Precision medicine system for hemodialysis patients based on multi-omics and big data

Abstract

The application relates to the fields of medical care informatics and precision medical technology, and discloses a precision medication system for hemodialysis patients based on multiple omics and big data, which comprises the following modules: a multi-source data acquisition module, which acquires multi-source heterogeneous data of patients and performs standardization and integration processing; a drug metabolism analysis module, which analyzes drug metabolism pathways and individual differences through multi-omics data, constructs a drug metabolism polymorphism model, and outputs gene metabolism capacity classification and biomarker risk labels; an initial medication decision module, which is used for constructing an initial medication decision model based on a federal optimization framework of momentum enhancement and robust aggregation mechanism, and outputting a personalized initial medication scheme for erythropoietin and intravenous iron agents; and a drug dose adjustment module, which is used for dynamically optimizing time nodes and dose adjustment opportunities of drug administration through an asynchronous federal reinforcement learning framework combined with a forward-looking parameter correction strategy. The application constructs an intelligent medication system with strong robustness and fine decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of healthcare informatics and precision medicine, and particularly relates to a precision medication system for hemodialysis patients based on multiomics and big data. BACKGROUND

[0002] Hemodialysis is the most common replacement therapy for patients with end-stage renal disease (ESRD), and its core goal is not only to remove toxins and regulate water and electrolyte balance, but also to correct anemia and mineral metabolism disorders. Erythropoietin (EPO) and intravenous iron are important drugs for correcting renal anemia, and there is significant individual variability in their dosing and frequency, which needs to be dynamically adjusted to avoid exacerbation of anemia or iron overload.

[0003] However, current medication regimens for dialysis patients rely heavily on human experience or fixed rules, ignoring individual differences in gene expression, metabolic capacity, inflammation levels, and drug responsiveness among patients, leading to fluctuating efficacy and high complication risk, which can result in drug waste, poor efficacy (such as inadequate correction of anemia), and increased risk of drug-related adverse reactions (such as hypertension, thrombosis, and iron overload-related organ damage). At the same time, due to the privacy, multi-source heterogeneity, and distributed characteristics of medical data, traditional centralized models are difficult to implement large-scale, high-quality patient data modeling.

[0004] Therefore, there is an urgent need to introduce a multiomics and big data-driven intelligent medication strategy model to build a robust and finely-tuned intelligent medication system while ensuring data privacy. SUMMARY

[0005] The present application aims to solve the problems of "one-size-fits-all" solutions, model fragility, and strategy incompatibility in current dialysis patient medication decisions, and proposes a precision medication regimen that integrates multiomics analysis, big data learning, and artificial intelligence decision optimization. Specifically, a precision medication system for hemodialysis patients based on multiomics and big data is provided.

[0006] To achieve the above-mentioned purpose, the following technical solutions are implemented:

[0007] A precision medication system for hemodialysis patients based on multiomics and big data, comprising:

[0008] A multi-source data acquisition module for acquiring multi-source heterogeneous data of patients and performing standardization and integration processing to build a patient-level data warehouse; wherein the multi-source heterogeneous data includes basic clinical data, multiomics detection data, and real-time physiological and laboratory indicators;

[0009] A drug metabolism analysis module is configured to analyze drug metabolism pathways and individual differences through multi-omics data, construct a drug metabolism polymorphism model for predicting patient response to drugs, and output gene metabolism capacity classification and biomarker risk labels.

[0010] The drug metabolism analysis module performs operations including:

[0011] Based on the patient's gene expression profile, metabolite concentration, and proteomic information, the in vivo metabolism pathway of the key drug is analyzed to generate individualized pharmacokinetic parameters, including drug half-life and peak concentration.

[0012] Through drug metabolism-related gene variation analysis, biomarkers are identified and metabolic capacity classification indicators are established.

[0013] The metabolic capacity classification indicators and biomarker risk labels are integrated into the feature space of federated learning to constrain the dose recommendation range of the initial drug use model.

[0014] An initial drug use decision module is configured to construct an initial drug use decision model based on a federated optimization framework with momentum enhancement and robust aggregation mechanism, and output personalized initial drug use schemes for erythropoietin and intravenous iron agents.

[0015] The construction of the initial drug use decision model includes:

[0016] An improved robust aggregation rule is constructed, and three types of aggregation methods are used to aggregate and then weightedly fuse the local model gradients, thereby constructing an improved robust aggregation function, including:

[0017] The three types of aggregation results obtained by geometric center aggregation, coordinate pruning mean aggregation, and norm filtering aggregation are linearly combined according to the preset weights to obtain the global update direction, and the improved robust aggregation function is obtained :

[0018]

[0019] wherein, : the final robust aggregation function output, representing the global drug strategy update direction constructed in the current round, used to adjust the model parameters; : the local gradient vector of the node ; n: the total number of participating nodes, i.e., the number of medical institutions participating in federated learning; : the weighting coefficient of geometric center aggregation ; : the weighting coefficient of coordinate pruning mean aggregation ; : the weighting coefficient of norm filtering aggregation ;

[0020] Based on the improved robust aggregation function, momentum enhancement and two-step bucket aggregation strategy are adopted to smooth the gradient update, and the nearest neighbor hybrid mechanism is combined to defend against abnormal nodes.

[0021] According to the patient characteristics of each medical node, the historical momentum gradient and the current gradient of the trusted neighbor node, the improved robust aggregation function is used to filter abnormal updates, and the parameter update amplitude is adjusted according to the dynamically decaying learning rate. Finally, the global drug parameter vector is output, including drug category recommendation weight and dose adjustment coefficient. The drug dose adjustment module is used to dynamically optimize the time node and dose adjustment opportunity of drug administration through the asynchronous federated reinforcement learning framework combined with the forward-looking parameter correction strategy.

[0022] Further, the operations performed by the drug metabolism analysis module include:

[0023] Based on the patient's gene expression profile, metabolite concentration and proteomic information, the in vivo metabolism path of the key drug is analyzed, and individualized pharmacokinetic parameters are generated, including drug half-life and peak concentration.

[0024] Through drug metabolism related gene variation analysis, biomarkers are identified and metabolic capacity grading indicators are established.

[0025] The metabolic capacity grading indicators and biomarker risk labels are integrated into the feature space of federated learning to constrain the dose recommendation range of the initial drug administration model.

[0026] Further, the construction of the initial drug administration decision model includes:

[0027] An improved robust aggregation rule is constructed, which aggregates the local model gradient based on three types of aggregation methods and then weighted fusion, thereby constructing an improved robust aggregation function.

[0028] Based on the improved robust aggregation function, momentum enhancement and two-step bucket aggregation strategy are adopted to smooth the gradient update, and the nearest neighbor hybrid mechanism is combined to defend against abnormal nodes.

[0029] According to the patient characteristics of each medical node, the historical momentum gradient and the current gradient of the trusted neighbor node, the improved robust aggregation function is used to filter abnormal updates, and the parameter update amplitude is adjusted according to the dynamically decaying learning rate. Finally, the global drug parameter vector is output, including drug category recommendation weight and dose adjustment coefficient.

[0030] Further, the construction of the improved robust aggregation rule based on the three types of aggregation methods to aggregate the local model gradient and then weighted fusion to construct the improved robust aggregation function includes:

[0031] Geometric center aggregation: select the gradient vector with the minimum distance to the majority node's drug gradient;

[0032] Coordinate pruning mean aggregation: independently sort and prune each coordinate by coordinate pruning mean, and take the average after pruning the extreme value;

[0033] Norm filtering aggregation: calculate the L2 norm of the gradient vector uploaded by each node, sort by size, and aggregate after pruning the node with the largest gradient change amplitude;

[0034] The above three types of aggregation results, i.e., geometric center aggregation, coordinate pruning mean aggregation, and norm filtering aggregation, are linearly combined according to a preset weight to obtain an improved robust aggregation function.

[0035] Further, the federated optimization framework based on momentum enhancement and robust aggregation mechanism comprises:

[0036] The Polyak momentum mechanism is introduced to suppress local fluctuations by weighted average of historical gradients;

[0037] A two-step bucket aggregation strategy is adopted to smooth the gradient update, and a nearest neighbor mixing mechanism is combined to defend against abnormal nodes; wherein the two-step bucket aggregation strategy comprises: first, randomly grouping nodes to calculate the group consensus, and then performing robust aggregation on the inter-group results.

[0038] Further, the two-step bucket aggregation strategy is adopted to smooth the gradient update, and the nearest neighbor mixing mechanism is combined to defend against abnormal nodes, comprising:

[0039] The medical nodes are grouped, and the medical nodes are randomly divided into multiple buckets, each bucket containing a number of nodes;

[0040] The average drug gradient of each node in each bucket is calculated to form a bucket consensus vector;

[0041] Robust aggregation is performed on the bucket consensus vector to realize inter-group aggregation and generate a global update direction;

[0042] Based on the nearest neighbor mixing mechanism, neighbors are selected, and for each node, a plurality of neighbor nodes with the most similar feature space are selected;

[0043] The drug gradients of the neighbor nodes are aggregated to form a final update strategy.

[0044] Further, the initial drug decision model also adopts a decentralized robust mechanism, specifically comprising:

[0045] In the edge medical computing scenario, a decentralized optimization strategy is adopted, gradient information is exchanged between nodes and neighbor nodes based on an asynchronous communication mechanism, a time stamp verification is introduced to filter out gradient data that is not updated in time, and the Byzantine fault tolerance capability is maintained in the sparse communication topology, allowing no more than 1 / 3 of the nodes to fail.

[0046] Further, the initial medication decision module adjusts the global parameters based on the momentum gradient mean of the trusted neighbor nodes, and generates a personalized medication parameter vector to guide the initial dose and type selection of erythropoietin and iron agents.

[0047] The initial medication decision module adjusts the global parameters based on the momentum gradient mean of the trusted neighbor nodes, and generates a personalized medication parameter vector, specifically including:

[0048] For each medical node, aggregate the historical momentum gradient and current gradient of its trusted neighbor nodes;

[0049] Process the aggregation results of all nodes through a weighted robust aggregation function;

[0050] Dynamically adjust the learning rate to generate an updated medication parameter vector, which contains drug category recommendation weights or dose coefficients.

[0051] Further, the drug dose adjustment module uses the dose adjustment rules based on real-time hemoglobin (Hb), ferritin (SF) and transferrin saturation (TSAT) indicators as clinical prior knowledge to construct the safety constraint boundary of the reinforcement learning reward function; wherein the dose adjustment rules are:

[0052] When hemoglobin is lower than 100g / L and ferritin is lower than 200μg / L, and transferrin saturation is lower than 20%, start intravenous iron supplementation and give erythropoietin 50-150U / kg / week;

[0053] When hemoglobin is not less than 130g / L and ferritin is not less than 300μg / L, and transferrin saturation is not less than 30%, suspend erythropoietin and iron therapy;

[0054] Combine blood drug concentration and electrolyte level to correct the dose.

[0055] Further, the drug dose adjustment module dynamically optimizes the time node and dose adjustment opportunity of drug administration through an asynchronous federated reinforcement learning framework combined with a forward-looking parameter correction strategy, including:

[0056] Generate a forward-looking correction parameter based on the historical trend of strategy parameter changes and the degree of delay, dynamically adjust the compensation intensity through an adaptive factor, and ensure the consistency of clinical efficacy in an asynchronous environment;

[0057] According to the probability derivative of the administration action in the trajectory and the reward discount, the local policy gradient is calculated, the server updates the global parameters after normalizing and momentum fusing the received delayed gradient, and the updated administration policy is dynamically issued to each node without waiting for all nodes to synchronize.

[0058] Compared with the prior art, the application has the following beneficial effects:

[0059] 1. The improved decentralized Byzantine robust learning method (including robust aggregation function F, momentum enhancement, two-step bucket aggregation, and nearest neighbor hybrid mechanism) is proposed, a decentralized federated learning framework is constructed on the basis of ensuring the privacy of patient data of each medical node, stable fusion of cross-node model parameters is realized through a robust optimization strategy, the nonlinear relationship between patient individual clinical characteristics and EPO / iron agent dosage is effectively mined, and the individuality and safety of initial drug recommendation are improved.

[0060] 2. The improved robust aggregation rule is proposed, the multi-robust aggregation mechanism integrating geometric center aggregation, coordinate pruning mean (CwTM) and comparative gradient elimination (CGE) is constructed for abnormal nodes or false data that may exist in medical data, potential malicious or distorted information is filtered out from multiple dimensions, and the fault tolerance and generalization ability of the model in the real medical environment are significantly enhanced.

[0061] 3. The application proposes a federated optimization framework based on momentum enhancement and robust aggregation mechanism, introduces the Polyak momentum mechanism to smooth local gradient oscillation, combines dynamic sampling strategy and two-step aggregation structure (bucketing + secondary aggregation), effectively deals with the challenges of sampling differences between dialysis centers, uneven computing power and delayed updates, and realizes stable and fast model convergence.

[0062] 4. The application proposes a two-step aggregation and nearest neighbor hybrid strategy (NNM), which dilutes local abnormal nodes by first performing intra-group aggregation, and further improves the robustness and distribution adaptation ability of the system in complex communication networks or data heterogeneous scenarios by eliminating isolated abnormal values through nearest neighbor geometric mixing.

[0063] 5. The application proposes an intelligent optimization framework based on asynchronous federated reinforcement learning (AFedPG), constructs an asynchronous optimization mechanism driven by a policy gradient, realizes non-synchronous model updating between different dialysis nodes, automatically optimizes the drug administration time node and adjustment timing, and is especially suitable for strategy modeling tasks considering EPO half-life, dialysis frequency, Hb dynamic response and other time sequence characteristics.

[0064] 6、The application proposes a forward-looking parameter correction strategy, aiming at the strategy lag problem caused by node upload delay, designs a delay adaptive parameter prediction mechanism to compensate for the historical deviation in parameter update, avoid strategy update instability or misjudgment, and realize high robustness strategy control in asynchronous environment.

[0065] It should be understood that the content described in the summary section is not intended to limit or important features of the embodiments of the application, nor to limit the scope of the application. Other features of the application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0066] The above and other features, advantages and aspects of embodiments of the present application will become more apparent upon reading the following detailed description in conjunction with the accompanying drawings, in which:

[0067] Figure 1 is a module schematic diagram of the precision medication system for hemodialysis patients based on multi-omics and big data of the embodiments of the application;

[0068] Figure 2 is an architecture schematic diagram of the precision medication system for hemodialysis patients based on multi-omics and big data of the embodiments of the application. DETAILED DESCRIPTION

[0069] In order to make the objects, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described clearly and completely below in conjunction with the drawings of the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0070] In addition, the term "and / or" in this paper is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents an "or" relationship between the front and rear associated objects.

[0071] Figure 1 is a module schematic diagram of the precision medication system for hemodialysis patients based on multi-omics and big data of the embodiments of the application; Figure 2 is an architecture schematic diagram of the precision medication system for hemodialysis patients based on multi-omics and big data of the embodiments of the application. As Figure 1 and Figure 2As shown, a precision medication system for hemodialysis patients based on multi-omics and big data 100 includes:

[0072] A multi-source data acquisition module 110 is configured to acquire multi-source heterogeneous data of a patient and perform standardization and integration processing, and to construct a patient-level data warehouse; wherein the multi-source heterogeneous data includes basic clinical data, multi-omics detection data, and real-time physiological and test indicators; the real-time physiological and test indicators include but are not limited to blood drug concentration (such as EPO, iron agent), electrolyte level (potassium, calcium, phosphorus, etc.), inflammatory factor (such as C-reactive protein, interleukin-6, etc.), hemoglobin (Hb), ferritin (SF), transferrin saturation (TSAT), and other dynamic monitoring indicators.

[0073] 1. Multi-dimensional clinical data acquisition:

[0074] Basic clinical data: collect patient age, weight, underlying diseases (such as diabetes, hypertension, etc.), dialysis parameters (dialysis frequency, time, mode, etc.), and medication habits (previous medication history, dosage, frequency, etc.).

[0075] Multi-omics detection data: obtain patient gene expression data through gene sequencing, detect metabolite levels (such as amino acids, fatty acids, small molecule metabolites, etc.), proteomics data (such as inflammatory factors, growth factors, etc.), and analyze drug metabolism-related gene polymorphisms (such as cytochrome P450 enzyme genes).

[0076] Real-time physiological and test indicators: dynamically monitor blood drug concentration (such as erythropoietin EPO, iron agent, etc.), electrolyte level (potassium, calcium, phosphorus, etc.), inflammatory factor (C-reactive protein, interleukin-6, etc.), and test indicators such as hemoglobin (Hb), ferritin (SF), and transferrin saturation (TSAT).

[0077] 2. Data standardization and integration:

[0078] (1) Standardization processing of multi-source heterogeneous data:

[0079] With the help of big data processing technology, pre-process the clinical information, multi-omics detection results, and real-time monitoring data from different systems and devices, including missing value filling, outlier removal, redundant data elimination, and time series alignment. Through unified variable naming conventions, dimension units (such as Hb concentration g / L, TSAT percentage, etc.), and coding systems, solve the inconsistency of different hospital information systems in data format and acquisition frequency, and ensure the consistency and comparability of data input for subsequent model training.

[0080] (2) Construct a patient-level data warehouse to realize deep data fusion:

[0081] Based on the structured data model, a unified data warehouse architecture is constructed to deeply integrate basic clinical data (such as age, gender, and complications), multi-omics data (gene expression, proteome, and metabolome), and dynamic monitoring indicators (EPO concentration, electrolyte changes, and inflammatory factors). By establishing a unique patient identifier, data silos are broken down, forming a real-time updateable and long-term traceable individualized data archive, providing a solid data foundation for the accurate prediction and continuous optimization of subsequent medication models.

[0082] The drug metabolism analysis module 120 is used to analyze drug metabolism pathways and individual differences through multi-omics data, construct a drug metabolism polymorphism model for predicting patient response to drugs, output gene metabolism ability classification, and biomarker risk labels;

[0083] The drug metabolism analysis module 120 is based on multi-omics drug metabolism analysis, aiming to provide medical prior knowledge for the model and constrain the dose recommendation range.

[0084] 1. Drug metabolism pathway modeling and individual difference prediction:

[0085] By combining patient gene expression profiles, metabolite concentrations, and proteome information, the in vivo metabolism pathways of key drugs (such as erythropoietin EPO and different types of iron agents) are analyzed to generate individualized pharmacokinetic parameters, including drug half-life and peak concentration. For example, the decomposition metabolism mechanism of EPO in the liver and the transport process of iron agents in the intestinal absorption and storage are analyzed, and by simulating the individual's in vivo drug concentration change process after administration, the drug efficacy duration and response intensity are evaluated, and the metabolic differences of different patients under the same dose are predicted.

[0086] 2. Establishing a drug metabolism polymorphism model to identify key biomarkers:

[0087] Based on the analysis of variations in drug metabolism-related genes such as CYP450 enzymes and transport proteins (such as TFR and FPN), a polymorphism model is constructed to predict patient response to drugs. By combining metabolite levels and clinical response indicators (such as Hb change speed and ferritin storage capacity), biomarkers with significant effects on drug efficacy (such as TSAT below 20% indicating potential iron utilization disorders and SF above 500 ng / mL indicating iron overload risk) are screened, and metabolic ability classification indicators (fast, medium, and slow metabolizers) are established to provide scientific basis for initial medication dose recommendation and individualized adjustment strategies.

[0088] The final metabolic ability classification indicators and biomarker risk labels are integrated into the feature space of federated learning to constrain the dose recommendation range of the initial medication decision module 130.

[0089] The initial drug decision module 130 is used to build an initial drug decision model based on a momentum-enhanced and robust aggregation mechanism federal optimization framework, and output a personalized initial drug regimen including erythropoietin and intravenous iron agent;

[0090] In order to realize the accurate initial recommendation of erythropoietin (EPO), iron agent and other drugs, the initial drug decision module 130 constructs a big data driven initial drug decision model based on the clinical data of nearly a thousand dialysis patients and multiple omics characteristics. The model uses an improved decentralized Byzantine robust learning method to mine the nonlinear mapping relationship between individual characteristics and drug dosage under the premise of protecting data privacy, and provides a basis for subsequent dynamic drug regulation (drug type, dosage, and administration time).

[0091] Further, the construction of the initial drug decision model includes:

[0092] An improved robust aggregation rule is constructed, and three types of aggregation methods are used to aggregate and then weightedly fuse the local model gradients, thereby constructing an improved robust aggregation function;

[0093] Based on the improved robust aggregation function, a momentum-enhanced and two-step bucket aggregation strategy is used to smooth the gradient update, and a nearest neighbor hybrid mechanism is used to defend against abnormal nodes;

[0094] According to the patient characteristics of each medical node, the historical momentum gradient and the current gradient of the trusted neighbor node, the improved robust aggregation function is used to filter abnormal updates, and the parameter update amplitude is adjusted according to the dynamically decaying learning rate, and finally the global drug parameter vector is output, including drug type recommendation weight and dosage adjustment coefficient.

[0095] Further, the construction of the initial drug decision model includes:

[0096] S1, an improved robust aggregation rule is constructed, and three types of aggregation methods are used to aggregate and then weightedly fuse the local model gradients, thereby constructing an improved robust aggregation function;

[0097] S1.1 Geometric center aggregation: select the gradient vector with the smallest distance to the majority node drug gradient;

[0098] This method is used to select the "most trusted" local model gradient (such as the drug adjustment direction calculated by a hospital according to the patient's Hb, TSAT, SF and other characteristics) in multiple medical nodes to resist the upload of incorrect gradients by abnormal nodes.

[0099] Select the gradient with the smallest distance to the nearest neighbor node (Krum), resist outlying gradients through local distance measurement, suitable for scenarios where node communication is complete, and the collapse point can reach 1 / 2. The Krum selection rule is:

[0100]

[0101] where, : the local gradient vector of node ; n: the total number of participating nodes, i.e., the number of medical institutions participating in federated learning; f: the maximum tolerable number of "abnormal or malicious" nodes, indicating the number of nodes allowed to upload abnormal data (such as abnormally high or low drug recommendations); : the local gradient vector of node ; and the Euclidean distance square between the local gradients of node , indicating the degree of difference in drug adjustment recommendations in the parameter space. : the nearest neighbors of node , indicating the most trusted set of nodes closest to in distance among all nodes.

[0102] By selecting gradients close to the majority of nodes, outliers caused by poor data quality, extreme samples, or communication errors are excluded, thereby improving the stability of the initial drug model (such as the EPO dose mapping function).

[0103] S1.2 Coordinate pruning mean aggregation: sort and prune each coordinate independently using coordinate pruning mean, and take the average after removing extreme values;

[0104] This method is used to handle high-dimensional gradient scenarios (such as EPO dose adjustment factors, iron absorption sensitivity, etc.), and extreme values are removed coordinate by coordinate to improve robustness.

[0105] Sort and prune extreme values for each coordinate independently using coordinate pruning mean (CwTM) , and take the average of the middle values, suitable for high-dimensional gradient scenarios where each dimension is independent. The aggregation rule is as follows:

[0106]

[0107] where, : the value of the kth coordinate in the local gradient vector calculated by the ith node; k: the dimension index of the gradient vector, indicating a specific medical feature dimension (such as the weight of a group of features affecting EPO metabolism); : the index of the node ranked i-th on the k-th coordinate after sorting, i.e., the i-th intermediate value position after sorting the k-th parameter; : the number of intermediate nodes to be retained, removing the first maximum values and the last minimum values to obtain the number of trusted nodes for averaging.

[0108] By sorting each model parameter dimension separately, removing the minimum and maximum extreme values, and then calculating the average, the specific dimension is prevented from being disturbed by extreme values. It is suitable for situations where there is no strong correlation between dimensions, such as input dimensions containing multiple omics, especially when each parameter dimension in the model represents different biological indicators or individual characteristics (such as CYP450 expression, liver function index, electrolyte abnormality marker) affecting medication decision-making. It effectively removes extreme values caused by local measurement errors or individual patient abnormal behavior.

[0109] S1.3 Norm filtering aggregation: Calculate the L2 norm of the gradient vector uploaded by each node, sort by size, remove the node with the largest gradient change amplitude, and then aggregate;

[0110] This method prevents extreme local medication strategy updates from disrupting the global model by removing the node with the largest gradient change amplitude. It is suitable for filtering large deviation model updates caused by incorrect omics measurements or data abnormalities. The aggregation process is as follows:

[0111] 1. First, calculate the L2 norm of the gradient vector uploaded by each node, and sort by size:

[0112]

[0113] : Local gradient L2 norm of the gradient vector uploaded by each node, indicating the change amplitude of the strategy of that center. The larger the value, the greater the deviation of its medication recommendation from the current global strategy. For example, a center may give a patient with an abnormally high level of secondary symptoms an overdose of EPO strategy, which may cause the model update direction to deviate seriously from the mean. In this case, its norm is usually large.

[0114] 2. Then take the gradient of the nodes with the smallest norm, and calculate their average:

[0115]

[0116] S1.4 Weighted fusion: Linearly combine the three types of aggregation results obtained by the above geometric center aggregation, coordinate pruning mean aggregation, and norm filtering aggregation (Krum, CwTM, CGE) according to the preset weight to obtain the global update direction, and obtain the improved robust aggregation function :

[0117]

[0118] where, : The final robust aggregation function output represents the global medication strategy update direction constructed in the current round, which is used to adjust the model parameters (such as the weights in the EPO dose recommendation function). Its input is the vector set to be aggregated ; Geometric center aggregation The weighting coefficients are such that a higher weight indicates overall consistency in the direction of the system's trust gradient. : Coordinate trimming mean The weighting coefficients are suitable for fine processing of high-dimensional features (such as multi-omics data); Norm filtering aggregation The weighting coefficients are such that a higher weight indicates that the system pays more attention to eliminating "aggressive nodes" with large gradient fluctuations. : Ensure that the combination of the three aggregation methods (Krum, CwTM, CGE) is a normalized linear fusion. , , The weights can be set based on experience or dynamically adjusted based on historical aggregation effects, or set in the early stages of model training and fine-tuned on the validation set. Preferably, the three weights are set to fixed empirical values ​​[0.4, 0.4, 0.2].

[0119] S2. Based on the improved robust aggregation function, momentum enhancement and two-step bucket aggregation strategy are used to smooth gradient updates, and the nearest neighbor hybrid mechanism is combined to defend against abnormal nodes.

[0120] In the distributed optimization process of building the initial medication model, it is necessary to address the problems of heterogeneity of multi-center medical data, instability of node communication, and interference from abnormal patient samples in local models. Therefore, this paper proposes a federated optimization framework based on momentum enhancement and robust aggregation mechanisms, combining dynamic sampling, a two-layer aggregation structure, and a decentralized robust mechanism to improve model stability and accuracy.

[0121] S2.1 Constructing a Momentum Enhancement and Robust Aggregation Mechanism

[0122] In real-world medical scenarios, due to differences in the quality of patient data collected at each node (such as different hospitals), local stochastic gradients... This will exhibit high variance and instability. To enhance model robustness, the Polyak momentum method is introduced to perform weighted smoothing of local updates, suppressing local fluctuations through weighted averaging of historical gradients.

[0123]

[0124] : The t-th momentum vector of node i, used to accumulate historical directions and alleviate gradient oscillations; The momentum decay coefficient (typically β≈0.9) controls the degree to which historical information is retained. : is a node iThe local stochastic gradient calculated from the current mini-batch samples is specifically the local stochastic gradient calculated by client i in round t, which reflects the current feedback of local patient data to the model.

[0125] By using the momentum mechanism, not only is the noise of local fluctuating gradients suppressed, but the convergence performance under non-convex loss structures is also improved.

[0126] An improved robust aggregation function is used during global model updates. Extracting reliable information from multiple client momentum:

[0127]

[0128] : Current global model parameters; Improved robust aggregation functions See step S1 for details; : Learning rate (0.001~0.01, decay).

[0129] S2.2 Constructing a Two-Step Aggregation and Nearest Neighbor Hybrid Strategy

[0130] When dealing with a large number of nodes (e.g., in multi-center hospitals), a two-layer aggregation strategy is introduced to further defend against malicious local updates. This strategy involves a two-step bucketing aggregation approach to smooth gradient updates and a nearest neighbor hybrid mechanism to defend against anomalous nodes. The two-step bucketing aggregation strategy first randomly groups nodes to calculate intra-group consensus, and then performs robust aggregation on inter-group results. Specifically:

[0131] First, the medical nodes are grouped using bucketing, randomly dividing the n clients into S buckets (Bucket(1), ..., Bucket(S)), each bucket containing approximately Each node (client);

[0132] Then, for each bucket s By aggregating within groups and averaging within buckets, the average medication gradient (local mean) of each node within a bucket is calculated, forming the bucket consensus vector:

[0133]

[0134] The local gradient vector uploaded by node i represents the medication adjustment strategy calculated by the medical institution based on the omics data of local dialysis patients (such as CYP450 expression, electrolyte levels, EPO responsiveness, etc.). : No. s The set of nodes contained in each bucket; : No. sThe average result of all node local strategies in a bucket, used to represent the consensus opinion of the group on the current medication strategy; b: the number of nodes in each group, the calculation formula is , which represents the number of dialysis centers included in each group; n: the total number of nodes, i.e. the number of all dialysis medical institutions participating in modeling; S: the number of buckets, dividing the n nodes into S buckets, used to prevent a small number of abnormal nodes from interfering with the overall update result; such as grouping by hospital region, grouping by equipment supplier type, or stratifying and bucketing by patient characteristics, etc.

[0135] represents that the client i prepares to upload the information participating in aggregation in the tth round. In the basic framework, , in the first step of the two-step bucketing, the group computing bucket consensus vector .

[0136] Finally, the S bucket consensus vector is applied to the robust aggregation function F to perform robust aggregation on the bucket consensus vector to realize inter-group aggregation (i.e. secondary aggregation, re-aggregation of the bucket mean), and the final global aggregation result is obtained:

[0137]

[0138] Among them, : improved robust aggregation function, see step S1.4 for details, which further aggregates the "bucket mean " of each bucket to ensure that the final output strategy can resist the influence of extreme values between buckets. In inter-group aggregation, the Krum, CwTM, CGE operations inside are based on the vector , and the dimension of is the same as the dimension, when the input of is (bucket consensus vector), n is replaced by S (bucket number).

[0139] The two-step bucketing strategy aims to dilute the influence of potential abnormal nodes through the first layer (in-bucket) aggregation, and further enhance the defense capability against malicious bucket consensus or calculation errors through the second layer (inter-bucket) robust aggregation F. The two-step aggregation strategy above dilutes the influence of malicious nodes through grouping, and is suitable for scenarios with a large number of nodes. The design of itself is universal, regardless of whether the input is a node vector or a bucket consensus vector, the calculation logic of the three methods inside is unchanged (sorting, pruning, filtering, etc. operations based on the input vector set).

[0140] Nearest-neighbor mixing (NNM) removes abnormal gradients by a "geometric proximity" measure and constructs a more robust local view aggregation. The neighbors are filtered by the nearest-neighbor mixing mechanism, and multiple neighbor nodes that are most similar in the feature space are selected for each node; the drug gradients of the neighbor nodes are aggregated to form the final update strategy. The feature space refers to the client local model parameter vector space or the gradient / momentum vector space uploaded by the client. Specifically:

[0141] On each node i, based on the Euclidean distance (or cosine similarity) of the uploaded gradient vector of node i and the gradient vectors of other nodes, the k' closest neighbor nodes (k' is a preset constant, for example, 5-10) are selected, and local mixing is performed:

[0142]

[0143] : the local strategy gradient vector of node j, representing the local model parameter update vector (such as gradient or momentum ) of neighbor node j, indicating the drug adjustment recommendation (such as the adjustment direction of EPO or iron agent dose) calculated by the dialysis center based on its own patient population (such as Hb level, TSAT, secondary syndrome factor, CYP450 gene expression, etc.); : the nearest neighbor node set of node i, consisting of the other nodes with the smallest geometric distance or cosine similarity in the feature space (such as drug sensitivity weight) with the strategy vector of node i; representing a group of dialysis centers most similar to node i in patient composition, pathological characteristics or treatment mode; the similarity calculation can be based on "Hb-TSTAT-SF three-dimensional feature", "multi-omics transformation expression factor", "individual dose response trend" vector, and the similarity between node i and node j is measured by calculating the Euclidean distance (or cosine similarity) of the gradient / momentum vector to be uploaded this round and . : the final aggregated strategy of node i, that is, the local drug adjustment direction obtained by node i after averaging the recommendations of the nearest neighbor nodes.

[0144] S2.3 Construction of asynchronous and decentralized architecture optimization

[0145] In the edge medical computing scenario, in order to alleviate the centralized bottleneck and single-point failure risk of the server, a decentralized optimization strategy is adopted. The algorithm maintains the Byzantine fault tolerance capability in the sparse communication topology through the asynchronous communication based on coordinate descent, combined with local neighbor aggregation and robust mechanism, allowing no more than 1 / 3 nodes to fail. It is suitable for deployment on telemedicine terminals or low-power edge devices. In the decentralized communication topology, each node iOnly exchange information with its topological neighbors or a few nearest neighbors selected according to feature similarity, and perform local NNM aggregation (i.e. calculate the average of neighbor information) to realize local model update. This mechanism uses the similarity between nodes for collaborative learning and suppresses the influence of a few abnormal neighbors through local aggregation.

[0146] Byzantine fault tolerance: Under the decentralized asynchronous federated optimization framework, by combining local NNM aggregation (selecting trusted neighbors and averaging) and / or local robust aggregation rules (such as using only Krum or CwTM to aggregate neighbor information), and the Gossip-style information dissemination protocol, the system can still guarantee that the model parameters of honest nodes converge to the consensus solution when at most R nodes (R < n / 3) fail (including crash or malicious behavior).

[0147] In high-latency networks or asynchronous node environments (such as mobile terminals collecting data for uploading), traditional synchronous aggregation is susceptible to lag. Therefore, a delay gradient filtering and timestamp mechanism is introduced, based on an asynchronous communication mechanism, in which nodes only exchange gradient information with neighbor nodes, making asynchronous stochastic gradient descent more robust. By aggregating only timestamp-legal and update-trusted gradient vectors, introducing timestamp verification, and filtering out outdated gradient data, long-term outdated or tampered node information can be effectively filtered, making it suitable for medical systems with high communication costs and unstable device online frequency in actual deployment.

[0148] Timestamp verification: Each gradient update information is accompanied by a strictly increasing global or local logical timestamp. The server (or receiving node) maintains an expected timestamp window (such as [T-ΔT, T], T is the current round time). Only handle updates with timestamps within the window, discard outdated (< T - ΔT) or future (> T) updates, and set ΔT according to the maximum expected delay of the network.

[0149] S3, according to the patient characteristics of each medical node, the historical momentum gradient and current gradient of the trusted neighbor node, filter abnormal updates through an improved robust aggregation function, and adjust the update amplitude according to the dynamically decaying learning rate, finally output the global medication parameter vector, including the drug category recommendation weight and the dose adjustment coefficient.

[0150] Further, the initial medication decision module 130 adjusts the global parameters based on the momentum gradient mean of trusted neighbor nodes and generates a personalized medication parameter vector to guide the initial dose and type selection of erythropoietin and iron agents, including: for each medical node, aggregate the historical momentum gradient and current gradient of its trusted neighbor nodes; process the aggregation results of all nodes through a weighted robust aggregation function; dynamically adjust the learning rate to generate an updated medication parameter vector, which contains drug category recommendation weights or dose coefficients. The initial medication model output is specifically represented as follows:

[0151] A robust joint model is constructed through distributed data (such as patient characteristics and initial diagnosis and treatment recommendations of each medical node), and a reasonable and personalized initial medication plan is finally generated:

[0152]

[0153] : Global initial medication parameter vector (such as drug category recommendation weight or dose coefficient) for guiding initial medication recommendations for patients; : Current iteration learning rate to control model convergence rate; : Local random gradient of the jth medical node in the tth round of calculation, from local patient data (such as symptoms, physiological indicators, and drug responses); : Momentum vector of the jth node in the last round, recording the previous direction trend, used to smooth the local gradient; β : Momentum decay factor to control the retention level of historical gradients, usually set to 0.9; : The i-th node's nearest neighbors (based on patient distribution similarity, such as medication patterns or disease portraits) for filtering potential malicious nodes; : Number of trusted neighbors of node i; : Improved robust aggregation function for robust aggregation of "neighbor mean momentum gradient" of all nodes to avoid extreme updates that deviate from medical consensus.

[0154] The global medication parameter vector θ is the weight parameter of the global prediction model obtained through federated learning optimization. The model receives input features (including basic clinical data, multi-omics features, biomarker labels, real-time indicators, etc.), outputs prediction results, and the prediction results include: recommended drug category (such as EPO type, iron agent type) probability distribution weight and / or recommended dose adjustment coefficient (such as a proportion factor relative to the standard dose or a specific dose value). The initial medication decision module generates a personalized initial medication plan based on the prediction results of the model.

[0155] For example, in an embodiment, the initial drug decision module 130 simultaneously outputs multiple types of drug parameters such as EPO, iron agent, vitamin D, phosphorus binding agent, etc., covering the entire path of anemia treatment, and the coefficient output is compatible with individualized calculation (such as combined with body weight, meal times): (1) EPO drug dose coefficient: drug type: recombinant human erythropoietin (such as epoetin alpha, darbepoetin alpha, methoxy polyethylene glycol epoetin beta); parameter output form: EPO_dose_coefficient = 0.85 (indicating that the recommended dose for this patient is 85% of the regular dose), combined with patient body weight to calculate the actual dose: recommended dose = standard dose (such as 100 U / kg) × dose coefficient × current body weight. For example: if the patient weighs 60 kg, the actual dose = 100 × 0.85 × 60 = 5100 U / week. (2) Intravenous iron agent type weight and dose coefficient: among them, iron agent type weight: sucrose iron corresponds to a weight of 0.92, sodium gluconate iron corresponds to a weight of 0.15, and carboxymaltose iron corresponds to a weight of 0.08. Iron agent dose adjustment parameter: 1.2 (indicating that the dose needs to be increased by 20%), actual dose = standard dose (such as 100 mg / week) × dose coefficient → 120 mg / week. (3) Auxiliary drug intervention parameters (for complications risk): active vitamin D (such as calcitriol, paricalcitol): vitamin D_activated flag = 0.75 (> 0.5 triggers drug administration); recommended dose: dynamically adjusted based on calcium-phosphorus product (Ca×P), such as: when the flag > 0.7 and Ca×P > 55 mg 2 / dL 2 : give calcitriol 0.25 μg / time, 3 times a week; phosphorus binding agent (such as sevelamer, lanthanum carbonate): phosphorus binding agent_dose_coefficient = 0.6 (60% of the regular dose), administration rule: taken with meals, dose = 500 mg × coefficient × daily meal times. (4) Anticoagulant drug adjustment parameter (for dialysis process): low molecular weight heparin (such as enoxaparin): anticoagulant_dose_coefficient = 1.3 (30% increase in dose during dialysis), adjustment basis: high potassium (potassium > 5.5 mmol / L) activates the coefficient, actual dose: standard dose 4000 U × 1.3 = 5200 U / dialysis.

[0156] The drug dose adjustment module 140 is used to dynamically optimize the time node and dose adjustment opportunity of drug administration through an asynchronous federated reinforcement learning framework combined with a forward-looking parameter correction strategy.

[0157] 1. The dose adjustment rule based on real-time hemoglobin (Hb), ferritin (SF), and transferrin saturation (TSAT) real-time indicators as clinical prior knowledge is used to construct the safety constraint boundary (such as ) of the reinforcement learning reward function ; wherein the dose adjustment rule is:

[0158] According to the real-time hemoglobin (Hb), ferritin (SF) and transferrin saturation (TSAT) index data of the patient, the drug dose is dynamically adjusted according to the following rules:

[0159] When the hemoglobin is lower than 100g / L and the ferritin is lower than 200μg / L, and the transferrin saturation is lower than 20%, start intravenous iron supplementation (sucrose iron 100mg / week) and give erythropoietin EPO 50-150U / kg / week;

[0160] When the hemoglobin is not lower than 130g / L and the ferritin is not lower than 300μg / L, and the transferrin saturation is not lower than 30%, suspend erythropoietin EPO and iron therapy;

[0161] Combined with blood drug concentration and electrolyte level, the drug dose is corrected (such as adjusting the dose of anticoagulant drugs when there is hyperkalemia).

[0162] The rules based on Hb / SF / TSAT thresholds are the basis for the design of clinical prior knowledge or reward functions.

[0163] 2. Optimization of drug administration nodes by artificial intelligence algorithm

[0164] To improve the individualization, dynamics and safety of drug administration strategies, an intelligent optimization framework based on asynchronous federated reinforcement learning (AFedPG) is constructed to optimize the time nodes and dose adjustment opportunities of drug administration, especially for the fine regulation of drugs with half-life related drugs such as erythropoietin (EPO) and intravenous iron (such as sucrose iron).

[0165] Further, the drug dose adjustment module 140 dynamically optimizes the time nodes and dose adjustment opportunities of drug administration through the asynchronous federated reinforcement learning framework combined with the forward-looking parameter correction strategy, including:

[0166] (1) Constructing a delay adaptive forward-looking mechanism: generate forward-looking correction parameters based on the historical trend of policy parameter changes and the degree of delay, and dynamically adjust the compensation intensity through an adaptive factor to ensure the consistency of clinical efficacy in an asynchronous environment;

[0167] Due to the feedback delay of different medical nodes (such as different dialysis centers), a delay compensation mechanism needs to be introduced to maintain the stability of training. In the kth global parameter update, a forward-looking parameter correction strategy is adopted, aiming to make a first-order extrapolation compensation for the historical deviation of the parameter update direction caused by the delay

[0168]

[0169] ​: The parameter after prediction correction, used to offset the error of the drug administration strategy caused by the delay of the node upload, to ensure reasonable adjustment within the pharmacokinetic time limit. : The global drug administration strategy parameter issued by the server at the kth update, representing the unified decision strategy of all nodes (such as the frequency of weekly drug administration). : The global strategy parameter of the previous round of update, used to estimate the trend of the change of the drug administration strategy. : Represents the direction of the last parameter update. : The adaptive factor related to the delay , reflecting the influence of data upload on strategy optimization, 0 < α ≤ 1, , λ is a hyperparameter that controls the decay rate, controlling the strength of delay compensation, usually λ = 0.1, which can be adjusted through cross-validation. The larger , the smaller. : The degree of delay of historical information relied on by the current strategy calculation (such as the update time lag of laboratory data). : Bias compensation coefficient, dynamically adjust the historical parameter change according to the delay. When is small, the compensation amplitude increases, amplifying the direction of the last update to predict the current update trend, trying to offset the strategy lag effect caused by the delay.

[0170] Through this mechanism, when the patient index collection lags, the predicted strategy can still be relied on to make approximately optimal decisions, effectively dealing with situations such as weekends or holidays without detection.

[0171] To further offset the second-order gradient bias caused by the delay, the following conditions are met:

[0172]

[0173] : The second derivative of the objective function, describing the sensitivity of the expected efficacy (such as Hb compliance rate) to the change of the drug administration strategy.

[0174] This formula ensures that the strategy correction of different medical points in an asynchronous environment can maintain clinical convergence and efficacy consistency, avoiding gradient bias caused by inconsistent update times of each node.

[0175] (2) Construct an asynchronous federated learning framework (AFedPG)

[0176] The training process of the administration node optimization includes: calculating the local policy gradient according to the probability derivative of the administration action in the trajectory and the reward discount; the server updates the global parameters after normalizing and momentum fusing the received delayed gradient; and the updated administration strategy is dynamically issued to each node without waiting for synchronization of all nodes. Specifically, the following three steps are included:

[0177] 1) Local calculation stage:

[0178] Definition: State (s) : includes the current and historical: hemoglobin (Hb), ferritin (SF), transferrin saturation (TSAT), blood concentration (such as EPO valley concentration), key electrolyte levels (potassium, calcium, phosphorus), inflammatory factors (such as CRP), liver and kidney function indicators, multiple group biomarker labels, and last medication record, etc.

[0179] Action (a) : defined as the adjustment decision of EPO dose and intravenous iron dose. It can be represented as a two-dimensional continuous action vector , ranging from [-MaxΔ, +MaxΔ]; or a discrete action set, such as {maintain, small increase EPO, large increase EPO, small decrease EPO, large decrease EPO,...}, covering independent and combined adjustments of EPO and iron agents.

[0180] Reward (r) : designed to balance treatment effectiveness and safety, for example

[0181]

[0182] where w1~w7 are weight coefficients, which can be optionally set as [1.0, 0.5, 0.3, 1.0, 1.0, 0.01, 0.01]; is the Hb target value, is the maximum SF value, is the maximum TSAT value, all of which are target values or safety thresholds set by the clinic, is an event indicator function (1 if it occurs, otherwise 0), where is a hypotension event (if related to administration), is a hyperkalemia event; represents the current administration dose, where is the EPO dosage (encouraging the minimum effective dose), is the iron dosage (encouraging the minimum effective dose); is a fixed reward for reaching the target. The specific value of the weight is determined by clinical expert experience or offline tuning.

[0183] ​​​Each medical node uses currently collected patient indicators (such as Hb, SF, TSAT) and historical strategies. Generate trajectory Calculate the corresponding policy gradient :

[0184]

[0185] in, Represents a node i The original strategy gradient is calculated based on the trajectory μ and parameter θ; μ: the patient's complete state-behavior-feedback trajectory, which records changes in patient indicators and medication behavior during continuous dialysis cycles; such as Hb value recorded once a week, whether EPO is given, corresponding Hb increase or side effects, etc. : The patient state vector at time p, which includes real-time physiological and laboratory indicators; In state The next action (decision) to be taken, such as "give". "Intravenous iron supplementation 100mg", "Stop iron supplementation", etc. : Under the current policy parameter θ, for the state Select Action The probability; represents a distributed probabilistic model for medication decisions, which gradually converges to the optimal dosing strategy as training progresses; The derivative of the strategy probability with respect to the parameters is used to guide the direction of parameter updates; when a certain type of patient condition produces a good therapeutic effect, its probability of occurrence will be amplified. In state Next action The immediate rewards obtained can be defined as: Hb levels reaching the target (positive), adverse reactions occurring (negative), and inflammation suppression effects, etc. : Reward discount factor (0.95~0.99) to control the trade-off between short-term and long-term effects (e.g., choose 0.95); avoid focusing only on the current Hb index while ignoring long-term iron overload or cardiovascular risks.

[0186] When the state When the safety boundary is violated (e.g., Hb < 80 g / L), mandatory setting is required. A large negative value, such as -10, ensures that the policy network avoids dangerous decision areas.

[0187] Policy Network Implemented using a three-layer fully connected neural network: Input layer: state Dimensions (e.g., 256); Hidden layer: 128 neurons, ReLU activation; Output layer: Discrete action: Softmax output each action probability; Continuous action: Tanh activation output [-1, 1] dose adjustment coefficient.

[0188] 2) Server update phase:

[0189] Server receives delayed gradient from different nodes , normalized update according to the following formula:

[0190]

[0191] : Learning rate (0.0001~0.001, decay), control the adjustment range of the drug strategy every round, decay over time. : Delayed gradient received by the server, from the node calculation result of the th iteration, reflecting the gradient direction of the historical strategy. : L2 norm of the gradient, to avoid large gradient causing policy fluctuation.

[0192] Before performing parameter update , optionally, apply the forward-looking parameter correction strategy to the current parameter or momentum state d to compensate for the delay of the information on which the update is based.

[0193] To improve stability, introduce momentum mechanism to construct fusion gradient direction:

[0194]

[0195] : Historical aggregated gradient, used to suppress violent fluctuations in drug adjustment. : The policy gradient corresponding to the current patient trajectory, used to update the current policy direction, see the expression above . The server maintains a global momentum state d, when receiving the delayed original policy gradient from node i, then uses the updated d to perform parameter update.

[0196] 3) Downlink policy push:

[0197] The server pushes the updated policy parameters to each node to realize closed-loop control, which does not depend on all nodes to synchronize, and can avoid the problem of "slowest node dragging down training".

[0198] The asynchronous federated learning framework (AFedPG) is the core optimization means of the present application. The drug dosage adjustment module 140 dynamically calculates and outputs the optimal drug dosage adjustment decision (including the drug dosage, the drug administration time node, and the adjustment timing) according to the real-time monitoring indicators (Hb, SF, TSAT, etc.) through the intelligent optimization framework based on the asynchronous federated reinforcement learning (AFedPG).

[0199] 3. Strategy modification in advance driven by risk prediction

[0200] The complication risk prediction model is embedded in the training framework, taking inflammatory factors (such as CRP), calcium and phosphorus levels, electrolytes, etc. as input to predict the likelihood of cardiovascular events and metabolic disorders. When the predicted risk increases, the EPO and iron drug administration intensity is automatically adjusted in advance, or auxiliary drugs (such as active vitamin D to inhibit calcium and phosphorus absorption) are introduced in a timely manner to achieve proactive strategy intervention.

[0201] It should be noted that the embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0202] It should also be noted that the relationship terms such as first and second in the embodiments of the present application are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or equipment including the element.

[0203] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the embodiments of the present application can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown in the embodiments of the present application, but will conform to the widest scope consistent with the principles and novel features disclosed in the embodiments of the present application.

Claims

1. A precision medication system for hemodialysis patients based on multi-omics and big data, characterized in that, The method comprises the following steps: A multi-source data acquisition module is configured to collect multi-source heterogeneous data of a patient and perform standardization and integration processing, and to construct a patient-level data warehouse; wherein the multi-source heterogeneous data comprises basic clinical data, multi-omics detection data, and real-time physiological and test indicators; A drug metabolism analysis module is configured to analyze drug metabolism pathways and individual differences through multi-omics data, construct a drug metabolism polymorphism model for predicting the ability of a patient to respond to a drug, and output a gene metabolism ability classification and a biomarker risk label; The drug metabolism analysis module performs operations including: Based on the gene expression profile, metabolite concentration, and proteomic information of the patient, the in vivo metabolism pathway of a key drug is analyzed to generate individualized pharmacokinetic parameters, including drug half-life and peak concentration; Through drug metabolism-related gene variation analysis, biomarkers are identified and a metabolism ability classification index is established; The metabolism ability classification index and the biomarker risk label are integrated into the feature space of federated learning to constrain the dose recommendation range of the initial drug use model; An initial drug use decision module is configured to construct an initial drug use decision model based on a federated optimization framework with momentum enhancement and robust aggregation mechanism, and to output an individualized initial drug use plan for erythropoietin and intravenous iron; The construction of the initial drug use decision model includes: An improved robust aggregation rule is constructed, and three types of aggregation methods are used to aggregate and then weightedly fuse the local model gradients, thereby constructing an improved robust aggregation function, including: The three types of aggregation results obtained by the geometric center aggregation, coordinate pruning mean aggregation and norm filtering aggregation are linearly combined according to preset weights to obtain a global update direction, and an improved robust aggregation function is obtained : wherein, : final robust aggregation function output, representing the global medication strategy update direction constructed in the current round, used to adjust the model parameters; : local gradient vector of the node ; n: total number of participating nodes, i.e. the number of medical institutions participating in federated learning; : weighted coefficient of geometric center aggregation ; : weighted coefficient of coordinate pruning mean aggregation ; : weighted coefficient of norm filtering aggregation ; Based on the improved robust aggregation function, a momentum enhancement and two-step bucket aggregation strategy are used to smooth the gradient update, and a nearest neighbor hybrid mechanism is used to defend against abnormal nodes; According to the patient characteristics of each medical node, the historical momentum gradient and the current gradient of the trusted neighbor node, the improved robust aggregation function is used to filter abnormal updates, and the parameter update amplitude is adjusted according to the dynamically decaying learning rate, and finally a global drug use parameter vector is output, including drug category recommendation weight and dose adjustment coefficient; A drug dose adjustment module is configured to dynamically optimize the time node and dose adjustment opportunity of drug administration through an asynchronous federated reinforcement learning framework combined with a forward-looking parameter correction strategy.

2. The multi-omics and big data-based precision medication system for hemodialysis patients according to claim 1, characterized in that, The improved robust aggregation rule is constructed based on three types of aggregation methods to aggregate and then weightedly fuse the local model gradients, thereby constructing an improved robust aggregation function, including: Geometric center aggregation: selecting the gradient vector with the smallest distance to the majority node drug gradient; Coordinate pruning mean aggregation: sorting and pruning each coordinate independently, and then taking the average after removing the extreme values; Norm filtering aggregation: calculate the L2 norm of each node uploaded gradient vector, sort by size, remove the node with the largest gradient change amplitude, and then aggregate; The three types of aggregation results, i.e., geometric center aggregation, coordinate pruning mean aggregation, and norm filtering aggregation, are linearly combined according to the preset weight to obtain the improved robust aggregation function. The federated optimization framework based on momentum enhancement and robust aggregation mechanism includes: 3.The multi-omics and big data-based precise medication system for hemodialysis patients of claim 2, wherein, Polyak momentum mechanism is introduced to suppress local fluctuations by weighted average of historical gradients; ​ ​ The two-step bucket aggregation strategy is adopted to smooth the gradient update and combined with the nearest neighbor hybrid mechanism to defend against abnormal nodes.

4. The multi-omics and big data-based precision medication system for hemodialysis patients according to claim 3, characterized in that, Wherein, The two-step bucket aggregation strategy is adopted to smooth the gradient update and combined with the nearest neighbor hybrid mechanism to defend against abnormal nodes. The medical nodes are grouped, and the medical nodes are randomly divided into multiple buckets, each bucket containing a number of nodes. The medication gradient average of each node in the bucket is calculated to form a bucket consensus vector. Robust aggregation is performed on the bucket consensus vector to realize inter-group aggregation and generate a global update direction. Based on the nearest neighbor hybrid mechanism, neighbors are selected for each node to select the most similar neighbor nodes in the feature space. The medication gradients of the neighbor nodes are aggregated to form the final update strategy.

5. The multi-omics and big data-based precision medication system for hemodialysis patients according to claim 4, characterized in that, Wherein, The initial medication decision model also adopts a decentralized robust mechanism, specifically including: In the edge medical computing scenario, a decentralized optimization strategy is adopted based on an asynchronous communication mechanism, in which nodes only exchange gradient information with neighbor nodes; a time stamp verification is introduced to filter out gradient data that has not been updated for a long time; and a Byzantine fault tolerance capability is maintained in the sparse communication topology, allowing no more than 1 / 3 of the nodes to fail. 6.The multi-omics and big data-based precise medication system for hemodialysis patients of claim 5, wherein, Wherein, The initial medication decision module adjusts the global parameters based on the momentum gradient mean of the trusted neighbor nodes and generates a personalized medication parameter vector to guide the selection of the initial dose and type of erythropoietin and iron agent; Wherein, the initial medication decision module adjusts the global parameters based on the momentum gradient mean of the trusted neighbor nodes and generates a personalized medication parameter vector, specifically including: For each medical node, the historical momentum gradient and the current gradient of its trusted neighbor nodes are aggregated. The aggregation results of all nodes are processed by a weighted robust aggregation function. The learning rate is dynamically adjusted to generate an updated medication parameter vector, which includes drug category recommendation weights or dose coefficients. 7.The multi-omics and big data-based precise medication system for hemodialysis patients of claim 1, wherein, The medication dose adjustment module uses the dose adjustment rules based on real-time hemoglobin, ferritin and transferrin saturation indexes as clinical prior knowledge to construct a safety constraint boundary for the reinforcement learning reward function; wherein the dose adjustment rules are: When hemoglobin is lower than 100g / L and ferritin is lower than 200μg / L, and transferrin saturation is lower than 20%, start intravenous iron supplementation and give erythropoietin 50-150U / kg / week; When hemoglobin is not less than 130g / L and ferritin is not less than 300μg / L, and transferrin saturation is not less than 30%, suspend erythropoietin and iron therapy; Combined with blood drug concentration and electrolyte level, the dose is corrected. 8.The multi-omics and big data-based precise medication system for hemodialysis patients of claim 1, wherein, The medication dose adjustment module dynamically optimizes the time node and dose adjustment opportunity of drug administration through an asynchronous federated reinforcement learning framework combined with a forward-looking parameter correction strategy, including: Based on the historical strategy parameter change trend and delay degree, generate a forward-looking correction parameter, and dynamically adjust the compensation intensity through an adaptive factor to ensure the consistency of clinical efficacy in an asynchronous environment; According to the probability derivative of the administration action in the trajectory and the reward discount, a local policy gradient is calculated; the server updates global parameters after normalizing and momentum fusing the received delayed gradients; and the updated administration policy is dynamically issued to each node without waiting for synchronization of all nodes.

Images