Data detection and model training method and device, equipment and storage medium

Abstract

The invention provides a data detection and model training method and device, equipment and a storage medium, which can be applied to various fields of medical diagnosis, data detection and the like. The method comprises the following steps: acquiring target data and Q candidate data; respectively extracting first feature information of the target data and the Q candidate data; selecting at least one candidate data related to the target data from the Q candidate data based on the first feature information of the target data and the first feature information of the Q candidate data; on the basis of the first feature information of the at least one piece of candidate data, supplementing the first feature information of the target data to obtain supplemented feature information of the target data; and performing detection processing on the supplemented feature information of the target data to obtain a detection result of the target data. The features of similar objects are integrated, the feature information of the target object is enriched, and the data detection accuracy is improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technology, and in particular to a model detection and model training method, apparatus, device and storage medium. Background Art

[0002] With the rapid development of artificial intelligence (AI) and computer technology, various data detection models have been widely used in some fields. In the medical field, accurate data detection can improve medical diagnosis. For example, early cancer detection can significantly improve survival rates and reduce treatment costs, which has attracted widespread attention in academia and industry.

[0003] Current medical tests, such as early cancer detection, are usually based on specific cancer biomarkers or image analysis methods. However, the results of current detection methods are not accurate enough. Summary of the invention

[0004] The present application provides a data detection and model training method, apparatus, device and storage medium, which can improve the accuracy of data detection, for example, can obtain diagnostic reference results for various cancers.

[0005] In a first aspect, the present application provides a data detection method, comprising:

[0006] Obtain target data to be detected and Q candidate data, and extract first feature information of the target data and the Q candidate data respectively, wherein the target data is the state value of L biomarkers of the target object, and the candidate data is the state value of L biomarkers of other objects different from the target object, and Q and L are both positive integers greater than 1;

[0007] Based on the first feature information of the target data and the first feature information of the Q candidate data, selecting at least one candidate data related to the target data from the Q candidate data;

[0008] Based on the first feature information of the at least one candidate data, supplement the first feature information of the target data to obtain supplemented feature information of the target data;

[0009] The supplemented characteristic information of the target data is detected and processed to obtain a detection result of the target data.

[0010] In a second aspect, the present application provides a model training method, comprising:

[0011] Obtaining a training sample and Q candidate samples, and extracting first feature information of the training sample and the Q candidate samples respectively through a retrieval module in a data detection model, wherein the training sample is a state value of L biomarkers of a first subject, and the candidate samples are a state value of L biomarkers of other subjects different from the first subject, and both Q and L are positive integers greater than 1;

[0012] Based on the first feature information of the training sample and the first feature information of the Q candidate samples, selecting at least one candidate sample related to the training sample from the Q candidate samples;

[0013] Based on the first feature information of the at least one candidate sample, supplement the first feature information of the training sample to obtain supplemented feature information of the training sample;

[0014] Performing detection processing on the supplemented feature information of the training sample through the detection module in the data detection model to obtain the detection result of the training sample;

[0015] Based on the detection results of the training samples, the loss of the data detection model is determined, and the data detection model is trained based on the loss.

[0016] In a third aspect, the present application provides a data detection device, comprising:

[0017] an extraction unit, configured to obtain target data to be detected and Q candidate data, and extract first feature information of the target data and the Q candidate data respectively, wherein the target data is the state values ​​of L biomarkers of the target object, and the candidate data is the state values ​​of L biomarkers of other objects different from the target object, and both Q and L are positive integers greater than 1;

[0018] A selection unit, configured to select at least one candidate data related to the target data from the Q candidate data based on the first feature information of the target data and the first feature information of the Q candidate data;

[0019] A supplementing unit, configured to supplement the first feature information of the target data based on the first feature information of the at least one candidate data to obtain supplemented feature information of the target data;

[0020] The detection unit is used to perform detection processing on the supplemented characteristic information of the target data to obtain the detection result of the target data.

[0021] In some embodiments, the L biomarkers include L gene mutations, and the detection results include probability reference values ​​that the target data belongs to multiple cancer classifications.

[0022] In some embodiments, the extraction unit is specifically used to extract the jth first feature information of the target data and the Q candidate data respectively through the jth retrieval unit among the R retrieval units, and the data detection model includes a retrieval module, the retrieval module includes the R retrieval units, R is a positive integer greater than 1, and j is a positive integer less than or equal to R; the selection unit is specifically used to select Kj candidate data related to the target data from the Q candidate data based on the jth first feature information of the target data and the jth first feature information of the Q candidate data, where Kj is a positive integer; the supplementation unit is specifically used to supplement the jth first feature information of the target data based on the jth first feature information of the Kj candidate data, so as to obtain the jth supplemented feature information of the target data under the jth retrieval unit; the detection unit is specifically used to perform detection processing on the R supplemented feature information of the target data under the R retrieval units through the detection module in the data detection model, so as to obtain the detection result of the target data.

[0023] In some embodiments, the j-th retrieval unit includes a j-th encoder and an extraction unit, which is specifically configured to extract the j-th first feature information of the target data and the Q candidate data respectively through the j-th encoder.

[0024] In some embodiments, the selection unit is specifically used to determine the similarity between the target data and each of the Q candidate data based on the j-th first feature information of the target data and the j-th first feature information of the Q candidate data; and based on the similarity, select the Kj candidate data from the Q candidate data.

[0025] In some embodiments, the selection unit is specifically used to project the jth first feature information of the qth candidate data among the Q candidate data using the first projection matrix to obtain the jth second feature information of the qth candidate data, where q is a positive integer less than or equal to Q; use the first projection matrix to project the jth first feature information of the target data to obtain the jth second feature information of the target data; determine the distance between the jth second feature information of the target data and the jth second feature information of the qth candidate data, and determine the similarity between the target data and the qth candidate data based on the distance.

[0026] In some embodiments, the supplementing unit is specifically used to determine the distance between the j-th second feature information of the target data and the j-th second feature information of the k-th candidate data for the k-th candidate data among the Kj candidate data, and based on the distance, determine the first supplementary information of the k-th candidate data for the target data, where k is a positive integer less than or equal to K; based on the first supplementary information of the target data respectively provided by the Kj candidate data, the j-th first feature information of the target data is supplemented to determine the j-th supplemented feature information of the target data.

[0027] In some embodiments, the supplementing unit is specifically used to perform linear activation processing on the distance and then perform random inactivation processing, and perform lossless linear transformation on the random inactivation processing result to obtain the second supplementary information of the kth candidate data to the target data; based on the second supplementary information of the kth candidate data to the target data and the label information of the kth candidate data, determine the first supplementary information of the kth candidate data to the target data.

[0028] In some embodiments, the supplementing unit is specifically used to use a second projection matrix to project the label information of the kth candidate data to obtain the label projection information of the kth candidate data; and determine the sum of the second supplementary information of the kth candidate data to the target data and the label projection information of the kth candidate data as the first supplementary information of the kth candidate data to the target data.

[0029] In some embodiments, the supplementing unit is specifically used to determine the weight of the kth candidate data among the Kj candidate data based on the similarity between the target data and the kth candidate data; based on the weights of the Kj candidate data, perform weighted processing on the first supplementary information of the Kj candidate data to the target data to obtain weighted supplementary information; based on the jth first feature information of the target data and the weighted supplementary information, determine the jth supplemented feature information of the target data.

[0030] In some embodiments, the detection module includes a hybrid expert module, which includes a gating unit and E expert units. The detection unit is specifically used to obtain S input tokens based on the supplemented feature information of the target data under the R retrieval units, where S is a positive integer greater than 1; for the sth input token among the S input tokens, the probability value of the sth input token being assigned to each expert unit in the E expert units is determined through the gating unit, where E is a positive integer greater than 1, and s is a positive integer less than or equal to S; based on the probability value, T expert units for processing the sth input token are selected from the E expert units, and the sth input token is processed by the T expert units respectively to obtain fused feature information of the sth input token, where T is a positive integer less than E; based on the fused feature information of the S input tokens, the mixed feature information of the target data is obtained; based on the mixed feature information, the target data is detected and processed to obtain the detection result of the target data.

[0031] In some embodiments, the detection unit is specifically used to perform feature fusion processing on each input token assigned to the expert unit for each of the T expert units, so as to obtain output feature information of each input token, wherein each input token includes the sth input token; determine the weights corresponding to the T expert units respectively based on the probability value of the sth input token being assigned to each of the T expert units; and perform weighted processing on the output feature information of the sth input token respectively output by the T expert units based on the weights corresponding to the T expert units, so as to obtain fused feature information of the sth input token.

[0032] In some embodiments, the detection unit is specifically used to take the supplemented feature information of the target data under each of the R retrieval units as a row of a first matrix to obtain a first matrix including R rows; and take each column of the first matrix as an input token to obtain the S input tokens.

[0033] In a fourth aspect, an embodiment of the present application provides a model training device, including:

[0034] an extraction unit, configured to obtain a training sample and Q candidate samples, and respectively extract first feature information of the training sample and the Q candidate samples through a retrieval module in a data detection model, wherein the training sample is a state value of L biomarkers of a first subject, and the candidate samples are state values ​​of L biomarkers of other subjects different from the first subject, and both Q and L are positive integers greater than 1;

[0035] A selection unit, configured to select at least one candidate sample related to the training sample from the Q candidate samples based on the first feature information of the training sample and the first feature information of the Q candidate samples;

[0036] A supplementing unit, configured to supplement the first feature information of the training sample based on the first feature information of the at least one candidate sample to obtain supplemented feature information of the training sample;

[0037] A detection unit, configured to perform detection processing on the supplemented feature information of the training sample through a detection module in the data detection model to obtain a detection result of the training sample;

[0038] A training unit is used to determine the loss of the data detection model based on the detection results of the training samples, and train the data detection model based on the loss.

[0039] In some embodiments, the retrieval module includes R retrieval units, where R is a positive integer greater than 1; an extraction unit, specifically used to extract the jth first feature information of the training sample and the Q candidate samples respectively through the jth retrieval unit among the R retrieval units, where j is a positive integer less than or equal to R; a selection unit, specifically used to select Kj candidate samples related to the training sample from the Q candidate samples based on the jth first feature information of the training sample and the jth first feature information of the Q candidate samples; a supplementation unit, specifically used to supplement the jth first feature information of the training sample based on the jth first feature information of the Kj candidate samples, to obtain the jth supplemented feature information of the training sample under the jth retrieval unit; a detection unit, specifically used to perform detection processing on the R supplemented feature information of the training sample under the R retrieval units through the detection module, to obtain the detection result of the training sample.

[0040] In some embodiments, the detection module includes a hybrid expert module, which includes a gating unit and E expert units. The detection unit is specifically used to obtain S input tokens based on the supplemented feature information of the target data under the R retrieval units, where S is a positive integer greater than 1; for the sth input token among the S input tokens, the probability value of the sth input token being assigned to each expert unit in the E expert units is determined through the gating unit, where E is a positive integer greater than 1, and s is a positive integer less than or equal to S; based on the probability value, T expert units for processing the sth input token are selected from the E expert units, and the sth input token is processed by the T expert units respectively to obtain fused feature information of the sth input token, where T is a positive integer less than E; based on the fused feature information of the S input tokens, the mixed feature information of the target data is obtained; based on the mixed feature information, the target data is detected and processed to obtain the detection result of the target data.

[0041] In some embodiments, the training unit is specifically used to determine a first loss based on the detection results of the training samples and the label information of the training samples; determine a second loss based on the number of input labels assigned to each of the E expert units and the assignment probability value; and determine the loss of the data detection model based on the first loss and the second loss.

[0042] In a fifth aspect, the present application provides an electronic device, comprising a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory to execute the method of the first aspect or the second aspect.

[0043] In a sixth aspect, a chip is provided for implementing the method in each implementation of the first aspect. Specifically, the chip includes: a processor for calling and running a computer program from a memory, so that a device equipped with the chip executes the method in the first or second aspect.

[0044] In a seventh aspect, a computer-readable storage medium is provided for storing a computer program, wherein the computer program enables a computer to execute the method of the first aspect or the second aspect.

[0045] In an eighth aspect, a computer program product is provided, comprising computer program instructions, wherein the computer program instructions enable a computer to execute the method of the first aspect or the second aspect.

[0046] In a ninth aspect, a computer program is provided, which, when executed on a computer, enables the computer to execute the method of the first aspect or the second aspect.

[0047] In summary, the present application performs data detection through a data detection model, which includes a retrieval module and a detection module. Specifically, during data detection, the target data to be detected and Q candidate data are first obtained. The first feature information of the target data and the Q candidate data are extracted respectively, and the target data is the state value of L biomarkers of the target object; then, based on the first feature information of the target data and the first feature information of the Q candidate data, at least one candidate data related to the target data is selected from the Q candidate data, and then based on the first feature information of the at least one candidate data, the first feature information of the target data is supplemented to obtain the supplemented feature information of the target data; finally, the supplemented feature information of the target data is detected and processed to obtain the detection result of the target data. It can be seen from this that the embodiment of the present application selects at least one candidate data related to the target data from Q candidate data, and uses the first feature information of the at least one candidate data to supplement the first feature information of the target data, integrates the features of similar objects, and obtains the supplemented feature information of the target data. When the supplemented feature information of the target data is used for data detection, multiple different classification detections can be achieved, for example, auxiliary detection of multiple different types of early cancers can be achieved, and the detection accuracy can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0049] Figure 1 A schematic diagram of an implementation environment of a data detection method provided in an embodiment of the present application;

[0050] Figure 2 A flowchart of a model training method provided in one embodiment of the present application;

[0051] Figure 3 is a schematic diagram of a data detection model;

[0052] Figure 4 is a schematic diagram of a training sample and a candidate sample;

[0053] Figure 5 A schematic diagram of a retrieval model;

[0054] Fig. 6A A schematic diagram of selecting relevant candidate samples based on similarity;

[0055] Figure 6B A schematic diagram for supplementing the features of training samples based on selected candidate samples;

[0056] Figure 7 A schematic diagram of using a detection module to perform data detection;

[0057] Figure 8 is a schematic diagram of a retrieval module including R detection units;

[0058] Fig. 9 A flowchart of a model training method provided in one embodiment of the present application;

[0059] Fig.10 A schematic diagram of a retrieval unit;

[0060] Fig.11 A schematic diagram of using the detection module to process R supplemented feature information;

[0061] Fig.12 A schematic diagram of a network structure of a detection module;

[0062] Fig.13 A schematic diagram of using a hybrid expert module for data detection;

[0063] Fig.14 A flowchart of a data detection method provided in one embodiment of the present application;

[0064] Fig.15 A flowchart of a data detection method provided in one embodiment of the present application;

[0065] Fig.16 A flowchart of a data detection method provided in one embodiment of the present application;

[0066] Fig.17 It is a schematic diagram of the overall process of data detection;

[0067] Fig.18 is a schematic block diagram of a data detection device provided in an embodiment of the present application;

[0068] Fig.19 is a schematic block diagram of a model training device provided in one embodiment of the present application;

[0069] Fig. 20 It is a schematic block diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0070] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0071] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present application described here can be implemented in an order other than those illustrated or described here. In an embodiment of the present invention, "B corresponding to A" means that B is associated with A. In one implementation, B can be determined according to A. However, it should also be understood that determining B according to A does not mean that B is determined only according to A, but B can also be determined according to A and / or other information. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or server that includes a series of steps or units does not have to be limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices. In the description of the present application, unless otherwise specified, "multiple" refers to two or more than two.

[0072] The data detection method provided in the embodiment of the present application can be applied to various fields such as medical diagnosis and data detection. It can realize the detection of multiple dimensions of target data and improve the detection accuracy. For example, it can obtain detection reference results of multiple early cancers, thereby improving the detection accuracy of early cancers.

[0073] In order to facilitate understanding of the embodiments of the present application, the relevant concepts involved in the embodiments of the present application are first introduced:

[0074] In medical testing, for example, early cancer detection can significantly improve survival rates and reduce treatment costs, and therefore has attracted widespread attention in academia and industry. At present, early cancer detection methods mainly include: detection methods based on specific cancer biomarkers and detection methods based on image analysis. Among them, the detection method based on specific cancer biomarkers is a medical test based on known cancer-related biomarkers. Specific biomarkers currently available for cancer detection include prostate-specific antigen (PSA) and for prostate cancer and CA-125 for ovarian cancer. Image analysis-based methods use various imaging techniques, such as low-dose CT for lung cancer detection, colonoscopy for colorectal cancer detection, cervical smear examination for cervical cancer detection, etc., to perform auxiliary detection of related cancers.

[0075] However, the above-mentioned detection methods based on specific cancer biomarkers are limited to specific types of cancers, and the relevant biomarkers of these cancers are clearly known. They lack sensitivity and specificity and cannot detect cancers related to unknown cancer biomarkers. In addition, the imaging analysis method can only target specific cancers and cannot be used for convenient and low-cost early cancer screening, nor can it perform pan-cancer analysis. In addition, some cancers do not have obvious imaging features in the early stages, and the detection rate using imaging-based analysis methods is low.

[0076] In order to solve the above technical problems, the embodiment of the present application proposes a data detection method. During data detection, the target data to be detected and Q candidate data are first obtained. The first feature information of the target data and the Q candidate data are extracted respectively, and the target data is the state value of L biomarkers of the target object; then, based on the first feature information of the target data and the first feature information of the Q candidate data, at least one candidate data related to the target data is selected from the Q candidate data, and then based on the first feature information of the at least one candidate data, the first feature information of the target data is supplemented to obtain the supplemented feature information of the target data; finally, the supplemented feature information of the target data is detected and processed to obtain the detection result of the target data. It can be seen that the embodiment of the present application selects at least one candidate data related to the target data from the Q candidate data, and uses the first feature information of the at least one candidate data to supplement the first feature information of the target data, integrates the features of similar objects, and obtains the supplemented feature information of the target data. When the supplemented feature information of the target data is used for data detection, multiple different classifications of detection can be achieved, for example, multiple different types of early cancer detection reference results can be obtained, and the detection accuracy can be improved.

[0077] The following is an introduction to the implementation environment of the data detection method provided in the embodiment of the present application.

[0078] Figure 1 A schematic diagram of an implementation environment of a data detection method provided in an embodiment of the present application, such as Figure 1 As shown, the implementation environment includes: a terminal device 101 and a server 102.

[0079] The terminal device 101 is connected to the server 102 via a wired or wireless connection.

[0080] In some embodiments, the terminal device 101 is installed with a client of the data detection system, and the server 102 can be understood as the server or backend of the data detection system. An object (e.g., a user) can interact with the client of the data detection system installed on the terminal device 101. The server 102 stores a data detection model, which is used for data detection, such as for early cancer screening.

[0081] In some embodiments, before performing data detection, the embodiment of the present application first trains the data detection model. Here, the process of training the data detection model by server 102 is introduced. Specifically, the object can specify a training data set to the server 102 through a terminal device. Server 102 obtains training samples and Q candidate samples from the training data set. In an embodiment of the present application, the training samples and candidate samples are both state values ​​of M biomarkers (such as gene mutations) of different objects. Taking the biomarker as a gene mutation as an example, the state of the gene mutation includes two states of mutation and non-mutation. For example, for a certain gene, if the gene mutates, the state value of the gene mutation is 1, and if the gene does not mutate, the state value of the gene mutation is 0. The server 102 extracts the first feature information of the training sample and Q candidate samples respectively through the retrieval module in the data detection model, where the training sample is the state value of L biomarkers of the first object, and the candidate sample is the state value of L biomarkers of other objects different from the first object, and Q and M are both positive integers greater than 1; based on the first feature information of the training sample and the first feature information of the Q candidate samples, at least one candidate sample related to the training sample is selected from the Q candidate samples; based on the first feature information of the at least one candidate sample, the first feature information of the training sample is supplemented to obtain the supplemented feature information of the training sample; through the detection module in the data detection model, the supplemented feature information of the training sample is detected and processed to obtain the detection result of the training sample; based on the detection result of the training sample, the loss of the data detection model is determined, and the data detection model is trained based on the loss.

[0082] In some embodiments, the above-mentioned data detection model can also be trained through the terminal device 101, and the embodiments of the present application do not limit this.

[0083] In an embodiment of the present application, after the data detection model is trained, the trained data detection model is used to execute the data detection method of the embodiment of the present application.

[0084] In some embodiments, the data detection method provided in the embodiments of the present application is jointly performed by the terminal device 101 and the server 102. For example, the subject starts the client of the data detection system on the terminal device 101, and the client displays the input interface of the data to be detected. The subject enters the target data to be processed in the input interface, and the target data is the state value of L biomarkers of the first subject, such as the state value of gene mutation in L of the first subject. The terminal device 101 sends the target data to be detected to the server 102. The server 102 loads the candidate data set and obtains Q candidate data from the candidate data set. Next, the first feature information of the target data and the Q candidate data are extracted respectively, for example, the server 102 extracts the first feature information of the target data and the Q candidate data respectively through the retrieval module in the data detection model, the target data is the state value of L biomarkers of the target object, the candidate data is the state value of L biomarkers of other objects different from the target object, and Q and L are both positive integers greater than 1; based on the first feature information of the target data and the first feature information of the Q candidate data, at least one candidate data related to the target data is selected from the Q candidate data; based on the first feature information of the at least one candidate data, the first feature information of the target data is supplemented to obtain the supplemented feature information of the target data; the supplemented feature information of the target data is detected and processed to obtain the detection result of the target data, for example, the supplemented feature information of the target data is detected and processed through the detection module in the data detection model to obtain the detection result of the target data. Next, the server 102 sends the detection result of the target data (for example, the predicted value that the target data belongs to the preset classification) to the terminal device 101 for display. It can be seen from this that the embodiment of the present application selects at least one candidate data related to the target data from Q candidate data, and uses the first feature information of the at least one candidate data to supplement the first feature information of the target data, integrates the features of similar objects, and obtains the supplemented feature information of the target data, and then uses the supplemented feature information of the target data to perform data detection, which can realize multiple classification detections and improve the detection accuracy of the target data.

[0085] In some embodiments, the data detection method provided in the embodiments of the present application is executed by the terminal device 101. At this time, the terminal device 101 can be any terminal device that can provide data detection services. The server 102 can be any server that can provide data download services. In one example, the server 102 stores a trained data detection model. The terminal device 101 is installed with a client of the data detection system, and the client can request the data detection model from the server 102 based on the request of the object. For example, the terminal device 101 displays an input box for the data to be processed, and the object enters the target data to be processed in the input interface. The terminal device 101 responds to the object's input operation on the target data and sends a request to the server 102 to request the data detection model. The server 102 sends the trained data detection model to the terminal device 101. The terminal device 101 loads the candidate data set and selects Q candidate data from the candidate data set. Then, the first feature information of the target data and the Q candidate data is extracted respectively, for example, through the retrieval module in the data detection model, the first feature information of the target data and the Q candidate data is extracted respectively, the target data is the state value of L biomarkers of the target object, and Q and L are both positive integers greater than 1; based on the first feature information of the target data and the Q candidate data, at least one candidate data related to the target data is selected from the Q candidate data; based on the first feature information of the at least one candidate data, the first feature information of the target data is supplemented to obtain the supplemented feature information of the target data; the supplemented feature information of the target data is detected and processed to obtain the detection result of the target data, for example, through the detection module in the data detection model, the supplemented feature information of the target data is detected and processed to obtain the detection result of the target data. Then, the terminal device 101 displays the detection result of the target data to the object.

[0086] In some embodiments, the terminal device 101 includes but is not limited to: desktop computers, laptop computers, smart phones, tablet computers, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart car devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. Terminal devices are often equipped with display devices, which can also be displays, display screens, touch screens, etc., and touch screens can also be touch screens, touch panels, etc.

[0087] In some embodiments, the above-mentioned server 102 may be one or more. When there are multiple servers, there are at least two servers for providing different services, and / or there are at least two servers for providing the same service, such as providing the same service in a load balancing manner, which is not limited in the embodiments of the present application. Among them, the above-mentioned server may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and industrial intelligent platforms. The server can also become a node of the blockchain.

[0088] It should be noted that the implementation environment of the embodiments of the present application includes but is not limited to Figure 1 shown.

[0089] The technical solutions of the embodiments of the present application are described in detail below through some embodiments. The following embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0090] First combine Figure 2 , the training process of the data detection model of the embodiment of the present application is introduced.

[0091] Figure 2 The flowchart of the model training method provided in one embodiment of the present application is shown in FIG. The execution subject of the embodiment of the present application may be a device having a model training function, such as a model training device. The model training device may be the above-mentioned Figure 1 The server 102 or terminal device 101 shown may also be Figure 1 The system is composed of the server 102 or 101 shown. For the convenience of description, the following embodiments take electronic devices as an example to illustrate the method of the embodiments of the present application.

[0092] like Figure 2 As shown, the model training method of the embodiment of the present application includes:

[0093] S101, obtaining a training sample and Q candidate samples, and respectively extracting first feature information of the training sample and the Q candidate samples through a retrieval module in a data detection model.

[0094] The training samples are the state values ​​of L biomarkers of the first subject, and the candidate samples are the state values ​​of L biomarkers of other subjects different from the first subject. For example, for the i-th candidate sample among Q candidate samples, the i-th candidate sample can be understood as the state value of L biomarkers of the i-th subject, the i-th subject is a different subject from the first subject, Q and L are both positive integers greater than 1, and i is a positive integer less than or equal to Q.

[0095] It should be noted that the data used in the embodiments of the present application and the process of obtaining these data are in compliance with the provisions of relevant laws and regulations.

[0096] The data detection model of the embodiment of the present application can be applied to different fields to detect data in different fields. For example, it can be applied to the biological and medical fields to detect data related to the biological and medical fields. For example, it can be used for auxiliary detection of early cancer.

[0097] In an embodiment of the present application, when a data detection model is used in different fields, the data detection model can be trained using training samples of the corresponding fields.

[0098] In the embodiment of the present application, before training the data detection model, the electronic device first obtains a training data set, which includes N training samples and can be expressed as {S 1 , S 2 , …, S N Each training sample contains the state values ​​of L possible biomarkers of an object. For example, the i-th training sample includes the state values ​​of L biomarkers of the i-th object, and the i-th training sample can be expressed as in represents the state value of the first biomarker of the i-th object, represents the state value of the second biomarker of the i-th object, and so on. Expresses the status value of the Lth biomarker of the ith subject.

[0099] Exemplarily, the N training samples included in the training data set may be as shown in Table 1:

[0100] Table 1

[0101]

[0102] Each row in Table 1 above represents a training sample. For example, the first row in Table 1 above is the training sample S 1 , the training sample S 1 Includes the status values ​​of L biomarkers of subject 1 The second line in Table 1 above is the training sample S2 , the training sample S 2 Includes the status values ​​of L biomarkers of subject 2 Exemplarily, the training data set of the embodiment of the present application may constitute a matrix of N rows and L columns, wherein each row of the matrix is ​​a training sample.

[0103] The present application example does not limit the specific types of the above L biomarkers.

[0104] In one possible case, the above-mentioned L kinds of biomarkers may include gene mutations, wherein the state value of gene mutations may be understood as the state value of whether a gene has mutated. At this time, the state value includes two values, namely, mutation and no mutation. Optionally, 1 may be used to indicate mutation, and 0 may be used to indicate no mutation. Of course, 1 may also be used to indicate no mutation, and 0 may be used to indicate mutation. The embodiments of the present application do not limit the specific types and quantities of gene mutations. Exemplarily, gene mutations may include gene mutations such as ABL1, ACVR1, and AKT1. Among them, ABL1 is a proto-oncogene, ACVR1 encodes a gene for activin type I receptor, and AKT1 is a thymoma virus proto-oncogene.

[0105] In a possible case, the L biomarkers mentioned above may also include serological markers and methylation status, etc. At this time, the status value of the biomarker may be determined according to specific actual conditions, such as the detection value of a specific biomarker, etc.

[0106] In the embodiment of the present application, label information needs to be determined for the above N training samples. For each training sample, the label information of the training sample indicates the true value of the training sample belonging to the preset category. Assuming that the preset category includes C categories, the label information of the N training samples can be defined as {Y 1 , Y 2 , …, Y N}, where Y i ∈{0, ..., C}, C represents the number of preset categories, and the specific value of C varies according to different training data sets.

[0107] In some embodiments, the L biomarkers include L gene mutations, and the corresponding preset classifications include multiple cancer classifications.

[0108] The data detection model of the embodiment of the present application attempts to construct a mapping function Y i =F(S i), multiple preset classifications can be detected based on the state values ​​of L biomarkers of any object. For example, when the above preset classification is multiple cancer categories, the embodiment of the present application is used to develop a highly sensitive and specific multi-cancer early detection (MCED) model, and the above biomarkers may involve biomarkers, gene mutations and other biomarkers.

[0109] In an embodiment of the present application, the training process of the data detection model using each training sample in the above-mentioned training data set is basically the same. For the sake of ease of description, a training sample is taken as an example to introduce the model training process of an embodiment of the present application.

[0110] In an embodiment of the present application, for each training sample in a training data set, when the data detection model is trained using the training sample, it is necessary to obtain Q candidate samples corresponding to the training sample.

[0111] The embodiment of the present application does not limit the specific selection method for obtaining the Q candidate samples corresponding to the training sample.

[0112] In a possible implementation, for each training sample among the N training samples, for example, the i-th training sample S i The electronic device removes the training sample S from the N training samples included in the training data set. i From the N-1 training samples other than i For example, if Q is less than N-1, Q training samples can be randomly selected from the N-1 training samples as the training sample S i For example, if Q is equal to N-1, then the training data set can be divided into the training sample S i The remaining N-1 training samples are taken as the training sample S i The corresponding Q candidate samples.

[0113] In a possible implementation, when training the data detection model, multiple rounds of training are performed, and in each round of training, a batch of training samples are selected from N training samples for the current round of training. i It can be understood as any training sample in the current batch of training samples. At this time, the Q candidate samples include all the training samples in the current batch except the training sample S i All samples except

[0114] In a possible implementation, the electronic device may also load other data sets, which include multiple candidate samples, where different candidate samples include state values ​​of L biomarkers of different objects. In this way, the electronic device may select Q candidate samples corresponding to the training sample from the data set.

[0115] In an embodiment of the present application, after the electronic device obtains the training sample and the Q candidate samples corresponding to the training sample, it uses the training sample and the Q candidate samples to train the data detection model.

[0116] like Figure 3 As shown, the data detection model of the embodiment of the present application includes a retrieval module and a detection module, wherein the main function of the retrieval module is to select at least one candidate sample related to the training sample from Q candidate samples, and use the at least one candidate sample to supplement the features of the training sample. At this time, the training sample integrates the features of other related candidate samples, so that when the supplementary features of the training sample are used for subsequent data detection, the accuracy of data detection can be improved. The detection module is used to detect and process the supplementary features of the training sample.

[0117] In some embodiments, for ease of description, the currently used training sample can be represented as S input , the S input The training sample S includes the state values ​​of L biomarkers of a certain object, for example, the first object. input The corresponding Q candidate samples can be expressed as S 1 candidate , S 2 candidate …S Q candidate . The Q candidate samples include the state values ​​of L biomarkers of other Q objects different from the first object. Exemplarily, Figure 4 It shows that when L is equal to 6 and Q is equal to 5, the training sample S input And a schematic diagram of 5 candidate samples, where Figure 4 Each small square in the training samples and candidate samples represents an element value, that is, the state value of a biomarker of a certain object. The different colors of the small squares represent different state values. For example, if the biomarker is a gene mutation, if the small square is dark, it means that the corresponding gene has mutated, and if the small square is light, it means that the gene has not mutated. The opposite is also true, that is, if the small square is dark, it means that the corresponding gene has not mutated, and if the small square is light, it means that the gene has mutated.

[0118] like Figure 4As shown, the electronic device inputs the training sample and Q candidate samples into a retrieval module, and the retrieval module extracts the first feature information of the training sample and the first feature information of each candidate sample respectively.

[0119] In the embodiment of the present application, the specific manner in which the retrieval module extracts the first feature information of the training sample and the Q candidate samples includes at least the following manners:

[0120] In mode 1, the retrieval module does not consider the Q candidate samples when extracting the first feature information of the training sample, and does not consider the other candidate samples in the training sample and the Q candidate samples when extracting the first feature information of each candidate sample in the Q candidate samples. This mode 1 can be understood as the electronic device inputting the training sample and the Q candidate samples, i.e., the Q+1 samples, into the retrieval module one by one for feature extraction, and in the feature extraction process, no feature interaction and fusion is performed between these samples.

[0121] In mode 2, the retrieval module considers the Q candidate samples when extracting the first feature information of the training sample, and does not consider the training sample and other candidate samples in the Q candidate samples when extracting the first feature information of each candidate sample in the Q candidate samples. This mode 2 can be understood as the electronic device inputting the training sample and the Q candidate samples (Q+1 samples) into the retrieval module for feature extraction. During the feature extraction process, the features of these samples are interactively fused.

[0122] The embodiment of the present application does not limit the specific network structure of the retrieval module.

[0123] In one example, if Figure 5 As shown, the retrieval module includes an encoder, which is used for feature extraction. In this way, the electronic device can input the training sample and Q candidate samples into the encoder for feature extraction to obtain the first feature information of the training sample and the first feature information of each candidate sample in the Q candidate samples.

[0124] Exemplarily, the process of the electronic device extracting the first feature information of the training sample and Q candidate samples through the encoder can be expressed as shown in the following formula (1):

[0125] h input =E(S input )

[0126] h i candidate =E(S i candidate ),i∈Q candidate (1)

[0127] Where E represents the encoder, h input ∈R1×d Represents the training sample S input The first characteristic information, h i candidate ∈R 1xd Represents the i-th candidate sample S i candidate The first characteristic information, Q candidate represents a candidate sample set consisting of Q candidate samples. d represents the hidden dimension, that is, the first feature information h of the training sample input is a d-dimensional vector. Similarly, the first feature information h of the i-th candidate sample i candidate is also a d-dimensional vector.

[0128] In an embodiment of the present application, the electronic device obtains a training sample and Q candidate samples corresponding to the training sample based on the above steps, and extracts the first feature information of the training sample and the first feature information of each of the Q candidate samples through the retrieval module in the data detection model, and then executes the following step S102.

[0129] S102: Based on the first feature information of the training sample and the first feature information of the Q candidate samples, select at least one candidate sample related to the training sample from the Q candidate samples.

[0130] In an embodiment of the present application, in order to improve the accuracy of data detection, at least one candidate sample that is the same as the training sample is selected from the Q candidate samples, and then the features of the at least one candidate sample are used to supplement the features of the training sample. In this way, the features of the first object included in the training sample also include features of other objects that are similar to the first object, thereby making the features of the first object richer and more comprehensive. In this way, when data detection is performed based on the rich and comprehensive feature information of the first object, the accuracy of data detection can be improved.

[0131] The specific process of the electronic device selecting at least one candidate sample related to the training sample from Q candidate samples is introduced below.

[0132] In an embodiment of the present application, the electronic device selects at least one candidate sample related to the training sample from the Q candidate samples based on the first feature information of the training sample and the first feature information of each of the Q candidate samples.

[0133] In some embodiments, the electronic device clusters the first feature information of the training sample and the first feature information of the Q candidate samples, and determines the candidate samples among the Q training samples that belong to the same cluster as the training sample as at least one candidate sample related to the training sample.

[0134] In some embodiments, the above S102 includes the following steps S102-A and S102-B:

[0135] S102-A, determining the similarity between the training sample and each of the Q candidate samples based on the first feature information of the training sample and the first feature information of the Q candidate samples;

[0136] S102-B. Based on the similarity, select at least one candidate sample from the Q candidate samples.

[0137] In this implementation, for each of the Q candidate samples, the similarity between the candidate sample and the training sample is determined based on the first feature information of the candidate sample and the first feature information of the training sample.

[0138] The embodiment of the present application does not limit the specific manner in which the electronic device determines the similarity between the candidate sample and the training sample based on the first feature information of the candidate sample and the first feature information of the training sample.

[0139] In one possible implementation, the electronic device uses a distance calculation method such as cosine distance or L2 distance to calculate the distance between the first feature information of the candidate sample and the first feature information of the training sample, and then determines the similarity between the candidate sample and the training sample based on the distance. The larger the distance, the less similar the training sample is to the candidate sample, and the smaller the corresponding similarity value. The smaller the distance, the more similar the training sample is to the candidate sample, and the larger the corresponding similarity value. In one example, the reciprocal of the distance between the first feature information of the candidate sample and the first feature information of the training sample can be determined as the similarity between the candidate sample and the training sample.

[0140] In a possible implementation, the above S102-A may include the following steps S102-A1 to S102-A3:

[0141] S102-A1, for the qth candidate sample among the Q candidate samples, use the first projection matrix to project the first feature information of the qth candidate sample to obtain the second feature information of the qth candidate sample, where q is a positive integer less than or equal to Q;

[0142] S102-A2, projecting the first feature information of the training sample using the first projection matrix to obtain the second feature information of the training sample;

[0143] S102-A3. Determine the distance between the second feature information of the training sample and the second feature information of the j-th candidate sample as the similarity between the training sample and the q-th candidate sample.

[0144] In this implementation, when calculating the similarity between the candidate sample and the training sample, a projection matrix is ​​used to project the first feature information of the candidate sample and the training sample to obtain the second feature information of the candidate sample and the training sample, and then the second feature information is used to calculate the similarity between the candidate sample and the training sample.

[0145] Specifically, for the qth candidate sample among the Q candidate samples, the first projection matrix W is used k The first feature information of the qth candidate sample is projected. For the convenience of description, the projected first feature information is recorded as the second feature information of the qth candidate sample. Similarly, the first projection matrix W is used k The first feature information of the training sample is projected to obtain the second feature information of the training sample. Next, the distance between the second feature information of the training sample and the second feature information of the qth candidate sample is calculated, and the distance is determined as the similarity between the training sample and the qth candidate sample.

[0146] In one example, the electronic device uses the L2 distance calculation method to determine the distance between the second feature information of the training sample and the second feature information of the qth candidate sample. At this time, the calculation formula of the similarity is shown in formula (2):

[0147]

[0148] in, represents the similarity between the second feature information of the training sample and the second feature information of the qth candidate sample, W k ∈R d×d is the first projection matrix, is the first feature information of the qth candidate sample. k (h input ) is the second feature information of the training sample, is the second feature information of the qth candidate sample.

[0149] like Fig. 6A As shown, the electronic device can determine the similarity between each of the Q candidate samples and the training sample based on the above method. Then, the electronic device selects at least one candidate sample related to the training sample from the Q training samples based on the similarity between each of the Q candidate samples and the training sample. In one example, the electronic device selects at least one candidate sample with the greatest similarity to the training sample (e.g., the top k candidate samples with the greatest similarity shown in FIG. 6 ) from the Q candidate samples as at least one candidate sample related to the training sample.

[0150] S103: Supplement the first feature information of the training sample based on the first feature information of at least one candidate sample to obtain supplemented feature information of the training sample.

[0151] In an embodiment of the present application, supplementing the first feature information of the training sample can be understood as expanding the first feature information of the training sample to enrich the first feature information of the training sample, so that the supplemented first feature information not only includes the feature information of the first object but also includes the feature information of other objects corresponding to the candidate sample.

[0152] The embodiment of the present application does not limit the specific method in which the electronic device supplements the first feature information of the training sample based on the first feature information of at least one candidate sample to obtain the supplemented feature information of the training sample.

[0153] In some embodiments, the electronic device may fuse the first feature information of the training sample with the first feature information of the at least one candidate sample to obtain the supplemented feature information of the training sample. For example, based on the similarity between the at least one candidate sample and the training sample, the weight of the at least one candidate sample is determined, and then based on the weight, the first feature information of the training sample is weighted with the first feature information of the at least one candidate sample to obtain the supplemented feature information of the training sample.

[0154] In some embodiments, the above S103 includes the following steps S103-A and S103-B:

[0155] S103-A, for a k-th candidate sample among at least one candidate sample, determine a distance between the second feature information of the training sample and the second feature information of the k-th candidate sample, and based on the distance, determine first supplementary information of the k-th candidate sample for the training sample, where k is a positive integer less than or equal to K;

[0156] S103-B: Supplement the first feature information of the training sample based on the first supplementary information of the training sample by at least one candidate sample, and determine the supplemented feature information of the training sample.

[0157] In this implementation, the distance between the training sample and at least one candidate sample is used as a supplementary feature in the hidden layer. Specifically, for the kth candidate sample in the at least one candidate sample, the distance between the second feature information of the training sample and the second feature information of the kth candidate sample is determined, and based on the distance, the first supplementary information of the kth candidate sample to the training sample is determined.

[0158] The embodiment of the present application does not limit the specific method of determining the first supplementary information of the kth candidate sample to the training sample based on the distance.

[0159] In some embodiments, the electronic device may determine the distance between the second feature information of the training sample and the second feature information of the kth candidate sample as the first supplementary information of the kth candidate sample to the training sample.

[0160] In some embodiments, the electronic device performs activation processing and then random deactivation processing on the distance between the second feature information of the training sample and the second feature information of the kth candidate sample, and performs lossless linear transformation on the random deactivation processing result to obtain the second supplementary information of the kth candidate sample to the training sample. Then, the second supplementary information of the kth candidate sample to the training sample is determined as the first supplementary information of the kth candidate sample to the training sample.

[0161] Exemplarily, the electronic device uses the following formula (3) to determine the second supplementary information:

[0162]

[0163] is the kth candidate sample, is the L1 distance between the second feature information of the training sample and the second feature information of the kth candidate sample, W T (·)=LinearWithoutBias(Dropout(ReLU(Linear(·)))), Linear(·) is linear processing, ReLU is the activation function, Dropout(·) is random dropout processing, and LinearWithoutBias(·) is lossless linear transformation.

[0164] It should be noted that the above-mentioned method of processing the distance between the second feature information of the training sample and the second feature information of the k-th candidate sample to obtain the second supplementary information of the k-th candidate sample to the training sample is only an example. The electronic device can also process the distance between the second feature information of the training sample and the second feature information of the k-th candidate sample in other ways to obtain the second supplementary information of the k-th candidate sample to the training sample.

[0165] In some embodiments, since at least one candidate sample is similar to the training sample, the label information of at least one candidate sample can also provide additional information about the characteristics of the training sample. Based on this, determining the first supplementary information of the kth candidate sample to the training sample based on the distance in S103-A includes the following steps S103-A1 and S103-A2:

[0166] S103-A1, performing linear activation processing and then random deactivation processing on the distance between the second feature information of the training sample and the second feature information of the k-th candidate sample, and performing lossless linear transformation on the random deactivation processing result to obtain the second supplementary information of the k-th candidate sample for the training sample;

[0167] S103-A2 determines the first supplementary information of the kth candidate sample for the training sample based on the second supplementary information of the kth candidate sample for the training sample and the label information of the kth candidate sample.

[0168] Among them, the distance between the second feature information of the training sample and the second feature information of the k-th candidate sample is subjected to linear activation processing and then random inactivation processing, and the random inactivation processing result is losslessly linearly transformed to obtain the second supplementary information of the k-th candidate sample for the training sample. The specific process can refer to the description of the above embodiment and will not be repeated here.

[0169] Next, the electronic device determines the first supplementary information of the kth candidate sample to the training sample based on the second supplementary information of the kth candidate sample to the training sample and the label information of the kth candidate sample.

[0170] In a possible implementation, the sum of the second supplementary information of the kth candidate sample of the electronic device for the training sample and the label information of the kth candidate sample is determined as the first supplementary information of the kth candidate sample for the training sample.

[0171] In one possible implementation, the label information of the kth candidate sample is projected using a second projection matrix to obtain the label projection information of the kth candidate sample; the sum of the second supplementary information of the kth candidate sample to the training sample and the label projection information of the kth candidate sample is determined as the first supplementary information of the kth candidate sample to the training sample.

[0172] Exemplarily, the electronic device determines the first supplementary information of the kth candidate sample to the training sample by the following formula (4):

[0173]

[0174] W v ∈R d×d is the second projection matrix, Y k is the label information of the kth candidate sample.

[0175] The above introduces the process of determining the first supplementary information of the kth candidate sample in at least one candidate sample for the training sample by the electronic device. Referring to the above steps, the electronic device can determine the first supplementary information of each candidate sample in at least one candidate sample for the training sample, and then perform the above step S103-B to supplement the first feature information of the training sample based on the first supplementary information of the at least one candidate sample for the training sample, and determine the supplemented feature information of the training sample.

[0176] The embodiment of the present application does not limit the specific method of supplementing the first feature information of the training sample based on the first supplementary information of the training sample of at least one candidate sample and determining the supplemented feature information of the training sample.

[0177] In some embodiments, the electronic device adds the sum or average of the first supplementary information of the training sample of at least one candidate sample to the first feature information of the training sample to obtain the supplemented feature information of the training sample.

[0178] In some embodiments, the above S103-B includes the following steps S103-B1 to S103-B3:

[0179] S103-B1, for a k-th candidate sample among the at least one candidate sample, determining a weight of the k-th candidate sample based on a similarity between the training sample and the k-th candidate sample;

[0180] S103-B2, based on the weight of at least one candidate sample, weighting the first supplementary information of the at least one candidate sample on the training sample to obtain weighted supplementary information;

[0181] S103-B3: Determine supplemented feature information of the training sample based on the first feature information of the training sample and the weighted supplemented information.

[0182] In this implementation, if Figure 6B As shown, for each candidate sample in at least one candidate sample, for example, the kth candidate sample, a weight of the kth candidate sample is determined based on the similarity between the kth candidate sample and the training sample, wherein the weight is positively correlated with the similarity.

[0183] The embodiment of the present application does not limit the specific manner in which the electronic device determines the weight of the kth candidate sample based on the similarity sim between the kth candidate sample and the training sample. For example, the similarity sim between the kth candidate sample and the training sample is determined as the weight of the kth candidate sample.

[0184] In a possible implementation, as shown in FIG6 , the similarity sim between the kth candidate sample and the training sample is normalized to obtain the weight of the kth candidate sample.

[0185] For example, the electronic device determines the weight of the kth candidate sample by the following formula (5):

[0186] P k =Softmax(Sim(h input ,h k candidate )) (5)

[0187] Among them, P k is the weight of the kth candidate sample, Sim(h input ,h k candidate ) is the similarity between the training sample and the kth candidate sample, and Softmax is the normalization function.

[0188] Based on the above steps, the electronic device can determine the weight of each candidate sample in at least one candidate sample, such as Figure 6B As shown, based on the weight of at least one candidate sample, the first supplementary information of the training sample of at least one candidate sample is weighted to obtain weighted supplementary information. Then, based on the first feature information of the training sample and the weighted supplementary information, the supplemented feature information of the training sample is determined. For example, as shown in FIG6 , the sum of the first feature information of the training sample and the weighted supplementary information is determined as the supplemented feature information of the training sample.

[0189] Exemplarily, the electronic device may determine the supplemented feature information of the training sample by the following formula (6):

[0190]

[0191] Among them, h is the supplemented feature information of the training sample, h input is the first feature information of the training sample, P k is the weight of the kth candidate sample in at least one candidate sample (it can also be called the normalized similarity), is the first feature information of the kth candidate sample, Y k is the label information of the kth candidate sample, It is the supplementary information of the kth candidate sample to the training sample. It is the weighted supplementary information.

[0192] S104: Perform detection processing on the supplemented feature information of the training sample through the detection module in the data detection model to obtain the detection result of the training sample.

[0193] In the embodiment of the present application, the electronic device determines the supplemented feature information of the training sample based on the above steps, such as Figure 7 As shown, the supplemented feature information of the training sample is input into the detection module for data detection processing to obtain the detection result of the training sample.

[0194] S105: Determine the loss of the data detection model based on the detection results of the training samples, and train the data detection model based on the loss.

[0195] As can be seen from the above, when the training sample is used to train the data detection model, the embodiment of the present application determines the label information of the training sample. In this way, the electronic device inputs the training sample into the data detection model for data detection, obtains the detection result of the training sample output by the data detection model, and then determines the loss of the data detection model based on the deviation between the detection result of the training sample and the label information of the training sample. The embodiment of the present application does not limit the type of loss function used by the electronic device to determine the loss of the data detection model based on the detection result of the training sample and the label information of the training sample. Then, based on the loss, the parameters in the data detection model are updated. Then, a training sample is reselected from the training data set, and the data detection model with updated parameters is iteratively trained, and the training is repeated many times until the model training end condition is reached. The model training end condition may include the number of training times reaching a preset number of times, and / or the loss of the model reaches a preset value.

[0196] The model training method provided in the embodiment of the present application trains a data detection model, and the data detection model includes a retrieval module and a detection module. During model training, first obtain a training sample and Q candidate samples. Through the retrieval module in the data detection model, extract the first feature information of the training sample and the Q candidate samples respectively, the training sample is the state value of L biomarkers of the first object, and the candidate sample is the state value of L biomarkers of other objects different from the first object; then, based on the first feature information of the training sample and the first feature information of the Q candidate samples, select at least one candidate sample related to the training sample from the Q candidate samples, and then based on the first feature information of the at least one candidate sample, supplement the first feature information of the training sample to obtain the supplemented feature information of the training sample; finally, through the detection module in the data detection model, the supplemented feature information of the training sample is detected and processed to obtain the detection result of the training sample, and then based on the detection result of the training sample, determine the loss of the data detection model, and train the data detection model based on the loss. That is, in an embodiment of the present application, at least one candidate sample related to the training sample is selected from Q candidate samples, and the first feature information of the at least one candidate sample is used to supplement the first feature information of the training sample, and the features of similar objects are integrated to obtain the supplemented feature information of the training sample. When the supplemented feature information of the training sample is used for data detection, multiple different classification detections can be achieved. For example, detection reference results for multiple different types of early cancer can be obtained, and the accuracy of medical detection can be improved.

[0197] The above introduces the overall training process of the data detection model. In some embodiments, in order to further improve the accuracy of data detection, the training samples are retrieved multiple times to use different linear projections of the training samples to the d dimension to further improve the extraction of the features of the training samples. Figure 8 As shown, the retrieval module of the embodiment of the present application includes R detection units. Fig. 9 The model training method of the embodiment of the present application is further introduced.

[0198] Fig. 9 A flowchart of a model training method provided in one embodiment of the present application.

[0199] like Fig. 9 As shown, the training process of the data detection model in the embodiment of the present application includes the following steps:

[0200] S201. Obtain training samples and Q candidate samples.

[0201] The training samples are the state values ​​of L biomarkers of a first object, and the candidate samples are the state values ​​of L biomarkers of other objects different from the first object.

[0202] The specific implementation process of the above S201 can refer to the relevant description of the above S101, which will not be repeated here.

[0203] S202. For the j-th retrieval unit among the R retrieval units, extract the j-th first feature information of the training sample and the Q candidate samples respectively through the j-th retrieval unit.

[0204] like Figure 8 As shown, the retrieval module of the embodiment of the present application includes R retrieval units, which run in parallel. The R retrieval units process the input training samples and Q candidate samples separately, and finally obtain R supplementary feature information of the training samples.

[0205] In the embodiment of the present application, each of the R detection units has the same processing process for the training sample and the Q candidate samples corresponding to the training sample. For the sake of ease of description, the jth detection unit is taken as an example for explanation.

[0206] In the embodiment of the present application, the data processing process of each of the R detection units is consistent with the data processing process of the retrieval module in the above embodiment, and the relevant specific process can refer to the description of the above embodiment.

[0207] like Figure 8 As shown, the electronic device inputs the training sample and the Q candidate samples into R detection units respectively. For each detection unit, for example, the jth detection unit, the jth detection unit extracts the first feature information of the training sample and the first feature information of each candidate sample in the Q candidate samples respectively. For the convenience of description, the first feature information of the training sample extracted by the jth detection unit is recorded as the jth first feature information of the training sample. Similarly, the first feature information of the candidate sample extracted by the jth detection unit is recorded as the jth first feature information of the candidate sample.

[0208] In the embodiment of the present application, the specific manner in which the jth retrieval unit extracts the jth first feature information of the training sample and the jth first feature information of the Q candidate samples includes at least the following manners:

[0209] Method 1: When the j-th detection unit extracts the j-th first feature information of the training sample, it does not consider the Q candidate samples. At the same time, when extracting the j-th first feature information of each of the Q candidate samples, it does not consider the training sample and other candidate samples in the Q candidate samples.

[0210] In method 2, the j-th detection unit considers the Q candidate samples when extracting the j-th first feature information of the training sample, and at the same time, when extracting the j-th first feature information of each of the Q candidate samples, it does not consider the training sample and other candidate samples in the Q candidate samples.

[0211] The embodiment of the present application does not limit the specific network structure of the j-th retrieval unit.

[0212] In some embodiments, Fig.10 As shown, the j-th retrieval unit includes an encoder, which is denoted as the j-th encoder. Through the j-th retrieval unit, training samples (S input ) and Q candidate samples (S candidate ), comprising: extracting the j-th first feature information of the training sample and Q candidate samples respectively through the j-th encoder.

[0213] like Fig.10 As shown, each of the R retrieval units in the embodiment of the present application includes an encoder, wherein the encoders included in different retrieval units may be different. In this way, the input training samples and candidate samples are subjected to different linear projections to the feature space of d dimension through the different encoders respectively included in the R retrieval units. In each different retrieval process, different learnable encoders are used to encode the training samples and candidate samples, and then the candidate samples related to the training samples are identified from different angles, and the features of the related candidate samples are used to supplement the features of the training samples, thereby enriching the feature extraction of the training samples.

[0214] The embodiment of the present application does not limit the grid structure of the encoder.

[0215] In one example, the encoder included in the j-th retrieval unit may include a linear layer to project the training samples and candidate samples to a hidden dimension.

[0216] S203: Based on the j-th first feature information of the training sample and the j-th first feature information of the Q candidate samples, select Kj candidate samples related to the training sample from the Q candidate samples.

[0217] In the embodiment of the present application, for each of the R retrieval units, for example, the jth retrieval unit, the electronic device extracts the jth first feature information of the training sample and the jth first feature information of the candidate sample through the jth retrieval unit. Then, based on the jth first feature information of the training sample and the jth first feature information of the Q candidate samples, Kj candidate samples related to the training sample are selected from the Q candidate samples.

[0218] It should be noted that, for different retrieval units, the number of corresponding Kj candidate samples may be different, and the selected candidate samples may also be different.

[0219] The embodiment of the present application does not limit the specific manner in which the electronic device selects Kj candidate samples related to the training sample from the Q candidate samples based on the j-th first feature information of the training sample and the j-th first feature information of the Q candidate samples.

[0220] In some embodiments, the electronic device clusters the j-th first feature information of the training sample and the j-th first feature information of Q candidate samples, and determines the candidate samples among the Q training samples that belong to the same cluster as the training samples as Kj candidate samples related to the training samples.

[0221] In some embodiments, the above S203 includes the following steps S203-A and S203-B:

[0222] S203-A, determining the similarity between the training sample and each of the Q candidate samples based on the j-th first feature information of the training sample and the j-th first feature information of the Q candidate samples;

[0223] S203-B, based on the similarity, select Kj candidate samples from the Q candidate samples.

[0224] The embodiment of the present application does not limit the specific manner in which the electronic device determines the similarity between the candidate sample and the training sample based on the j-th first feature information of the candidate sample and the j-th first feature information of the training sample.

[0225] In a possible implementation, the electronic device uses a distance calculation method such as cosine distance or L2 distance to calculate the distance between the jth first feature information of the candidate sample and the jth first feature information of the training sample, and then determines the similarity between the candidate sample and the training sample based on the distance. In one example, the reciprocal of the distance between the jth first feature information of the candidate sample and the jth first feature information of the training sample can be determined as the similarity between the candidate sample and the training sample.

[0226] In some embodiments, the above S203-A includes the following steps S203-A1 to S203-A3:

[0227] S203-A1. For the qth candidate sample among the Q candidate samples, use the first projection matrix to project the jth first feature information of the qth candidate sample to obtain the jth second feature information of the qth candidate sample, where q is a positive integer less than or equal to Q;

[0228] S203-A2, projecting the j-th first feature information of the training sample using the first projection matrix to obtain the j-th second feature information of the training sample;

[0229] S203-A3, determining the distance between the j-th second feature information of the training sample and the j-th second feature information of the q-th candidate sample, and determining the similarity between the training sample and the q-th candidate sample based on the distance.

[0230] Specifically, for the qth candidate sample among the Q candidate samples, the first projection matrix W is used k The jth first feature information of the qth candidate sample is projected. For the convenience of description, the projected jth first feature information is recorded as the jth second feature information of the qth candidate sample. Similarly, the first projection matrix W is used k The jth first feature information of the training sample is projected to obtain the jth second feature information of the training sample. Next, the distance between the jth second feature information of the training sample and the jth second feature information of the qth candidate sample is calculated, and the distance is determined as the similarity between the training sample and the qth candidate sample.

[0231] In one example, the electronic device uses an L2 distance calculation method to determine the distance between the jth second feature information of the training sample and the jth second feature information of the qth candidate sample. Specifically, the method shown in the above formula (2) can be referred to.

[0232] Based on the above method, the electronic device can determine the similarity between each of the Q candidate samples and the training sample.

[0233] Next, the electronic device selects Kj candidate samples related to the training sample from the Q training samples based on the similarity between each of the Q candidate samples and the training sample.

[0234] In one example, the electronic device selects Kj candidate samples having the greatest similarity with the training sample from the Q candidate samples as the Kj candidate samples related to the training sample.

[0235] S204 . Supplement the jth first feature information of the training sample based on the jth first feature information of the Kj candidate samples to obtain the jth supplemented feature information of the training sample in the jth retrieval unit.

[0236] In an embodiment of the present application, for the j-th retrieval unit among the R retrieval units, the electronic device can select a training sample from the Q candidate samples based on the above steps. The electronic device can supplement the first feature information of the training sample based on the first feature information of at least one candidate sample under the j-th retrieval unit. The specific method of obtaining the supplemented feature information of the training sample is not limited.

[0237] In some embodiments, the electronic device may fuse the jth first feature information of the training sample with the jth first feature information of the Kj candidate samples to obtain the jth supplemented feature information of the training sample. For example, based on the similarity between the Kj candidate samples and the training sample, the weight of each candidate sample in the Kj candidate samples is determined, and then based on the weight, the jth first feature information of the training sample is weighted with the jth first feature information of the Kj candidate samples to obtain the jth supplemented feature information of the training sample.

[0238] In some embodiments, the above S204 includes the following steps S204-A and S204-B:

[0239] S204-A, for the kth candidate sample among the Kj candidate samples, determine the distance between the jth second feature information of the training sample and the jth second feature information of the kth candidate sample, and based on the distance, determine the first supplementary information of the kth candidate sample for the training sample, where k is a positive integer less than or equal to K;

[0240] S204-B, based on the Kj candidate samples, respectively supplement the first supplementary information of the training sample and supplement the j-th first feature information of the training sample to determine the j-th supplemented feature information of the training sample.

[0241] In this implementation, the distance between the training sample and the Kj candidate samples is used as the jth supplementary feature in the hidden layer. Specifically, for the kth candidate sample among the Kj candidate samples, the distance between the jth second feature information of the training sample and the jth second feature information of the kth candidate sample is determined, and based on the distance, the first supplementary information of the kth candidate sample to the training sample is determined.

[0242] The embodiment of the present application does not limit the specific method of determining the first supplementary information of the kth candidate sample to the training sample based on the distance.

[0243] In some embodiments, the electronic device may determine the distance between the j-th second feature information of the training sample and the j-th second feature information of the k-th candidate sample as the first supplementary information of the k-th candidate sample to the training sample.

[0244] In some embodiments, the electronic device performs activation processing and then random deactivation processing on the distance between the j-th second feature information of the training sample and the j-th second feature information of the k-th candidate sample, and performs lossless linear transformation on the random deactivation processing result to obtain the second supplementary information of the k-th candidate sample to the training sample. Then, the second supplementary information of the k-th candidate sample to the training sample is determined as the first supplementary information of the k-th candidate sample to the training sample.

[0245] Exemplarily, the electronic device uses the above formula (3) to determine the second supplementary information, and then determines the second supplementary information as the first supplementary information of the kth candidate sample for the training sample.

[0246] In some embodiments, since the Kj candidate samples are similar to the training samples, the label information of the Kj candidate samples can also provide additional information about the characteristics of the training samples. Based on this, the above S204-A determines the first supplementary information of the kth candidate sample to the training sample based on the distance, including the following steps S204-A1 and S204-A2:

[0247] S204-A1, performing linear activation processing on the distance and then performing random deactivation processing, and performing lossless linear transformation on the random deactivation processing result to obtain the second supplementary information of the k-th candidate sample to the training sample;

[0248] S204-A2: Determine the first supplementary information of the kth candidate sample for the training sample based on the second supplementary information of the kth candidate sample for the training sample and the label information of the kth candidate sample.

[0249] Among them, the distance between the j-th second feature information of the training sample and the j-th second feature information of the k-th candidate sample is linearly activated and then randomly deactivated, and the random deactivation result is losslessly linearly transformed to obtain the second supplementary information of the k-th candidate sample for the training sample. The specific process can be referred to the description of the above embodiment and will not be repeated here.

[0250] Next, the electronic device determines the first supplementary information of the kth candidate sample to the training sample based on the second supplementary information of the kth candidate sample to the training sample and the label information of the kth candidate sample.

[0251] In a possible implementation, the sum of the second supplementary information of the kth candidate sample of the electronic device for the training sample and the label information of the kth candidate sample is determined as the first supplementary information of the kth candidate sample for the training sample.

[0252] In one possible implementation, the label information of the kth candidate sample is projected using a second projection matrix to obtain the label projection information of the kth candidate sample; the sum of the second supplementary information of the kth candidate sample to the training sample and the label projection information of the kth candidate sample is determined as the first supplementary information of the kth candidate sample to the training sample.

[0253] The above describes the process of determining the first supplementary information of the kth candidate sample among Kj candidate samples for the training sample by the electronic device. Referring to the above steps, the electronic device can determine the first supplementary information of each candidate sample among the Kj candidate samples for the training sample, and then execute the above step S204-B to supplement the jth first feature information of the training sample based on the first supplementary information of the Kj candidate samples for the training sample, and determine the jth supplemented feature information of the training sample.

[0254] The embodiment of the present application does not limit the specific method of supplementing the j-th first feature information of the training sample based on the first supplementary information of the training sample based on Kj candidate samples and determining the j-th supplemented feature information of the training sample.

[0255] In some embodiments, the electronic device adds the sum or average of the first supplementary information of the training sample of the Kj candidate samples to the j-th first feature information of the training sample to obtain the j-th supplemented feature information of the training sample.

[0256] In some embodiments, the above S204-B includes the following steps S204-B1 to S204-B3:

[0257] S204-B1, for the kth candidate sample among the Kj candidate samples, determine the weight of the kth candidate sample based on the similarity between the training sample and the kth candidate sample;

[0258] S204-B2, based on the weights of the Kj candidate samples, weighting the first supplementary information of the Kj candidate samples on the training samples to obtain weighted supplementary information;

[0259] S204-B3: Determine j-th supplemented feature information of the training sample based on the j-th first feature information of the training sample and the weighted supplemented information.

[0260] In this implementation, for each candidate sample among the Kj candidate samples, for example, the kth candidate sample, a weight of the kth candidate sample is determined based on the similarity between the kth candidate sample and the training sample, wherein the weight is positively correlated with the similarity.

[0261] The embodiment of the present application does not limit the specific manner in which the electronic device determines the weight of the kth candidate sample based on the similarity sim between the kth candidate sample and the training sample. For example, the similarity sim between the kth candidate sample and the training sample is determined as the weight of the kth candidate sample.

[0262] In a possible implementation, the similarity sim between the kth candidate sample and the training sample is normalized to obtain the weight of the kth candidate sample. Exemplarily, the weight of each of the Kj candidate samples can be determined by the method shown in the above formula (5).

[0263] Based on the above steps, the electronic device can determine the weight of each of the Kj candidate samples, and based on the weights of the Kj candidate samples, perform weighted processing on the first supplementary information of the training sample by the Kj candidate samples to obtain weighted supplementary information. Then, based on the jth first feature information of the training sample and the weighted supplementary information, determine the jth supplemented feature information of the training sample. For example, the sum of the jth first feature information of the training sample and the weighted supplementary information is determined as the jth supplemented feature information of the training sample.

[0264] The above takes the jth retrieval unit among the R retrieval units as an example to introduce the specific process of determining the jth supplemented feature information of the training sample through the jth retrieval unit. Referring to the above method, the electronic device can determine the supplemented feature information of the training sample corresponding to each retrieval unit in the R retrieval units, and then obtain R supplemented feature information of the training sample, recorded as h 1 、h 2 …h j ……h r , where h j It is the j-th supplemented feature information of the training sample corresponding to the j-th retrieval unit.

[0265] S205 . Perform detection processing on the R supplemented feature information of the training samples under the R retrieval units respectively through the detection module to obtain the detection results of the training samples.

[0266] like Fig.11 As shown, the embodiment of the present application uses a detection module on the electronic device to detect and process R supplemented feature information of the training samples under R retrieval units respectively, to obtain the detection results of the training samples.

[0267] In some embodiments, the electronic device performs a fusion process (such as addition, concatenation, multiplication, etc.) on the R supplemented feature information of the training sample, and performs a detection process on the fused feature information through the inspection module to obtain a detection result of the training sample.

[0268] In some embodiments, Fig.12 As shown, the detection module includes a hybrid expert module, and the hybrid expert module includes a gating unit and E expert units. The gating unit is used to determine which top-n probability expert unit each input token is assigned to. At this time, the above S205 includes the following steps S205-A to S205-C:

[0269] S205-A. Based on the supplemented feature information of the training samples under R retrieval units, S input tokens are obtained, where S is a positive integer greater than 1.

[0270] In the embodiment of the present application, for the jth retrieval unit among the R retrieval units, the jth supplemented feature information of the training sample corresponding to the jth detection unit includes the supplemented feature information of L biomarkers, and the jth supplemented feature information includes d elements, that is, a vector with a length of d. In this way, the supplemented feature information of the training samples under the R retrieval units can form a matrix.

[0271] In some embodiments, the supplemented feature information of the training sample in each of the R retrieval units is taken as a row of the first matrix to obtain a first matrix including R rows; each column of the first matrix is ​​taken as an input token to obtain S input tokens, and if the dimension of the supplemented feature information under each unit is d, then S is equal to d.

[0272] Exemplarily, the first matrix H can be expressed as: H = Concat (h 1 ,h 2 ,...,h r ) T , where H∈R d×r , Cancat represents concatenation, each element in the d dimension can be regarded as an input token (also called a tag) in the subsequent mixture of experts (MoE) module, and the letter A is used to represent the input token in the embodiment of the present application. The r dimension is embedded as its hidden layer.

[0273] S205-B. For the sth input token among the S input tokens, determine the probability value of the sth input token being assigned to each of the E expert units through the gating unit.

[0274] Wherein, E is a positive integer greater than 1, and s is a positive integer less than or equal to S.

[0275] Exemplarily, the electronic device may determine the probability value of each input token A being assigned to each expert unit by the following formula (7):

[0276] G=H·W Gate (7)

[0277] Among them, W gate is the learnable parameter matrix, G is the probability value of the input token being sent to each expert unit, and H is the first matrix.

[0278] For example, Fig.13 As shown, the Fig.13 Take the example of input token number S equal to 6 and expert unit number E equal to 4. The electronic device determines the probability value of these 6 input tokens A1, A2, A3, A4, A5 and A6 being assigned to each of the 4 expert units through the gating unit. For example, the probability value of input token A1 being assigned to the 4 expert units is in is the probability value assigned to the first expert unit e1 for processing by the input token A1, and so on. The probability value assigned to the input token A1 for processing by the fourth expert unit e4.

[0279] S205-C. Based on the probability value, select T expert units for processing the sth input token from the E expert units, and process the sth input token respectively by the T expert units to obtain the fusion feature information of the sth input token.

[0280] Wherein, T is a positive integer less than E.

[0281] Based on the above steps, for the sth input token among the S input tokens, the electronic device can determine the probability value of the sth input token being assigned to each of the E expert units. Then, based on the probability value, T expert units for processing the sth input token are selected from the E expert units. That is, T expert units with the largest probability values ​​are selected from the E expert units.

[0282] Exemplarily, the electronic device may determine T expert units for processing the sth input token by the following formula (8):

[0283]

[0284] Among them, G s is the probability value of the E expert units to which the sth input token is assigned, and topN(·) represents the probability of each input token (d) in the expert dimension through W Gate The position of the Nth element with the highest probability after that. These T top experts will receive the sth input token, topT_indices s are T expert units used to process the sth input token.

[0285] Continue to refer to Fig.13, taking the sth input token as the first input token A1 as an example, Fig.13 As shown, the probability value assigned to the first expert unit e1 is and the probability value assigned to the second expert unit e2 Highest, so it can be determined that the first input token A1 is distributed to the first expert unit e1 and the second expert unit e2, and T is equal to 2 at this time.

[0286] In actual operation, some expert units may receive a large amount of input tokens, while other experts only receive a small amount of input tokens. This phenomenon is called unbalanced expert load. This situation is not only contrary to the initial design intention of the data detection model of the embodiment of the present application, but also affects the computing efficiency (for example, causing uneven communication load in distributed training). Based on this, the embodiment of the present application sets a fixed capacity value C (i.e. the maximum number of tags that each expert unit can accept) for each expert unit. If the number of input tokens currently received by a certain expert unit has exceeded its capacity, the expert unit will no longer receive input tokens. At this time, redundant input tokens are called overflow.

[0287] Based on the above steps, after the electronic device selects T expert units from P expert units for processing the sth input token, it distributes the sth input token to the T expert units for separate processing to obtain the fused feature information of the sth input token.

[0288] The embodiment of the present application does not limit the specific manner in which the electronic device processes the sth input token through T expert units respectively to obtain the fused feature information of the sth input token.

[0289] In some embodiments, the electronic device inputs the sth input token into T expert units respectively, obtains the feature information obtained by the T expert units processing the sth input token respectively, and then obtains T feature information. Then, the T feature information is fused, for example, the T feature information is added, or the average value of the T feature information is determined as the fused feature information of the sth input token.

[0290] In some embodiments, the above S205-C processes the sth input token by T expert units respectively to obtain the fusion feature information of the sth input token, including the following steps S205-C1 to S205-C3:

[0291] S205-C1, for each of the T expert units, perform feature fusion processing on each input token assigned to the expert unit through the expert unit to obtain output feature information of each input token, wherein each input token includes the sth input token;

[0292] S205-C2, based on the probability value of the sth input token being assigned to each of the T expert units, determine the weights corresponding to the T expert units respectively;

[0293] S205-C3. Based on the weights corresponding to the T expert units, weighted processing is performed on the output feature information of the s-th input token output by the T expert units to obtain fused feature information of the s-th input token.

[0294] In this implementation, for each of the T expert units, the electronic device performs feature fusion processing on each input token assigned to the expert unit through the expert unit to obtain output feature information of each input token, wherein the input tokens include the sth input token.

[0295] The embodiment of the present application does not limit the specific network structure of the expert unit. In one example, the expert unit includes two linear transformations and an activation function, such as a ReLU activation function, or other types of activation functions.

[0296] In some embodiments, each of the E expert units included in the data detection model of the embodiment of the present application has a unified structure, rather than different structures specially designed for certain specific classifications (such as cancer types), so that the data detection model proposed in the embodiment of the present application can be extended to other data sets.

[0297] In the embodiment of the present application, the electronic device also determines the weights corresponding to the T expert units respectively based on the probability value of the sth input token being assigned to each of the T expert units.

[0298] Exemplarily, the electronic device determines the weights corresponding to the expert units respectively through the following formula (9):

[0299] W = Softmax(TopT(G)) (9)

[0300] In this example, the electronic device uses the Softmax function to normalize the probability value of the sth input token being assigned to the T expert units, and can derive the weight of each input token being assigned to the T expert units.

[0301] Next, based on the weights corresponding to the T expert units, the output feature information of the s-th input token output by the T expert units is weighted to obtain the fused feature information of the s-th input token.

[0302] The above takes the sth input token among the S input tokens as an example to introduce the process of determining the fused feature information of the sth input token. Referring to this method, the electronic device can determine the fused feature information of each input token among the S input tokens.

[0303] Exemplarily, the electronic device determines the fusion feature information of the sth input token by the following formula (10):

[0304]

[0305] in, Its shape is the same as the first matrix H, Output feature information of the sth input token output by the expert unit.

[0306] Continue to refer to Fig.13 As shown, taking the sth input token as the first input token A1 as an example, the first input token is assigned to the first expert unit e1 and the second expert unit e2, wherein the weight of the first expert unit e1 processing the first input token A1 is The weight of the second expert unit e2 processing the first input token A1 is Thus, the weight matrix corresponding to the first input token A1 is The first expert unit processes the first input token A1 and obtains the output feature information A1 1. The second expert unit processes the first input token A1 and obtains the output feature information A12. At this time, the output feature information corresponding to the first input token A1 is In this way, the fused feature information of the first input token A1

[0307] Referring to the above method, the electronic device can determine the fusion feature information of each of the S input tokens. Then, the following step S205-D is performed.

[0308] S205-D: Based on the fused feature information of the S input tokens, obtain the mixed feature information of the training sample.

[0309] The embodiment of the present application does not limit the specific manner in which the electronic device obtains the mixed feature information of the training sample based on the fused feature information of S input tokens.

[0310] In some embodiments, the average value or sum of the fused feature information of the S input tokens is determined as the mixed feature information of the training sample.

[0311] In some embodiments, the fused feature information of the S input tokens is concatenated to obtain mixed feature information of the training sample.

[0312] S205-E: Based on the mixed feature information, the training samples are detected and processed to obtain detection results of the training samples.

[0313] The embodiment of the present application does not limit the specific manner in which the electronic device performs detection processing on the training samples based on the mixed feature information to obtain the detection results of the training samples.

[0314] In some embodiments, the mixed feature information of the training samples is classified to obtain the detection results of the training samples.

[0315] In some embodiments, a linear layer is used to perform a linear transformation on the mixed feature information of the training sample to reduce the dimension and aggregate the features. Subsequently, a simple linear classification head is used to map the hidden layer to the final output dimension. The probability is then normalized to generate the final output, predicting the probability value of the training sample belonging to the preset classification, and then obtaining the detection result of the training sample.

[0316] Exemplarily, the electronic device determines the detection result of the training sample by the following formula (11):

[0317] p y =Softmax(FC head (Linear(h MoE ))) (11)

[0318] Among them, p y is the detection result of the training sample, that is, the probability value of the training sample belonging to the preset classification, Linear is the linear layer, FC head is a linear classification head and Softmax is a normalization function.

[0319] S206: Determine the loss of the data detection model based on the detection results of the training samples, and train the data detection model based on the loss.

[0320] In some embodiments, the electronic device determines a first loss based on the detection results of the training samples and the label information of the training samples, and further determines the first loss as the loss of the data detection model.

[0321] Exemplarily, the electronic device determines the first loss L by the following formula (12): main (p y , y):

[0322]

[0323] Among them, n class is the number of preset categories, y i is the true probability that the training sample belongs to the i-th category in the preset classification, The probability value that the training sample predicted by the data detection model belongs to the i-th category.

[0324] It should be noted that electronic devices use the classification cross entropy loss function L main In addition to using θ to calculate the first loss, other loss functions may also be used to calculate the first loss, and the embodiments of the present application do not limit this.

[0325] In some embodiments, the above S206 includes the following steps S206-A to S206-C:

[0326] S206-A, determining a first loss based on the detection result of the training sample and the label information of the training sample;

[0327] S206-B, determining a second loss based on the number of input labels assigned to each of the E expert units and the assignment probability value;

[0328] S206-C. Determine the loss of the data detection model based on the first loss and the second loss.

[0329] In this implementation, in order to achieve uniform distribution among expert units, prevent uneven distribution of input tokens, and avoid discarding too many input tokens due to overlap, the embodiment of the present application introduces a balanced loss function L balance This loss function aims to minimize the proportion of input tokens c received by each expert unit e / S squared difference. However, due to c e is obtained through the top-T process, which is not differentiable. To solve this problem, the present embodiment uses the average probability m of all tokens received by each expert unit e As a differentiable alternative, replace the square with m e (c e / S). In this way, the embodiment of the present application can optimize the loss function by gradient descent, effectively retain the original concept, solve the problem of non-differentiability, and ensure the balanced distribution of input tokens allocated by the expert unit.

[0330] Exemplarily, the electronic device can determine the second loss L by the following formula (13): balance :

[0331]

[0332] Among them, E is the number of expert units, c e is the number of input tokens assigned to the e-th expert unit among the E expert units, S is the total number of input tokens, m e is the average probability of all tokens assigned to the e-th expert unit.

[0333] After the electronic device determines the first loss and the second loss based on the above steps, it determines the loss of the data detection model based on the first loss and the second loss.

[0334] In one example, the electronic device determines the sum of the first loss and the second loss as the loss of the data detection model.

[0335] In one example, the electronic device multiplies the second loss by the balance coefficient and adds the result to the first loss to obtain the loss of the data detection model.

[0336] Exemplarily, the electronic device determines the loss L of the data detection model according to the following formula (14):

[0337] L=L main +α*L balance (14)

[0338] Here, α is a coefficient used for load balancing.

[0339] Next, the electronic device adjusts the parameters in the data detection model based on the above loss. In this way, multiple rounds of training are performed to obtain a trained data detection model.

[0340] The model training method provided by the embodiment of the present application, the data detection model includes a retrieval module and a detection module, and the data detection model includes R detection units. The R first feature information of the training sample and the R first feature information of each training sample in the Q candidate samples are extracted through the R detection units. For each detection unit, for example, the jth detection unit, based on the jth first feature information of the training sample and the jth first feature information of the Q training samples, Kj candidate samples related to the training sample are selected from the Q candidate samples. Then, based on the jth first feature information of the Kj candidate samples, the jth first feature information of the training sample is supplemented to obtain the jth supplemented feature information of the training sample. By analogy, the R supplemented feature information of the training sample can be obtained. Then, through the detection module in the data detection model, the R supplemented feature information of the training sample is detected and processed to obtain the detection result of the training sample, for example, through the mixed expert module, the R supplemented feature information of the training sample is mixed and processed and then data detection is performed to obtain the detection result of the training sample. Then, based on the detection results of the training samples, the loss of the data detection model is determined, and the data detection model is trained based on the loss. It can be seen that by retrieving the features of similar objects multiple times from different linear transformation spaces, the R supplemented feature information of the training samples includes richer features, and then when the R supplemented feature information of the training samples is used for data detection, multiple different classification detections can be achieved, for example, multiple different types of early cancer detection reference results can be obtained, and the accuracy of medical detection can be improved.

[0341] The above introduces the training process of the data detection model. The following introduces the data detection process based on the trained data detection model.

[0342] Fig.14 The flowchart of the data detection method provided in one embodiment of the present application is shown in FIG. The execution subject of the embodiment of the present application may be a device having a data detection function, such as a data detection device. The data detection device may be the above-mentioned Figure 1 The server 102 or terminal device 101 shown may also be Figure 1 The system is composed of the server 102 or 101 shown. For the convenience of description, the following embodiments take electronic devices as an example to illustrate the method of the embodiments of the present application.

[0343] like Fig.14 As shown, the data detection method of the embodiment of the present application includes:

[0344] S301, obtaining target data to be detected and Q candidate data, and extracting first feature information of the target data and the Q candidate data respectively.

[0345] The target data are the state values ​​of L biomarkers of the target object, the candidate data are the state values ​​of L biomarkers of other objects different from the target object, and Q and L are both positive integers greater than 1.

[0346] In an embodiment of the present application, a candidate data set is involved, and the candidate data set includes multiple different candidate data. When it is necessary to detect the target data, the electronic device can obtain Q candidate data corresponding to the target data from the candidate data set. For example, Q candidate data are randomly selected from the candidate data set.

[0347] The embodiment of the present application does not limit the specific manner in which the electronic device extracts the first feature information of the target data and the first feature information of the Q candidate data. For example, the first feature information of the target data and the first feature information of the Q candidate data are extracted through a trained feature extraction model.

[0348] In some embodiments, the electronic device extracts the first feature information of the target data and the first feature information of the Q candidate data through the retrieval module in the above-mentioned data detection model.

[0349] It should be noted that the specific implementation process of the above S301 is basically the same as the implementation process of the above S101. Please refer to the relevant description of the above S101 and will not repeat it here.

[0350] S302: Based on the first feature information of the target data and the first feature information of the Q candidate data, select at least one candidate data related to the target data from the Q candidate data.

[0351] The specific implementation process of the above S302 is basically the same as the implementation process of the above S102. Please refer to the relevant description of the above S102 and will not be repeated here.

[0352] S303: Supplement the first feature information of the target data based on the first feature information of at least one candidate data to obtain supplemented feature information of the target data.

[0353] In an embodiment of the present application, supplementing the first characteristic information of the target data can be understood as expanding the first characteristic information of the target data to enrich the first characteristic information of the target data, so that the supplemented first characteristic information includes not only the characteristic information of the target object but also the characteristic information of other objects corresponding to the candidate data.

[0354] The specific implementation process of the above S303 is basically the same as the implementation process of the above S103. Please refer to the relevant description of the above S103 and will not be repeated here.

[0355] S304: Detect and process the supplemented characteristic information of the target data to obtain a detection result of the target data.

[0356] The embodiment of the present application does not limit the specific method of the electronic device detecting and processing the supplemented feature information of the target data to obtain the detection result of the target data. For example, the detection result of the target data is obtained by detecting and processing the supplemented feature information of the target data through a trained data detection model.

[0357] In some embodiments, the electronic device detects and processes the supplemented feature information of the target data through the detection module in the above-mentioned data detection model to obtain the detection result of the target data.

[0358] It should be noted that if the data detection method of the embodiment of the present application is applied to disease diagnosis, the detection result obtained by the embodiment of the present application is a reference result of the disease, not a direct result of disease diagnosis. For example, when the data detection method of the embodiment of the present application is applied to the detection of early cancer, the embodiment of the present application can obtain a detection reference result of whether the target object belongs to multiple types of early cancer.

[0359] The specific implementation process of the above S304 is basically the same as the implementation process of the above S104. Please refer to the relevant description of the above S104 and will not be repeated here.

[0360] The data detection method provided by the embodiment of the present application, when detecting data, first obtains the target data to be detected, and Q candidate data. The first feature information of the target data and the Q candidate data is extracted respectively, and the target data is the state value of L biomarkers of the target object; then, based on the first feature information of the target data and the first feature information of the Q candidate data, at least one candidate data related to the target data is selected from the Q candidate data, and then based on the first feature information of the at least one candidate data, the first feature information of the target data is supplemented to obtain the supplemented feature information of the target data; finally, the supplemented feature information of the target data is detected and processed to obtain the detection result of the target data. It can be seen that the embodiment of the present application selects at least one candidate data related to the target data from the Q candidate data, and uses the first feature information of the at least one candidate data to supplement the first feature information of the target data, integrates the features of similar objects, obtains the supplemented feature information of the target data, and then uses the supplemented feature information of the target data for data detection, which can realize the detection of multiple different classifications, for example, the detection reference results of multiple different types of early cancer can be obtained, and the detection accuracy can be improved.

[0361] The above describes the data detection process involved in the embodiments of the present application. Figure 8 As shown, the retrieval module of the embodiment of the present application includes R detection units.

[0362] Fig.15 The following is a flow chart of a data detection method provided in one embodiment of the present application. Fig.15 As shown, the data detection process of the embodiment of the present application includes the following steps:

[0363] S401, obtaining target data and Q candidate data.

[0364] The specific implementation process of the above S401 is consistent with the above S301. Please refer to the relevant description of the above S301 and will not be repeated here.

[0365] S402 . For the j-th retrieval unit among the R retrieval units, extract the j-th first feature information of the target data and the Q candidate data respectively through the j-th retrieval unit.

[0366] Wherein, j is a positive integer less than or equal to R.

[0367] For example, through the j-th encoder, the j-th first feature information of the target data and the Q candidate data are extracted respectively.

[0368] The specific implementation process of the above S402 can refer to the relevant description of the above S202, which will not be repeated here.

[0369] S403 . Based on the j-th first feature information of the target data and the j-th first feature information of the Q candidate data, select Kj candidate data related to the target data from the Q candidate data.

[0370] For example, based on the j-th first feature information of the target data and the j-th first feature information of the Q candidate data, the similarity between the target data and each of the Q candidate data is determined; based on the similarity, Kj candidate data are selected from the Q candidate data.

[0371] In some embodiments, the above-mentioned method of determining the similarity between the target data and each of the Q candidate data based on the j-th first feature information of the target data and the j-th first feature information of the Q candidate data includes: for the q-th candidate data among the Q candidate data, using the first projection matrix to project the j-th first feature information of the q-th candidate data to obtain the j-th second feature information of the q-th candidate data, where q is a positive integer less than or equal to Q; using the first projection matrix to project the j-th first feature information of the target data to obtain the j-th second feature information of the target data; determining the distance between the j-th second feature information of the target data and the j-th second feature information of the q-th candidate data, and determining the similarity between the target data and the q-th candidate data based on the distance.

[0372] The specific implementation process of the above S403 can refer to the relevant description of the above S203, which will not be repeated here.

[0373] S404. Supplement the j-th first feature information of the target data based on the j-th first feature information of the Kj candidate data to obtain the j-th supplemented feature information of the target data under the j-th retrieval unit.

[0374] For example, for the kth candidate data among Kj candidate data, determine the distance between the jth second feature information of the target data and the jth second feature information of the kth candidate data, and based on the distance, determine the first supplementary information of the kth candidate data to the target data, where k is a positive integer less than or equal to K; based on the first supplementary information of the target data by the Kj candidate data, respectively, supplement the jth first feature information of the target data to determine the jth supplemented feature information of the target data.

[0375] In some embodiments, the above-mentioned determination of the first supplementary information of the kth candidate data to the target data based on the distance includes: performing linear activation processing on the distance and then performing random deactivation processing, and performing lossless linear transformation on the result of the random deactivation processing to obtain the second supplementary information of the kth candidate data to the target data; based on the second supplementary information of the kth candidate data to the target data and the label information of the kth candidate data, determining the first supplementary information of the kth candidate data to the target data.

[0376] In some embodiments, the above-mentioned determining the first supplementary information of the kth candidate data to the target data based on the second supplementary information of the kth candidate data to the target data and the label information of the kth candidate data includes: using a second projection matrix to project the label information of the kth candidate data to obtain the label projection information of the kth candidate data; determining the sum of the second supplementary information of the kth candidate data to the target data and the label projection information of the kth candidate data as the first supplementary information of the kth candidate data to the target data.

[0377] In some embodiments, the above-mentioned method of supplementing the first supplementary information of the target data based on the Kj candidate data respectively and the j-th first feature information of the target data to determine the j-th supplemented feature information of the target data includes: for the k-th candidate data among the Kj candidate data, based on the similarity between the target data and the k-th candidate data, determining the weight of the k-th candidate data; based on the weights of the Kj candidate data, weighting the first supplementary information of the target data by the Kj candidate data to obtain weighted supplementary information; based on the j-th first feature information of the target data and the weighted supplementary information, determining the j-th supplemented feature information of the target data.

[0378] The specific implementation process of the above S404 can refer to the relevant description of the above S204, which will not be repeated here.

[0379] S405 , through the detection module, the R supplemented feature information of the target data under the R retrieval units are detected and processed to obtain the detection result of the target data.

[0380] In some embodiments, the detection module includes a hybrid expert module, the hybrid expert module includes a gating unit and E expert units, and the detection module performs detection processing on the R supplemented feature information of the target data under the R retrieval units to obtain the detection result of the target data, including:

[0381] Based on the supplemented feature information of the target data in the R retrieval units, S input tokens are obtained, where S is a positive integer greater than 1. For example, the supplemented feature information of the target data in each of the R retrieval units is used as a row of the first matrix to obtain a first matrix including R rows; each column of the first matrix is ​​used as an input token to obtain S input tokens.

[0382] Next, for the sth input token among the S input tokens, the probability value of the sth input token being assigned to each of the E expert units is determined through the gating unit, where E is a positive integer greater than 1, and s is a positive integer less than or equal to S; based on the probability value, T expert units for processing the sth input token are selected from the E expert units, and the sth input token is processed by the T expert units respectively to obtain the fusion feature information of the sth input token, where T is a positive integer less than E. For example, for each of the T expert units, the expert unit performs feature fusion processing on each input token assigned to the expert unit to obtain the output feature information of each input token, and each input token includes the sth input token; based on the probability value of the sth input token being assigned to each of the T expert units, the weights corresponding to the T expert units are determined; based on the weights corresponding to the T expert units, the output feature information of the sth input token output by the T expert units is weighted to obtain the fusion feature information of the sth input token.

[0383] Then, based on the fused feature information of the S input tokens, the mixed feature information of the target data is obtained; based on the mixed feature information, the target data is detected and processed to obtain the detection result of the target data.

[0384] The specific implementation process of the above S405 can refer to the relevant description of the above S205, which will not be repeated here.

[0385] In the embodiment of the present application, multiple searches are performed from Q candidate data, and at least one candidate data related to the target data is obtained in each search, and then the features of the target data are supplemented based on the features of the candidate data obtained in each search, so that the features of the target data are supplemented differently in different linear spaces, thereby enriching the feature information of the target data. In this way, when data detection is performed based on the rich feature information of the target data, the detection accuracy of the target data can be improved.

[0386] The following example uses L biomarkers including L gene mutations, and the test results including the probability reference values ​​of the target data belonging to multiple cancer classifications. Fig.14 The data detection process of the embodiment of the present application is introduced.

[0387] Fig.16 The following is a flow chart of a data detection method provided in one embodiment of the present application. Fig.16 As shown, the data detection method of the embodiment of the present application includes the following steps:

[0388] S501, obtaining target data to be detected and Q candidate data.

[0389] The target data are the state values ​​of L gene mutations of the target object, the candidate data are the state values ​​of L gene mutations of other objects different from the target object, and Q and L are both positive integers greater than 1.

[0390] For example, Fig.17 As shown, the first dotted circle includes the status values ​​of L gene mutations of the target object, and the L gene mutations include ABL1, ACVR1, AKT1, etc., and the remaining dotted circles include the status values ​​of L gene mutations of other objects.

[0391] S502 . For the j-th retrieval unit among the R retrieval units, extract the j-th first feature information of the target data and the Q candidate data respectively through the j-th retrieval unit.

[0392] The specific implementation process of the above S502 can refer to the relevant description of the above S202, which will not be repeated here.

[0393] S503 . Based on the j-th first feature information of the target data and the j-th first feature information of the Q candidate data, select Kj candidate data related to the target data from the Q candidate data.

[0394] The specific implementation process of the above S503 can refer to the relevant description of the above S203, which will not be repeated here.

[0395] S504 . Supplement the j-th first feature information of the target data based on the j-th first feature information of the Kj candidate data to obtain the j-th supplemented feature information of the target data in the j-th retrieval unit.

[0396] like Fig.17 As shown, the electronic device inputs the target data and Q candidate data into R detection units respectively, and each detection unit processes the target data and Q candidate data, and finally obtains R supplemented feature information of the target data, such as h1, h2...hr. The R supplemented feature information is concatenated to obtain the first matrix H.

[0397] The specific implementation process of the above S504 can refer to the relevant description of the above S204, which will not be repeated here.

[0398] S505 , through the detection module, the R supplemented feature information of the target data under the R retrieval units are detected and processed to obtain a probability reference value that the target data belongs to a preset cancer classification.

[0399] In the embodiment of the present application, the probability reference value that the target data belongs to a preset cancer classification can be understood as a detection reference result that the target data belongs to a preset cancer classification, so that relevant personnel can diagnose the target object based on the detection reference result.

[0400] In some embodiments, Fig.17 As shown, the detection module includes a hybrid expert module, and the hybrid expert module includes a gating unit and E expert units. Through the detection module, the R supplemented feature information of the target data under the R retrieval units are detected and processed to obtain a probability reference value that the target data belongs to a preset cancer classification, including: based on the supplemented feature information of the target data under the R retrieval units, S input tokens are obtained, S is a positive integer greater than 1; for the sth input token among the S input tokens, the probability value of the sth input token being assigned to each expert unit in the E expert units is determined through the gating unit, E is a positive integer greater than 1, and s is a positive integer less than or equal to S; based on the probability value, T expert units for processing the sth input token are selected from the E expert units, and the sth input token is processed by the T expert units respectively to obtain fused feature information of the sth input token, T is a positive integer less than E; based on the fused feature information of the S input tokens, the mixed feature information of the target data is obtained; as shown Fig.17 As shown, based on the mixed feature information, the target data is classified and detected to obtain probability reference values ​​that the target data belongs to multiple cancer classifications.

[0401] The specific implementation process of the above S505 can refer to the relevant description of the above S205, which will not be repeated here.

[0402] The embodiment of the present application uses a hybrid expert module to extract the relationship between different mutations and take interpretability into consideration, which helps to discover potential unknown mutations of specific cancer types. The embodiment of the present application starts from the two dimensions of patient dimension and gene mutation status, learns the deep correlation between different patients and different gene mutation states, realizes the detection of different types of cancer, and obtains the detection reference results of target data belonging to multiple different types of cancer, thereby improving the detection range and detection accuracy of early cancer.

[0403] Combination of the above Figures 2 to 17 , describes in detail the method embodiment of the present application, and the following is combined with Fig.18 and Fig.19 , describe in detail the device embodiments of the present application.

[0404] Fig.18 It is a schematic block diagram of a data detection device provided in one embodiment of the present application.

[0405] like Fig.18 As shown, the data detection device 10 comprises:

[0406] The extraction unit 11 is used to obtain target data to be detected and Q candidate data, and extract first feature information of the target data and the Q candidate data respectively, wherein the target data is the state value of L biomarkers of the target object, and the candidate data is the state value of L biomarkers of other objects different from the target object, and Q and L are both positive integers greater than 1;

[0407] A selection unit 12, configured to select at least one candidate data related to the target data from the Q candidate data based on the first feature information of the target data and the first feature information of the Q candidate data;

[0408] A supplementing unit 13, configured to supplement the first feature information of the target data based on the first feature information of the at least one candidate data to obtain supplemented feature information of the target data;

[0409] The detection unit 14 is used to perform detection processing on the supplemented characteristic information of the target data to obtain a detection result of the target data.

[0410] In some embodiments, the L biomarkers include L gene mutations, and the detection results include probability reference values ​​that the target data belongs to multiple cancer classifications.

[0411] In some embodiments, the extraction unit 11 is specifically used to extract the jth first feature information of the target data and the Q candidate data respectively through the jth retrieval unit among the R retrieval units, and the data detection model includes a retrieval module, and the retrieval module includes the R retrieval units, where R is a positive integer greater than 1, and j is a positive integer less than or equal to R; the selection unit 12 is specifically used to select Kj candidate data related to the target data from the Q candidate data based on the jth first feature information of the target data and the jth first feature information of the Q candidate data, where Kj is a positive integer; the supplementation unit 13 is specifically used to supplement the jth first feature information of the target data based on the jth first feature information of the Kj candidate data, so as to obtain the jth supplemented feature information of the target data under the jth retrieval unit; the detection unit 14 is specifically used to perform detection processing on the R supplemented feature information of the target data under the R retrieval units through the detection module in the data detection model, so as to obtain the detection result of the target data.

[0412] In some embodiments, the j-th retrieval unit includes a j-th encoder, and the extraction unit 11 is specifically configured to extract the j-th first feature information of the target data and the Q candidate data respectively through the j-th encoder.

[0413] In some embodiments, the selection unit 12 is specifically used to determine the similarity between the target data and each of the Q candidate data based on the j-th first feature information of the target data and the j-th first feature information of the Q candidate data; and based on the similarity, select the Kj candidate data from the Q candidate data.

[0414] In some embodiments, the selection unit 12 is specifically used to project the jth first feature information of the qth candidate data among the Q candidate data using the first projection matrix to obtain the jth second feature information of the qth candidate data, where q is a positive integer less than or equal to Q; use the first projection matrix to project the jth first feature information of the target data to obtain the jth second feature information of the target data; determine the distance between the jth second feature information of the target data and the jth second feature information of the qth candidate data, and determine the similarity between the target data and the qth candidate data based on the distance.

[0415] In some embodiments, the supplementing unit 13 is specifically used to determine the distance between the j-th second feature information of the target data and the j-th second feature information of the k-th candidate data for the k-th candidate data among the Kj candidate data, and based on the distance, determine the first supplementary information of the k-th candidate data for the target data, where k is a positive integer less than or equal to K; based on the first supplementary information of the target data respectively provided by the Kj candidate data, the j-th first feature information of the target data is supplemented to determine the j-th supplemented feature information of the target data.

[0416] In some embodiments, the supplementing unit 13 is specifically used to perform linear activation processing on the distance and then perform random inactivation processing, and perform lossless linear transformation on the random inactivation processing result to obtain the second supplementary information of the kth candidate data with respect to the target data; based on the second supplementary information of the kth candidate data with respect to the target data and the label information of the kth candidate data, determine the first supplementary information of the kth candidate data with respect to the target data.

[0417] In some embodiments, the supplementing unit 13 is specifically used to use a second projection matrix to project the label information of the kth candidate data to obtain the label projection information of the kth candidate data; and determine the sum of the second supplementary information of the kth candidate data to the target data and the label projection information of the kth candidate data as the first supplementary information of the kth candidate data to the target data.

[0418] In some embodiments, the supplementing unit 13 is specifically used to determine the weight of the kth candidate data among the Kj candidate data based on the similarity between the target data and the kth candidate data; based on the weights of the Kj candidate data, perform weighted processing on the first supplementary information of the Kj candidate data to the target data to obtain weighted supplementary information; based on the jth first feature information of the target data and the weighted supplementary information, determine the jth supplemented feature information of the target data.

[0419] In some embodiments, the detection module includes a hybrid expert module, which includes a gating unit and E expert units. The detection unit 14 is specifically used to obtain S input tokens based on the supplemented feature information of the target data under the R retrieval units, where S is a positive integer greater than 1; for the sth input token among the S input tokens, determine the probability value of the sth input token being assigned to each expert unit in the E expert units through the gating unit, where E is a positive integer greater than 1, and s is a positive integer less than or equal to S; based on the probability value, select T expert units from the E expert units for processing the sth input token, and process the sth input token through the T expert units to obtain fused feature information of the sth input token, where T is a positive integer less than E; based on the fused feature information of the S input tokens, obtain the mixed feature information of the target data; based on the mixed feature information, perform detection processing on the target data to obtain the detection result of the target data.

[0420] In some embodiments, the detection unit 14 is specifically used to perform feature fusion processing on each input token assigned to the expert unit for each of the T expert units, so as to obtain output feature information of each input token, wherein each input token includes the sth input token; determine the weights corresponding to the T expert units respectively based on the probability value of the sth input token being assigned to each of the T expert units; and perform weighted processing on the output feature information of the sth input token respectively output by the T expert units based on the weights corresponding to the T expert units, so as to obtain fused feature information of the sth input token.

[0421] In some embodiments, the detection unit 14 is specifically configured to use the supplemented feature information of the target data in each of the R retrieval units as a row of a first matrix to obtain a first matrix including R rows; and use each column of the first matrix as an input token to obtain the S input tokens.

[0422] It should be understood that the device embodiment and the method embodiment may correspond to each other, and similar descriptions may refer to the method embodiment. To avoid repetition, they will not be described here. Specifically, Fig.18 The device shown can execute the embodiment of the above-mentioned data detection method, and the above-mentioned and other operations and / or functions of each module in the device are respectively for realizing the above-mentioned method embodiment, which will not be described in detail here for the sake of brevity.

[0423] Fig.19 It is a schematic block diagram of a model training device provided in one embodiment of the present application.

[0424] like Fig.19 As shown, the model training device 20 includes:

[0425] An extraction unit 21 is used to obtain a training sample and Q candidate samples, and respectively extract first feature information of the training sample and the Q candidate samples through a retrieval module in a data detection model, wherein the training sample is a state value of L biomarkers of a first subject, and the candidate sample is a state value of L biomarkers of other subjects different from the first subject, and both Q and L are positive integers greater than 1;

[0426] A selection unit 22, configured to select at least one candidate sample related to the training sample from the Q candidate samples based on the first feature information of the training sample and the first feature information of the Q candidate samples;

[0427] A supplementing unit 23, configured to supplement the first feature information of the training sample based on the first feature information of the at least one candidate sample to obtain supplemented feature information of the training sample;

[0428] A detection unit 24, configured to perform detection processing on the supplemented feature information of the training sample through a detection module in the data detection model to obtain a detection result of the training sample;

[0429] The training unit 25 is used to determine the loss of the data detection model based on the detection results of the training samples, and train the data detection model based on the loss.

[0430] In some embodiments, the retrieval module includes R retrieval units, where R is a positive integer greater than 1; an extraction unit 21, specifically used to extract the jth first feature information of the training sample and the Q candidate samples through the jth retrieval unit among the R retrieval units, where j is a positive integer less than or equal to R; a selection unit 22, specifically used to select Kj candidate samples related to the training sample from the Q candidate samples based on the jth first feature information of the training sample and the jth first feature information of the Q candidate samples; a supplementation unit 23, specifically used to supplement the jth first feature information of the training sample based on the jth first feature information of the Kj candidate samples, to obtain the jth supplemented feature information of the training sample under the jth retrieval unit; a detection unit 24, specifically used to perform detection processing on the R supplemented feature information of the training sample under the R retrieval units through the detection module, to obtain the detection result of the training sample.

[0431] In some embodiments, the detection module includes a hybrid expert module, which includes a gating unit and E expert units. The detection unit 24 is specifically used to obtain S input tokens based on the supplemented feature information of the target data under the R retrieval units, where S is a positive integer greater than 1; for the sth input token among the S input tokens, determine the probability value of the sth input token being assigned to each expert unit in the E expert units through the gating unit, where E is a positive integer greater than 1, and s is a positive integer less than or equal to S; based on the probability value, select T expert units from the E expert units for processing the sth input token, and process the sth input token through the T expert units to obtain fused feature information of the sth input token, where T is a positive integer less than E; based on the fused feature information of the S input tokens, obtain the mixed feature information of the target data; based on the mixed feature information, perform detection processing on the target data to obtain the detection result of the target data.

[0432] In some embodiments, the training unit 25 is specifically used to determine a first loss based on the detection results of the training samples and the label information of the training samples; determine a second loss based on the number of input labels assigned to each of the E expert units and the assignment probability value; and determine the loss of the data detection model based on the first loss and the second loss.

[0433] It should be understood that the device embodiment and the method embodiment may correspond to each other, and similar descriptions may refer to the method embodiment. To avoid repetition, they will not be described here. Specifically, Fig.19 The device shown can execute the above-mentioned model training method embodiment, and the aforementioned and other operations and / or functions of each module in the device are respectively for realizing the above-mentioned method embodiment, which will not be repeated here for the sake of brevity.

[0434] The above describes the device of the embodiment of the present application from the perspective of the functional module in conjunction with the accompanying drawings. It should be understood that the functional module can be implemented in hardware form, can be implemented by instructions in software form, and can also be implemented by a combination of hardware and software modules. Specifically, the steps of the method embodiment in the embodiment of the present application can be completed by the hardware integrated logic circuit and / or software form instructions in the processor, and the steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to perform, or a combination of hardware and software modules in the decoding processor to perform. Optionally, the software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads the information in the memory, and completes the steps in the above method embodiment in conjunction with its hardware.

[0435] Fig. 20 It is a schematic block diagram of an electronic device provided in an embodiment of the present application, and the electronic device may be the above-mentioned terminal device or server.

[0436] like Fig. 20 As shown, the electronic device 40 may include:

[0437] The memory 41 and the processor 42, the memory 41 is used to store the computer program 43 and transmit the program code 43 to the processor 42. In other words, the processor 42 can call and run the computer program 43 from the memory 41 to implement the method in the embodiment of the present application.

[0438] For example, the processor 42 may be configured to execute the steps in the above method according to the instructions in the computer program 43 .

[0439] In some embodiments of the present application, the processor 42 may include but is not limited to:

[0440] General-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, etc.

[0441] In some embodiments of the present application, the memory 41 includes but is not limited to:

[0442] Volatile memory and / or non-volatile memory. Among them, the non-volatile memory can be read-only memory (ROM), programmable ROM (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM) or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct RAM bus random access memory (DR RAM).

[0443] In some embodiments of the present application, the computer program 43 may be divided into one or more modules, which are stored in the memory 41 and executed by the processor 42 to complete the method for recording pages provided by the present application. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program 43 in the electronic device.

[0444] like Fig. 20 As shown, the electronic device 40 may further include:

[0445] The transceiver 44 may be connected to the processor 42 or the memory 41 .

[0446] The processor 42 may control the transceiver 44 to communicate with other devices, specifically, to send information or data to other devices, or to receive information or data sent by other devices. The transceiver 44 may include a transmitter and a receiver. The transceiver 44 may further include an antenna, and the number of antennas may be one or more.

[0447] It should be understood that the various components in the electronic device 40 are connected via a bus system, wherein the bus system includes not only a data bus but also a power bus, a control bus and a status signal bus.

[0448] According to one aspect of the present application, a computer storage medium is provided, on which a computer program is stored, and when the computer program is executed by a computer, the computer can perform the method of the above method embodiment. In other words, the present application embodiment also provides a computer program product containing instructions, and when the instructions are executed by a computer, the computer can perform the method of the above method embodiment.

[0449] According to another aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method of the above method embodiment.

[0450] In other words, when implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a digital video disc (digital video disc, DVD)), or a semiconductor medium (e.g., a solid state disk (solid state disk, SSD)), etc.

[0451] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0452] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the module is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0453] The modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. For example, each functional module in each embodiment of the present application may be integrated into a processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0454] The above contents are only specific implementation methods of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A data detection method, characterized in that: include: Obtain target data to be detected and Q candidate data, and extract first feature information of the target data and the Q candidate data respectively, wherein the target data is the state value of L biomarkers of the target object, and the candidate data is the state value of L biomarkers of other objects different from the target object, and Q and L are both positive integers greater than 1; Based on the first feature information of the target data and the first feature information of the Q candidate data, selecting at least one candidate data related to the target data from the Q candidate data; Based on the first feature information of the at least one candidate data, supplement the first feature information of the target data to obtain supplemented feature information of the target data; The supplemented characteristic information of the target data is detected and processed to obtain a detection result of the target data.

2. The method according to claim 1, characterized in that: The L biomarkers include L gene mutations, and the detection results include probability reference values ​​that the target data belongs to multiple cancer classifications.

3. The method according to claim 1 or 2, characterized in that: The extracting the first feature information of the target data and the Q candidate data respectively includes: for a j-th retrieval unit among the R retrieval units, extracting the j-th first feature information of the target data and the Q candidate data respectively through the j-th retrieval unit, the data detection model includes a retrieval module, the retrieval module includes the R retrieval units, R is a positive integer greater than 1, and j is a positive integer less than or equal to R; The selecting, based on the first feature information of the target data and the first feature information of the Q candidate data, at least one candidate data related to the target data from the Q candidate data comprises: selecting, based on the j-th first feature information of the target data and the j-th first feature information of the Q candidate data, Kj candidate data related to the target data from the Q candidate data, where Kj is a positive integer; The step of supplementing the first feature information of the target data based on the first feature information of the at least one candidate data to obtain supplemented feature information of the target data includes: supplementing the jth first feature information of the target data based on the jth first feature information of the Kj candidate data to obtain the jth supplemented feature information of the target data under the jth retrieval unit; The detecting and processing the supplemented characteristic information of the target data to obtain the detection result of the target data includes: using the detection module in the data detection model, detecting and processing the R supplemented characteristic information of the target data under the R retrieval units respectively to obtain the detection result of the target data.

4. The method according to claim 3, characterized in that The j-th retrieval unit includes a j-th encoder, and the j-th first feature information of the target data and the Q candidate data are respectively extracted by the j-th retrieval unit, including: The j-th first feature information of the target data and the Q candidate data are extracted respectively through the j-th encoder.

5. The method according to claim 3, characterized in that: The selecting Kj candidate data related to the target data from the Q candidate data based on the jth first feature information of the target data extracted by the jth retrieval unit and the jth first feature information of the Q candidate data comprises: Determine, based on the j-th first feature information of the target data and the j-th first feature information of the Q candidate data, a similarity between the target data and each of the Q candidate data; Based on the similarity, the Kj candidate data are selected from the Q candidate data.

6. The method according to claim 5, characterized in that The determining, based on the j-th first feature information of the target data and the j-th first feature information of the Q candidate data, the similarity between the target data and each of the Q candidate data comprises: For the qth candidate data among the Q candidate data, use the first projection matrix to project the jth first feature information of the qth candidate data to obtain the jth second feature information of the qth candidate data, where q is a positive integer less than or equal to Q; Projecting the j-th first feature information of the target data using the first projection matrix to obtain the j-th second feature information of the target data; A distance between the j-th second feature information of the target data and the j-th second feature information of the q-th candidate data is determined, and a similarity between the target data and the q-th candidate data is determined based on the distance.

7. The method according to claim 3, characterized in that The method of supplementing the jth first feature information of the target data based on the jth first feature information of the Kj candidate data to obtain the jth supplemented feature information of the target data under the jth retrieval unit includes: For the kth candidate data among the Kj candidate data, determine the distance between the jth second feature information of the target data and the jth second feature information of the kth candidate data, and based on the distance, determine the first supplementary information of the kth candidate data to the target data, where k is a positive integer less than or equal to K; Based on the first supplementary information of the target data respectively based on the Kj candidate data, the j-th first feature information of the target data is supplemented to determine the j-th supplemented feature information of the target data.

8. The method according to claim 7, characterized in that The determining, based on the distance, first supplementary information of the kth candidate data to the target data includes: Performing a linear activation process on the distance and then a random deactivation process, and performing a lossless linear transformation on the random deactivation process result to obtain second supplementary information of the k-th candidate data to the target data; Based on the second supplementary information of the kth candidate data with respect to the target data and the label information of the kth candidate data, the first supplementary information of the kth candidate data with respect to the target data is determined.

9. The method according to claim 8, characterized in that The determining, based on the second supplementary information of the kth candidate data to the target data and the label information of the kth candidate data, the first supplementary information of the kth candidate data to the target data comprises: Projecting the label information of the k-th candidate data using a second projection matrix to obtain label projection information of the k-th candidate data; The sum of the second supplementary information of the k-th candidate data with respect to the target data and the label projection information of the k-th candidate data is determined as the first supplementary information of the k-th candidate data with respect to the target data.

10. The method according to claim 7, characterized in that The method of supplementing the j-th first feature information of the target data based on the Kj candidate data to determine the j-th supplemented feature information of the target data includes: For a kth candidate data among the Kj candidate data, based on the similarity between the target data and the kth candidate data, determining a weight of the kth candidate data; Based on the weights of the Kj candidate data, weighted processing is performed on the first supplementary information of the target data by the Kj candidate data to obtain weighted supplementary information; Based on the j-th first feature information of the target data and the weighted supplementary information, the j-th supplemented feature information of the target data is determined.

11. The method according to claim 3, characterized in that The detection module includes a hybrid expert module, and the hybrid expert module includes a gating unit and E expert units. The detection module performs detection processing on the R supplemented feature information of the target data under the R retrieval units to obtain the detection result of the target data, including: Based on the supplemented feature information of the target data under the R search units, S input tokens are obtained, where S is a positive integer greater than 1; For an s-th input token among the S input tokens, determining, by the gating unit, a probability value of the s-th input token being assigned to each of the E expert units, where E is a positive integer greater than 1, and s is a positive integer less than or equal to S; Based on the probability value, T expert units are selected from the E expert units for processing the s-th input token, and the s-th input token is processed by the T expert units respectively to obtain fused feature information of the s-th input token, where T is a positive integer less than E; Based on the fused feature information of the S input tokens, obtaining mixed feature information of the target data; Based on the mixed feature information, the target data is detected and processed to obtain a detection result of the target data.

12. The method according to claim 11, characterized in that The step of processing the s-th input token by the T expert units respectively to obtain fused feature information of the s-th input token includes: For each of the T expert units, the expert unit performs feature fusion processing on each input token assigned to the expert unit to obtain output feature information of each input token, wherein each input token includes the sth input token; Determine the weights corresponding to the T expert units respectively based on the probability value that the sth input token is assigned to each of the T expert units; Based on the weights corresponding to the T expert units, the output feature information of the s-th input token output by the T expert units is weighted to obtain the fused feature information of the s-th input token.

13. The method according to claim 11, characterized in that The method of obtaining S input tokens based on the supplemented feature information of the target data under the R search units includes: Taking the supplemented feature information of the target data in each of the R retrieval units as a row of the first matrix, obtaining a first matrix including R rows; Each column of the first matrix is ​​used as an input token to obtain the S input tokens.

14. A model training method, characterized in that: include: Obtaining a training sample and Q candidate samples, and extracting first feature information of the training sample and the Q candidate samples respectively through a retrieval module in a data detection model, wherein the training sample is a state value of L biomarkers of a first subject, and the candidate samples are a state value of L biomarkers of other subjects different from the first subject, and both Q and L are positive integers greater than 1; Based on the first feature information of the training sample and the first feature information of the Q candidate samples, selecting at least one candidate sample related to the training sample from the Q candidate samples; Based on the first feature information of the at least one candidate sample, supplement the first feature information of the training sample to obtain supplemented feature information of the training sample; Performing detection processing on the supplemented feature information of the training sample through the detection module in the data detection model to obtain the detection result of the training sample; Based on the detection results of the training samples, the loss of the data detection model is determined, and the data detection model is trained based on the loss.

15. The method according to claim 14, characterized in that The retrieval module includes R retrieval units, where R is a positive integer greater than 1; The extracting the first feature information of the training sample and the Q candidate samples respectively through the retrieval module in the data detection model includes: for the j-th retrieval unit among the R retrieval units, extracting the j-th first feature information of the training sample and the Q candidate samples respectively through the j-th retrieval unit, where j is a positive integer less than or equal to R; The selecting, based on the first feature information of the training sample and the first feature information of the Q candidate samples, at least one candidate sample related to the training sample from the Q candidate samples comprises: selecting, based on the j-th first feature information of the training sample and the j-th first feature information of the Q candidate samples, Kj candidate samples related to the training sample from the Q candidate samples; The step of supplementing the first feature information of the training sample based on the first feature information of the at least one candidate sample to obtain supplemented feature information of the training sample includes: supplementing the jth first feature information of the training sample based on the jth first feature information of the Kj candidate samples to obtain the jth supplemented feature information of the training sample under the jth retrieval unit; The detection module in the data detection model detects and processes the supplemented feature information of the training sample to obtain the detection result of the training sample, including: detecting and processing the R supplemented feature information of the training sample under the R retrieval units respectively through the detection module to obtain the detection result of the training sample.

16. The method according to claim 15, characterized in that The detection module includes a hybrid expert module, the hybrid expert module includes a gating unit and E expert units, and the loss of the data detection model is determined based on the detection result of the training sample, including: Determine a first loss based on the detection result of the training sample and the label information of the training sample; Determine a second loss based on the number of input labels and the assignment probability value assigned to each of the E expert units; Based on the first loss and the second loss, a loss of the data detection model is determined.

17. A data detection device, characterized in that: include: an extraction unit, configured to obtain target data to be detected and Q candidate data, and extract first feature information of the target data and the Q candidate data respectively, wherein the target data is the state values ​​of L biomarkers of the target object, and the candidate data is the state values ​​of L biomarkers of other objects different from the target object, and both Q and L are positive integers greater than 1; A selection unit, configured to select at least one candidate data related to the target data from the Q candidate data based on the first feature information of the target data and the first feature information of the Q candidate data; A supplementing unit, configured to supplement the first feature information of the target data based on the first feature information of the at least one candidate data to obtain supplemented feature information of the target data; The detection unit is used to perform detection processing on the supplemented characteristic information of the target data to obtain the detection result of the target data.

18. A model training device, characterized in that: include: an extraction unit, configured to obtain a training sample and Q candidate samples, and respectively extract first feature information of the training sample and the Q candidate samples through a retrieval module in a data detection model, wherein the training sample is a state value of L biomarkers of a first subject, and the candidate samples are state values ​​of L biomarkers of other subjects different from the first subject, and both Q and L are positive integers greater than 1; A selection unit, configured to select at least one candidate sample related to the training sample from the Q candidate samples based on the first feature information of the training sample and the first feature information of the Q candidate samples; A supplementing unit, configured to supplement the first feature information of the training sample based on the first feature information of the at least one candidate sample to obtain supplemented feature information of the training sample; A detection unit, configured to perform detection processing on the supplemented feature information of the training sample through a detection module in the data detection model to obtain a detection result of the training sample; A training unit is used to determine the loss of the data detection model based on the detection results of the training samples, and train the data detection model based on the loss.

19. An electronic device, characterized in that: including a processor and a memory; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the method according to any one of claims 1 to 13 or 14 to 16.

20. A computer-readable storage medium, characterized in that: For storing computer programs; The computer program enables a computer to execute the method according to any one of claims 1 to 13 or 14 to 16 above.

Images