visualization and execution feedback.

[0100] It should be understood that Figure 2 The modules and references described in the document Figure 1 The steps described in the text correspond to those in the method described above. Therefore, the operations, features, and corresponding technical effects described above also apply to the method described in the text. Figure 2 The various modules in the document will not be described in detail here.

[0101] In other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the multi-risk early warning intervention method for dialysis patients in any of the above method embodiments.

[0102] In one embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, which are configured as follows:

[0103] Acquire multi-source time-series monitoring data, which includes at least dialysis machine operating parameters, continuous blood pressure signals, continuous blood glucose signals, and patient medical history data from the hospital information system;

[0104] Based on neural differential equations, continuous time state modeling and spatiotemporal alignment are performed on the multi-source time-series monitoring data to obtain time-synchronized multimodal physiological feature sequences.

[0105] The time-synchronized multimodal physiological feature sequence is input into a causal graph neural network for multi-risk joint inference, and the joint distribution of volume state level, probability of hypotension and probability of hypoglycemia is output.

[0106] Based on the optimal intervention strategy model of deep reinforcement learning, a structured intervention instruction sequence containing intervention action type, execution timing and dose parameters is generated according to the current joint distribution and individual patient status.

[0107] The structured intervention instruction sequence is sent to the nurse station terminal for visualization and execution feedback.

[0108] Computer-readable storage media may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application program required for at least one function; the data storage area may store data created based on the use of the dialysis patient multi-risk warning intervention system, etc. Furthermore, the computer-readable storage medium may include high-speed random access memory, and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely configured relative to a processor, which can be connected to the dialysis patient multi-risk warning intervention system via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0109] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 3 As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 3 Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the multi-risk early warning intervention method for dialysis patients described in the above embodiment. The input device 330 can receive input digital or character information and generate key signal inputs related to user settings and function control of the multi-risk early warning intervention system for dialysis patients. The output device 340 may include a display screen or other display device.

[0110] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0111] In one implementation, the above-described electronic device is used in a multi-risk early warning intervention system for dialysis patients, serving as a client, and includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

[0112] Acquire multi-source time-series monitoring data, which includes at least dialysis machine operating parameters, continuous blood pressure signals, continuous blood glucose signals, and patient medical history data from the hospital information system;

[0113] Based on neural differential equations, continuous time state modeling and spatiotemporal alignment are performed on the multi-source time-series monitoring data to obtain time-synchronized multimodal physiological feature sequences.

[0114] The time-synchronized multimodal physiological feature sequence is input into a causal graph neural network for multi-risk joint inference, and the joint distribution of volume state level, probability of hypotension and probability of hypoglycemia is output.

[0115] Based on the optimal intervention strategy model of deep reinforcement learning, a structured intervention instruction sequence containing intervention action type, execution timing and dose parameters is generated according to the current joint distribution and individual patient status.

[0116] The structured intervention instruction sequence is sent to the nurse station terminal for visualization and execution feedback.

[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-risk early warning intervention method for dialysis patients, characterized in that, include: Acquire multi-source time-series monitoring data, which includes at least dialysis machine operating parameters, continuous blood pressure signals, continuous blood glucose signals, and patient medical history data from the hospital information system; Based on neural differential equations, continuous time state modeling and spatiotemporal alignment are performed on the multi-source time-series monitoring data to obtain time-synchronized multimodal physiological feature sequences. The time-synchronized multimodal physiological feature sequence is input into a causal graph neural network for multi-risk joint inference, and the joint distribution of volume state level, probability of hypotension and probability of hypoglycemia is output. Based on the optimal intervention strategy model of deep reinforcement learning, a structured intervention instruction sequence containing intervention action type, execution timing and dose parameters is generated according to the current joint distribution and individual patient status. The structured intervention instruction sequence is sent to the nurse station terminal for visualization and execution feedback.

2. A multi-risk early warning intervention method for dialysis patients according to claim 1, characterized in that, The continuous-time state modeling and spatiotemporal alignment of the multi-source time-series monitoring data based on neural differential equations includes: Defining latent variable states Ordinary differential equations controlled by neural network parameters evolve over time: ， In the formula, For deep neural networks, For any observation of sensor data at time t; State estimation at any missing sampling time is achieved through backpropagation training using the adjoint method. The training loss function is: ， In the formula, For the observation time The true hidden state; An attention mechanism is introduced to dynamically weight the reliability of different sensors. The attention weights are calculated as follows: ， 。 3. The method for multi-risk early warning intervention for dialysis patients according to claim 1, characterized in that, The causal graph neural network includes: The temporal causal graph construction module uses volume, blood pressure, blood sugar, heart rate, and ultrafiltration rate as nodes. It learns the causal direction between nodes from the data through a causal structure learning algorithm to obtain the prior structure of the causal graph. Graph convolutional layers are used to update node representations: ， In the formula, For nodes The causal parent node, For activation function, , The weight matrix is ​​a learnable weight matrix; Temporal causal convolutional layers are used to capture causal effects across time steps. The output of the causal convolution is: ， In the formula, For causal convolution kernel, The expansion rate; The output layer is used to output the joint distribution of capacity state level, probability of hypotension, and probability of hypoglycemia. 。 4. The multi-risk early warning intervention method for dialysis patients according to claim 1, characterized in that, The probability of hypoglycemia is predicted using a denoised diffusion probability model, specifically including: Using the blood glucose sequence, heart rate sequence, and conditional information from a preset time window as input, Gaussian noise is gradually added to the actual blood glucose trajectory through a forward noise addition process. The forward process is defined as follows: ， In the formula, The noise variance is pre-defined for scheduling; The neural network is trained through a reverse denoising process. Predict noise, and optimize the objective function as follows: ， In the formula, For conditional information, ; After generating multiple future blood glucose trajectories, the probability of hypoglycemia is calculated as follows: ， In the formula, Low blood sugar threshold, This represents the number of sampling trajectories.

5. The multi-risk early warning intervention method for dialysis patients according to claim 1, characterized in that, The optimal intervention strategy model based on deep reinforcement learning is trained using the soft actor-critic algorithm or the proximal policy optimization algorithm. The state space is defined as follows: ， The action space consists of discrete action combinations, including: no action, glucose injection, fluid resuscitation, pause ultrafiltration, calling the doctor, and combined actions; The reward function is defined as: ， in, For preset weighting coefficients, This is the ultrafiltration rate adjustment amount; The policy network uses a Transformer structure, and the action probability distribution output is as follows: ， In the formula, For temperature coefficient, This is the action value function.

6. The multi-risk early warning intervention method for dialysis patients according to claim 1, characterized in that, The generation of a structured intervention instruction sequence containing intervention action type, execution timing, and dosage parameters also includes: An asymmetric decision tree with hypoglycemia as the root node was used for intervention grading. If there is a high risk of hypoglycemia and a high risk of hypotension, a three-level intervention instruction sequence will be output, and the mandatory timing is: intravenous glucose injection, saline infusion, suspension of ultrafiltration, elevation of lower limbs and notification of doctor; If there is a high risk of hypoglycemia and a moderate or low risk of hypotension, then the decision to implement a secondary or tertiary intervention is based on the absolute blood glucose level. The decision function is as follows: ， If the risk of hypoglycemia is moderate, then a primary or secondary intervention should be initiated based on the patient's volume status. If the risk of hypoglycemia is low, then a level 3 or level 1 intervention should be initiated based on the risk of hypotension.

7. The multi-risk early warning intervention method for dialysis patients according to claim 1, characterized in that, It also includes a continuous learning mechanism based on Bayesian online variational inference: Treating the model parameters as random variables, and assuming that the prior distribution of the parameters is Gaussian: ， After each batch of new data arrives, the approximate posterior is updated using online variational Bayesian methods: ， In the formula, This is the learning rate decay coefficient; Using the diagonal Gaussian approximation The parameter update rule is as follows: ， ， In the formula, This is the step size hyperparameter.

8. A multi-risk early warning and intervention system for dialysis patients, characterized in that, include: The acquisition module is configured to acquire multi-source time-series monitoring data, which includes at least dialysis machine operating parameters, continuous blood pressure signals, continuous blood glucose signals, and patient medical history data from the hospital information system. The modeling module is configured to perform continuous time state modeling and spatiotemporal alignment on the multi-source time-series monitoring data based on neural differential equations, so as to obtain a time-synchronized multimodal physiological feature sequence. The output module is configured to input the time-synchronized multimodal physiological feature sequence into a causal graph neural network for multi-risk joint inference, and output the joint distribution of volume state level, probability of hypotension, and probability of hypoglycemia. The generation module is configured as an optimal intervention strategy model based on deep reinforcement learning. Based on the current joint distribution and the individual patient status, it generates a structured sequence of intervention instructions containing the type of intervention action, execution timing, and dosage parameters. The sending module is configured to send the structured intervention instruction sequence to the nurse station terminal for visualization and execution feedback.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method according to any one of claims 1 to 7.

Images