Matching method and device for clinical drug test patients and computer equipment

Abstract

The application discloses a kind of matching method, device and computer equipment of clinical drug test patient, it is related to computer technical field, can solve the technical problem of the matching of clinical drug test patient at present, workload is big, efficiency is relatively low, and it is prone to cause error accumulation and information loss. Among them, the method comprises: obtaining patient information of test patient and medicine test entry condition information, patient information includes first numerical information and medical information, medicine test entry condition information includes second numerical information and description information;First numerical information and second numerical information are numerically matched, and first matching result is obtained;Medical information and description information are preprocessed, and the embedded matching of medical information and description information after preprocessing is carried out, and second matching result is obtained;According to first matching result and second matching result, the final matching result of test patient and medicine test is generated.

Description

TECHNICAL FIELD

[0001] The present application relates to the computer technical field, and particularly relates to a clinical drug test patient matching method and device and a computer device. BACKGROUND

[0002] Clinical drug innovation research and development develops rapidly under the encouragement of China's policy, and a difficult problem is how to find suitable patients for innovative drugs in the clinical trial stage of the drugs. Different drugs for the test have different requirements for the patients, such as requiring patients not to use drugs containing gemcitabine in the past treatment; patients with hepatitis B can be allowed, but the quantitative hepatitis B cannot exceed 2000 IU / mL, and the like. How to find suitable patients for drug tests under the conditions of numerous unstructured language descriptions and numerical descriptions (drug test entry and exclusion conditions) is a task with huge workload and prone to errors.

[0003] The entry and exclusion conditions of the drug test are generally described by unstructured natural language, and the patient information (age, diagnosis result, surgical treatment, drug treatment, etc.) is structured information. How to match the two, a conventional idea is to use common NLP technologies such as word segmentation, NER, etc. to preprocess and structure the sentence description of the entry and exclusion conditions, and then match with the patient information. However, this method is processed separately in steps, resulting in a large workload, low efficiency, and prone to error accumulation and information loss. SUMMARY

[0004] Therefore, the present application provides a clinical drug test patient matching method, device and computer device, which can solve the technical problems of large workload, low efficiency, and prone to error accumulation and information loss in the matching of clinical drug test patients.

[0005] According to one aspect of the present application, a clinical drug test patient matching method is provided, which comprises:

[0006] Obtaining patient information of a test patient and drug test entry and exclusion condition information, the patient information comprising first numerical information and medical information, and the drug test entry and exclusion condition information comprising second numerical information and description information;

[0007] Numerical matching the first numerical information and the second numerical information to obtain a first matching result;

[0008] Preprocessing the medical information and the description information, and embedding matching the preprocessed medical information and the description information to obtain a second matching result;

[0009] generate a final matching result of the trial patient and the medical trial according to the first matching result and the second matching result.

[0010] Optionally, the numerical matching of the first numerical type information and the second numerical type information to obtain a first matching result comprises:

[0011] The first numerical type information and the second numerical type information are respectively subjected to three-element extraction, and patient information triplets corresponding to the first numerical type information and enrollment and exclusion condition information triplets corresponding to the second numerical type information are generated based on the three elements, wherein the patient information triplets comprise numerical values of patient data items, units of the patient data items, and meanings of the patient data items, and the enrollment and exclusion condition information triplets comprise numerical value conditions of preset data items, units of the preset data items, and meanings of the preset data items.

[0012] Based on the units of the preset data items, the meanings of the preset data items, the units of the patient data items, and the meanings of the patient data items, target patient data items matching the preset data items are screened from the patient information triplets.

[0013] The numerical values of the target patient data items are compared with the numerical value conditions of the preset data items to obtain the first matching result.

[0014] Optionally, the preprocessing of the visit information and the description information comprises:

[0015] The field values of each field in the visit information are subjected to one-hot encoding processing to obtain a high-dimensional binary vector, and a preset dimension reduction algorithm is used to perform low-dimensional representation learning on the high-dimensional binary vector to obtain a first feature vector corresponding to the visit information.

[0016] The description information is input into a pre-trained language representation model to obtain a second feature vector corresponding to the description information.

[0017] The first feature vector and the second feature vector are subjected to normalization processing so that the first feature vector and the second feature vector are in the same feature space dimension.

[0018] Optionally, the language representation model is a BERT language model, and before the description information is input into the pre-trained language representation model to obtain the second feature vector corresponding to the description information, the method further comprises:

[0019] Obtaining sample description information and a sample feature vector corresponding to the sample description information.

[0020] After the sample description information is randomly subjected to partial data covering, the initial BERT language model is input to output a training result.

[0021] calculating a loss value of the training result and the sample feature vector;

[0022] if the loss value meets the model convergence requirement, determining the initial BERT language model as the language representation model;

[0023] if the loss value does not meet the model convergence requirement, updating the model parameters of the initial BERT language model, and iteratively training the updated initial BERT language model until the loss value meets the model convergence requirement, and determining the initial BERT language model meeting the model convergence requirement as the language representation model.

[0024] Optionally, the embedding matching of the preprocessed diagnosis information and the description information to obtain a second matching result comprises:

[0025] inputting the first feature vector and the second feature vector in the same feature space dimension into a neural network with a depth of 2 to obtain a matching score;

[0026] comparing the matching score with a preset threshold to obtain a second matching result.

[0027] Optionally, before the preprocessing of the diagnosis information and the description information and the embedding matching of the preprocessed diagnosis information and the description information to obtain a second matching result, the method further comprises:

[0028] determining whether the trial patient and the medical trial are successfully matched according to the first matching result;

[0029] if yes, preprocessing the diagnosis information and the description information, embedding matching the preprocessed diagnosis information and the description information to obtain a second matching result, and generating a final matching result of the trial patient and the medical trial according to the first matching result and the second matching result;

[0030] if no, outputting a final matching result of the trial patient and the medical trial being unsuccessfully matched.

[0031] Optionally, the generating of the final matching result of the trial patient and the medical trial according to the first matching result and the second matching result comprises:

[0032] if the first matching result and the second matching result are determined to be successfully matched, outputting a final matching result of the trial patient and the medical trial being successfully matched;

[0033] If it is judged that at least one of the first matching result and the second matching result is a matching failure, a final matching result of matching failure of the trial patient and the medical trial is output.

[0034] According to another aspect of the present application, a device for matching a clinical drug trial patient is provided, and the device comprises:

[0035] An acquisition module is configured to acquire patient information of a trial patient and information of drug trial admission and exclusion conditions, wherein the patient information comprises first numerical information and visit information, and the information of drug trial admission and exclusion conditions comprises second numerical information and description information;

[0036] A first matching module is configured to perform numerical matching on the first numerical information and the second numerical information to obtain a first matching result.

[0037] A second matching module is configured to perform preprocessing on the visit information and the description information, and perform embedding matching on the preprocessed visit information and the description information to obtain a second matching result.

[0038] A generation module is configured to generate a final matching result of the trial patient and the medical trial according to the first matching result and the second matching result.

[0039] Optionally, the first matching module comprises a first generation unit, a screening unit, and a first comparison unit.

[0040] The first generation unit is configured to perform three-element extraction on the first numerical information and the second numerical information respectively, and generate a patient information triple corresponding to the first numerical information and an admission and exclusion condition information triple corresponding to the second numerical information based on the three elements, wherein the patient information triple comprises a numerical value of a patient data item, a unit of the patient data item, and a meaning of the patient data item, and the admission and exclusion condition information triple comprises a numerical value condition of a preset data item, a unit of the preset data item, and a meaning of the preset data item.

[0041] The screening unit is configured to screen a target patient data item matching the preset data item from the patient information triple based on the unit of the preset data item, the meaning of the preset data item, the unit of the patient data item, and the meaning of the patient data item.

[0042] The first comparison unit is configured to compare the numerical value of the target patient data item with the numerical value condition of the preset data item to obtain a first matching result.

[0043] Optionally, the second matching module comprises a first processing unit, a first input unit, and a second processing unit.

[0044] The first processing unit is configured to perform one-hot encoding processing on a field value of each field in the medical information to obtain a high-dimensional binary vector, and perform low-dimensional representation learning on the high-dimensional binary vector by using a preset dimension reduction algorithm to obtain a first feature vector corresponding to the medical information.

[0045] The first input unit is configured to input the description information into a pre-trained language representation model to obtain a second feature vector corresponding to the description information.

[0046] The second processing unit is configured to perform normalization processing on the first feature vector and the second feature vector to make the first feature vector and the second feature vector in the same feature space dimension.

[0047] Optionally, the device further comprises a training module.

[0048] The training module is configured to obtain sample description information and a sample feature vector corresponding to the sample description information, input the sample description information after being randomly partially covered into an initial BERT language model to output a training result, calculate a loss value of the training result and the sample feature vector, determine the initial BERT language model as the language representation model if the loss value meets a model convergence requirement, update model parameters of the initial BERT language model if the loss value does not meet the model convergence requirement, and perform iterative training on the updated initial BERT language model until the loss value meets the model convergence requirement, and determine the initial BERT language model meeting the model convergence requirement as the language representation model.

[0049] Optionally, the second matching module comprises a second input unit and a second comparison unit.

[0050] The second input unit is configured to input the first feature vector and the second feature vector in the same feature space dimension into a neural network with a depth of 2 to obtain a matching score.

[0051] The second comparison unit is configured to compare the matching score with a preset threshold to obtain a second matching result.

[0052] Optionally, the device further comprises:

[0053] A judgment module is configured to judge whether the test patient and the medical test are successfully matched according to the first matching result.

[0054] If not, the output module is configured to output the final matching result that the trial patient fails to match the medical trial.

[0055] The output module is configured to output the final matching result that the trial patient fails to match the medical trial if not.

[0056] Optionally, the generation module comprises a second generation unit and a third generation unit.

[0057] The second generation unit is configured to generate the final matching result that the trial patient succeeds to match the medical trial if it is determined that the first matching result and the second matching result are matching successes.

[0058] The third generation unit is configured to generate the final matching result that the trial patient fails to match the medical trial if it is determined that at least one of the first matching result and the second matching result is a matching failure.

[0059] According to yet another aspect of the present application, there is provided a storage medium having a computer program stored thereon, the program being executed by a processor to implement the matching method for a clinical drug trial patient.

[0060] According to still another aspect of the present application, there is provided a computer device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, the processor implementing the matching method for a clinical drug trial patient when executing the program.

[0061] The present application provides a matching method, device and computer device for a clinical drug trial patient. The patient information of the trial patient and the drug trial entry and exclusion condition information are first acquired. The patient information comprises first numerical information and medical information, and the drug trial entry and exclusion condition information comprises second numerical information and description information. Then, the first numerical information and the second numerical information are subjected to numerical matching to obtain a first matching result. The medical information and the description information are preprocessed, and the preprocessed medical information and the description information are subjected to embedding matching to obtain a second matching result. Finally, the final matching result of the trial patient and the medical trial is generated according to the first matching result and the second matching result. Through the technical solution in the present application, the drug trial entry and exclusion condition and the patient information in patient recruitment can be intelligently matched by using a deep neural network, and end-to-end training can be performed to achieve global optimization. The traditional operations such as word segmentation and relationship extraction that are prone to cumulative errors are avoided, a higher matching precision can be obtained, and the effect of precise matching of drug and patient is achieved.

[0062] The technical solutions of the present application will be further described in detail below with the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0063] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the application and, together with the description, serve to explain the principles of the application.

[0064] The present application can be more clearly understood and appreciated from the following detailed description, taken in conjunction with the accompanying drawings, in which:

[0065] Figure 1 A flow chart of a matching method for a clinical drug test patient is shown according to an embodiment of the present application;

[0066] Figure 2 A flow chart of another matching method for a clinical drug test patient is shown according to an embodiment of the present application;

[0067] Figure 3 A principle diagram of a matching method for a clinical drug test patient is shown according to an embodiment of the present application;

[0068] Figure 4 A structure diagram of a matching device for a clinical drug test patient is shown according to an embodiment of the present application;

[0069] Figure 5 A structure diagram of another matching device for a clinical drug test patient is shown according to an embodiment of the present application;

[0070] Figure 6 An entity structure diagram of a computer device is shown according to an embodiment of the present application. DETAILED DESCRIPTION

[0071] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement, numerical expressions, and values of the components and steps set forth in these embodiments are not limiting to the scope of the present application unless otherwise specifically stated.

[0072] It should be understood, of course, that the various embodiments of the present application are not limited to the illustrations described above, but can vary the steps of the methods and the components and parameters of the devices described and / or shown.

[0073] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way limiting to the scope of the application or its applications or uses.

[0074] Techniques, methods, and apparatus known to those of ordinary skill in the relevant art can not be discussed in detail herein. However, where appropriate, such techniques, methods, and apparatus can be considered part of the present application.

[0075] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0076] Embodiments of this invention can be applied to computer systems / servers that can operate with a wide range of other general-purpose or special-purpose computing system environments or configurations. Examples of well-known computing systems, environments, and / or configurations suitable for use with computer systems / servers include, but are not limited to: personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, and distributed cloud computing environments that include any of the above systems, etc.

[0077] Computer systems / servers can be described in the general context of computer system executable instructions (such as program modules) executed by the computer system. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., which perform specific tasks or implement specific abstract data types. Computer systems / servers can be implemented in distributed cloud computing environments, where tasks are performed by remote processing devices linked through a communication network. In distributed cloud computing environments, program modules can reside on local or remote computing system storage media, including storage devices.

[0078] This invention provides a method for matching patients in clinical drug trials, such as... Figure 1 As shown, the method includes:

[0079] 101. Obtain patient information for the trial patients and drug trial inclusion / exclusion criteria information. Patient information includes first-value information and medical history. Drug trial inclusion / exclusion criteria information includes second-value information and descriptive information.

[0080] In specific application scenarios, patient information can be obtained through medical platforms or by collecting information from patients via questionnaires. This patient information includes basic information and medical records. Basic information includes disease name, subtype, gender, age, patient ID number, and daily activity level. Medical records include treatment methods received, treatment duration, medications received, laboratory test results, disease condition information, and complication information. Numerical information such as patient age, ID number, and laboratory data (e.g., platelet count) cannot be well represented in semantic embedding. Therefore, a matching mode can be designed to run in parallel with textual information for numerical information. Based on this, the patient information can be divided into two categories: primary numerical information and medical visit information.

[0081] Drug trial inclusion and exclusion criteria information includes inclusion criteria and exclusion criteria. Factors that allow participation in a clinical drug trial are called "inclusion criteria"; factors that disqualify a person from participating are called "exclusion criteria." It is important to note that establishing "inclusion / exclusion criteria" is not intended to reject individuals from participating in clinical drug trials, but rather to determine their suitability for participation and ensure their safety. When obtaining drug trial inclusion and exclusion criteria information, intelligent analysis of clinical drug trial documents can extract relevant information to obtain the corresponding inclusion and exclusion criteria for that clinical drug trial. This inclusion and exclusion criteria information may include: age, gender, disease type and stage, treatment history, and other medical conditions. Only patients who meet the inclusion and exclusion criteria can participate in clinical trials. Corresponding to the patient information of the trial patients, in order to achieve parallel matching and detection of the first numerical information and medical information in the patient information, the drug trial inclusion and exclusion criteria information can also be divided into second numerical information and descriptive information. The second numerical information is the numerical limitation condition for each data item when the "inclusion criteria" are met, which is used to match with the first numerical information. The descriptive information is the requirement for the patient's condition for the trial drug when the "inclusion criteria" are met, such as requiring that the patient has not used drugs containing gemcitabine in previous treatments, and allowing the patient to have hepatitis B, but the hepatitis B quantitative amount cannot exceed 2000 IU / mL, etc.

[0082] The implementing entity of this application can be a system used to support patient matching in clinical drug trials. It can be configured on the client or server side. After obtaining the patient information of the trial patient and the drug trial inclusion / exclusion criteria information, it can perform numerical matching between the first numerical information in the patient information and the second numerical information in the drug trial inclusion / exclusion criteria information to obtain a first matching result; then, it can preprocess the medical information in the patient information and the descriptive information in the drug trial inclusion / exclusion criteria information, and embed the preprocessed medical information and descriptive information for matching to obtain a second matching result; finally, it can generate the final matching result between the trial patient and the drug trial based on the first matching result and the second matching result.

[0083] 102. Perform numerical matching between the first numerical information and the second numerical information to obtain the first matching result.

[0084] In this embodiment, for numerical information, a numerical matching module can be designed independently. The numerical matching module is used to perform numerical matching between the first numerical information and the second numerical information. Specifically, the numerical information can be encoded into a triple (numerical value, unit, meaning). In the in-exclusion description statement, the numerical information can be simply located by describing the "unit". The extracted numerical information is compared with the numerical information with the same unit in the patient information to further obtain the first matching result.

[0085] In a specific application scenario, as a preferred mode, after obtaining the first matching result, in order to improve the matching efficiency, it can be determined whether the subsequent matching process needs to be performed based on the first matching result: when the first matching result is a matching success, the subsequent matching process of the embodiment steps 103 to 104 can be continued; when the first matching result is a matching failure, it will inevitably lead to the final matching result of the trial patient and the medical trial being a matching failure. Therefore, in order to improve the matching efficiency and save the matching resources, the subsequent matching process can no longer be continued, and the final matching result of the trial patient and the medical trial being a matching failure is output.

[0086] 103、Preprocess the clinic information and the description information, and perform embedding matching on the preprocessed clinic information and the description information to obtain a second matching result.

[0087] In a specific application scenario, since the patient information and the drug trial entry and exclusion condition information are obtained from the patient and the clinical drug trial document respectively, when describing the disease, the two may use different description methods for the same medical name. The drug trial entry and exclusion condition information is generally described by unstructured natural language, while the patient information (age, diagnosis result, surgical treatment, drug treatment, etc.) is structured information. Directly comparing and analyzing the clinic information in the patient information and the description information in the drug trial entry and exclusion condition information may lead to a lower matching accuracy. Therefore, the clinic information and the description information need to be preprocessed first to unify them into the same feature space dimension. The preprocessing can include embedding representation processing of the clinic information and the description information, obtaining of a feature vector, and semantic alignment processing of the feature vectors in different feature space dimensions, so as to normalize the feature vectors corresponding to the clinic information and the description information into the same feature space dimension. Correspondingly, the embedding matching can be performed on the feature vectors corresponding to the clinic information and the description information in the same feature space dimension to obtain a second matching result.

[0088] 104、Generating a final matching result of the trial patient and the medical trial according to the first matching result and the second matching result.

[0089] In a specific application scenario, when the first matching result and the second matching result are both matching successes, it can be determined that the final matching result of the trial patient and the medical trial is a matching success. When there is any or all of the first matching result and the second matching result being a non-matching, it can be determined that the final matching result of the trial patient and the medical trial is a matching failure. In this regard, the matching algorithm of the drug trial and the patient can be designed as follows:

[0090] Input: N is the number of pairs in the training set;

[0091]

[0092]

[0093] Output: The prediction result of whether the patient information of the test patient matches the drug test entry condition.

[0094] The matching method for a clinical drug test patient provided by the embodiment of the present application can first acquire patient information of a test patient and drug test entry condition information, the patient information including first numerical information and medical information, and the drug test entry condition information including second numerical information and description information, then perform numerical matching on the first numerical information and the second numerical information to obtain a first matching result, and perform preprocessing on the medical information and the description information, and then perform embedding matching on the preprocessed medical information and the description information to obtain a second matching result, and finally generate a final matching result of the test patient and the drug test according to the first matching result and the second matching result. Through the technical solution in the present application, the drug test entry condition in patient recruitment and the patient information can be intelligently matched by using a deep neural network, end-to-end training can be performed, global optimization can be performed, traditional operations such as word segmentation and relationship extraction that are prone to cause cumulative errors can be avoided, a higher matching accuracy can be obtained, and the effect of precise matching of patients and drugs can be achieved.

[0095] Further, in order to better illustrate the above matching process of the clinical drug test patient, as a refinement and expansion of the above embodiment, the embodiment of the present application provides another matching method for a clinical drug test patient, as shown in Figure 2 The method comprises the following steps:

[0096] 201, acquiring patient information of a test patient and drug test entry condition information, the patient information including first numerical information and medical information, and the drug test entry condition information including second numerical information and description information.

[0097] For the execution process of the embodiment, refer to the related description in the embodiment step 101, which will not be repeated here.

[0098] 202, performing numerical matching on the first numerical information and the second numerical information to obtain a first matching result.

[0099] For the embodiment, in a specific application scenario, when the first numerical information and the second numerical information are matched, the numerical information can be encoded into a triple (value, unit, meaning), and in the description statement, the value information can be positioned by the unit, and the extracted value information can be compared with the value information of the same unit in the patient information to further obtain the first matching result. Correspondingly, as a preferred manner, the embodiment step 202 can specifically include: performing three-element extraction on the first numerical information and the second numerical information respectively, and generating a patient information triple corresponding to the first numerical information and an admission condition information triple corresponding to the second numerical information based on the three elements, wherein the patient information triple includes a value of a patient data item, a unit of the patient data item, and a meaning of the patient data item, and the admission condition information triple includes a value condition of a preset data item, a unit of the preset data item, and a meaning of the preset data item; based on the unit of the preset data item, the meaning of the preset data item, the unit of the patient data item, and the meaning of the patient data item, the target patient data item matching the preset data item is screened out in the patient information triple; the value of the target patient data item is compared with the value condition of the preset data item to obtain the first matching result. If it is judged that the values of all the target patient data items match the value conditions of the corresponding preset data items, it is determined that the first matching result is a matching success; otherwise, it is determined that the first matching result is a matching failure.

[0100] Correspondingly, after obtaining the first matching result, to improve the efficiency of matching, it can be determined based on the first matching result whether the subsequent matching process needs to be performed: when the first matching result is a matching success, the subsequent embodiment steps 203 to 207 can be continued to perform the matching process; when the first matching result is a matching failure, it will inevitably lead to a final matching result of the test patient and the medical test being a matching failure. Therefore, to improve the matching efficiency and save matching resources, the subsequent matching process can no longer be continued, and the final matching result of the test patient and the medical test being a matching failure is output. Correspondingly, the embodiment step can specifically further include: judging whether the test patient and the medical test are a matching success according to the first matching result; if yes, the visit information and the description information are preprocessed, the preprocessed visit information and the description information are embedded and matched to obtain a second matching result; the final matching result of the test patient and the medical test is generated according to the first matching result and the second matching result; if not, the final matching result of the test patient and the medical test being a matching failure is output.

[0101] 203. The field values of each field in the visit information are subjected to one-hot encoding processing to obtain a high-dimensional binary vector, and a preset dimension reduction algorithm is used to perform low-dimensional representation learning on the high-dimensional binary vector to obtain a first feature vector corresponding to the visit information.

[0102] In a specific application scenario, since the medical information is generally structured information, such as including diagnosed diseases, treatment plans (surgery, drugs), etc., in order to facilitate matching with drug test entry and exit condition information, it is necessary to first perform embedding representation processing on the medical information to obtain a first feature vector. When performing embedding representation processing on the medical information, as an optional solution, One-Hot Encoding can be used to encode the field value of each field, the encodings of multiple fields are concatenated into a high-dimensional binary vector, and then a preset dimension reduction algorithm is used to perform low-dimensional representation learning on the high-dimensional binary vector to obtain patient information embedding representation.

[0103] In a matching type machine learning algorithm, the calculation of the distance between features or the calculation of the similarity is very important. The commonly used distance or similarity calculation is the similarity calculation in the Euclidean space, and the calculation of the cosine similarity is based on the Euclidean space. Therefore, for the present embodiment, One-Hot Encoding can be used to expand the value of the discrete feature corresponding to the medical information to the Euclidean space, and a certain value of the discrete feature corresponds to a certain point in the Euclidean space. Using One-Hot Encoding for the discrete feature will make the distance calculation between features more reasonable. One-Hot Encoding is a method of using an N-bit status register to encode N states, each state has its own independent register bit, and at any time, only one bit is valid. That is, only one bit is 1 and the rest are zero. Through One-Hot Encoding processing, a high-dimensional binary vector corresponding to the medical information can be obtained.

[0104] However, the high-dimensional binary vectors identified above contain multiple feature variables. While these multiple feature variables undoubtedly provide rich information for research and application, they also increase the workload of data computation to some extent. Furthermore, redundant feature variables are prone to exist, and correlations may exist between multiple feature variables, thus increasing the complexity of problem analysis. Therefore, to address this, dimensionality reduction processing can be performed on the high-dimensional binary vectors. This involves retaining the most important feature variables from the high-dimensional binary vectors while removing noise and unimportant feature variables, thereby improving both data processing speed and quality. Specifically, as an optional approach, a preset dimensionality reduction algorithm can be used to learn the low-dimensional representation of the high-dimensional binary vectors to obtain the first feature vector corresponding to the medical information. The preset dimensionality reduction algorithm can include, but is not limited to, any one of the following: MultiDimensional Scaling (MDS), Isometric Mapping (ISOMAP), Principal Component Analysis (PCA), and Linear Discriminant Analysis (LDA). The specific algorithm can be selected based on the actual application scenario, and no specific limitations are imposed here.

[0105] 204. Input the descriptive information into the pre-trained language representation model to obtain the second feature vector corresponding to the descriptive information.

[0106] The language representation model can be the BERT language model, where BERT (Bidirectional Encoder Representation Transformer) represents a bidirectional encoder representation from the Transformer. BERT aims to pre-train deep bidirectional representations by jointly adjusting the left and right contexts in all layers. BERT uses WordPieces to segment the input. WordPieces use double-byte encoding to split words into pieces. Double-byte encoding reduces the influence of tense on the same word, separating the meaning and tense of a word, effectively reducing the vocabulary size. Secondly, special characters [CLS] and [SEP] are introduced when embedding sentences. The first label of each sentence sequence is always [CLS]. For classification tasks, the output of the Transformer corresponding to this label can be used for the overall representation of the sequence. The role of [SEP] is to distinguish the sentence pairs that are packed together. Next, BERT adds an A embedding to each label of the first sentence and an E embedding to each label of the second sentence. Finally, a position embedding vector is trained for each position in the input.

[0107] In a specific application scenario, before performing the steps of the embodiment, the language representation model also needs to be pre-trained. Correspondingly, as a preferred mode, the steps of the embodiment can further include: obtaining sample description information and a sample feature vector corresponding to the sample description information; inputting the initial BERT language model after randomly performing partial data covering on the sample description information to output a training result; calculating a loss value of the training result and the sample feature vector; if the loss value meets the model convergence requirement, determining the initial BERT language model as the language representation model; if the loss value does not meet the model convergence requirement, updating the model parameters of the initial BERT language model, and iteratively training the updated initial BERT language model until the loss value meets the model convergence requirement, and determining the initial BERT language model meeting the model convergence requirement as the language representation model.

[0108] Further, the description information can be input into the pre-trained language representation model, and the language representation model is used to analyze the description information to output a second feature vector corresponding to the description information.

[0109] 205. Normalize the first feature vector and the second feature vector so that the first feature vector and the second feature vector are in the same feature space dimension.

[0110] In a specific application scenario, since the patient information and the drug trial inclusion and exclusion condition information are obtained from the patient and the clinical drug trial document respectively, when describing the condition, the two may use different description methods for the same medical name. The drug trial inclusion and exclusion condition information is generally described in unstructured natural language, while the patient information (age, diagnosis result, surgical treatment, drug treatment, etc.) is structured information. Directly comparing and analyzing the patient information and the description information in the drug trial inclusion and exclusion condition information may result in low matching accuracy. Therefore, the patient information and the description information need to be normalized first to unify them into the same feature space dimension.

[0111] For the embodiment, the first feature vector and the second feature vector can be normalized based on a shallow neural network with an attention mechanism. Let the first feature vector corresponding to the patient information be the second feature vector corresponding to the description information in the drug trial inclusion and exclusion condition information be The attention weight is as follows:

[0112]

[0113] Based on the attention weight, the first feature vector and the second feature vector after semantic alignment (normalization) can be calculated as follows:

[0114]

[0115]

[0116] 206. Input the first feature vector and the second feature vector, which are in the same feature space dimension, into a neural network with a depth of 2 to obtain a matching score. Compare the matching score with a preset threshold to obtain a second matching result.

[0117] In this embodiment, after obtaining the normalized first and second feature vectors based on the above steps, the normalized first and second feature vectors can be further input into a neural network of depth 2. The neural network of depth 2 outputs a matching score between the patient's medical information and the information described in the drug trial inclusion / exclusion criteria. The matching score is then compared with a preset threshold to obtain a second matching result. The matching score and the preset threshold are values ​​greater than 0 and less than 1. The value of the preset threshold can be set according to the actual application scenario; the closer the preset threshold is to 1, the higher the matching accuracy of the second matching result.

[0118] The neural network of depth 2 can include fully connected layers and a softmax output layer. The first and second feature vectors of the same feature space dimension can be used to calculate the matching score through the fully connected layer and the softmax output layer. The corresponding formula is described as follows:

[0119] x i =σ(W a [u i ,β i ])

[0120] x j =σ(W a [v j α j ])

[0121] y = softmax(W f [x i x j ]+b f )

[0122] 207. Generate the final matching results between the trial patients and the drug trial based on the first matching results and the second matching results.

[0123] In specific application scenarios, the drug trial and patient matching algorithm described in step 104 of the embodiment can be used, and the final matching result between the trial patient and the drug trial can be determined based on the first matching result and the second matching result. Accordingly, as a preferred embodiment, step 207 may specifically include: if the first matching result and the second matching result are determined to be a successful match, then output the final matching result of the successful match between the trial patient and the drug trial; if at least one of the first matching result and the second matching result is determined to be a failed match, then output the final matching result of the failed match between the trial patient and the drug trial.

[0124] To fully illustrate the technical solutions in this application, the appendix to the specification is provided below. Figure 3 A schematic diagram illustrating the principle of patient matching in a clinical drug trial is provided to fully explain the implementation process of this application: Patient information is divided into a first numerical category and medical visit information; the drug trial inclusion / exclusion criteria information is divided into a second numerical category and descriptive information; numerical matching is performed between the first and second numerical categories to obtain a first matching result; one-hot encoding is performed on the field value of each field in the medical visit information to obtain a high-dimensional binary vector; and MLP is used to learn a low-dimensional representation of the high-dimensional binary vector to obtain the first feature vector corresponding to the medical visit information; the descriptive information is input into a pre-trained BERT language model to obtain the descriptive information... The first and second feature vectors are input into the semantic alignment module to normalize them so that they are in the same feature space dimension. The first and second feature vectors in the same feature space dimension are then input into a neural network of depth 2 for embedding matching to obtain a second matching result. Finally, the first feature vector and the second matching result are input into the fusion prediction module, which uses the drug trial and patient matching algorithm in the fusion prediction module to perform logical discrimination of the final matching result between the trial patient and the drug trial based on the first feature vector and the second matching result.

[0125] By the matching method of the clinical drug test patient, patient information of the test patient and drug test entry condition information can be obtained first, the patient information includes first numerical information and medical information, and the drug test entry condition information includes second numerical information and description information; then the first numerical information and the second numerical information are matched in value to obtain a first matching result; and the medical information and the description information are preprocessed, and the preprocessed medical information and the description information are embedded and matched to obtain a second matching result; finally, the final matching result of the test patient and the medical test is generated according to the first matching result and the second matching result. Through the technical solution in the present application, the drug test entry condition in patient recruitment and patient information can be intelligently matched by using a deep neural network, the structured step of the entry condition is omitted in the matching process, an embedding representation space is learned using an end-to-end learning framework, the patient and the drug test are matched in the aligned embedding representation space, global optimization can be performed through end-to-end training, the traditional operations such as word segmentation and relationship extraction that are prone to cause cumulative errors are avoided, higher matching accuracy can be obtained, and the effect of precise matching of patients and drugs is realized.

[0126] Further, as an implementation of the method shown in the above Figure 1 , the embodiment of the present application provides a kind of matching device of clinical drug test patient, as shown in the above Figure 4 , the device includes: acquisition module 31, first matching module 32, second matching module 33 and generation module 34.

[0127] Acquisition module 31 can be used to obtain the patient information of the test patient and the drug test entry condition information, the patient information includes first numerical information and medical information, and the drug test entry condition information includes second numerical information and description information;

[0128] First matching module 32 can be used to match the first numerical information and the second numerical information in value, and obtain a first matching result;

[0129] Second matching module 33 can be used to preprocess the medical information and the description information, and embed and match the preprocessed medical information and the description information to obtain a second matching result;

[0130] Generation module 34 can be used to generate the final matching result of the test patient and the medical test according to the first matching result and the second matching result.

[0131] In specific application scenarios, as shown in the above Figure 5 , first matching module 32 includes: first generation unit 321, screening unit 322, first comparison unit 323;

[0132] The first generation unit 321 can be used to extract three elements from the first numerical information and the second numerical information respectively, and generate a patient information triplet corresponding to the first numerical information and an ingress / exgress condition information triplet corresponding to the second numerical information based on the three elements. The patient information triplet includes the value of the patient data item, the unit of the patient data item, and the meaning of the patient data item. The ingress / exgress condition information triplet includes the numerical condition of the preset data item, the unit of the preset data item, and the meaning of the preset data item.

[0133] The filtering unit 322 can be used to filter target patient data items that match preset data items in the patient information triplet based on preset data item units, preset data item meanings, patient data item units, and patient data item meanings;

[0134] The first comparison unit 323 can be used to compare the value of the target patient data item with the value conditions of the preset data item to obtain the first matching result.

[0135] In specific application scenarios, such as Figure 5 As shown, the second matching module 33 includes: a first processing unit 331, a first input unit 332, and a second processing unit 333;

[0136] The first processing unit 331 can be used to perform one-hot encoding on the field value of each field in the medical information to obtain a high-dimensional binary vector, and use a preset dimensionality reduction algorithm to perform low-dimensional representation learning on the high-dimensional binary vector to obtain the first feature vector corresponding to the medical information.

[0137] The first input unit 332 can be used to input descriptive information into a pre-trained language representation model to obtain the second feature vector corresponding to the descriptive information.

[0138] The second processing unit 333 can be used to normalize the first feature vector and the second feature vector so that the first feature vector and the second feature vector are in the same feature space dimension.

[0139] In specific application scenarios, such as Figure 5 As shown, the device also includes: a training module 35;

[0140] The training module 35 can be used to obtain sample description information and a sample feature vector corresponding to the sample description information; input the initial BERT language model after randomly performing partial data covering on the sample description information, to output a training result; calculate a loss value of the training result and the sample feature vector; if the loss value meets a model convergence requirement, determine the initial BERT language model as the language representation model; if the loss value does not meet the model convergence requirement, update model parameters of the initial BERT language model, and iteratively train the updated initial BERT language model until the loss value meets the model convergence requirement, and determine the initial BERT language model meeting the model convergence requirement as the language representation model.

[0141] In a specific application scenario, as shown in Figure 5 The second matching module further includes a second input unit 334 and a second comparison unit 335.

[0142] The second input unit 334 can be used to input the first feature vector and the second feature vector in the same feature space dimension into a neural network with a depth of 2, to obtain a matching score.

[0143] The second comparison unit 335 can be used to compare the matching score with a preset threshold, to obtain a second matching result.

[0144] In a specific application scenario, as shown in Figure 5 The device further includes a judgment module 36 and an output module 37.

[0145] The judgment module 36 can be used to judge whether the test patient and the medical test are successfully matched according to the first matching result.

[0146] If yes, the second matching module 33 is used to pre-process the medical information and the description information, to perform embedding matching on the pre-processed medical information and the description information, to obtain a second matching result; and the generation module 34 is used to generate a final matching result of the test patient and the medical test according to the first matching result and the second matching result.

[0147] The output module 37 can be used to output a final matching result of the test patient and the medical test failing to be matched, if no.

[0148] In a specific application scenario, as shown in Figure 5 The generation module 34 includes a second generation unit 341 and a third generation unit 342.

[0149] The second generation unit 341 can be used to generate a final matching result of the test patient and the medical test successfully matched, if the first matching result and the second matching result are judged to be successfully matched.

[0150] The third generating unit 342 is configured to generate a final matching result of the trial patient and the medical trial if it is determined that at least one of the first matching result and the second matching result is a matching failure.

[0151] Based on the method as shown in Figure 1 and Figure 2 , correspondingly, the embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the method as shown in Figures 1 to 2 .

[0152] Based on the method as shown in Figure 1 and the device as shown in Figure 4 , the embodiment of the present application further provides an entity structure diagram of a computer device, as shown in Figure 6 , the computer device comprises a processor 41, a memory 42, and a computer program stored in the memory 42 and executable on the processor, wherein the memory 42 and the processor 41 are arranged on a bus 43, and the processor 41 implements the method as shown in Figures 1 to 2 when executing the program.

[0153] Through the technical solution of the present application, the patient information of the trial patient and the drug trial entry and exclusion condition information can be acquired first, the patient information comprises first numerical information and medical information, the drug trial entry and exclusion condition information comprises second numerical information and description information; then the first numerical information and the second numerical information are subjected to numerical matching to obtain a first matching result; and the medical information and the description information are preprocessed, and the preprocessed medical information and the description information are subjected to embedding matching to obtain a second matching result; finally, a final matching result of the trial patient and the medical trial is generated according to the first matching result and the second matching result. Through the technical solution in the present application, the drug trial entry and exclusion condition in patient recruitment and the patient information can be intelligently matched by using a deep neural network, end-to-end training can be performed, global optimization can be performed, traditional operations such as word segmentation and relation extraction which are prone to cause error accumulation can be avoided, higher matching precision can be obtained, and the effect of precise matching of drugs and patients can be achieved.

[0154] Each of the embodiments in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be understood by mutual reference. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be understood by referring to the part of the method embodiment.

[0155] The methods and systems of the present application can be implemented in a number of ways. For example, the methods and systems of the present application can be implemented via software, hardware, firmware, or any combination of software, hardware, and firmware. The above described order of steps for the methods is merely illustrative, and the steps of the methods of the present application are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, the present application can also be implemented as a program recorded on a recording medium, which includes machine readable instructions for implementing the methods according to the present application. Thus, the present application also covers recording media storing programs for executing the methods according to the present application.

[0156] The description of the present application is given for the sake of exemplification and description and is not intended to limit the application to the specific form described. Many modifications and variations are possible in light of the above teachings. The implementation of those steps described previously can be implemented in a variety of ways. The elements and implementation of the present application can also be implemented in a variety of ways. The above described embodiments are merely descriptive in nature and, thus, are not intended to limit the application as described herein.

Claims

1. A method of matching patients for a clinical drug trial, characterized by, The method comprises the following steps: obtaining patient information of a test patient and medical test entry and exit condition information, wherein the patient information comprises first numerical information and medical history information, and the medical test entry and exit condition information comprises second numerical information and description information; performing numerical matching on the first numerical information and the second numerical information to obtain a first matching result, comprising: performing three-element extraction on the first numerical information and the second numerical information respectively, and generating a patient information triple corresponding to the first numerical information and an entry and exit condition information triple corresponding to the second numerical information based on the three elements, wherein the patient information triple comprises a numerical value of a patient data item, a unit of the patient data item, and a meaning of the patient data item, and the entry and exit condition information triple comprises a numerical condition of a preset data item, a unit of the preset data item, and a meaning of the preset data item; filtering a target patient data item matching the preset data item from the patient information triple based on the unit of the preset data item, the meaning of the preset data item, and the unit of the patient data item and the meaning of the patient data item; comparing the numerical value of the target patient data item with the numerical condition of the preset data item to obtain the first matching result; determining whether the test patient and the medical test match successfully according to the first matching result; if yes, pre-processing the medical history information and the description information, and performing embedding matching on the pre-processed medical history information and the description information to obtain a second matching result, wherein the pre-processing of the medical history information and the description information comprises: performing one-hot encoding processing on the field value of each field in the medical history information to obtain a high-dimensional binary vector, and performing low-dimensional representation learning on the high-dimensional binary vector by using a preset dimension reduction algorithm to obtain a first feature vector corresponding to the medical history information; inputting the description information into a pre-trained language representation model to obtain a second feature vector corresponding to the description information; performing normalization processing on the first feature vector and the second feature vector to make them in the same feature space dimension; generating a final matching result of the test patient and the medical test according to the first matching result and the second matching result, comprising: if it is determined that the first matching result and the second matching result are matching success, outputting a final matching result that the test patient and the medical test match successfully; if it is determined that at least one of the first matching result and the second matching result is matching failure, outputting a final matching result that the test patient and the medical test match failure; if no, outputting a final matching result that the test patient and the medical test match failure.

2. The method of claim 1, wherein, The language representation model is a BERT language model, and before inputting the description information into the pre-trained language representation model to obtain the second feature vector corresponding to the description information, the method further comprises: obtaining sample description information and a sample feature vector corresponding to the sample description information; The sample description information is randomly partially data covered, and an initial BERT language model is input to output a training result; Loss value of the training result and the sample feature vector is calculated; If the loss value meets the model convergence requirement, the initial BERT language model is determined as the language representation model; If the loss value does not meet the model convergence requirement, the model parameters of the initial BERT language model are updated, and the updated initial BERT language model is iteratively trained until the loss value meets the model convergence requirement, and the initial BERT language model meeting the model convergence requirement is determined as the language representation model.

3. The method of claim 1, wherein, The preprocessed consultation information and the description information are embedded and matched to obtain a second matching result, including: The first feature vector and the second feature vector in the same feature space dimension are input into a neural network with a depth of 2 to obtain a matching score; The matching score is compared with a preset threshold to obtain a second matching result.

4. A matching device for clinical drug trial patients, characterized by, Including: An acquisition module is configured to acquire patient information of a test patient and drug test admission condition information, the patient information including first numerical information and consultation information, and the drug test admission condition information including second numerical information and description information; A first matching module is configured to perform numerical matching on the first numerical information and the second numerical information to obtain a first matching result, and includes a first generation unit, a screening unit, and a first comparison unit. The first generation unit is configured to perform three-element extraction on the first numerical information and the second numerical information, respectively, and generate patient information triples corresponding to the first numerical information and admission condition information triples corresponding to the second numerical information based on the three elements. The patient information triples include numerical values of patient data items, units of the patient data items, and meanings of the patient data items, and the admission condition information triples include numerical value conditions of preset data items, units of the preset data items, and meanings of the preset data items. The screening unit is configured to screen target patient data items matching the preset data items from the patient information triples based on the units of the preset data items, the meanings of the preset data items, the units of the patient data items, and the meanings of the patient data items. The first comparison unit is configured to compare numerical values of the target patient data items with numerical value conditions of the preset data items to obtain a first matching result. A judgment module is configured to determine whether the test patient and the medical test are successfully matched according to the first matching result. If yes, the second matching module is used to preprocess the medical information and the description information, and to embed and match the preprocessed medical information and the description information to obtain a second matching result. The second matching module comprises a first processing unit, a first input unit and a second processing unit. The first processing unit is configured to perform one-hot encoding processing on a field value of each field in the medical information to obtain a high-dimensional binary vector, and to perform low-dimensional representation learning on the high-dimensional binary vector by using a preset dimension reduction algorithm to obtain a first feature vector corresponding to the medical information. The first input unit is configured to input the description information into a pre-trained language representation model to obtain a second feature vector corresponding to the description information. The second processing unit is configured to perform normalization processing on the first feature vector and the second feature vector to make the first feature vector and the second feature vector in the same feature space dimension. A generation module is configured to generate a final matching result of the trial patient and the medical trial according to the first matching result and the second matching result. The generation module comprises a second generation unit and a third generation unit. The second generation unit is configured to generate a final matching result of the trial patient and the medical trial if it is determined that the first matching result and the second matching result are both successful. The third generation unit is configured to generate a final matching result of the trial patient and the medical trial if it is determined that at least one of the first matching result and the second matching result is unsuccessful. An output module is configured to output a final matching result of the trial patient and the medical trial if no.

5. The apparatus of claim 4, wherein, The device further comprises a training module. The training module is configured to obtain sample description information and a sample feature vector corresponding to the sample description information, input the sample description information after random partial data masking into an initial BERT language model to output a training result, calculate a loss value of the training result and the sample feature vector, determine the initial BERT language model as the language representation model if the loss value meets a model convergence requirement, update model parameters of the initial BERT language model if the loss value does not meet the model convergence requirement, and perform iterative training on the updated initial BERT language model until the loss value meets the model convergence requirement, and determine the initial BERT language model meeting the model convergence requirement as the language representation model.

6. The apparatus of claim 4, wherein, The second matching module further comprises a second input unit and a second comparison unit. The second input unit is configured to input the first feature vector and the second feature vector in the same feature space dimension into a neural network with a depth of 2 to obtain a matching score. The second comparison unit is configured to compare the matching score with a preset threshold to obtain a second matching result.

7. A storage medium having stored thereon a computer program, characterized in that The program is executed by a processor to implement the matching method of the clinical drug trial patient in any one of claims 1 to 3.

8. A computer device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, The processor implements the matching method for clinical drug test patients in any one of claims 1 to 3 when executing the program.

Images