Chronic heart failure patient water intake dynamic regulation and control method based on machine learning

Abstract

The invention provides a chronic heart failure patient water intake dynamic regulation and control method based on machine learning, relates to the field of dynamic regulation and control, and improves the individualized level and clinical practicability of water intake management. The method comprises the following steps: firstly, constructing a standard time sequence data set through physiological parameter acquisition and time sequence alignment processing; thirdly, constructing a risk prediction model by adopting a time sequence fusion encoder trained by a meta-learning framework, realizing rapid individualized adaptation, and outputting a liquid retention risk probability value and a basic water intake interval; and finally, based on the risk probability value, in combination with a patient comfort index and a renal function safety index, performing multi-objective decision in the basic water intake interval through adaptive weighted optimization and a reinforcement learning adjustment mechanism, and generating a comprehensive recommended water intake. By integrating the meta-learning framework, the time sequence fusion encoder and the multi-target collaborative optimization mechanism, the accuracy, the safety and the individualized level of water intake management of the chronic heart failure patient are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of dynamic regulation, specifically a method for dynamic regulation of water intake in patients with chronic heart failure based on machine learning. Background Technology

[0002] Daily fluid intake management is a crucial aspect of disease control for patients with chronic heart failure. Inappropriate fluid intake can directly lead to fluid retention, causing acute symptoms such as dyspnea and edema, and even increasing the risk of readmission. Traditional fluid intake management methods mainly rely on static clinical guideline recommendations (such as 1500 ml per day) or physicians' individual experience. This approach fails to consider the dynamic fluctuations in patients' physiological state, individual differences, and potential conflicts between different management goals (such as risk control, comfort, and renal function protection). This results in rigid management strategies that are difficult to implement in complex and ever-changing real-world scenarios, limiting both patient compliance and management effectiveness.

[0003] Existing technologies have introduced data-driven approaches, collecting patient physiological data and using machine learning models for risk prediction, achieving an initial shift from static rules to dynamic assessment. However, existing methods still have significant shortcomings: First, the models lack effective individualized adaptation mechanisms, resulting in a "cold start" problem when dealing with new patients, requiring a large amount of labeled data to achieve usable accuracy, making it difficult to quickly apply to individuals. Second, the models typically perform simple concatenation of multidimensional physiological parameters, failing to deeply explore the dynamic correlations and interactions between different parameters over time, and have limited ability to capture early, weak pathological signals. In addition, most existing solutions focus on a single risk prediction objective, failing to incorporate key factors such as patient comfort and renal function safety into a unified optimization framework, making it impossible to make comprehensive optimal decisions under multiple constraints, thus limiting their clinical applicability. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes a machine learning-based method for dynamic regulation of water intake in patients with chronic heart failure.

[0005] Therefore, the technical solution adopted by the present invention is as follows:

[0006] A machine learning-based method for dynamic regulation of water intake in patients with chronic heart failure, comprising:

[0007] S1: Collect patients' physiological parameter data and perform time window alignment and resampling to obtain the patients' standard time series dataset;

[0008] S2: A risk prediction model is constructed using a temporal fusion encoder trained with a meta-learning framework. The patient's input dataset is input into the risk prediction model to obtain the probability value of fluid retention risk and the corresponding baseline water intake range. The model is configured to adjust the initial parameters based on new standard time-series data using a built-in gradient update algorithm.

[0009] S3: Based on the probability value of fluid retention risk, combined with preset patient comfort indicators and renal function safety indicators, and according to the baseline water intake range, multi-objective optimization is performed to finally output a comprehensive recommended water intake.

[0010] Furthermore, the physiological parameters include weight, heart rate, urine output, blood pressure, and blood oxygen saturation.

[0011] The time window alignment and resampling are based on a unified time window, which aligns the physiological parameter data to the target time point, and then the standard time series dataset is obtained through standardization processing.

[0012] Furthermore, the risk prediction model employs an attention mechanism to calculate the weights between standard time-series data at each target time point, expressed as:

[0013]

[0014] in, Indicates attention weight; Represents an exponential function; Represents the similarity function; Represents any time point in a standard time series dataset Sum the indexed similarities; Indicates the first Query vectors at each time point; Indicates the first Key vectors at each time point; Represents any time point in a standard time series dataset The key vector;

[0015] Based on the attention weights, a context-aware vector is obtained, represented as:

[0016]

[0017] in, Indicates the first Context-aware vectors at each time point; This represents the total number of time points in a standard time series dataset. Indicates the first A vector of values ​​at each time point.

[0018] Furthermore, the training of the model's meta-learning framework includes an inner loop and an outer loop.

[0019] The specific process of the inner loop is as follows:

[0020] 1) Construct meta-tasks, including support sets and query sets;

[0021] 2) Based on the support set of the aforementioned meta-task, the initial parameters of the model are adjusted through gradient updates, as follows:

[0022]

[0023] in, Indicates the initial parameters after model adjustment; Indicates the learning rate of the inner loop; Represents the gradient operator; This represents the loss function computed on the support set;

[0024] The outer loop obtains the adjusted initial parameters through the inner loop, and updates the model's initial parameters based on the query set of the same meta-task, as follows:

[0025]

[0026] in, This indicates all meta-tasks Expected value; This represents the loss function calculated on the query set.

[0027] Furthermore, the support set for the meta-task is constructed based on the standard time-series data and input into the trained risk prediction model. After adjustment by the gradient update algorithm in the model's internal loop, personalized model parameters are obtained.

[0028] Based on the personalized model parameters and the standard time-series dataset, the time-series fusion encoder of the risk prediction model generates a context-aware vector through an attention mechanism and inputs it into the prediction layer to obtain the probability value of liquid retention risk. and the corresponding basic water intake range .

[0029] Furthermore, the comprehensive recommended water intake is within the basic water intake range.

[0030] The weights of the objectives are dynamically adjusted based on an adaptive weighted optimization algorithm. These objectives include the probability value of liquid retention risk. Patient comfort index and kidney function safety indicators , is represented as:

[0031]

[0032] in, Indicate target At the moment Dynamic weights; Represents a mapping function;

[0033] Based on the aforementioned dynamic weights, reinforcement learning adjustments are performed, specifically as follows:

[0034] 1) Construct a framework for reinforcement learning regulation mechanisms, including state space, action space, and reward function;

[0035] 2) State Space This includes the probability value of fluid retention risk, patient comfort indicators, and renal function safety indicators;

[0036] 3) Action Space Adjusted water intake;

[0037] 4) Reward Function Represented as:

[0038]

[0039] The comprehensive recommended water intake is finally obtained through the adaptive weighted optimization algorithm and reinforcement learning adjustment mechanism.

[0040] Compared with the prior art, the advantages of the present invention are as follows:

[0041] 1. This invention employs a meta-learning training framework, enabling the model to quickly fine-tune parameters based on a small amount of data from new patients. This solves the data scarcity and cold start problems faced by traditional machine learning models when applied to new individuals, and achieves a rapid leap from a general model to a personalized, dedicated model.

[0042] 2. By introducing a temporal fusion encoder based on an attention mechanism, this invention can dynamically capture the complex interactions and dependencies of multiple physiological parameters in the time dimension, thereby more sensitively identifying early and weak signals of fluid retention and significantly improving the timeliness and accuracy of risk prediction.

[0043] 3. This invention constructs a multi-objective optimization strategy that integrates the risk of fluid retention, patient comfort, and renal function safety. Through adaptive weight allocation and reinforcement learning adjustment mechanisms, it ensures medical safety while taking into account the humanization of treatment and patient compliance, and outputs more personalized water intake recommendations that are more in line with complex clinical realities. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating the dynamic regulation of water intake in patients with chronic heart failure according to the present invention.

[0046] Figure 2 This is a flowchart illustrating the generation process of the probability value of liquid retention risk and the basic water intake range for this invention.

[0047] Figure 3 The following is a flowchart illustrating the comprehensive recommended water intake generation process of this invention. Detailed Implementation

[0048] To achieve the above objectives, this invention provides a method for dynamic regulation of water intake in patients with chronic heart failure based on machine learning. Please refer to [link to relevant documentation]. Figures 1-3 The method includes:

[0049] S1: Collect patients' physiological parameter data and perform time window alignment and resampling to obtain the patients' standard time-series dataset.

[0050] In this step, the core physiological parameters of patients with chronic heart failure are first collected in real time using smart devices or health monitoring systems. The collected physiological parameters include the following key data: the patient's weight, heart rate, urine output, blood pressure, and blood oxygen saturation.

[0051] Patient weight data is used to monitor changes in fluid retention, as weight gain or loss directly affects adjustments to water intake; heart rate data reflects cardiac workload, and changes in heart rate may indicate fluid retention or changes in circulatory system pressure; urine output data reflects the kidneys' ability to excrete fluid, and decreased urine output may indicate fluid retention, requiring adjustments to water intake; blood pressure data is used to assess changes in cardiovascular system pressure, and blood pressure fluctuations are often closely related to fluid balance; blood oxygen saturation data is used to monitor the patient's oxygenation, and low blood oxygen may indicate signs of edema or worsening heart failure. These physiological parameters are collected regularly through wearable devices, medical sensors, or hospital monitoring systems and transmitted to a data processing center for further analysis.

[0052] Considering that the collection frequencies of different physiological parameters may vary, it is necessary to uniformly process the collected data to ensure temporal consistency. Processing methods include time window alignment and resampling.

[0053] First, a unified time window is set; in this embodiment, a 1-hour time window is set. For various physiological parameters, resampling is performed according to the sampling frequency (e.g., weight once daily, heart rate once per minute) to ensure that all data have corresponding values ​​at the same time point. , , , and For different types of physiological parameters, the sampling time steps are as follows: , , , and Various physiological parameter data are transformed into data with a unified time window through interpolation or mean resampling methods. For the corresponding time-series dataset, through time window alignment and resampling, the data points of each physiological parameter will be uniformly aligned to the target time point to ensure data consistency.

[0054] To eliminate inconsistencies between different physiological parameters due to different dimensions, all collected data needs to be standardized. Data standardization ensures that physiological parameters can be compared and analyzed on the same scale.

[0055] After time window alignment, resampling, and standardization, all physiological parameter data were transformed into a standard time-series dataset. .

[0056] All collected data, including physiological parameter data and standard time-series datasets, must be stored in a database and backed up regularly. The database should support efficient querying and real-time processing to ensure data integrity and consistency.

[0057] S2: A risk prediction model is constructed using a temporal fusion encoder trained with a meta-learning framework. The patient's input dataset is input into the risk prediction model to obtain the probability value of fluid retention risk and the corresponding baseline water intake range. The model is configured to adjust parameters based on new standard time-series data using a built-in gradient update algorithm.

[0058] Traditional physiological data processing methods often involve concatenating different physiological parameter data together for processing. However, due to the heterogeneous sources of these data, direct concatenation often fails to fully uncover the deeper relationships between various parameters, leading to the neglect of potential physiological signals.

[0059] In this step, a risk prediction model is constructed using a temporal fusion encoder. First, different types of physiological parameter data are processed separately. Then, cross-attention calculation is performed along the temporal dimension to capture subtle early signs of physiological deterioration.

[0060] First, the model maps standard time-series data for each target time point to corresponding feature representations through an embedding layer. To capture the dynamic dependencies between different time points, the model employs an attention mechanism to calculate the association weights between time points.

[0061] For the first in the time series At each time point, the model generates a query vector. For the first in the time series At each time point, a key vector is generated. and a value vector The calculation of attention weights depends on the similarity between the query vector and the key vector. The similarity function is defined as a dot product operation. Then, the similarity is normalized using the Softmax function to obtain the attention weights, expressed as:

[0062]

[0063] in, This represents the attention weights, with weight values ​​ranging from 1 to 2. The higher the weight, the higher the number of... The time series data at the nth time point is used for the nth The more important the time series data at each point in time; This represents an exponential function used to convert similarity scores into positive numbers and amplify differences, ensuring the non-negativity and distinguishability of the weights; Represents the similarity function; This represents each time point in the standard time series dataset. The exponential similarity is summed, which performs normalization and ensures the probability distribution characteristics of the attention weights.

[0064] The value vectors are weighted and summed based on the calculated attention weights to generate a context-aware vector, represented as:

[0065]

[0066] in, Indicates the first Context-aware vectors at each time point; This represents the total number of time points in the standard time series dataset. The calculation of attention weights enables the model to dynamically focus on key physiological changes in the time series. For example, when weight increases rapidly, the model will automatically increase its attention to the heart rate and blood pressure data in the same period, thereby more accurately capturing the early signs of fluid retention.

[0067] Finally, the value vectors of all time points are weighted and summed according to the attention weights to obtain a new feature representation of each time point after being enhanced by contextual information. These representations are then aggregated into a fused feature vector of the entire standard time series dataset.

[0068] To overcome the challenges of large individual differences and limited labeled data in medical data, the risk prediction model is trained using a meta-learning framework. This framework simulates the learning paradigm, enabling the model to quickly adapt to new patients. By pre-training the model on a large population dataset, it learns how to learn. When encountering new patients, it can be quickly adapted with only a few days' worth of data. This allows for the training of a model with rapid learning capabilities, rather than just training the model itself. It achieves a rapid leap from general to specific models, solving the pain points of individualization and cold start problems.

[0069] During the training phase, the meta-learning framework comprises a two-layer learning process: an inner loop and an outer loop.

[0070] The goal of the inner loop is to enable the model to quickly learn and adapt to a specific meta-task. The process of constructing a meta-task is as follows: from a large historical patient database, a complete set of data for a certain period of time for a patient is randomly extracted. This patient's data is divided into a support set and a query set. The support set simulates the initial small amount of data provided by the new patient, such as the data from the first 3-5 days. The query set simulates the data generated by the patient later to verify the model's adaptation effect, such as the data from the 4th to the 7th day. In this way, each meta-task represents a complete learning scenario, that is, given the initial data of a historical patient (now playing the role of a new patient) (support set), the model is required to accurately predict its subsequent state after learning (query set).

[0071] During the inner loop, the model will start from the initial parameters. Initially, only the support set of this meta-task is used. Performing one or more gradient updates is represented as:

[0072]

[0073] in, This represents the individualized parameters after the model gradient update; Indicates the learning rate of the inner loop; Represents the gradient operator; This represents the loss function calculated on the support set, which combines binary cross-entropy (used for risk prediction) and mean squared error (used for water intake interval regression); the inner loop process is like a student doing an open-book exercise, referring to limited materials, i.e. the support set, to quickly find a solution for this specific question, i.e., individualized parameters;

[0074] The goal of the outer loop is to evaluate and optimize the model's performance after the rapid adaptation of the inner loop. The outer loop does not optimize the model's performance on a specific meta-task, but rather optimizes the model's initial parameters. This allows the model to acquire individualized parameters through a few updates in the inner loop, regardless of the meta-task it faces.

[0075] The outer loop process uses individualized parameters obtained from the inner loop, within the same meta-task query set. The test is performed, and the loss is calculated; then, this loss is backpropagated to update the initial parameters, expressed as:

[0076]

[0077] in, This indicates all meta-tasks The expected value, i.e., the average loss; This represents the loss function computed on the query set; the outer loop optimizes the model's initial parameters across all tasks in this way, enabling the model to quickly and powerfully adapt to any new task, i.e., new patients.

[0078] Through the meta-learning framework, when the model encounters a brand new patient in a real environment, it can quickly complete the inner loop adaptation based on the initial parameters and a small amount of standard time-series data of the patient in the past few days, and immediately provide a highly personalized and accurate prediction model. This perfectly solves the problems of personalization and cold start in machine learning applications in the medical field.

[0079] When new standard time-series data is input into the risk prediction model, an individualized risk prediction model is first obtained through an inner loop. Then, the model's time-series fusion encoder generates context-aware feature representations through an attention mechanism. These features are aggregated and input into the prediction layer, outputting two key clinical indicators: the probability value of fluid retention risk. and the corresponding basic water intake range Liquid retention risk probability value The range of values ​​is The basal water intake range is the safe water intake range based on the individual physiological state of the patient.

[0080] To achieve continuous and accurate management, the risk prediction model supports dynamic fine-tuning. It periodically (e.g., weekly) uses newly generated standard time-series datasets of patients as support sets to initiate gradient update algorithms, allowing the model parameters to adapt to more individual characteristics. This mechanism ensures that the model can track changes in the patient's physiological state and achieve truly individualized and adaptive prediction.

[0081] S3: Based on the fluid retention risk probability value, combined with preset patient comfort indicators and renal function safety indicators, and according to the baseline water intake range, a multi-objective optimization algorithm is executed to finally output a comprehensive recommended water intake.

[0082] This step transforms the predictions of the machine learning model into a comprehensive recommended water intake that is clinically feasible and takes into account multiple objectives. The key is to introduce multi-objective optimization theory to optimize patient comfort and renal function stability while ensuring medical safety, thereby overcoming the limitations of single-objective decision-making.

[0083] The baseline water intake range obtained in step two defines the search space for decision variables in the multi-objective optimization problem. This means the final comprehensive recommended water intake must fall within this range, ensuring an individualized safety benchmark for the recommended value.

[0084] Multi-objective optimization theory considers three parallel optimization metrics that may conflict clinically: the probability value of fluid retention risk, patient comfort metrics, and renal function safety metrics.

[0085] Patient comfort index To maximize patient comfort and treatment adherence, specifically by minimizing the sensation of thirst, a thirst index is introduced based on the patient's daily self-reported thirst scale or indirectly estimated from physiological indicators such as salivary osmolality.

[0086] Kidney function safety indicators To maintain stable renal function and avoid prerenal damage caused by drastic fluctuations or inappropriate restriction of water intake, this indicator is quantified by the estimated change in glomerular filtration rate obtained by monitoring the dynamics of serum creatinine, aiming to minimize short-term fluctuations in renal function.

[0087] In traditional multi-objective optimization algorithms, the weights of the objective function are usually static. However, in fluid management for patients with chronic heart failure, the relative importance of fluid retention risk, comfort, and renal safety changes with physiological state. Therefore, an adaptive weighted optimization algorithm is introduced, allowing the weights in the objective function to be dynamically adjusted based on the patient's real-time physiological data, expressed as:

[0088]

[0089] in, Indicate target At the moment Dynamic weights; This represents a mapping function that dynamically adjusts the weight of each target based on the patient's physiological state. In this embodiment, when the risk of fluid retention increases, the mapping function increases. The value of this value emphasizes the control of the risk of fluid retention; while in cases of impaired renal function, increasing Kidney safety should be given top priority.

[0090] Building upon adaptive weighted optimization, a reinforcement learning regulation mechanism is introduced. Reinforcement learning constructs a Markov decision process (MDP) to treat the regulation of each patient's water intake as a dynamic optimization process.

[0091] The key link between reinforcement learning and adaptive weighted optimization lies in the fact that reinforcement learning can update the strategy based on the current adaptive weights, thereby making fluid intake regulation more in line with the patient's actual needs. Specifically, the reward function in reinforcement learning is closely integrated with the weighted optimization objective, reflecting the comprehensive impact of fluid retention risk, comfort, and renal function safety.

[0092] First, a framework for reinforcement learning regulation mechanisms is constructed, including state space, action space, and reward function.

[0093] state space This includes the probability value of fluid retention risk, patient comfort indicators, and renal function safety indicators.

[0094] Action space For adjusting the water intake, such as the increase in water intake for each adjustment,

[0095] reward function Defined based on the risk of fluid retention, comfort, and renal safety, the goal is to minimize the risk of fluid retention while optimizing patient comfort and renal safety as much as possible, expressed as:

[0096]

[0097] The synergistic effect between reinforcement learning and adaptive weighted optimization enables flexible responses to changes in patients' physiological states. Specifically, reinforcement learning learns through real-time feedback to adjust dynamic decision-making strategies for water intake, while adaptive weighted optimization ensures that reinforcement learning considers the dynamic importance of different objectives in each decision.

[0098] In this embodiment, when the patient has a high risk of fluid retention, the adaptive weighted optimization algorithm automatically increases the weight of the fluid retention risk. The reinforcement learning agent then makes decisions based on the new weights, tending to reduce water intake. When the patient's kidney function is abnormal, the adaptive weighted optimization algorithm increases the weight of kidney function safety, and the reinforcement learning agent prioritizes kidney function safety, thereby avoiding excessive water intake that could burden the kidneys.

[0099] Through this synergistic effect, the risk of fluid retention and renal function safety are optimized to the greatest extent possible without sacrificing patient comfort.

[0100] Finally, through the aforementioned adaptive weighted optimization and reinforcement learning adjustment mechanism, a comprehensive recommended water intake is obtained, which is a personalized water intake recommendation value. This value not only considers the minimization of the risk of fluid retention, but also takes into account the patient's comfort and renal function safety, and can provide the most appropriate water adjustment suggestions in the dynamic changes of the patient's physiological state.

[0101] The proposed method for dynamic regulation of water intake in patients with chronic heart failure based on machine learning constructs a complete individualized closed-loop management approach by integrating a meta-learning framework, a temporal fusion encoder, and a multi-objective collaborative optimization mechanism. This method first performs temporal alignment and standardization on multi-source physiological parameters, then utilizes an attention-based temporal fusion encoder to mine the dynamic dependencies between physiological parameters, and then uses meta-learning to achieve rapid individualized model adaptation, accurately predicting the risk of fluid retention and the baseline water intake range. Finally, it combines adaptive weighted optimization and reinforcement learning adjustment mechanisms to generate a comprehensive recommended water intake under multiple clinical objective constraints.

[0102] In summary, this invention effectively solves the core challenges in managing water intake in patients with chronic heart failure through technological innovation, such as individualized adaptation, temporal feature mining, and multi-objective decision optimization. It achieves a leap from static general rules to dynamic individualized regulation, providing an intelligent, precise, and clinically practical solution for the long-term management of patients with chronic heart failure.

[0103] 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 machine learning-based method for dynamic regulation of water intake in patients with chronic heart failure, characterized in that, The method includes: S1: Collect patients' physiological parameter data and perform time window alignment and resampling to obtain the patients' standard time series dataset; S2: A risk prediction model is constructed using a temporal fusion encoder trained with a meta-learning framework. The patient's input dataset is input into the risk prediction model to obtain the probability value of fluid retention risk and the corresponding baseline water intake range. The model is configured to adjust the initial parameters based on new standard time-series data using a built-in gradient update algorithm. S3: Based on the probability value of fluid retention risk, combined with patient comfort index and renal function safety index, and according to the baseline water intake range, perform multi-objective optimization, and finally output a comprehensive recommended water intake.

2. The method for dynamic regulation of water intake in patients with chronic heart failure based on machine learning according to claim 1, characterized in that, The physiological parameters include weight, heart rate, urine output, blood pressure, and blood oxygen saturation. The time window alignment and resampling are based on a unified time window, which aligns the physiological parameter data to the target time point, and then the standard time series dataset is obtained through standardization processing.

3. The method for dynamic regulation of water intake in patients with chronic heart failure based on machine learning according to claim 2, characterized in that, The risk prediction model uses an attention mechanism to calculate the weights between standard time-series data at each target time point, expressed as: in, Indicates attention weight; Represents an exponential function; Represents the similarity function; Represents any time point in a standard time series dataset Sum the exponential similarities; Indicates the first Query vectors at each time point; Indicates the first Key vectors at each time point; Represents any time point in a standard time series dataset The key vector; Based on the attention weights, a context-aware vector is obtained, represented as: in, Indicates the first Context-aware vectors at each time point; This represents the total number of time points in a standard time series dataset. Indicates the first A vector of values ​​at each time point.

4. The method for dynamic regulation of water intake in patients with chronic heart failure based on machine learning according to claim 2, characterized in that, The training of the model's meta-learning framework includes an inner loop and an outer loop. The specific process of the inner loop is as follows: 1) Construct meta-tasks, including support sets and query sets; 2) Based on the support set of the aforementioned meta-task, the initial parameters of the model are adjusted through gradient updates, as follows: in, Indicates the initial parameters after model adjustment; Indicates the learning rate of the inner loop; Represents the gradient operator; This represents the loss function computed on the support set; The outer loop obtains the adjusted initial parameters through the inner loop, and updates the model's initial parameters based on the query set of the same meta-task, as follows: in, This indicates all meta-tasks Expected value; This represents the loss function calculated on the query set.

5. The method for dynamic regulation of water intake in patients with chronic heart failure based on machine learning according to claim 4, characterized in that, The support set for the meta-task is constructed based on the standard time series data and input into the trained risk prediction model. After adjustment by the gradient update algorithm in the model's internal loop, personalized model parameters are obtained. Based on the personalized model parameters and the standard time-series dataset, the time-series fusion encoder of the risk prediction model generates a context-aware vector through an attention mechanism and inputs it into the prediction layer to obtain the probability value of liquid retention risk. and the corresponding basic water intake range .

6. The method for dynamic regulation of water intake in patients with chronic heart failure based on machine learning according to claim 5, characterized in that, The comprehensive recommended water intake is within the basic water intake range. The weights of the objectives are dynamically adjusted based on an adaptive weighted optimization algorithm. These objectives include the probability value of liquid retention risk. Patient comfort index and kidney function safety indicators , is represented as: in, Indicate target At the moment Dynamic weights; Represents a mapping function; Based on the aforementioned dynamic weights, reinforcement learning adjustments are performed, specifically as follows: 1) Construct a framework for reinforcement learning regulation mechanisms, including state space, action space, and reward function; 2) State Space This includes the probability value of fluid retention risk, patient comfort indicators, and renal function safety indicators; 3) Action Space Adjusted water intake; 4) Reward Function Represented as: The comprehensive recommended water intake is finally obtained through the adaptive weighted optimization algorithm and reinforcement learning adjustment mechanism.

Images