Precise medication system for hemodialysis patient based on multiple omics and big data

Abstract

The invention relates to the technical field of medical care informatics and precision medical treatment, and discloses a hemodialysis patient precision medication system based on multiple omics and big data, and the system comprises a multi-source data collection module which collects the multi-source heterogeneous data of a patient and carries out the standardization and integration processing; the drug metabolism analysis module is used for analyzing drug metabolism paths and individual differences through multi-omics data, constructing a drug metabolism polymorphic model and outputting gene metabolism capability grades and biomarker risk tags; the initial medication decision-making module is used for constructing an initial medication decision-making model based on a federal optimization framework of a momentum enhancement and robust aggregation mechanism, and outputting a personalized initial medication scheme for erythropoietin and a vein iron agent; and the drug dosage adjustment module is used for dynamically optimizing the time node and dosage adjustment opportunity of drug administration through an asynchronous federal reinforcement learning framework in combination with a prospective parameter correction strategy. According to the invention, the intelligent medication system with strong robustness and fine decision is constructed.

Description

Technical Field

[0001] The present invention relates to the fields of healthcare informatics and precision medicine technology, and in particular to a precision medication system for hemodialysis patients based on multi-omics and big data. Background Art

[0002] Hemodialysis is the most common alternative treatment for patients with end-stage renal disease (ESRD). 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, important medications for correcting renal anemia, exhibit significant individual variability, requiring dynamic adjustment of dosage and frequency to avoid worsening anemia or iron overload.

[0003] However, current medication regimens for dialysis patients often rely on manual experience or fixed rules, ignoring individual differences in gene expression, metabolic capacity, inflammation levels, and drug responsiveness. This leads to fluctuating efficacy and a high risk of complications, potentially resulting in drug waste, poor efficacy (such as suboptimal anemia correction), and an increased risk of drug-related adverse reactions (such as hypertension, thrombosis, and iron overload-related organ damage). Furthermore, due to the privacy, multi-source heterogeneity, and distributed nature of medical data, traditional centralized models struggle to achieve large-scale, high-quality patient data modeling.

[0004] Therefore, there is an urgent need to introduce a multi-omics and big data-driven intelligent medication strategy model to build an intelligent medication system with strong robustness and precise decision-making while ensuring data privacy. Summary of the Invention

[0005] The present invention aims to address the current problems of "one-size-fits-all" solutions, model fragility, and strategy inadaptability in medication decision-making for dialysis patients. It proposes a precision medication solution that integrates multi-omics analysis, big data learning, and artificial intelligence decision optimization. Specifically, it provides a precision medication system for hemodialysis patients based on multi-omics and big data.

[0006] To achieve the above objectives, the following technical solutions are adopted: A precision medication system for hemodialysis patients based on multi-omics and big data, including: A multi-source data acquisition module is used to collect, standardize, and integrate multi-source heterogeneous data from patients to build a patient-level data warehouse; wherein the multi-source heterogeneous data includes: basic clinical data, multi-omics test data, and real-time physiological and laboratory indicators; The drug metabolism analysis module is used to analyze drug metabolism pathways and individual differences through multi-omics data, build drug metabolism polymorphism models to predict patients' ability to respond to drugs, and output gene metabolism capacity classification and biomarker risk labels; The operations performed by the drug metabolism analysis module include: Analyze the in vivo metabolic pathways of key drugs based on the patient's gene expression profile, metabolite concentration, and proteomic information to generate personalized pharmacokinetic parameters, including drug half-life and peak concentration; Identify biomarkers and establish metabolic capacity grading indicators through analysis of drug metabolism-related gene variations; The metabolic capacity grading index and biomarker risk labels are integrated into the feature space of federated learning to constrain the recommended dosage range of the initial medication model. The initial medication decision module is used to build an initial medication decision model based on a federated optimization framework with momentum enhancement and robust aggregation mechanism, and output a personalized initial medication plan including erythropoietin and intravenous iron; The construction of the initial medication decision model includes: An improved robust aggregation rule is constructed. Local model gradients are aggregated based on three types of aggregation methods and then weighted fusion is performed to construct an improved robust aggregation function, including: The three types of aggregation results obtained by the three aggregation methods of geometric center aggregation, coordinate trimmed mean aggregation and norm filtering aggregation are linearly combined according to the preset weights as the global update direction to obtain an improved robust aggregation function :

[0007] in, : The final robust aggregation function output represents the update direction of the global medication strategy constructed in the current round and is used to adjust the model parameters; :node The local gradient vector of n: the total number of participating nodes, that is, the number of medical institutions participating in federated learning; : Geometric center aggregation The weighting coefficient of : Coordinate trimmed mean aggregation The weighting coefficient of : Norm filtering aggregation The weighting coefficient of Based on the improved robust aggregation function, momentum enhancement and a two-step bucket aggregation strategy are used to smooth gradient updates, combined with a nearest neighbor hybrid mechanism to defend against abnormal nodes. Based on the patient characteristics of each medical node, the historical momentum gradient and current gradient of trusted neighbor nodes, an improved robust aggregation function is used to filter abnormal updates and the parameter update amplitude is adjusted according to the dynamically decayed learning rate. Finally, a global medication parameter vector is output, which includes the recommended weights of drug categories and dosage adjustment coefficients; the drug dosage adjustment module is used to dynamically optimize the time nodes of drug administration and the timing of dosage adjustment through an asynchronous federated reinforcement learning framework combined with a forward-looking parameter correction strategy.

[0008] Furthermore, the operations performed by the drug metabolism analysis module include: Analyze the in vivo metabolic pathways of key drugs based on the patient's gene expression profile, metabolite concentration, and proteomic information to generate personalized pharmacokinetic parameters, including drug half-life and peak concentration; Identify biomarkers and establish metabolic capacity grading indicators through analysis of drug metabolism-related gene variations; Among them, the metabolic capacity grading index and biomarker risk label are integrated into the feature space of federated learning to constrain the dosage recommendation range of the initial medication model.

[0009] Furthermore, the construction of the initial medication decision model includes: An improved robust aggregation rule is constructed. Local model gradients are aggregated based on three types of aggregation methods and then weighted fusion is performed to construct an improved robust aggregation function. Based on the improved robust aggregation function, momentum enhancement and a two-step bucket aggregation strategy are used to smooth gradient updates, combined with a nearest neighbor hybrid mechanism to defend against abnormal nodes. Based on the patient characteristics of each medical node, the historical momentum gradient and current gradient of trusted neighbor nodes, an improved robust aggregation function is used to filter abnormal updates and the parameter update amplitude is adjusted according to the dynamically decayed learning rate. Finally, a global medication parameter vector is output, which includes the recommended weights of drug categories and dosage adjustment coefficients.

[0010] Furthermore, the improved robust aggregation rule is constructed, and local model gradients are aggregated based on three types of aggregation methods and then weighted fusion is performed to construct an improved robust aggregation function, including: Geometric center aggregation: select the gradient vector with the smallest distance to the medication gradient of most nodes; Coordinate trimmed mean aggregation: Each coordinate is independently sorted by coordinate trimmed mean and the extreme values ​​are trimmed and then the average is taken; Norm filtering and aggregation: Calculate the L2 norm of the gradient vector uploaded by each node, sort it by size, and aggregate after removing the node with the largest gradient change; The three types of aggregation results, namely geometric center aggregation, coordinate trimmed mean aggregation and norm filtering aggregation, are linearly combined according to preset weights as the global update direction to obtain an improved robust aggregation function.

[0011] Furthermore, the federated optimization framework based on momentum enhancement and robust aggregation mechanism includes: The Polyak momentum mechanism is introduced to suppress local fluctuations through weighted averaging of historical gradients; A two-step bucket aggregation strategy is adopted to smooth gradient updates and combined with a nearest neighbor hybrid mechanism to defend against abnormal nodes; wherein, the two-step bucket aggregation strategy includes: first randomly grouping nodes to calculate the intra-group consensus, and then performing robust aggregation on the inter-group results.

[0012] Furthermore, the two-step bucket aggregation strategy is used to smooth the gradient update and combined with the nearest neighbor hybrid mechanism to defend against abnormal nodes, including: Group the medical nodes and randomly divide them into multiple buckets, each containing several nodes. Calculate the average medication gradient of the nodes in each bucket to form a bucket consensus vector; Perform robust aggregation on the bucket consensus vector to achieve inter-group aggregation and generate a global update direction; Neighbors are screened based on the nearest neighbor hybrid mechanism, selecting multiple neighbor nodes with the most similar feature spaces for each node; Aggregate the medication gradients of neighboring nodes to form the final update strategy.

[0013] Furthermore, 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 the asynchronous communication mechanism, nodes only exchange gradient information with neighboring nodes. Timestamp verification is introduced to filter out gradient data that has timed out and not been updated. Byzantine fault tolerance is maintained in a sparse communication topology, allowing no more than 1 / 3 of the nodes to fail.

[0014] Furthermore, the initial medication decision module performs global parameter adjustment 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 supplements; The initial medication decision module performs global parameter adjustment based on the momentum gradient mean of the trusted neighbor nodes and generates a personalized medication parameter vector, specifically including: For each medical node, aggregate the historical momentum gradient and current gradient of its trusted neighbor nodes; The aggregation results of all nodes are processed through weighted robust aggregation function; The learning rate is dynamically adjusted to generate an updated medication parameter vector, wherein the medication parameter vector includes a drug category recommendation weight or a dosage coefficient.

[0015] Furthermore, the drug dosage adjustment module uses dosage 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 dosage adjustment rules are: When hemoglobin is less than 100 g / L, ferritin is less than 200 μg / L, and transferrin saturation is less than 20%, intravenous iron supplementation and erythropoietin 50-150 U / kg / week are initiated; When hemoglobin is not less than 130 g / L, ferritin is not less than 300 μg / L, and transferrin saturation is not less than 30%, erythropoietin and iron therapy should be suspended; Adjust the dosage based on blood drug concentration and electrolyte levels.

[0016] Furthermore, the drug dosage adjustment module uses an asynchronous federated reinforcement learning framework and a forward-looking parameter correction strategy to dynamically optimize the timing of drug administration and dosage adjustment opportunities, including: Generate forward-looking correction parameters based on historical strategy parameter change trends and delay levels, and dynamically adjust compensation intensity through adaptive factors to ensure consistency of clinical efficacy in an asynchronous environment. The local policy gradient is calculated based on the probability derivative and reward discount of the drug administration action in the trajectory. The server normalizes and momentum-fuse the received delayed gradient and then updates the global parameters. The updated drug administration policy is dynamically distributed to each node without waiting for all nodes to synchronize.

[0017] Compared with the prior art, the present invention achieves the following beneficial effects: 1. This paper proposes an improved decentralized Byzantine robust learning method (including robust aggregation function F, momentum enhancement, two-step bucket aggregation, and nearest neighbor hybrid mechanism). On the basis of ensuring the privacy of patient data at each medical node, a decentralized federated learning framework is constructed. Through a robust optimization strategy, the stable fusion of cross-node model parameters is achieved, and the nonlinear relationship between individual clinical characteristics of patients and EPO / iron dosage is effectively explored, thereby improving the personalization and safety of initial medication recommendations.

[0018] 2. This paper proposes an improved robust aggregation rule. To address possible abnormal nodes or false positives in medical data, it constructs a multi-robust aggregation mechanism that integrates geometric center aggregation, coordinate trimmed mean (CwTM), and comparative gradient elimination (CGE). This mechanism filters out potentially malicious or distorted information from multiple dimensions, significantly enhancing the model's fault tolerance and generalization capabilities in real medical environments.

[0019] 3. This paper proposes a federated optimization framework based on momentum enhancement and robust aggregation mechanism, introduces the Polyak momentum mechanism to smooth local gradient oscillations, and combines a dynamic sampling strategy with a two-step aggregation structure (bucketing + secondary aggregation) to effectively address challenges such as sampling differences among dialysis centers, uneven computing power, and delayed updates, achieving stable and rapid model convergence.

[0020] 4. This paper proposes a two-step aggregation and nearest neighbor mixing strategy (NNM), which first dilutes local abnormal nodes through intra-group aggregation and then eliminates isolated outliers through nearest neighbor geometric mixing, further improving the system's robustness and distribution adaptability in complex communication networks or data heterogeneous scenarios.

[0021] 5. This paper proposes an intelligent optimization framework based on asynchronous federated reinforcement learning (AFedPG), constructs a policy gradient-driven asynchronous optimization mechanism, realizes asynchronous model updates between different dialysis nodes, and automatically optimizes the drug administration time nodes and adjustment timing. It is particularly suitable for strategy modeling tasks that consider temporal characteristics such as EPO half-life, dialysis frequency, and Hb dynamic response.

[0022] 6. The present invention proposes a forward-looking parameter correction strategy. To address the policy lag problem caused by node upload delay, a delay-adaptive parameter prediction mechanism is designed to compensate for historical deviations in parameter updates, avoid policy update instability or misjudgment, and achieve highly robust policy control in an asynchronous environment.

[0023] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The above and other features, advantages and aspects of the embodiments of the present invention will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present invention and do not constitute a limitation of the present invention. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which: Figure 1 Schematic diagram of a module of a multi-omics and big data-based precise medication system for hemodialysis patients according to an embodiment of the present invention; Figure 2 Schematic diagram of the architecture of a multi-omics and big data-based precision medication system for hemodialysis patients according to an embodiment of the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0026] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0027] Figure 1 Schematic diagram of a module of a multi-omics and big data-based precise medication system for hemodialysis patients according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the architecture of a precise medication system for hemodialysis patients based on multi-omics and big data according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, a multi-omics and big data-based precision medication system 100 for hemodialysis patients includes: The multi-source data acquisition module 110 is used to collect multi-source heterogeneous data of patients and perform standardization and integration processing to build 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; real-time physiological and test indicators include but are not limited to: blood drug concentration (such as EPO, iron supplements), electrolyte levels (potassium, calcium, phosphorus, etc.), inflammatory factors (such as C-reactive protein, interleukin-6, etc.), hemoglobin (Hb), ferritin (SF), transferrin saturation (TSAT) and other dynamic monitoring indicators.

[0028] 1. Multi-dimensional clinical data collection: 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.).

[0029] Multi-omics testing 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).

[0030] Real-time physiological and laboratory indicators: dynamically monitor blood drug concentration (such as erythropoietin EPO, iron agent, etc.), electrolyte level (potassium, calcium, phosphorus, etc.), inflammatory factors (C-reactive protein, interleukin-6, etc.), and laboratory indicators such as hemoglobin (Hb), ferritin (SF), and transferrin saturation (TSAT).

[0031] 2. Data standardization and integration: (1) Standardization processing of multi-source heterogeneous data: With the help of big data processing technology, pre-process the clinical information, multi-omics test 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 specification, dimension unit (such as Hb concentration g / L, TSAT percentage, etc.) and coding system, solve the inconsistency of different hospital information systems in data format and acquisition frequency, ensure the consistency and comparability of data input for subsequent model training.

[0032] (2) Construct patient-level data warehouse and realize deep data fusion: Based on the structured data model, build a unified data warehouse architecture, deeply integrate basic clinical data (such as age, gender, complications), multi-omics data (gene expression, proteome, metabolome), and dynamic monitoring indicators (EPO concentration, electrolyte changes, inflammatory factors, etc.). Through the establishment of a unique patient identifier, break down data silos and form a real-time updateable, long-term traceable individualized data archive, providing a solid data foundation for the accurate prediction and continuous optimization of subsequent drug models.

[0033] The drug metabolism analysis module 120 is used to analyze the drug metabolism pathway and individual differences through multi-omics data, construct a drug metabolism polymorphism model for predicting the patient's response to drugs, and output the genetic metabolism ability classification and biomarker risk label. 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.

[0034] 1. Drug metabolism pathway modeling and individual difference prediction: Combine the patient's gene expression profile, metabolite concentration, and proteome information to analyze the in vivo metabolism pathway of key drugs (such as erythropoietin EPO, different types of iron agents, etc.), generate individualized pharmacokinetic parameters, including drug half-life and peak concentration. For example, analyze the decomposition and metabolism mechanism of EPO in the liver, the transport process of iron agents in the intestinal absorption and storage, etc., simulate the individual's in vivo drug concentration change process after administration, evaluate its drug efficacy duration and response intensity, and predict the metabolic differences of different patients under the same dose.

[0035] 2. Establish drug metabolism polymorphism models and identify key biomarkers: Based on variation analysis of drug metabolism-related genes, such as CYP450 enzymes and transporters (e.g., TFR and FPN), a polymorphism model is constructed to predict a patient's drug response. By combining metabolite levels with clinical response indicators (e.g., rate of change in Hb and ferritin reserve capacity), biomarkers with significant impact on drug efficacy are screened (e.g., TSAT below 20% indicates potential iron utilization impairment, and SF above 500 ng / mL indicates risk of iron overload). This identification of biomarkers allows for the establishment of metabolic capacity grading indices (rapid, intermediate, or slow metabolizers), providing a scientific basis for initial dosing recommendations and individualized medication adjustment strategies.

[0036] The finally generated metabolic capacity grading index and biomarker risk label are integrated into the feature space of federated learning to constrain the dosage recommendation range of the initial medication decision module 130.

[0037] An initial medication decision module 130 is configured to construct an initial medication decision model based on a federated optimization framework of momentum enhancement and robust aggregation mechanism, and output a personalized initial medication plan including erythropoietin and intravenous iron; To provide accurate initial recommendations for medications such as erythropoietin (EPO) and iron supplements, this initial medication decision module 130 builds a big data-driven initial medication decision model based on clinical data and multi-omics features from nearly 1,000 dialysis patients. This model leverages an improved decentralized Byzantine robust learning approach to explore the nonlinear mapping relationship between individual characteristics and drug dosage, while ensuring data privacy. This provides a foundation for subsequent dynamic medication control (drug type, dosage, and timing).

[0038] Furthermore, the construction of the initial medication decision model includes: An improved robust aggregation rule is constructed. Local model gradients are aggregated based on three types of aggregation methods and then weighted fusion is performed to construct an improved robust aggregation function. Based on the improved robust aggregation function, momentum enhancement and a two-step bucket aggregation strategy are used to smooth gradient updates, combined with a nearest neighbor hybrid mechanism to defend against abnormal nodes. Based on the patient characteristics of each medical node, the historical momentum gradient and current gradient of trusted neighbor nodes, an improved robust aggregation function is used to filter abnormal updates and the parameter update amplitude is adjusted according to the dynamically decayed learning rate. Finally, a global medication parameter vector is output, which includes the recommended drug category weights and dosage adjustment coefficients.

[0039] Furthermore, the construction of the initial medication decision model includes: S1. Construct an improved robust aggregation rule. Based on three types of aggregation methods, local model gradients are aggregated separately and then weighted fused to construct an improved robust aggregation function. S1.1 Geometric center aggregation: select the gradient vector with the smallest distance to the medication gradient of most nodes; This method is used to select the "most credible" local model gradient among multiple medical nodes (such as the medication adjustment direction calculated by a hospital based on patient Hb, TSAT, SF and other characteristics) to resist the upload of erroneous gradients by abnormal nodes.

[0040] Select the gradient with the smallest distance to the nearest neighbor node (Krum), and resist outlier gradients through local distance measurement. It is suitable for scenarios with complete communication between nodes, and the collapse point can reach 1 / 2. The Krum selection rule is:

[0041] in, :node local gradient vector; n: the total number of participating nodes, that is, the number of medical institutions participating in federated learning; f: the maximum tolerable number of "abnormal or malicious" nodes, which indicates the number of nodes allowed to upload abnormal data (such as abnormally high or low medication recommendations); :node and nodes The squared Euclidean distance between local gradients indicates the degree of difference in their medication adjustment recommendations in parameter space; :node Recently neighbors, indicating that among all nodes, The closest and most trusted set of nodes.

[0042] By selecting gradients that are "close" to the opinions of the majority of nodes, outlier gradients that may be caused by poor data quality, extreme samples, or communication errors are excluded, thereby improving the stability of the initial medication model (such as the EPO dosage mapping function).

[0043] S1.2 Coordinate trimmed mean aggregation: Use coordinate trimmed mean to sort each coordinate independently and trim and remove extreme values ​​before taking the average; This method is used to process high-dimensional gradient scenarios (such as each dimension representing the EPO dose adjustment factor, iron absorption sensitivity, etc.), eliminate extreme values ​​coordinate by coordinate, and improve robustness.

[0044] Each coordinate is independently sorted by the coordinate trimmed mean (CwTM) And trim the extreme values ​​and keep the middle The average of the values ​​is suitable for scenarios with high-dimensional gradients and independent dimensions. The aggregation rules are as follows:

[0045] in, : the value of the kth coordinate in the local gradient vector calculated by the i-th node; k: the dimension index of the gradient vector, representing a specific medical feature dimension (such as the weight of a certain omics feature affecting EPO metabolism); The index of the node ranked i after sorting on the k-th coordinate, that is, the i-th middle value position after sorting the k-th parameter; : The number of intermediate nodes retained, before removing The maximum value and The minimum value is used to average the number of trusted nodes.

[0046] By sorting each model parameter dimension separately, removing the minimum and maximum extreme values, and then calculating the average, we can prevent specific dimensions from being perturbed by extreme values. This approach is suitable for scenarios where there is no strong correlation between dimensions, such as those involving multi-omics input dimensions. This is particularly true when each parameter dimension in the model represents a different biomarker or individual characteristic (such as CYP450 expression, liver function index, or electrolyte abnormality markers) that influence medication decisions. This effectively eliminates extreme values ​​caused by local measurement errors or abnormal behavior of individual patients.

[0047] S1.3 Norm filtering and aggregation: Calculate the L2 norm of the gradient vector uploaded by each node, sort it by size, and aggregate the nodes after removing the nodes with the largest gradient changes; This method prevents extreme local drug strategy updates from disrupting the global model by removing nodes with the "largest gradient change". It is suitable for filtering large deviation model updates caused by incorrect omics measurements or data anomalies. The aggregation process is as follows: 1. First, calculate the L2 norm of the gradient vector uploaded by each node and sort them by size:

[0048] : Local gradient The L2 norm of represents the magnitude of the center's policy change. Larger values ​​indicate a greater deviation from the current global policy. For example, if a center prescribes an excessive dose of EPO to a patient with an abnormally high level of secondary symptoms, this may cause the model update direction to deviate significantly from the mean, resulting in a large L2 norm.

[0049] 2. Then take the one with the smallest norm Node gradients, calculate their average value:

[0050] S1.4 Weighted Fusion: The three types of aggregation results obtained by the three aggregation methods (Krum, CwTM, CGE) mentioned above, namely geometric center aggregation, coordinate trimmed mean aggregation and norm filtering aggregation, are linearly combined according to preset weights as the global update direction to obtain an improved robust aggregation function. :

[0051] in, : The final robust aggregation function output represents the update direction of the global medication strategy constructed in the current round, which is used to adjust the model parameters (such as the weight in the EPO dose recommendation function). Its input is the vector set to be aggregated ; : Geometric center aggregation The weighting coefficient of , a high weight indicates the overall consistency of the system trust gradient direction; : Coordinate trimmed mean The weighting coefficient is suitable for fine processing of high-dimensional features (such as multi-omics data); : Norm filtering aggregation The weighting coefficient of , a high weight means that the system pays more attention to eliminating "radical nodes" with large gradient fluctuations; : Ensure that the combination of the three aggregation methods (Krum, CwTM, CGE) is a normalized linear fusion, , , They can be set based on experience or dynamically adjusted based on historical aggregation effects, or set at the initial stage 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].

[0052] S2. Based on the improved robust aggregation function, momentum enhancement and two-step bucket aggregation strategy are used to smooth gradient updates, combined with the nearest neighbor hybrid mechanism to defend against abnormal nodes; The distributed optimization process for building initial medication models faces challenges such as heterogeneous multi-center medical data, unstable node communication, and interference from anomalous patient samples in local models. Therefore, this paper proposes a federated optimization framework based on momentum enhancement and robust aggregation. This framework combines dynamic sampling, a two-layer aggregation structure, and a decentralized robust mechanism to improve model stability and accuracy.

[0053] S2.1 Constructing momentum enhancement and robust aggregation mechanisms In real medical scenarios, due to the differences in the quality of patient data collected by each node (such as different hospitals), local stochastic gradient It will show high variance and unstable characteristics. To enhance the robustness of the model, the Polyak momentum method is introduced to perform weighted smoothing on local updates, and local fluctuations are suppressed by weighted averaging of historical gradients:

[0054] : The t-th momentum vector of node i, used to accumulate historical directions and alleviate gradient oscillation; : Momentum decay coefficient (usually β≈0.9), which controls the degree of historical information retention; : is a node i The local stochastic gradient calculated for the current mini-batch sample, specifically the local stochastic gradient calculated by client i in round t, reflects the current feedback of the local patient data to the model.

[0055] Through the momentum mechanism, not only the noise of local fluctuation gradient is suppressed, but also the convergence performance under non-convex loss structure is improved.

[0056] Use improved robust aggregation functions when updating global models Extracting trusted information from multiple client momentum:

[0057] : current global model parameters; : Improved robust aggregation function , see step S1 for details; : learning rate (0.001~0.01, decay).

[0058] S2.2 Constructing a two-step aggregation and nearest neighbor hybrid strategy When there are many nodes (such as multi-center hospitals), in order to further defend against local malicious updates, a two-layer aggregation strategy is introduced, namely a two-step bucket aggregation strategy that smooths gradient updates and combines it with a nearest neighbor hybrid mechanism to defend against abnormal nodes. The two-step bucket aggregation strategy first randomly groups nodes to calculate the intra-group consensus, and then performs robust aggregation on the inter-group results. Specifically: First, the medical nodes are grouped by bucketing, and n clients are randomly divided into S buckets (Bucket(1), ..., Bucket(S)). Each bucket contains approximately nodes (clients); Then, for each bucket s , perform intra-bucket averaging through intra-group aggregation, calculate the average medication gradient (local mean) of the nodes in each bucket, and form a bucket consensus vector:

[0059] : The local gradient vector uploaded by node i, representing 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 reactivity, etc.); : No. s The set of nodes contained in the bucket; : No. s The average result of the local strategies of all nodes in the bucket is used to express the consensus of the group on the current medication strategy; b: the number of nodes in each group, calculated by the formula: , represents the number of dialysis centers included in each group; n: the total number of nodes, that is, the number of all dialysis medical institutions involved in the modeling; S: the number of buckets, n nodes are divided into S buckets to prevent a small number of abnormal nodes from interfering with the overall update results; for example, grouping by hospital region, grouping by equipment supplier type, or stratifying and bucketing by patient characteristics, etc.

[0060] Represents client i preparing to upload information for aggregation in round t. In the basic framework, , in the first step of the two-step bucketing, Grouped bucket consensus vectors .

[0061] Finally, yes S Bucket consensus vector Apply the robust aggregation function F to perform robust aggregation on the bucket consensus vector to achieve inter-group aggregation (i.e., secondary aggregation, re-aggregation of the bucket mean), and obtain the final global aggregation result. :

[0062] in, : Improved robust aggregation function, see step S1.4 for details, for each bucket's "bucket mean Aggregate again to ensure that the final output strategy can resist the influence of extreme values ​​between buckets. When aggregating between groups, the internal Krum, CwTM, and CGE operations are based on vector carried out, and The dimension is the same as the dimension, when The input is (bucket consensus vector), n is replaced by S (number of barrels).

[0063] This two-step bucketing strategy aims to dilute the impact of potential abnormal nodes through the first layer (intra-bucket) aggregation, and then further enhance the defense capability against malicious bucket consensus or calculation errors through the second layer (inter-bucket) robust aggregation F. The above two-step aggregation strategy dilutes the impact of malicious nodes by grouping, which is suitable for nodes with a large number of nodes. Larger scene. The design itself is universal. Regardless of whether the input is a node vector or a bucket consensus vector, the calculation logic of the three internal methods remains unchanged (sorting, pruning, filtering and other operations are performed based on the input vector set).

[0064] Nearest Neighbor Mixing (NNM) uses the "geometric proximity" metric to remove outliers and construct a more robust local perspective aggregation. This mechanism uses the nearest neighbor mixing mechanism to filter neighbors, selecting multiple neighboring nodes with the most similar feature spaces for each node. The gradients of these neighboring nodes are aggregated to form the final update strategy. The feature space refers to the client's local model parameter vector space or the gradient / momentum vector space uploaded by the client. Specifically: At each node i, based on the Euclidean distance (or cosine similarity) between its uploaded gradient vector and the gradient vectors of other nodes, the nearest k' neighbor nodes (k' is a preset constant, such as 5-10) are selected to perform local blending:

[0065] : The local policy gradient vector of node j, representing the local model parameter update vector of neighbor node j (such as gradient or momentum ), indicating the medication adjustment recommendations (e.g., the direction of adjusting the EPO or iron dosage) calculated by the dialysis center based on its own patient population (e.g., Hb level, TSAT, secondary symptom factors, CYP450 gene expression, etc.); : The set of nearest neighbor nodes of node i, consisting of nodes with the smallest geometric distance or cosine similarity to its strategy vector in the feature space (such as drug sensitivity weight) other nodes; represents a group of dialysis centers that are most similar to node i in terms of patient composition, pathological characteristics or treatment mode; similarity calculation can be based on vectors such as "Hb-TSTAT-SF three-dimensional features", "multi-omics translation expression factors", and "individual dose response trends". The similarity between node i and node j is calculated by calculating the gradient / momentum vector to be uploaded in this round. and It is measured by the Euclidean distance (or cosine similarity). : The final aggregation strategy of node i, that is, the local medication adjustment direction obtained by averaging the recommendations of the nearest neighbor nodes.

[0066] S2.3 Optimize asynchronous and decentralized architecture In the edge medical computing scenario, a decentralized optimization strategy is adopted to alleviate the server centralization bottleneck and single-point failure risk. The algorithm maintains Byzantine fault tolerance capability in sparse communication topology through asynchronous communication based on coordinate descent, combined with local neighbor aggregation and robust mechanism, allowing no more than 1 / 3 node failure. It is suitable for deployment on remote medical terminals or low-power edge devices. In the decentralized communication topology, each node i only exchanges information with its topological neighbors or a few nearest neighbors selected according to feature similarity, and performs NNM aggregation locally (i.e., calculates the average of neighbor information) to achieve local model update. This mechanism uses the similarity between nodes for collaborative learning and suppresses the influence of a small number of abnormal neighbors through local aggregation.

[0067] 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).

[0068] In high-latency networks or asynchronous node environments (such as mobile terminals collecting data for uploading), traditional synchronous aggregation is easily affected by lag. Therefore, delay gradient filtering and timestamp mechanism are introduced, based on asynchronous communication mechanism, 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, filtering outdated gradient data, it can effectively filter long-term outdated or tampered node information, suitable for medical systems with high communication cost and unstable device online frequency in actual deployment.

[0069] Timestamp verification: Each gradient update information is attached with 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 ΔT is set according to the maximum expected delay of the network.

[0070] 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 the improved robust aggregation function, and adjust the update amplitude according to the dynamically decaying learning rate, finally output the global drug parameter vector, including the drug category recommendation weight and the dose adjustment coefficient.

[0071] 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: 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:

[0072] : 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 previous 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 number of i-th node : The number of i-th node : 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.

[0073] The global medication parameter vector θ is the weight parameter of the global prediction model obtained by 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.

[0074] For example, in one embodiment, the initial medication decision module 130 simultaneously outputs parameters of multiple types of drugs, such as EPO, iron, vitamin D, and phosphate binders, covering the entire path of anemia treatment, and the coefficient output is compatible with individualized calculations (such as combined with body weight and meal frequency): (1) EPO drug dose coefficient: Drug type: recombinant human erythropoietin (such as epoetin α, darbepoetin α, methoxy polyethylene glycol epoetin β); parameter output format: EPO_dose coefficient = 0.85 (indicating that the recommended dose for this patient is 85% of the conventional dose), combined with the patient's weight to calculate the actual dose: recommended dose = standard dose (such as 100U / kg) × dose coefficient × current weight. For example: if the patient weighs 60kg, the actual dosage = 100 × 0.85 × 60 = 5100U / week. (2) Intravenous iron type weight and dose coefficient: Among them, the weight of the iron type: the weight corresponding to sucrose iron is 0.92, the weight corresponding to sodium gluconate iron is 0.15, and the weight corresponding to carboxymaltose iron is 0.08. Iron dosage adjustment parameter: 1.2 (indicates a 20% increase in dosage), actual dosage = standard dosage (e.g., 100 mg / week) × dosage coefficient → 120 mg / week. (3) Auxiliary drug intervention parameters (for complication risk): Active vitamin D (e.g., calcitriol, paricalcitol): Vitamin D activation marker = 0.75 (trigger administration when > 0.5); Recommended dosage: dynamically adjusted based on calcium-phosphorus product (Ca×P), e.g., when the marker > 0.7 and Ca×P > 55 mg 2 / dL 2 During dialysis: give calcitriol 0.25 μg / time, 3 times a week; Phosphate binders (such as sevelamer, lanthanum carbonate): Phosphate binder_dose coefficient = 0.6 (60% of the regular dose), dosing rule: take with meals, dose = 500 mg × coefficient × number of meals per day. (4) Anticoagulant adjustment parameters (for dialysis process): Low molecular weight heparin (such as enoxaparin): Anticoagulant_dose coefficient = 1.3 (increase the dose by 30% during dialysis), adjustment basis: activation coefficient when hyperkalemia (blood potassium > 5.5 mmol / L), actual dose: standard dose 4000 U × 1.3 = 5200 U / dialysis.

[0075] The drug dosage adjustment module 140 is used to dynamically optimize the time nodes of drug administration and the timing of dosage adjustment through an asynchronous federated reinforcement learning framework combined with a forward-looking parameter correction strategy.

[0076] 1. Dose adjustment rules based on real-time hemoglobin (Hb), ferritin (SF), and transferrin saturation (TSAT) indicators are used as clinical prior knowledge to construct reinforcement learning reward functions. The security constraint boundary (such as ); wherein the dosage adjustment rule is: Based on the patient's real-time hemoglobin (Hb), ferritin (SF), and transferrin saturation (TSAT) indicator data, the drug dosage is dynamically adjusted according to the following rules: When hemoglobin is less than 100 g / L, ferritin is less than 200 μg / L, and transferrin saturation is less than 20%, intravenous iron supplementation (iron sucrose 100 mg / week) and erythropoietin 50-150 U / kg / week are initiated; When hemoglobin is not less than 130 g / L, ferritin is not less than 300 μg / L, and transferrin saturation is not less than 30%, erythropoietin (EPO) and iron therapy should be suspended. Adjust the dosage based on blood drug concentration and electrolyte levels (e.g., adjust the anticoagulant dosage in case of hyperkalemia).

[0077] Rules based on Hb / SF / TSAT thresholds are the basis for clinical prior knowledge or reward function design.

[0078] 2. Optimization of drug delivery nodes using artificial intelligence algorithms To improve the individualization, dynamism, and safety of drug administration strategies, an intelligent optimization framework based on asynchronous federated reinforcement learning (AFedPG) was constructed to optimize the timing of drug administration and the timing of dose adjustments, especially for the fine-grained regulation of half-life-related drugs such as erythropoietin (EPO) and intravenous iron preparations (such as iron sucrose).

[0079] Furthermore, the drug dosage adjustment module 140 uses an asynchronous federated reinforcement learning framework and a forward-looking parameter correction strategy to dynamically optimize the timing of drug administration and the timing of dosage adjustment, including: (1) Constructing a delay-adaptive forward-looking mechanism: Generate forward-looking correction parameters based on the historical trend of strategy parameter changes and the degree of delay, and dynamically adjust the compensation intensity through adaptive factors to ensure the consistency of clinical efficacy in an asynchronous environment; Due to the feedback delay between different medical nodes (such as different dialysis centers), a delay compensation mechanism needs to be introduced to maintain training stability. In the kth global parameter update, a forward-looking parameter correction strategy is adopted to compensate for the delay. The historical deviation of the parameter update direction caused by the first-order extrapolation compensation is performed:

[0080] : Add predicted corrected parameters to offset the drug administration strategy errors caused by node upload delays and ensure reasonable adjustments within the pharmacokinetic time range. : The global medication policy parameter sent by the server at the kth update, which represents the unified decision-making strategy of all nodes (such as the weekly medication frequency). : The global strategy parameters updated in the previous round are used to estimate the changing trend of the drug administration strategy; Represents the direction of the most recent parameter update. : With delay The relevant adaptive factor reflects the impact of data upload on strategy optimization, 0 < α ≤ 1, ,λ is a hyperparameter that controls the decay rate and the delay compensation strength, usually set to λ=0.1, which can be adjusted through cross-validation; The bigger, The smaller. : The degree of delay in the historical information on which the current strategy is calculated (such as the update time lag of laboratory data). : Deviation compensation coefficient, dynamically adjusts historical parameter changes according to delay. When it is small, the compensation amplitude Increase, for the direction of the most recent update Zoom in to predict the current update trend and try to offset the strategic lag effect caused by delays.

[0081] Through this mechanism, when the collection of patient indicators is delayed, we can still rely on the prediction strategy to make approximately optimal decisions and effectively deal with situations such as not testing on weekends or holidays.

[0082] To further offset the second-order gradient deviation caused by the delay, the following conditions are met:

[0083] : The second-order derivative of the objective function, which characterizes the sensitivity of changes in drug administration strategy to the expected therapeutic effect (such as Hb target rate).

[0084] This formula ensures that the strategy correction of different medical points in an asynchronous environment can maintain clinical convergence and consistency of efficacy, and avoid gradient deviation caused by inconsistent update time of each node.

[0085] (2) Building an asynchronous federated learning framework (AFedPG) The training process for drug delivery node optimization includes: calculating local policy gradients based on the probability derivatives and reward discounts of drug delivery actions in the trajectory; the server normalizes and integrates the received delayed gradients with momentum before updating global parameters; and dynamically distributing the updated drug delivery strategy to each node without waiting for all nodes to synchronize. It is divided into the following three steps: 1) Local computing stage: Definition: Status( ): Contains the patient's current and historical: hemoglobin (Hb), ferritin (SF), transferrin saturation (TSAT), blood drug concentrations (such as EPO trough concentration), key electrolyte levels (potassium, calcium, phosphorus), inflammatory factors (such as CRP), liver and kidney function indicators, multi-omic biomarker labels, and last medication records, etc.

[0086] action( ): 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 discrete action sets, such as {maintain, small increase in EPO, large increase in EPO, small decrease in EPO, large decrease in EPO, ...}, covering independent and combined adjustments of EPO and iron.

[0087] award( ): The design needs to balance treatment efficacy and safety, e.g.

[0088] Among them, w1~w7 are weight coefficients, which can be set to [1.0, 0.5, 0.3, 1.0, 1.0, 0.01,0.01]; is the target Hb value, is the maximum value of SF, is the maximum value of TSAT, which is the target value or safety threshold set clinically. is the event indicator function (1 if it occurs, 0 otherwise), where For hypotensive events (if drug-related), For hyperkalemia events; Indicates the current dosage, where EPO dosage (minimum effective dose is encouraged), Iron dosage (minimum effective dose encouraged); A fixed reward for achieving the goal. The specific value of the weight is determined by clinical expert experience or offline tuning.

[0089] Each medical node is based on the currently collected patient indicators (such as Hb, SF, TSAT) and historical strategies Generate trajectories , calculate the corresponding policy gradient :

[0090] in, Representation node iThe original policy gradient is calculated based on the trajectory μ and the parameter θ; μ: The patient's complete state-behavior-feedback trajectory, which records the changes in patient indicators and medication behavior during the continuous dialysis cycle; for example, the Hb value recorded once a week, whether EPO was given, the corresponding Hb increase or side effects, etc. : The patient state vector at time p, including real-time physiological and laboratory indicators; : In state The drug administration action (decision) taken under ”, “Intravenous iron supplementation 100 mg”, “Suspend iron supplementation”, etc.; :Under the current policy parameter θ, for the state Select Action The probability of ; represents the distributed probability model of medication decision-making, which gradually converges to the optimal medication strategy with training; : The derivative of the strategy probability with respect to the parameter is used to guide the direction of parameter update; when a certain medication behavior produces a good therapeutic effect in a certain type of patient state, its probability of occurrence will be amplified; : In state Next action The immediate rewards obtained can be defined as: Hb target level (positive), adverse reactions (negative), inflammation suppression effect, etc. : Reward discount factor (0.95~0.99) to control the trade-off between short-term and long-term effects (such as choosing 0.95); avoid focusing only on current Hb indicators and ignoring long-term iron overload or cardiovascular risks.

[0091] When the status When the safety limit is violated (such as Hb < 80g / L), the A large negative value, such as -10, ensures that the policy network avoids dangerous decision areas.

[0092] Policy Network Implemented using a three-layer fully connected neural network: Input layer: state Dimension (such as 256 dimensions); hidden layer: 128 neurons, ReLU activation; output layer: discrete action: Softmax outputs the probability of each action; continuous action: Tanh activation outputs the dosage adjustment coefficient of [-1,1].

[0093] 2) Server update phase: The server receives delayed gradients from different nodes , normalized and updated according to the following formula:

[0094] : Learning rate (0.0001~0.001, decaying), controls the adjustment amplitude of the drug administration strategy in each round, and decays over time. : The delay gradient received by the server, from The node calculation result of the iteration reflects the gradient direction of the historical strategy. : L2 norm of the gradient to avoid policy fluctuations caused by large gradients.

[0095] During parameter update Before, optionally, for the current parameter Or the momentum state d applies the forward-looking parameter modification strategy to compensate for the delay in the information on which the update is based.

[0096] To improve stability, the momentum mechanism is introduced to construct the fusion gradient direction:

[0097] : Historical aggregation gradient, used to suppress drastic fluctuations in drug administration adjustments. : The policy gradient corresponding to the current patient trajectory is used to update the current policy direction. The expression is shown above The server maintains a global momentum state d and when it receives the delayed raw policy gradient from node i, it then uses the updated d for parameter update.

[0098] 3) Downward strategy push: The server will update the policy parameters It is sent to each node to achieve closed-loop control, which does not rely on the synchronization of all nodes and can avoid the problem of "the slowest node slowing down training".

[0099] The asynchronous federated reinforcement learning framework (AFedPG) is the core optimization tool of this invention. Based on real-time monitoring indicators (Hb, SF, TSAT, etc.), the drug dosage adjustment module 140 dynamically calculates and outputs the optimal drug dosage adjustment decision (including dosage, dosing time, and adjustment timing) through an intelligent optimization framework based on asynchronous federated reinforcement learning (AFedPG).

[0100] 3. Risk prediction-driven early strategy revision A complication risk prediction model embedded in the training framework uses inflammatory factors (such as CRP), calcium and phosphorus levels, and electrolytes as inputs to predict the likelihood of cardiovascular events and metabolic disorders. When the predicted risk increases, the dosage of EPO and iron supplements is automatically adjusted in advance, or adjunctive medications (such as active vitamin D to inhibit calcium and phosphorus absorption) are introduced as appropriate, enabling proactive strategic intervention.

[0101] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the methods.

[0102] It should also be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device comprising the elements.

[0103] The above description of the disclosed embodiments enables those 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 may 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 rather will conform to the widest scope consistent with the principles and novel features disclosed in the embodiments of the present application.

Claims

1. A precise medication system for hemodialysis patients based on multi-omics and big data, characterized by: The following steps are involved: A multi-source data acquisition module is used to collect, standardize, and integrate multi-source heterogeneous data from patients to build a patient-level data warehouse; wherein the multi-source heterogeneous data includes: basic clinical data, multi-omics test data, and real-time physiological and laboratory indicators; The drug metabolism analysis module is used to analyze drug metabolism pathways and individual differences through multi-omics data, build drug metabolism polymorphism models to predict patients' ability to respond to drugs, and output gene metabolism capacity classification and biomarker risk labels; The operations performed by the drug metabolism analysis module include: Analyze the in vivo metabolic pathways of key drugs based on the patient's gene expression profile, metabolite concentration, and proteomic information to generate personalized pharmacokinetic parameters, including drug half-life and peak concentration; Identify biomarkers and establish metabolic capacity grading indicators through analysis of drug metabolism-related gene variations; The metabolic capacity grading index and biomarker risk labels are integrated into the feature space of federated learning to constrain the recommended dosage range of the initial medication model. The initial medication decision module is used to build an initial medication decision model based on a federated optimization framework with momentum enhancement and robust aggregation mechanism, and output a personalized initial medication plan including erythropoietin and intravenous iron; The construction of the initial medication decision model includes: An improved robust aggregation rule is constructed. Local model gradients are aggregated based on three types of aggregation methods and then weighted fusion is performed to construct an improved robust aggregation function, including: The three types of aggregation results obtained by the three aggregation methods of geometric center aggregation, coordinate trimmed mean aggregation and norm filtering aggregation are linearly combined according to the preset weights as the global update direction to obtain an improved robust aggregation function : in, : The final robust aggregation function output represents the update direction of the global medication strategy constructed in the current round and is used to adjust the model parameters; :node The local gradient vector of n: the total number of participating nodes, that is, the number of medical institutions participating in federated learning; : Geometric center aggregation The weighting coefficient of : Coordinate trimmed mean aggregation The weighting coefficient of : Norm filtering aggregation The weighting coefficient of Based on the improved robust aggregation function, momentum enhancement and a two-step bucket aggregation strategy are used to smooth gradient updates, combined with a nearest neighbor hybrid mechanism to defend against abnormal nodes. Based on the patient characteristics of each medical node, the historical momentum gradient and current gradient of trusted neighbor nodes, an improved robust aggregation function is used to filter abnormal updates and the parameter update amplitude is adjusted according to the dynamically decayed learning rate. Finally, a global medication parameter vector is output, including the recommended weight of drug category and the dosage adjustment coefficient. The drug dosage adjustment module is used to dynamically optimize the timing of drug administration and dosage adjustment through an asynchronous federated reinforcement learning framework combined with a forward-looking parameter correction strategy.

2. The multi-omics and big data-based precise medication system for hemodialysis patients according to claim 1 is characterized in that: in, The improved robust aggregation rule is constructed by aggregating local model gradients based on three types of aggregation methods and then weighted fusion, thereby constructing an improved robust aggregation function, including: Geometric center aggregation: select the gradient vector with the smallest distance to the medication gradient of most nodes; Coordinate trimmed mean aggregation: Each coordinate is independently sorted by coordinate trimmed mean and the extreme values ​​are trimmed and then the average is taken; Norm filtering and aggregation: Calculate the L2 norm of the gradient vector uploaded by each node, sort it by size, and aggregate after removing the node with the largest gradient change; The three types of aggregation results, namely geometric center aggregation, coordinate trimmed mean aggregation and norm filtering aggregation, are linearly combined according to preset weights as the global update direction to obtain an improved robust aggregation function.

3. The multi-omics and big data-based precise medication system for hemodialysis patients according to claim 2 is characterized in that: in, The federated optimization framework based on momentum enhancement and robust aggregation mechanism includes: The Polyak momentum mechanism is introduced to suppress local fluctuations through weighted averaging of historical gradients; A two-step bucket aggregation strategy is adopted to smooth gradient updates and combined with a nearest neighbor hybrid mechanism to defend against abnormal nodes; wherein, the two-step bucket aggregation strategy includes: first randomly grouping nodes to calculate the intra-group consensus, and then performing robust aggregation on the inter-group results.

4. The multi-omics and big data-based precise medication system for hemodialysis patients according to claim 3 is characterized in that: in, The two-step bucket aggregation strategy is used to smooth gradient updates and combined with the nearest neighbor hybrid mechanism to defend against abnormal nodes, including: Group the medical nodes and randomly divide them into multiple buckets, each containing several nodes. Calculate the average medication gradient of the nodes in each bucket to form a bucket consensus vector; Perform robust aggregation on the bucket consensus vector to achieve inter-group aggregation and generate a global update direction; Neighbors are screened based on the nearest neighbor hybrid mechanism, selecting multiple neighbor nodes with the most similar feature spaces for each node; Aggregate the medication gradients of neighboring nodes to form the final update strategy.

5. The multi-omics and big data-based precise medication system for hemodialysis patients according to claim 4 is characterized in that: in, 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 the asynchronous communication mechanism, nodes only exchange gradient information with neighboring nodes. Timestamp verification is introduced to filter out gradient data that has timed out and not been updated. Byzantine fault tolerance is maintained in a 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 according to claim 5 is characterized in that: in, The initial medication decision module performs global parameter adjustment 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 supplements. The initial medication decision module performs global parameter adjustment based on the momentum gradient mean of the trusted neighbor nodes and generates a personalized medication parameter vector, specifically including: For each medical node, aggregate the historical momentum gradient and current gradient of its trusted neighbor nodes; The aggregation results of all nodes are processed through weighted robust aggregation function; The learning rate is dynamically adjusted to generate an updated medication parameter vector, wherein the medication parameter vector includes a drug category recommendation weight or a dosage coefficient.

7. The multi-omics and big data-based precise medication system for hemodialysis patients according to claim 1 is characterized in that: The drug dosage adjustment module uses dosage adjustment rules based on real-time hemoglobin, ferritin, and transferrin saturation indicators as clinical prior knowledge to construct the safety constraint boundary of the reinforcement learning reward function; wherein the dosage adjustment rules are: When hemoglobin is less than 100 g / L, ferritin is less than 200 μg / L, and transferrin saturation is less than 20%, intravenous iron supplementation and erythropoietin 50-150 U / kg / week are initiated; When hemoglobin is not less than 130 g / L, ferritin is not less than 300 μg / L, and transferrin saturation is not less than 30%, erythropoietin and iron therapy should be suspended; Adjust the dosage based on blood drug concentration and electrolyte levels.

8. The multi-omics and big data-based precise medication system for hemodialysis patients according to claim 1 is characterized in that: The drug dosage adjustment module uses an asynchronous federated reinforcement learning framework and a forward-looking parameter correction strategy to dynamically optimize the timing of drug administration and dosage adjustment opportunities. Generate forward-looking correction parameters based on historical strategy parameter change trends and delay levels, and dynamically adjust compensation intensity through adaptive factors to ensure consistency of clinical efficacy in an asynchronous environment. The local policy gradient is calculated based on the probability derivative and reward discount of the drug administration action in the trajectory. The server normalizes and momentum-fuse the received delayed gradient and then updates the global parameters. The updated drug administration policy is dynamically distributed to each node without waiting for all nodes to synchronize.

Images