Multi-terminal training method of acidosis identification model, electronic equipment and storage medium

Abstract

The invention provides a multi-terminal training method of an acidosis identification model, electronic equipment and a storage medium. The method comprises the following steps: receiving a global model sent by a central server node, wherein the global model is divided into a plurality of sub-modules according to the cause type of metabolic acidosis; according to the etiological type at which each secondary client node is skilled, distributing a corresponding sub-module to each secondary client node; distributing the distributed sub-modules to the corresponding secondary client nodes, so that each secondary client node performs local training on the received sub-modules by using the corresponding local dominant pathogenesis data; receiving sub-module parameters which are returned by each secondary client node and are subjected to local training; performing aggregation processing on each received sub-module parameter to obtain a local aggregation model parameter of the first-level client node; and the local aggregation model parameters are fed back to the central server node for updating the global model, so that the problems in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of acid-base imbalance detection, and in particular to a multi-end training method of an acidosis identification model, an electronic device and a storage medium. BACKGROUND

[0002] Metabolic acidosis is a common acid-base imbalance in clinical practice. The core features of this disease are the primary reduction of plasma bicarbonate (HCO3 - ) and arterial blood pH < 7.35. The main harm is the impairment of multiple system functions and high risk of severe illness, such as easy disturbance of energy metabolism in the nervous system, leading to consciousness disorders, lethargy, and even coma.

[0003] With the development of artificial intelligence technology, large models based on deep learning (such as Transformer and CNN architecture) can analyze blood gas analysis, biochemical indicators, electronic medical records and other clinical data to achieve automatic identification of metabolic acidosis, thereby providing support for clinical decision-making. Currently, the training of metabolic acidosis identification models generally uses a centralized training mode. In this centralized training mode, each medical institution (such as a hospital or a testing center) uploads the patient data (including blood gas indicators, etiology labels, and biochemical results) collected locally to a server, which integrates all the data and then trains the model uniformly.

[0004] However, the core limitation of the existing centralized training mode is that it cannot effectively adapt to the characteristics of the institutional-specific distribution of medical data. That is, due to differences in diagnosis and treatment range, regional disease spectrum, and other factors, different medical institutions naturally accumulate specialized advantage etiology data (for example, primary hospitals focus on gastrointestinal HCO3 - missing cases, and third-grade hospitals rich in lactic acidosis critical illness data). In the current centralized training mode, the server collects patient data from each medical institution and integrates them for unified training. This process can dilute or mask the unique information of such specialized advantage etiology data, which cannot effectively drive the model to learn the deep features of the etiology, resulting in a decline in the model's ability to identify the institution's specialized etiology, and ultimately leading to a decrease in overall recognition accuracy. SUMMARY

[0005] The present application provides a multi-end training method of an acidosis identification model, an electronic device and a storage medium to solve the problem of low model recognition accuracy caused by the centralized training mode in the prior art.

[0006] In one aspect, the present application provides a multi-end training method of an acidosis identification model. The model training system includes a central server node, a plurality of first client nodes connected to the central server node, and a plurality of second client nodes connected to each first client node. The method comprises: receiving the global model sent by the central server node, wherein the global model is divided into a plurality of independently trainable sub-modules according to the etiology types of metabolic acidosis; allocating the corresponding sub-modules to each secondary client node according to the etiology type that each connected secondary client node is good at; distributing the allocated sub-modules to the corresponding secondary client nodes, so that each secondary client node locally trains the received sub-modules using the corresponding local advantage etiology data; receiving the sub-module parameters returned by each secondary client node after local training; performing aggregation processing on each received sub-module parameter to obtain local aggregation model parameters of the primary client node; feeding back the local aggregation model parameters to the central server node for updating the global model.

[0007] Preferably, the etiology types of metabolic acidosis include high anion gap type and normal anion gap type; and, the global model is divided into a plurality of independently trainable sub-modules according to the etiology types of metabolic acidosis, specifically including: the global model is divided into high anion gap type and normal anion gap type according to the etiology types of metabolic acidosis.

[0008] Preferably, before each secondary client node performs local training, the method further comprises determining the local advantage etiology data of each secondary client node by the following way: determining the etiology type that the secondary client node is good at; from the local historical etiology data within the preset time node, filtering out the local historical etiology data corresponding to the etiology type as the local advantage etiology data of the secondary client node.

[0009] Preferably, the method further comprises: when distributing the allocated sub-modules to the corresponding secondary client nodes, determining the corresponding reference training duration of each secondary client node; monitoring whether each secondary client node returns the sub-module parameters within the corresponding reference training duration; if any secondary client node does not return the sub-module parameters within the corresponding reference training duration, sending a fault detection prompt to the central server node and / or the secondary client node that does not return the sub-module parameters.

[0010] Preferably, the corresponding reference training duration of each secondary client node is determined by the following way: determine a parameter scale of a sub-module allocated to the secondary client node, and a performance parameter and a current workload of the secondary client node; calculate a reference training duration of the secondary client node through a predefined duration calculation model based on the parameter scale, the performance parameter and the current workload; wherein the duration calculation model satisfies that the parameter scale is positively correlated with the reference training duration, the performance parameter is negatively correlated with the reference training duration, and the current workload is positively correlated with the reference training duration.

[0011] Preferably, after determining the reference training duration corresponding to each secondary client node respectively, the method further comprises: distributing the reference training duration corresponding to each secondary client node respectively to the corresponding secondary client node, so that the secondary client node compares a previous round training time with the reference training duration, and in the case that the previous round training time is greater than the reference training duration, filters a preset proportion of local advantage cause data from the local advantage cause data to locally train the received sub-module.

[0012] Preferably, the preset proportion is determined by a ratio of the reference training duration to the previous round training time.

[0013] Preferably, the received each sub-module parameter is aggregated to obtain a local aggregation model parameter of the primary client node, and specifically comprises: classifying the received each sub-module parameter according to a cause type corresponding to each sub-module parameter respectively; performing weighted average aggregation on the sub-module parameters of the same cause type after classification to obtain a plurality of sub-module aggregation parameters respectively corresponding to different cause types; combining the plurality of sub-module aggregation parameters to constitute the local aggregation model parameter of the primary client node.

[0014] In the second aspect, the application further provides an electronic device, comprising: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method provided in the embodiments of the application.

[0015] In the third aspect, the application further provides a computer readable storage medium storing a computer program, and the computer program is executed by a processor to implement the method provided in the embodiments of the application.

[0016] The multi-end training method of the acidosis recognition model provided in the embodiments of the present application is used, and the model training system comprises a central server node, a plurality of first client nodes connected with the central server node, and a plurality of second client nodes respectively connected with each first client node. The method comprises receiving a global model sent by the central server node, the global model being divided into a plurality of independently trainable sub-modules according to the cause types of metabolic acidosis, then respectively distributing corresponding sub-modules to each second client node according to the cause types that each connected second client node is good at, then distributing the distributed sub-modules to the corresponding second client nodes, so that each second client node uses the corresponding local advantage cause data to locally train the received sub-modules, then receiving the sub-module parameters returned by each second client node after local training, then performing aggregation processing on each received sub-module parameter to obtain local aggregation model parameter of the first client node, and then feeding back the local aggregation model parameter to the central server node, thereby updating the global model. The method divides the global model into a plurality of independently trainable sub-modules according to the cause types of metabolic acidosis, and distributes corresponding sub-modules to each second client node according to the cause types that each connected second client node is good at, so that the second client node can use the local advantage cause data to locally train the sub-module. This way can prevent the loss of unique information features caused by the integration of the local advantage cause data of different second client nodes, thereby effectively driving the model to learn the deep features of the cause, improving the generalization ability of the model, and finally improving the accuracy of model recognition. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0018] Figure 1 The structural schematic diagram of the model training system provided by the present application; Figure 2 The flowchart of the multi-end training method of the acidosis recognition model provided by the present application; Figure 3 The structural block diagram of the acidosis recognition model multi-end training device provided by the present application; Figure 4 The structural schematic diagram of the electronic device provided by the present application. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in conjunction with the drawings in the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0020] As described above, the large model based on deep learning has been able to realize the automatic identification of metabolic acidosis. The model currently generally adopts a centralized training mode. In the current centralized training mode, the server collects patient data of each medical institution and integrates and uniformly trains after integration. This process will cause the unique information of such specialized advantage etiology data to be diluted or covered in the integration, and thus cannot effectively drive the model to learn the deep features of the etiology, resulting in the decline of the recognition ability of the model to the institutional specialized etiology, and ultimately causing the overall recognition accuracy to be reduced.

[0021] Therefore, the embodiments of the present application provide a multi-end training method of an acidosis identification model, an electronic device and a storage medium, which can be used to solve the problems in the prior art. For ease of understanding, the embodiments of the present application can be first described as a whole. The present application draws on the inventive concept of multi-end training (Federated Learning, also known as federated learning). The architecture of the overall model training system is a three-level structure of "central server-primary client node-secondary client node" as shown in Figure 1 The central server node 10 is connected to a plurality of primary client nodes 11, and each primary client node 11 is connected to a plurality of secondary client nodes 12.

[0022] In this architecture, the central server node 10 is used to generate or initialize a global model, and the global model is distributed (to the primary client node 11), and then the aggregated parameters from the primary client node 11 are received to complete the iteration of the global model. The central server node 10 does not directly contact the original data in the model training process, so that the privacy of the medical data can be ensured. The primary client node 11 serves as an intermediate scheduling layer, responsible for receiving the global model, decomposing the task and distributing it to the subordinate secondary client nodes 12, and collecting the local training results of the secondary client nodes 12 and performing parameter aggregation. The secondary client node 12 is a terminal unit actually participating in the training, deployed in a medical institution or department with specific clinical expertise, and has the ability to store and process local advantage etiology data. The method provided by the embodiments of the present application will be further described below.

[0023] In actual application, the second client nodes 12 connected by the same first client node 11 can be various medical departments under the same medical structure, and the first client node 11 is the medical structure, thereby serving as an aggregation node for planning data of multiple departments; of course, the second client nodes 12 connected by the same first client node 11 can also be different medical institutions (such as community hospitals, county hospitals, and specialized hospitals) in the same geographical region (provinces, cities, etc.), and the first client node 11 is a client node of the region, thereby assuming the functions of gathering and transferring data of multiple medical institutions in the geographical region. In addition, considering that some large medical structures (such as chain hospitals), each branch can be regarded as the second client node 12, and the general hospital of the large medical structure can be regarded as the first client node 11, thereby planning the aggregation of data of each branch.

[0024] For example, as an implementation manner, the central server node can be a high-performance computing instance on a cloud service platform, the first client node can be a server device of a provincial medical data center, and the second client node can be a terminal workstation of each medical institution. For example, when the global model is divided into sub-modules for different etiology types, the second client node skilled in processing gastrointestinal HCO3 - The second client node (such as a medical information system terminal of a primary hospital) that loses cases receives and trains a corresponding sub-module, and optimizes using the local accumulated gastrointestinal related historical etiology data; at the same time, the second client node skilled in processing lactic acidosis critical data (such as a clinical workstation of a top three hospital) receives and trains another sub-module, and focuses on the local advantage data in its specialized field.

[0025] In the federated learning of the three-level structure in the present application, model training can be performed through multiple training cycles. For example, in the first training cycle, the central server node 10 distributes the global model of the first training cycle to each first client node 11, each first client node 11 respectively executes the method provided in the embodiments of the present application (which will be specifically described later), and finally feeds back the local aggregated model parameters to the central server node 10 for updating the global model; then the second training cycle is started, the central server node 10 distributes the global model of the second training cycle to each first client node 11, and so on. In this way, multiple training cycles are performed in sequence until the training termination condition is met. In each training cycle, the first client node 11 can execute the method provided in the embodiments of the present application.

[0026] For example, as an implementation manner, the central server node can be a high-performance computing instance on a cloud service platform, the first client node can be a server device of a provincial medical data center, and the second client node can be a terminal workstation of each medical institution. For example, when the global model is divided into sub-modules for different etiology types, the second client node skilled in processing gastrointestinal HCO3 Figure 2A specific flowchart of a multi-end training method of an acidosis identification model provided in the application is shown, which can be executed by any one of the first client nodes in the model training system, and the method comprises the following steps: Step S21: receiving the global model sent by the central server node, wherein the global model is divided into multiple independently trainable sub-modules according to the cause type of metabolic acidosis.

[0027] In actual application, a global model for acidosis identification can be generated or initialized by the central server node, which can adopt a Transformer or CNN architecture. Specifically, the global model is divided into multiple independently trainable sub-modules according to the cause type of metabolic acidosis, that is, the model structure is decomposed into independent components for different cause characteristics, so that these sub-modules are isolated from each other in the parameter space and can be independently trained in parallel on different devices without gradient interference.

[0028] In actual application, the implementation can be divided based on a pre-defined cause classification standard, for example, according to the specific cause types of lactic acidosis, ketoacidosis, HCO3 - loss, renal tubular acidosis, etc. Of course, the cause type can also include high anion gap type and normal anion gap type, so the global model is divided into high anion gap type sub-module and normal anion gap type sub-module. In an optional embodiment, the division granularity of the sub-module can also be further refined according to clinical needs, for example, according to specific causes rather than large cause types, such as setting up “diabetic ketoacidosis sub-module”, “alcoholic ketoacidosis sub-module”, etc., so as to realize more targeted sub-modules to further improve the recognition accuracy of the model for these specific causes.

[0029] Among them, the division method of the sub-modules in the global model can be realized by model pruning, attention mask, gating mechanism or modular network design, such as using the expert network in the MoE (Mixture of Experts) architecture as the sub-module, or guiding the input sample to the corresponding branch path through label-aware routing. In this way, each sub-module can maintain a certain knowledge transfer ability through the shared bottom feature extraction layer (such as convolutional backbone), while the upper decision part remains independent to ensure that it is not disturbed during training.

[0030] The central server node can send the global model to each connected primary client node after generating or initializing the global model, so that in this step S21, the primary client node can receive the global model sent by the central server node, such as obtaining the global model of the current round from the central server through a secure communication protocol.

[0031] Step S22: According to the cause type that each connected secondary client node is good at, respectively assign corresponding sub-modules to each secondary client node.

[0032] In this step S22, the primary client node can first determine the cause type that each secondary client node under it is good at, such as one way can determine the cause type that each secondary client node is good at through pre-configured mapping rules, for example, according to the specialized field declared by the secondary client node when registering to determine the cause type that it is good at, or through statistical analysis of the historical data distribution or professional type of the secondary client node, to determine the cause type that it is good at, for example, a secondary client node belongs to the endocrine department, then it is determined that it is good at high anion gap acidosis, if it belongs to the digestive department, it may be more skilled in dealing with diarrhea caused by normal anion gap acidosis; of course, in some dynamic scenarios, before this step S22, the primary client node can also send a query request to each secondary client node to request to obtain the cause type that it is good at.

[0033] After the primary client node determines the cause type that each secondary client node under it is good at, it can respectively assign corresponding sub-modules to each secondary client node according to the cause type that each secondary client node is good at, such as if the cause type that the secondary client node is good at is high anion gap type, it can assign the high anion gap type sub-module to the secondary client node, if the cause type that the secondary client node is good at is normal anion gap type, it can assign the normal anion gap type sub-module to the secondary client node, so as to realize special disease special training through this matching mode, so that the sub-module and its professional background of training data are more consistent.

[0034] Step S23: Distribute the assigned sub-modules to the corresponding secondary client nodes, so that each secondary client node uses the corresponding local advantage cause data to locally train the received sub-modules.

[0035] In actual application, in order to improve the security of data transmission, the secondary client node can send each sub-module parameter to the corresponding secondary client node through an encrypted transmission channel. The distributed content can include the network structure definition of the sub-module, the initialization weight and the necessary context information (such as the input format, the label coding rule), so as to ensure that the secondary client node can load and run after receiving it.

[0036] After receiving the corresponding sub-module, the secondary client node can use the corresponding local advantage etiology data to locally train the received sub-module. Specifically, the secondary client node can use the local advantage etiology data accumulated by itself and matched with the etiology type corresponding to the sub-module to train the sub-module in its local environment. For example, the secondary client node receiving the high anion gap type sub-module can use the local advantage etiology data matched with the high anion gap type, including the physiological indicators and laboratory examination results of patients with lactic acidosis, uremic acidosis and the like, to train the high anion gap type sub-module.

[0037] In the process of locally training the received sub-module by the secondary client node using the local advantage etiology data, the existing technical related model training method can be adopted, such as the standard machine learning process including forward propagation, loss calculation, backward propagation and parameter update. In this training process, there is no need to upload the original data to the central server node and the primary client node, only the updated parameters need to be fed back, so as to better protect the privacy.

[0038] Of course, by using the corresponding local advantage etiology data to locally train the received sub-module by the secondary client node, the unique information characteristics of the local advantage etiology data are retained compared with the current centralized training mode, which can effectively drive the model to learn the deep features of the etiology and improve the generalization ability of the model, and finally improve the accuracy of model recognition.

[0039] Of course, considering that each secondary client node has its own type of cause that is good at, but in the daily diagnosis and treatment process, both patients with the type of cause that is good at will be diagnosed and treated, thereby generating local advantage cause data, and patients with other types of causes will also be diagnosed and treated, so the data generated is non-local advantage cause data. Therefore, before each secondary client node performs local training, it usually also needs to determine its own local advantage cause data respectively. Specifically, the type of cause that the secondary client node is good at can be determined first (the specific method can refer to the above content), and after determining the type of cause that the secondary client node is good at, the local historical cause data corresponding to the type of cause can be screened from the local historical cause data in the preset time node as the local advantage cause data of the secondary client node. For example, if the secondary client node determines that the type of cause it is good at is high anion gap type, it can screen the local historical cause data corresponding to the high anion gap type from the local historical cause data in the preset time node as the local advantage cause data of the secondary client node. The preset time node can be 12 months, 24 months or other time nodes; the local historical cause data is obtained by the secondary client node in the preset time node when diagnosing and treating patients with different types of causes.

[0040] Step S24: receiving the sub-module parameters returned by each secondary client node after local training.

[0041] After each secondary client node uses the corresponding local advantage cause data to perform local training on the received sub-module, the secondary client nodes can return the sub-module parameters after local training to the primary client node, so that in this step S24, the primary client node can receive the sub-module parameters returned by each secondary client node after local training. The returned sub-module parameters can be a complete weight matrix, or an incremental gradient or a parameter update amount after differential privacy perturbation.

[0042] Step S25: performing aggregation processing on each received sub-module parameter to obtain the local aggregation model parameters of the primary client node.

[0043] After receiving the sub-module parameters returned by each of the subordinate secondary client nodes through the above step S24, the primary client node can further aggregate the received sub-module parameters to obtain the local aggregated model parameters of the primary client node in this step S25. In this application, considering that there are multiple sub-modules, these sub-module parameters cannot be directly aggregated, so the specific implementation of this step S25 can be that, first, according to the cause type corresponding to each sub-module parameter, the received sub-module parameters are classified, then the weighted average aggregation is performed on the sub-module parameters of the same cause type after classification, thereby obtaining multiple sub-module aggregated parameters corresponding to different cause types, and then the multiple sub-module aggregated parameters are combined to form the local aggregated model parameters of the primary client node. In the process of weighted average aggregation, the weights of each sub-module parameter can be the same or different, and can be specifically set according to actual needs, for example, the more data of the corresponding secondary client node (i.e. the data amount of the local dominant cause data), the greater the weight of its sub-module parameter.

[0044] For example, the global model includes sub-module 1, sub-module 2 and sub-module 3, and the primary client node has 7 subordinate secondary client nodes, which are secondary client node 1, secondary client node 2, …, and secondary client node 7. Among them, the secondary client node 1 and the secondary client node 2 are matched with the sub-module 1 because of the cause type they are good at, so the sub-module 1 is allocated to the secondary client node 1 and the secondary client node 2; similarly, the sub-module 2 is allocated to the secondary client node 3, the secondary client node 4 and the secondary client node 5; and the sub-module 3 is allocated to the secondary client node 6 and the secondary client node 7.

[0045] At this time, in the step S24, the secondary client nodes 1 to 7 return the sub-module parameters after local training, respectively referred to as sub-module parameters 1, sub-module parameters 2, …, and sub-module parameters 7. In this step S25, considering the differences between the sub-modules, the sub-module parameters cannot be directly aggregated. At this time, the received sub-module parameters can be classified according to the corresponding cause types, for example, the sub-module parameters 1 and 2 are classified as a class (referred to as the first class), the sub-module parameters 3, 4, and 5 are classified as a class (referred to as the second class), and the sub-module parameters 6 and 7 are classified as a class (referred to as the third class). Then, the sub-module parameters of the same cause type are aggregated by weighted average, for example, the sub-module parameters 1 and 2 of the first class are aggregated by weighted average to obtain a first sub-module aggregation parameter, the sub-module parameters 3, 4, and 5 of the second class are aggregated by weighted average to obtain a second sub-module aggregation parameter, and the sub-module parameters 6 and 7 of the third class are aggregated by weighted average to obtain a third sub-module aggregation parameter. Then, the multiple sub-module aggregation parameters are combined, that is, the first, second, and third sub-module aggregation parameters are combined to form a parameter package that is complete in structure and can be directly used to update the global model, as the local aggregation model parameter of the primary client node.

[0046] The multiple sub-module aggregation parameters are combined to splice or integrate the independent sub-module aggregation parameters according to the positions of the corresponding sub-modules in the original structure of the global model, to form a parameter package that is complete in structure and can be directly used to update the global model, as the local aggregation model parameter of the primary client node. The combination method can be implemented by parameter dictionary merging, layer sequence splicing, or modular loading, to ensure that the local aggregation model parameter output is topologically consistent with the global model sent by the central server.

[0047] Step S26: feeding back the local aggregation model parameter to the central server node for updating the global model.

[0048] After obtaining the local aggregation model parameter of the primary client node through the step S25, the local aggregation model parameter can be further fed back to the central server node in the step S26. After receiving the local aggregation model parameter, the central server node can be used to update the global model, for example, the central server aggregates the local aggregation parameters of the multiple primary client nodes by global aggregation (such as FedAvg) to generate a new global model and start the next round of training until the training termination condition is met.

[0049] In actual application, after receiving the local aggregation parameters of the plurality of primary client nodes, the central server can assign different weight coefficients to the local aggregation parameters of different primary client nodes when performing global aggregation on the local aggregation parameters of the plurality of primary client nodes. For example, the corresponding weight coefficients can be assigned according to the number of secondary client nodes under each primary client node, so that the weight coefficient of a primary client node with a larger number of secondary client nodes is larger, and the weight coefficient of a primary client node with a smaller number of secondary client nodes is smaller. Then, the local aggregation parameters of each primary client node are weighted and summed to realize the global aggregation, so as to generate a new global model and start the next round of training.

[0050] The multi-end training method of the acidosis identification model provided by the embodiment of the application is used. The model training system includes a central server node, a plurality of primary client nodes connected with the central server node, and a plurality of secondary client nodes connected with each primary client node respectively. The method includes receiving a global model sent by the central server node, the global model being divided into a plurality of independently trainable sub-modules according to the etiology types of metabolic acidosis, then according to the etiology types that each secondary client node connected is good at, respectively assigning corresponding sub-modules to each secondary client node, then distributing the assigned sub-modules to the corresponding secondary client nodes, so that each secondary client node uses the corresponding local advantage etiology data to locally train the received sub-modules, then receiving the sub-module parameters returned by each secondary client node after local training, then performing aggregation processing on the received each sub-module parameter to obtain the local aggregation model parameter of the primary client node, and then feeding back the local aggregation model parameter to the central server node, so as to update the global model. The method divides the global model into a plurality of independently trainable sub-modules according to the etiology types of metabolic acidosis, and assigns corresponding sub-modules to each secondary client node according to the etiology types that each secondary client node connected is good at, so that the secondary client node can use the local advantage etiology data to locally train the sub-module. This way can prevent the loss of unique information features caused by the integration of the local advantage etiology data of different secondary client nodes, and can effectively drive the model to learn the deep features of the etiology, improve the generalization ability of the model, and finally improve the accuracy of model identification.

[0051] It needs to be further explained that the secondary client node may fail during the local training of the received sub-module using the corresponding local advantage cause data, resulting in difficulty in returning the sub-module parameters to the primary client node in time. Therefore, the method provided by the embodiment of the present application can further provide a fault monitoring mechanism. The method can further include determining the reference training duration corresponding to each secondary client node when distributing the allocated sub-module to the corresponding secondary client node through the above-mentioned step S23, and then monitoring whether each secondary client node returns the sub-module parameters within the corresponding reference training duration. If all the secondary client nodes return the sub-module parameters, it means that no secondary client node fails. On the contrary, if any secondary client node does not return the sub-module parameters within the corresponding reference training duration, it means that the secondary client node may have failed. At this time, a fault detection prompt can be sent to the central server node and / or the secondary client node that does not return the sub-module parameters, so as to prompt the fault detection of the secondary client node.

[0052] In this fault detection method, how to determine the reference training duration corresponding to each secondary client node is crucial. For example, if the reference training duration is too small, it is easy to cause false judgment of failure. If the reference training duration is too large, it is easy to cause failure of the secondary client node to be detected in time, affecting the training efficiency of the model. In the embodiment of the present application, for each secondary client node, the reference training duration corresponding to the secondary client node can be determined by the following method: the parameter size of the sub-module allocated to the secondary client node, and the performance parameter and current workload of the secondary client node can be determined first. The parameter size refers to the number of trainable parameters or the model structure complexity of the sub-module allocated to the secondary client node for local training. In actual application, the parameter size can be in the form of neural network layer number, number of neurons per layer, convolution kernel size, total parameter quantity (in KB, MB), etc., which is used to quantify the computing intensity of the model training task. Therefore, a larger parameter size usually means higher computing demand and longer training time.

[0053] The performance parameter refers to the hardware processing capability index of the corresponding secondary client node, which usually includes CPU frequency, GPU model and computing power (such as TFLOPS), memory bandwidth, storage read-write speed, etc. Obviously, the secondary client node with higher performance parameter can complete more computing operations in unit time, and therefore has stronger training ability.

[0054] The current workload refers to the proportion of resources occupied by other computing tasks being executed by the secondary client node at the task distribution moment, which can include CPU utilization, memory occupancy, disk I / O frequency, network throughput, etc. In actual application, the higher the current workload of the secondary client node, the greater the possibility of squeezing the resources required for local training of the sub-module, thereby more easily leading to the extension of the training time.

[0055] After determining the parameter scale of the sub-module allocated to the secondary client node, and the performance parameter and the current workload of the secondary client node, the reference training duration of the secondary client node can be calculated based on the parameter scale, the performance parameter and the current workload, through a predefined duration calculation model. The duration calculation model needs to meet the following conditions: the parameter scale is positively correlated with the reference training duration, the performance parameter is negatively correlated with the reference training duration, and the current workload is positively correlated with the reference training duration. The duration calculation model can be, for example, as follows: T=k1×S / P+k2×L. In the formula, T is the calculated reference training duration; S is the parameter scale; P is the normalized performance score; L is the percentage of the current workload; k1 and k2 are preset adjustment coefficients (both greater than 0), which can be set according to actual needs in actual application. Of course, other duration calculation models can also be used according to actual needs.

[0056] Of course, after determining the reference training duration corresponding to each secondary client node, in addition to fault detection through the reference training duration, the method can further include distributing the reference training duration corresponding to each secondary client node to the corresponding secondary client node, so that the secondary client node compares the previous round training time with the reference training duration. At this time, if the previous round training time is less than or equal to the reference training duration, it means that the secondary client node has a relatively low possibility of delay in the process of local training in this round, so the local advantage cause data of the secondary client node can be directly used to train the received sub-module; otherwise, in the case where the previous round training time is greater than the reference training duration, it means that the secondary client node has a relatively high possibility of delay in the process of local training in this round. In order to avoid the delay caused by fault detection prompt (after sending the fault detection prompt, shutdown detection may be required, resulting in delay of model training), the preset proportion of local advantage cause data can be further selected from the local advantage cause data, and then the preset proportion of local advantage cause data is used to train the received sub-module. The preset proportion can be determined by the ratio of the reference training duration to the previous round training time, such as directly using the ratio as the preset proportion, and then reducing the data amount of training to reduce the time of this round of local training and the probability of fault detection prompt.

[0057] Based on the same inventive concept as the multi-end training method of the acidosis identification model provided in the embodiments of the present application, the embodiments of the present application can also provide a multi-end training device of an acidosis identification model. For the content in the device embodiments, reference can be made to the related content in the above method embodiments if it is not clear. The model training system includes a central server node, a plurality of primary client nodes connected with the central server node, and a plurality of secondary client nodes respectively connected with each primary client node. As shown in the figure, Figure 3 The specific structure diagram of the multi-end training device 30 (hereinafter referred to as device 30) of the acidosis identification model is shown. The device 30 includes a receiving unit 301, an allocation unit 302, a distribution training unit 303, a second receiving unit 304, an aggregation unit 305, and a feedback unit 306, wherein: The receiving unit 301 is configured to receive the global model sent by the central server node, wherein the global model is divided into a plurality of independently trainable sub-modules according to the cause types of metabolic acidosis; The allocation unit 302 is configured to allocate corresponding sub-modules to each secondary client node according to the cause types that each connected secondary client node is good at; The distribution training unit 303 is configured to distribute the allocated sub-modules to the corresponding secondary client nodes, so that each secondary client node locally trains the received sub-module by using the corresponding local advantage etiology data; The second receiving unit 304 is configured to receive the sub-module parameters returned by each secondary client node after local training. The aggregation unit 305 is configured to aggregate the received sub-module parameters to obtain local aggregation model parameters of the primary client node. The feedback unit 306 is configured to feed back the local aggregation model parameters to the central server node for updating the global model.

[0058] The device 30 provided in the embodiments of the present application can solve the problems in the prior art because the device 30 adopts the same inventive concept as the method provided in the embodiments of the present application. Therefore, the device 30 can solve the problems in the prior art, which will not be described here.

[0059] The etiology types of metabolic acidosis include high anion gap type and normal anion gap type. The global model is divided into a plurality of independently trainable sub-modules according to the etiology types of metabolic acidosis, and specifically includes: the global model is divided into high anion gap type and normal anion gap type according to the etiology types of metabolic acidosis.

[0060] Before the local training of each secondary client node, the method further includes determining the local advantage etiology data of each secondary client node by the following manner: Determining the etiology type in which the secondary client node is good at; From the local historical etiology data in the preset time node, the local historical etiology data corresponding to the etiology type is screened out as the local advantage etiology data of the secondary client node.

[0061] The device 30 can further include a fault detection unit configured to respectively allocate a corresponding reference training duration to each secondary client node when distributing the allocated sub-modules to the corresponding secondary client nodes; monitor whether each secondary client node returns the sub-module parameters within the corresponding reference training duration; and if any secondary client node does not return the sub-module parameters within the corresponding reference training duration, send a fault detection prompt to the central server node and / or the secondary client node that does not return the sub-module parameters.

[0062] The corresponding reference training duration is allocated to each secondary client node by the following manner: determine a parameter scale of a sub-module allocated to the secondary client node, and a performance parameter and a current workload of the secondary client node; calculate, based on the parameter scale, the performance parameter and the current workload, a reference training duration of the secondary client node through a predefined duration calculation model; wherein the duration calculation model satisfies that the parameter scale is positively correlated with the reference training duration, the performance parameter is negatively correlated with the reference training duration, and the current workload is positively correlated with the reference training duration.

[0063] wherein, after the respective reference training durations are respectively allocated to the secondary client nodes, the method further comprises: distributing the respective reference training durations of the secondary client nodes to the corresponding secondary client nodes, so that the secondary client nodes compare a previous round training duration with the reference training duration, and in the case that the previous round training duration is greater than the reference training duration, filter a preset proportion of local advantage cause data from the local advantage cause data to locally train the received sub-module.

[0064] wherein the preset proportion is determined by a ratio of the reference training duration to the previous round training duration.

[0065] wherein the received sub-module parameters are aggregated to obtain local aggregated model parameters of the primary client node, specifically comprising: classify the received sub-module parameters according to the respective cause types of the sub-module parameters; perform weighted average aggregation on the sub-module parameters of the same cause type after classification to obtain a plurality of sub-module aggregated parameters respectively corresponding to different cause types; combine the plurality of sub-module aggregated parameters to form the local aggregated model parameters of the primary client node.

[0066] Figure 4 An example of an entity structure diagram of an electronic device is shown as Figure 4As shown, the electronic device can include a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 complete communications with each other through the communications bus 440. The processor 410 can invoke a logical instruction in the memory 430 to execute a multi-end training method of an acidosis identification model provided by the embodiments of the present application. The model training system includes a central server node, a plurality of first client nodes connected to the central server node, and a plurality of second client nodes respectively connected to each first client node. The method includes receiving a global model sent by the central server node, the global model being divided into a plurality of independently trainable sub-modules according to the etiology type of metabolic acidosis, then respectively assigning each second client node with a corresponding sub-module according to the etiology type that each connected second client node is good at, then distributing the assigned sub-modules to the corresponding second client nodes, so that each second client node uses the corresponding local advantage etiology data to locally train the received sub-module, then receives the sub-module parameters returned by each second client node after local training, then aggregates each received sub-module parameter to obtain local aggregation model parameter of the first client node, and then feeds back the local aggregation model parameter to the central server node, thereby updating the global model. The method can prevent the loss of unique information features due to the integration of the local advantage etiology data of different second client nodes, thereby effectively driving the model to learn the depth features of the etiology, improving the generalization ability of the model, and ultimately improving the accuracy of model identification.

[0067] Obviously, since the processor 410 can invoke a logical instruction in the memory 430 to execute the method provided by the embodiments of the present application, the problems in the prior art can also be solved.

[0068] Further, the logic instructions in the memory 430 described above can be implemented in the form of software functional units and sold or used as standalone products, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0069] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the multi-end training method of the acidosis identification model provided in the embodiments of the present application. The model training system comprises a central server node, a plurality of first client nodes connected to the central server node, and a plurality of second client nodes respectively connected to each first client node. The method comprises receiving a global model sent by the central server node, the global model is divided into a plurality of independently trainable sub-modules according to the etiology type of metabolic acidosis, then according to the etiology type that each second client node connected is good at, respectively assigning corresponding sub-modules to each second client node, then distributing the assigned sub-modules to the corresponding second client nodes, so that each second client node uses the local advantage etiology data to locally train the received sub-modules, then receiving the sub-module parameters returned by each second client node after local training, then performing aggregation processing on each received sub-module parameter to obtain local aggregation model parameter of the first client node, and then feeding back the local aggregation model parameter to the central server node, thereby updating the global model. The method can prevent the loss of unique information features caused by the integration of the local advantage etiology data of different second client nodes, and can effectively drive the model to learn the deep features of the etiology, improve the generalization ability of the model, and finally improve the accuracy of model identification.

[0070] Obviously, since the computer program is executed by the processor, the computer can execute the method provided by the embodiments of the application, and thus the problems in the prior art can be solved.

[0071] In yet another aspect, the application further provides a non-transitory computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, performs the method provided by the embodiments of the application.

[0072] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiments according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0073] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some part of the embodiments.

[0074] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A multi-terminal training method for an acidosis identification model, characterized in that, The model training system includes a central server node, multiple primary client nodes connected to the central server node, and multiple secondary client nodes connected to each of the primary client nodes. The method includes: The system receives a global model sent by the central server node, wherein the global model is divided into multiple independently trainable sub-modules according to the etiological type of metabolic acidosis. Based on the types of diseases that each connected secondary client node is good at, the corresponding sub-modules are assigned to each secondary client node. The assigned sub-modules are distributed to the corresponding secondary client nodes, so that each secondary client node can use its local advantageous etiology data to perform local training on the received sub-modules. Receive the submodule parameters returned by each secondary client node after local training; The received parameters from each submodule are aggregated to obtain the local aggregated model parameters of the first-level client node; The local aggregated model parameters are fed back to the central server node for global model updates.

2. The method according to claim 1, characterized in that, The etiologies of metabolic acidosis include hyperanion gap type and normal anion gap type; and, The global model is divided into multiple independently trainable sub-modules based on the etiological type of metabolic acidosis, specifically including: the global model is divided into high anion gap type and normal anion gap type based on the etiological type of metabolic acidosis.

3. The method according to claim 1, characterized in that, Before each secondary client node performs local training, the method further includes determining the local dominant etiology data of each secondary client node in the following manner: Determine the etiological types that the secondary client nodes are good at; From the local historical etiology data within the preset time point, the local historical etiology data corresponding to the etiology type is selected as the local dominant etiology data of the secondary client node.

4. The method according to claim 1, characterized in that, The method further includes: When distributing the assigned sub-modules to the corresponding secondary client nodes, determine the reference training duration for each secondary client node; Monitor whether each secondary client node returns submodule parameters within the corresponding reference training duration; If any secondary client node fails to return the submodule parameters within the corresponding reference training time, a fault detection prompt is sent to the central server node and / or the secondary client node that failed to return the submodule parameters.

5. The method according to claim 4, characterized in that, For each secondary client node, the reference training duration is determined as follows: Determine the parameter size of the sub-modules allocated to the secondary client node, as well as the performance parameters and current workload of the secondary client node; Based on the parameter size, the performance parameters, and the current workload, the reference training duration of the secondary client node is calculated using a predefined duration calculation model; wherein, the duration calculation model satisfies the following: the parameter size is positively correlated with the reference training duration, the performance parameters are negatively correlated with the reference training duration, and the current workload is positively correlated with the reference training duration.

6. The method according to claim 4, characterized in that, After determining the reference training duration for each secondary client node, the method further includes: The reference training time corresponding to each secondary client node is distributed to the corresponding secondary client node, so that the secondary client node compares the previous round training time with the reference training time, and if the previous round training time is greater than the reference training time, it selects a preset proportion of local dominant etiology data from the local dominant etiology data to perform local training on the received sub-module.

7. The method according to claim 6, characterized in that, The preset ratio is determined by the ratio of the reference training duration to the training time of the previous round.

8. The method according to claim 1, characterized in that, The received parameters from each submodule are aggregated to obtain the local aggregated model parameters for the first-level client node, specifically including: The received parameters from each submodule are classified according to the etiological type corresponding to each parameter in each submodule. Weighted average aggregation of sub-module parameters for the same etiological type after classification is performed to obtain multiple aggregated parameters for sub-modules corresponding to different etiological types. The aggregation parameters of the multiple sub-modules are combined to form the local aggregation model parameters of the first-level client node.

9. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by a processor to implement the method described in any one of claims 1 to 8.

Images