Physical examination data processing method, device, equipment and storage medium

Abstract

The present invention relates to the field of artificial intelligence, and is applied to the field of smart medical care. A method, device, equipment and storage medium for processing physical examination data are disclosed, which are used to improve the accuracy of physical examination data analysis. The method for processing physical examination data includes: obtaining initial physical examination data, preprocessing the initial physical examination data to obtain preprocessed physical examination data; performing vector conversion on the preprocessed physical examination data to obtain an input vector; calling a preset target deep language model, performing department probability prediction on the input vector to obtain an initial prediction pair, sorting the initial prediction pair to obtain a sorting result, and extracting a target prediction pair according to the sorting result; calling a preset target twin network model, and performing loss function similarity calculation and department information similarity matching on the target prediction pair based on a preset department information database to obtain a target result. In addition, the present invention also relates to blockchain technology, and the target result can be stored in a blockchain node.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and in particular to a method, device, equipment and storage medium for processing physical examination data. Background Art

[0002] Health check-up, as a test or screening at the front end of medical care, is an important part of the medical field. The physical examination report generated after the physical examination refers to a document in a certain format generated based on the data of the body's reactions after the physical examination. The physical examination report is parsed and classified through the artificial intelligence AI model. According to the classification results, the corresponding hospital department of the user can be confirmed, thereby improving the user experience.

[0003] However, existing physical examination data analysis models are unable to identify details. For example, for routine blood examination data, only data such as white blood cells, red blood cells and hemoglobin are analyzed, and parameters such as the absolute value of monocytes are often unable to be analyzed, resulting in low accuracy in the analysis of physical examination data. Summary of the invention

[0004] The present invention provides a method, device, equipment and storage medium for processing physical examination data, which is used to call a target deep language model to perform probability prediction on an input vector to obtain an initial prediction pair, sort the initial prediction pair to obtain a sorting result, extract a target prediction pair according to the sorting result, call a target twin network model, calculate the similarity of the loss function of the target prediction pair and match the similarity of department information to obtain a target result, thereby improving the accuracy of physical examination data parsing.

[0005] The first aspect of the present invention provides a method for processing physical examination data, including: obtaining initial physical examination data, preprocessing the initial physical examination data to obtain preprocessed physical examination data; performing vector conversion on the preprocessed physical examination data to obtain an input vector, wherein the input vector includes multiple word vectors, multiple segment vectors, and multiple position vectors; calling a preset target deep language model, performing department probability prediction on the input vector to obtain an initial prediction pair, sorting the initial prediction pair to obtain a sorting result, and extracting a target prediction pair based on the sorting result; calling a preset target twin network model, performing loss function similarity calculation and department information similarity matching on the target prediction pair based on a preset department information database to obtain a target result.

[0006] Optionally, in a first implementation method of the first aspect of the present invention, the pre-processed physical examination data is vectorized to obtain an input vector, and the input vector includes multiple word vectors, multiple segment vectors and multiple position vectors, including: based on a preset tag embedding layer, the pre-processed physical examination data is double-byte encoded to obtain multiple word vectors; through a preset segment embedding layer, the multiple word vectors are embedded to obtain a segment vector corresponding to each word vector; through a preset position embedding layer, the multiple word vectors are embedded to obtain a position vector corresponding to each word vector; the multiple word vectors, the segment vector corresponding to each word vector and the position vector corresponding to each word vector are concatenated to obtain the input vector.

[0007] Optionally, in a second implementation method of the first aspect of the present invention, calling a preset target deep language model, predicting the department probability of the input vector to obtain an initial prediction pair, sorting the initial prediction pair to obtain a sorting result, and extracting a target prediction pair according to the sorting result includes: calling a preset target deep language model, predicting the department probability corresponding to the input vector to obtain an initial prediction pair, the initial prediction pair including the input vector and multiple department prediction probabilities corresponding to the input vector; sorting the initial prediction pairs in order of the multiple department prediction probabilities from large to small to obtain a sorting result, extracting a preset number of prediction pairs from the sorting result, and obtaining a target prediction pair.

[0008] Optionally, in a third implementation method of the first aspect of the present invention, calling a preset target twin network model, performing loss function similarity calculation and department information similarity matching on the target prediction pair based on a preset department information database, and obtaining a target result includes: calling a preset target twin network model, respectively calculating the loss function values ​​of the input vector in the target prediction pair and the multiple department prediction probabilities corresponding to the input vector, to obtain the input vector loss function value and the department probability loss function value, and calculating the similarity between the input vector loss function value and the department probability loss function value based on a preset department information database to obtain a prediction pair similarity value; sorting the target prediction pairs in order of the prediction pair similarity values ​​from large to small to obtain a prediction pair sequence, and determining the target prediction pair ranked first in the prediction pair sequence as the target result, and the target result is used to indicate the target department information.

[0009] Optionally, in a fourth implementation manner of the first aspect of the present invention, before obtaining the initial physical examination data, preprocessing the initial physical examination data, and obtaining the preprocessed physical examination data, the method also includes: obtaining a model training vector, and training a preset initial deep language model and a preset initial twin network model based on the model training vector, respectively, to obtain a target deep language model and a target twin network model.

[0010] Optionally, in a fifth implementation manner of the first aspect of the present invention, the acquiring of the model training vector, and the training of the preset initial deep language model and the preset initial twin network model based on the model training vector to obtain the target deep language model and the target twin network model include: acquiring the model training vector, calling the preset deep language model, and performing department probability prediction on the model training vector to obtain multiple training prediction pairs, wherein the multiple training prediction pairs include the model training vector and multiple training prediction probabilities corresponding to the model training vector; optimizing the preset deep language model through the multiple training prediction pairs to obtain the target deep language model; calling the preset twin network model, performing feature extraction on the multiple training prediction pairs to obtain a training department label, wherein the training department label is used to indicate the department information corresponding to the model training vector; optimizing the preset twin network model through the training department label to obtain the target twin network model.

[0011] Optionally, in a sixth implementation manner of the first aspect of the present invention, before obtaining the initial physical examination data, preprocessing the initial physical examination data, and obtaining the preprocessed physical examination data, the method also includes: obtaining historical data, using a preset anonymization algorithm to desensitize the historical data to obtain desensitized data, wherein the historical data includes historical physical examination data and department data corresponding to the historical physical examination data; loading the desensitized data into a preset database construction template to obtain a department information database.

[0012] The second aspect of the present invention provides a device for processing physical examination data, including: an acquisition module, used to acquire initial physical examination data, preprocess the initial physical examination data, and obtain preprocessed physical examination data; a conversion module, used to perform vector conversion on the preprocessed physical examination data to obtain an input vector, wherein the input vector includes multiple word vectors, multiple segment vectors, and multiple position vectors; a prediction module, used to call a preset target deep language model, perform department probability prediction on the input vector, obtain an initial prediction pair, sort the initial prediction pair, obtain a sorting result, and extract a target prediction pair according to the sorting result; a parsing module, used to call a preset target twin network model, perform loss function similarity calculation and department information similarity matching on the target prediction pair based on a preset department information database, and obtain a target result.

[0013] Optionally, in a first implementation manner of the second aspect of the present invention, the conversion module includes: an encoding unit, used to perform double-byte encoding on the preprocessed physical examination data based on a preset tag embedding layer to obtain multiple word vectors; a first embedding unit, used to embed the multiple word vectors through a preset segment embedding layer to obtain a segment vector corresponding to each word vector; a second embedding unit, used to embed the multiple word vectors through a preset position embedding layer to obtain a position vector corresponding to each word vector; a splicing unit, used to splice the multiple word vectors, the segment vector corresponding to each word vector, and the position vector corresponding to each word vector to obtain an input vector.

[0014] Optionally, in a second implementation method of the second aspect of the present invention, the prediction module includes: a first prediction unit, used to call a preset target deep language model to predict the department probability corresponding to the input vector to obtain an initial prediction pair, wherein the initial prediction pair includes the input vector and multiple department prediction probabilities corresponding to the input vector; a sorting unit, used to sort the initial prediction pair in order of the multiple department prediction probabilities from large to small to obtain a sorting result, extract a preset number of prediction pairs from the sorting result, and obtain a target prediction pair.

[0015] Optionally, in a third implementation method of the second aspect of the present invention, the parsing module includes: a similarity calculation unit, used to call a preset target twin network model, and respectively calculate the loss function values ​​of the input vector in the target prediction pair and the multiple department prediction probabilities corresponding to the input vector, to obtain the input vector loss function value and the department probability loss function value, and based on a preset department information database, calculate the similarity between the input vector loss function value and the department probability loss function value to obtain a prediction pair similarity value; a determination unit, used to sort the target prediction pairs in order of the prediction pair similarity values ​​from large to small, to obtain a prediction pair sequence, and determine the target prediction pair ranked first in the prediction pair sequence as the target result, and the target result is used to indicate the target department information.

[0016] Optionally, in a fourth implementation method of the second aspect of the present invention, before the acquisition module, the physical examination data processing device also includes: a training module, used to obtain a model training vector, and based on the model training vector, respectively train a preset initial deep language model and a preset initial twin network model to obtain a target deep language model and a target twin network model.

[0017] Optionally, in a fifth implementation manner of the second aspect of the present invention, the training module includes: a second prediction unit, used to obtain a model training vector, call a preset deep language model, perform department probability prediction on the model training vector, and obtain multiple training prediction pairs, wherein the multiple training prediction pairs include the model training vector and multiple training prediction probabilities corresponding to the model training vector; a first optimization unit, used to optimize the preset deep language model through the multiple training prediction pairs to obtain a target deep language model; a feature extraction unit, used to call a preset twin network model, perform feature extraction on the multiple training prediction pairs, and obtain a training department label, wherein the training department label is used to indicate the department information corresponding to the model training vector; a second optimization unit, used to optimize the preset twin network model through the training department label to obtain a target twin network model.

[0018] Optionally, in a sixth implementation manner of the second aspect of the present invention, before the acquisition module, the physical examination data processing device also includes a construction module, and the construction module includes: a desensitizing unit, used to acquire historical data, and use a preset anonymization algorithm to desensitize the historical data to obtain desensitized data, wherein the historical data includes historical physical examination data and department data corresponding to the historical physical examination data; a loading unit, used to load the desensitized data into a preset database construction template to obtain a department information database.

[0019] The third aspect of the present invention provides a physical examination data processing device, comprising: a memory and at least one processor, wherein the memory stores a computer program; the at least one processor calls the computer program in the memory so that the physical examination data processing device executes the above-mentioned physical examination data processing method.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned method for processing physical examination data.

[0021] In the technical solution provided by the present invention, initial physical examination data is obtained, the initial physical examination data is preprocessed to obtain preprocessed physical examination data; the preprocessed physical examination data is vectorized to obtain an input vector, the input vector includes multiple word vectors, multiple segment vectors and multiple position vectors; a preset target deep language model is called to perform department probability prediction on the input vector to obtain an initial prediction pair, the initial prediction pair is sorted to obtain a sorting result, and a target prediction pair is extracted according to the sorting result; a preset target twin network model is called to perform similarity calculation of the loss function and similarity matching of the department information on the target prediction pair based on a preset department information database to obtain a target result. In an embodiment of the present invention, by calling the target deep language model, probabilistically predicting the input vector to obtain an initial prediction pair, sorting the initial prediction pair to obtain a sorting result, extracting the target prediction pair according to the sorting result, calling the target twin network model, performing similarity calculation of the loss function and similarity matching of the department information on the target prediction pair to obtain the target result, the accuracy of physical examination data parsing is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of an embodiment of a method for processing physical examination data in an embodiment of the present invention;

[0023] Figure 2 A schematic diagram of another embodiment of a method for processing physical examination data in an embodiment of the present invention;

[0024] Figure 3 A schematic diagram of an embodiment of a device for processing physical examination data in an embodiment of the present invention;

[0025] Figure 4 It is a schematic diagram of another embodiment of the device for processing physical examination data in an embodiment of the present invention;

[0026] Figure 5 The figure is a schematic diagram of an embodiment of a device for processing physical examination data in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The embodiments of the present invention provide a method, apparatus, device and storage medium for processing physical examination data, which are used to perform probability prediction on input vectors by calling a target deep language model to obtain initial prediction pairs, sort the initial prediction pairs to obtain sorting results, extract target prediction pairs based on the sorting results, call a target twin network model, calculate the similarity of loss functions of the target prediction pairs and match the similarity of department information to obtain target results, thereby improving the accuracy of physical examination data parsing.

[0028] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention 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 described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily 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.

[0029] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the method for processing physical examination data in an embodiment of the present invention includes:

[0030] 101. Acquire initial physical examination data, and preprocess the initial physical examination data to obtain preprocessed physical examination data.

[0031] It is understandable that the execution subject of the present invention may be a device for processing physical examination data, or may be a terminal or a server, which is not specifically limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0032] The server obtains initial physical examination data, preprocesses the initial physical examination data, and obtains preprocessed physical examination data. The patient is tested and the initial physical examination data is generated by the preset medical device. The medical device uploads the initial physical examination data to the medical cloud. The server extracts the corresponding initial physical examination data from the medical cloud according to the patient information. The initial physical examination data is obtained by crawlers. The initial physical examination data in this embodiment are all authorized by the user. The initial physical examination data includes multiple project names and the inspection data corresponding to each project name. The initial physical examination data can be stored in the patient's corresponding health record. For example: the project name is blood routine, and the inspection data corresponding to the blood routine includes but is not limited to white blood cell count, red blood cell count, hemoglobin determination, mean red blood cell volume and platelet count, etc., or the project name is urine routine, and the inspection data corresponding to the urine routine includes but is not limited to urine glucose, urine protein and urine ketone bodies, etc. The process of preprocessing the initial physical examination data mainly includes missing value filling, outlier filtering and duplicate value filtering. Among them, missing value filling can be multiple interpolation or hot card filling, outlier filtering mainly uses the outlier detection algorithm z-score to identify outliers and delete them, and duplicate value filtering is to deduplicate duplicate values ​​to obtain preprocessed physical examination data.

[0033] Before step 101, the server obtains a model training vector, and trains a preset deep language model and a preset twin network model based on the model training vector to obtain a target deep language model and a target twin network model. The model training vector in this embodiment is obtained by concatenating the word vector, segment vector and position vector corresponding to the model training data. The preset deep language model is a BERT (bidirectional encoder representations from transformers) model. When training the BERT model, the three processed vectors need to be added and input into the BERT network. The BERT model learns the correspondence between the model training data and the corresponding department, and outputs a plurality of prediction probabilities (i.e., training prediction probabilities) of the corresponding departments. The model training vector is respectively combined with a plurality of training prediction probabilities corresponding to the model training vector to obtain a plurality of training prediction pairs. Through the plurality of training prediction pairs, the parameters of the BERT model are adjusted and optimized to obtain a target deep language model. The parameter adjustment includes calculating the model loss function to update the hyperparameters of the model. The hyperparameters of the model may be a batch size or a model learning rate. The preset twin network model is divided into two inputs, input1 and input2. The model training vectors in multiple training prediction pairs are input through input1, and multiple training prediction probabilities are input through input2. The feature extractor used by the twin network model is the text classifier textcnn. The text classifier is used to extract features from the input multiple training prediction pairs to obtain the training department labels. The twin network model parameters are updated based on the training department labels to obtain the target twin network model. The parameter update includes calculating the model loss function to update the model's hyperparameters, which can be the model learning rate.

[0034] 102. Perform vector conversion on the preprocessed physical examination data to obtain an input vector, where the input vector includes multiple word vectors, multiple segment vectors, and multiple position vectors.

[0035] The server performs vector conversion on the preprocessed physical examination data to obtain an input vector, which includes multiple word vectors, multiple segment vectors, and multiple position vectors. The process of vector conversion on the preprocessed physical examination data mainly includes passing the preprocessed physical examination data through a preset tag embedding layer, a preset segment embedding layer, and a preset position embedding layer, respectively, and converting the preprocessed physical examination data into the input form of the BERT model (i.e., a deep language model) (i.e., a word vector, a segment vector, and a position vector). In the tag embedding layer, the preprocessed physical examination data is double-byte encoded by the word cutting model wordpiece to obtain multiple word vectors. The segment embedding layer has only two vector representations, which are used to represent the contextual relationship of the sentence. The first vector assigns 0 to each word in the first sentence, and the second vector assigns 1 to each word in the second sentence. If the input has only one sentence, its segment vector is all 0. The position vector represents the corresponding position of each word in the sentence. Multiple word vectors, the segment vector corresponding to each word vector, and the position vector corresponding to each word vector are concatenated to finally obtain the input vector.

[0036] 103. Call the preset target deep language model, perform department probability prediction on the input vector, obtain initial prediction pairs, sort the initial prediction pairs to obtain sorting results, and extract the target prediction pairs according to the sorting results.

[0037] The server calls the preset target deep language model, predicts the department probability of the input vector, obtains the initial prediction pair, sorts the initial prediction pair, obtains the sorting result, and extracts the target prediction pair based on the sorting result. During the prediction, the initial physical examination data is converted into the vector form required by the BERT model (i.e., the input vector), and input into the trained BERT model. The model will output the probability values ​​corresponding to different departments, and combine the input vector with different department prediction probabilities to obtain the initial prediction pair. The initial prediction pair includes the input vector and multiple department prediction probabilities corresponding to the input vector. The server sorts the initial prediction pairs according to the department prediction probability from large to small, and selects the top five prediction pairs in the sorting result as the target prediction pairs. The target deep language model can be applied to a variety of different fields, such as speech recognition, medical diagnosis, application testing, etc.

[0038] 104. Call the preset target twin network model, calculate the similarity of the loss function and match the similarity of the department information for the target prediction pair based on the preset department information database, and obtain the target result.

[0039] The server calls the preset target twin network model, and calculates the similarity of the loss function and matches the similarity of the department information for the target prediction pair based on the preset department information database to obtain the target result. The target twin network model is divided into two inputs, input1 and input2. The input vector in the target prediction pair is input through input1, and multiple department prediction probabilities are input through input2, forming 5 pairs of data input twin networks for model prediction. By calculating the input vector loss function value and the department probability loss function value, the target twin network model calculates the similarity of the input vector loss function value and the department probability loss function value to obtain the prediction pair similarity value. The target prediction pair corresponding to the maximum value in the prediction pair similarity value is determined as the target result. After obtaining the target result, the patient can go to the corresponding department for registration and treatment, thereby improving the consultation efficiency.

[0040] In an embodiment of the present invention, by calling the target deep language model, probabilistic prediction is performed on the input vector to obtain an initial prediction pair, the initial prediction pair is sorted to obtain the sorting result, the target prediction pair is extracted according to the sorting result, and the target twin network model is called to calculate the similarity of the loss function and match the similarity of the department information for the target prediction pair to obtain the target result, thereby improving the accuracy of physical examination data analysis. This solution can be applied to the field of smart medical care, thereby promoting the construction of smart cities.

[0041] See also Figure 2 Another embodiment of the method for processing physical examination data in the embodiment of the present invention includes:

[0042] 201. Acquire initial physical examination data, and preprocess the initial physical examination data to obtain preprocessed physical examination data.

[0043] The server obtains initial physical examination data, preprocesses the initial physical examination data, and obtains preprocessed physical examination data. Specifically, the server obtains initial physical examination data, and the initial physical examination data includes multiple project names and inspection data corresponding to each project name; the server performs missing value completion, outlier filtering and duplicate value filtering on the initial physical examination data to obtain preprocessed physical examination data. The initial physical examination data is obtained through a crawler. The initial physical examination data in this embodiment are all authorized by the user. The initial physical examination data includes multiple project names and inspection data corresponding to each project name. For example: the project name is blood routine, and the inspection data corresponding to the blood routine includes but is not limited to white blood cell count, red blood cell count, hemoglobin determination, mean corpuscular volume and platelet count, etc., or the project name is urine routine, and the inspection data corresponding to the urine routine includes but is not limited to urine glucose, urine protein and urine ketone bodies, etc. The process of preprocessing the initial physical examination data mainly includes missing value filling, outlier filtering and duplicate value filtering. Among them, missing value filling can be multiple interpolation or hot card filling, outlier filtering mainly uses the outlier detection algorithm z-score to identify outliers and delete them, and duplicate value filtering is to deduplicate duplicate values ​​to obtain preprocessed physical examination data.

[0044] 202. Based on the preset tag embedding layer, the preprocessed physical examination data is double-byte encoded to obtain multiple word vectors.

[0045] The server performs double-byte encoding on the pre-processed physical examination data based on the preset tag embedding layer to obtain multiple word vectors. In the preset tag embedding layer, each word in the pre-processed physical examination data is converted into a 768-dimensional vector representation. The pre-processed physical examination data is processed through the word-slicing wordpiece model. The [CLS] tag is inserted before the sentence as a classification tag, and the [SEP] is inserted at the end of the sentence as an end tag to obtain multiple word vectors. One of the main implementation methods of wordpiece is called double-byte encoding (byte-pair encoding, BPE). For example, the pre-processed physical examination data includes "accompanied by headaches, fatigue, and inattention", which is divided into "accompanied by / headaches / fatigue / appears / inattention ...

[0046] 203. Through the preset segment embedding layer, multiple word vectors are embedded to obtain a segment vector corresponding to each word vector.

[0047] The server embeds multiple word vectors through the preset segment embedding layer to obtain the segment vector corresponding to each word vector. The output of the preset segment embedding layer (i.e., the segment vector corresponding to each word vector) has only two vector representations. The segment vector is used to distinguish different sentences in the context. The first vector representation is to assign 0 to each word in the first sentence, and the second vector representation is to assign 1 to each word in the second sentence. If there is only one sentence in the input, then its segment vector is all 0. For example, "accompanied by headache, fatigue, and inattention" contains two sentences, so the segment vectors corresponding to the word vectors "accompanied by / with / headache / fatigue / fatigue" are all 0, which can be obtained by E A It means that the segment vectors corresponding to the word vectors of "appearing / not / attention / concentrated" are all 1, which can be expressed by E B It means that the vector dimension corresponding to each segment vector is also 768 dimensions.

[0048] 204. Embed multiple word vectors through a preset position embedding layer to obtain a position vector corresponding to each word vector.

[0049] The server embeds multiple word vectors through a preset position embedding layer to obtain the position vector corresponding to each word vector. The position embedding layer is used to obtain the position of each word in the sentence. Therefore, the position vector is represented by the vector corresponding to the position of each word in the sentence. For example, the position vector corresponding to "with headache and fatigue" can be represented as "E0, E1, E2, E3, E4 and E5" respectively.

[0050] 205. Concatenate multiple word vectors, the segment vector corresponding to each word vector, and the position vector corresponding to each word vector to obtain an input vector.

[0051] The server concatenates multiple word vectors, the segment vector corresponding to each word vector, and the position vector corresponding to each word vector to obtain an input vector. The server adds each word vector, the segment vector corresponding to each word vector, and the position vector corresponding to each word vector to achieve concatenation, and finally obtains an N*768-dimensional input vector, where N represents the corresponding sentence length.

[0052] 206. Call the preset target deep language model, perform department probability prediction on the input vector, obtain an initial prediction pair, sort the initial prediction pair to obtain a sorting result, and extract the target prediction pair according to the sorting result.

[0053] The server calls the preset target deep language model, predicts the department probability of the input vector, obtains an initial prediction pair, sorts the initial prediction pair, obtains the sorting result, and extracts the target prediction pair according to the sorting result. Specifically, the server calls the preset target deep language model, predicts the department probability corresponding to the input vector, and obtains an initial prediction pair, the initial prediction pair includes the input vector and multiple department prediction probabilities corresponding to the input vector; the server sorts the initial prediction pairs in order of multiple department prediction probabilities from large to small, obtains the sorting result, extracts a preset number of prediction pairs from the sorting result, and obtains the target prediction pair.

[0054] The server converts the initial physical examination data into the vector form required by the BERT model (i.e., input vector), and inputs it into the trained BERT model. The model will output the probability values ​​corresponding to different departments, and combine the input vectors with different department prediction probabilities to obtain the initial prediction pair. The initial prediction pair includes the input vector and multiple department prediction probabilities corresponding to the input vector. The initial prediction pairs are sorted from large to small according to the department prediction probabilities, and the top five prediction pairs in the sorting results are selected as the target prediction pairs. For example, the information corresponding to the input vector is "hip dysplasia", and the corresponding department prediction probabilities are " Department of Gastroenterology 40%", "Department of Obstetrics and Gynecology 50%", "Department of Rehabilitation 80%", "Department of Orthopedics 90%", "Department of Ultrasound 20%", and "Department of Cardiovascular Medicine 10%" are sorted in descending order, and the top five prediction pairs are extracted, namely "Hip Dysplasia-Orthopedics 90%", "Hip Dysplasia-Rehabilitation 80%", "Hip Dysplasia-Obstetrics and Gynecology 50%", "Hip Dysplasia-Gastroenterology 40%" and "Hip Dysplasia-Ultrasound 20%", and the target prediction pairs are obtained. In this embodiment, it can be considered that the accuracy of the top five BERT model rankings is 100%.

[0055] 207. Call the preset target twin network model, calculate the similarity of the loss function and match the similarity of the department information for the target prediction pair based on the preset department information database, and obtain the target result.

[0056] The server calls the preset target twin network model, and performs similarity calculation of the loss function and similarity matching of the department information on the target prediction pair based on the preset department information database to obtain the target result. Specifically, the server calls the preset target twin network model, performs similarity calculation on the target prediction pair and the preset department information database to obtain the prediction pair similarity value; the server sorts the target prediction pairs in descending order of the prediction pair similarity values ​​to obtain a prediction pair sequence, and determines the target prediction pair ranked first in the prediction pair sequence as the target result, and the target result is used to indicate the target department information.

[0057] The target twin network model is divided into two inputs, input1 and input2. The input vector in the target prediction pair is input through input1, and the prediction probabilities of multiple departments are input through input2, forming 5 pairs of data input twin networks for model prediction. By calculating the input vector loss function value and the department probability loss function value, the target twin network model performs similarity calculation on the input vector loss function value and the department probability loss function value to obtain the prediction pair similarity value. The target prediction pairs are sorted based on the prediction pair similarity value to obtain the prediction pair sequence, and the target prediction pair corresponding to the maximum value in the prediction pair similarity value (that is, the target prediction pair ranked first in the prediction pair sequence) is confirmed. The target result is determined as the target result. For example, after similarity calculation, the five groups of data in the target prediction pair are determined to have the highest similarity with orthopedics for "hip dysplasia", and the target result is finally determined to be "orthopedics". In this embodiment, the target BERT model and the target twin network model are connected in series for model prediction. The high-level semantic representation ability of the target BERT model and the general input and output layer processing input vectors, as well as the powerful lexical semantic similarity analysis ability of the twin network, are combined to further predict the target prediction pair output by the target BERT model and output the target result. After obtaining the target result, the patient can go to the corresponding department for registration and treatment, thereby improving the consultation efficiency.

[0058] In an embodiment of the present invention, by calling the target deep language model, probabilistic prediction is performed on the input vector to obtain an initial prediction pair, the initial prediction pair is sorted to obtain the sorting result, the target prediction pair is extracted according to the sorting result, and the target twin network model is called to calculate the similarity of the loss function and match the similarity of the department information for the target prediction pair to obtain the target result, thereby improving the accuracy of physical examination data analysis. This solution can be applied to the field of smart medical care, thereby promoting the construction of smart cities.

[0059] The above describes the method for processing physical examination data in the embodiment of the present invention. The following describes the device for processing physical examination data in the embodiment of the present invention. Figure 3 , an embodiment of a device for processing physical examination data in an embodiment of the present invention includes:

[0060] The acquisition module 301 is used to acquire initial physical examination data and pre-process the initial physical examination data to obtain pre-processed physical examination data;

[0061] A conversion module 302 is used to perform vector conversion on the pre-processed physical examination data to obtain an input vector, where the input vector includes multiple word vectors, multiple segment vectors and multiple position vectors;

[0062] The prediction module 303 is used to call a preset target deep language model, perform department probability prediction on the input vector, obtain an initial prediction pair, sort the initial prediction pair to obtain a sorting result, and extract a target prediction pair according to the sorting result;

[0063] The parsing module 304 is used to call the preset target twin network model, perform similarity calculation of the loss function and similarity matching of the department information on the target prediction pair based on the preset department information database, and obtain the target result.

[0064] In an embodiment of the present invention, by calling the target deep language model, probabilistic prediction is performed on the input vector to obtain an initial prediction pair, the initial prediction pair is sorted to obtain the sorting result, the target prediction pair is extracted according to the sorting result, and the target twin network model is called to calculate the similarity of the loss function and match the similarity of the department information for the target prediction pair to obtain the target result, thereby improving the accuracy of physical examination data analysis. This solution can be applied to the field of smart medical care, thereby promoting the construction of smart cities.

[0065] See also Figure 4 Another embodiment of the device for processing physical examination data in the embodiment of the present invention includes:

[0066] The acquisition module 301 is used to acquire initial physical examination data and pre-process the initial physical examination data to obtain pre-processed physical examination data;

[0067] A conversion module 302 is used to perform vector conversion on the pre-processed physical examination data to obtain an input vector, where the input vector includes multiple word vectors, multiple segment vectors and multiple position vectors;

[0068] The conversion module 302 specifically includes:

[0069] The encoding unit 3021 is used to perform double-byte encoding on the pre-processed physical examination data based on a preset tag embedding layer to obtain multiple word vectors;

[0070] A first embedding unit 3022, configured to embed the multiple word vectors through a preset segment embedding layer to obtain a segment vector corresponding to each word vector;

[0071] A second embedding unit 3023 is used to embed the multiple word vectors through a preset position embedding layer to obtain a position vector corresponding to each word vector;

[0072] A concatenation unit 3024 is used to concatenate multiple word vectors, a segment vector corresponding to each word vector, and a position vector corresponding to each word vector to obtain an input vector;

[0073] The prediction module 303 is used to call a preset target deep language model, perform department probability prediction on the input vector, obtain an initial prediction pair, sort the initial prediction pair to obtain a sorting result, and extract a target prediction pair according to the sorting result;

[0074] The parsing module 304 is used to call the preset target twin network model, perform similarity calculation of the loss function and similarity matching of the department information on the target prediction pair based on the preset department information database, and obtain the target result.

[0075] Optionally, the prediction module 303 includes:

[0076] The first prediction unit 3031 is used to call a preset target deep language model to predict the probability of the department corresponding to the input vector to obtain an initial prediction pair, where the initial prediction pair includes the input vector and a plurality of department prediction probabilities corresponding to the input vector;

[0077] The sorting unit 3032 is used to sort the initial prediction pairs according to the order of the prediction probabilities of multiple departments from large to small, obtain the sorting results, extract a preset number of prediction pairs from the sorting results, and obtain the target prediction pairs.

[0078] Optionally, the parsing module 304 includes:

[0079] A similarity calculation unit 3041 is used to call a preset target twin network model, respectively calculate the loss function values ​​of the input vector in the target prediction pair and the prediction probabilities of multiple departments corresponding to the input vector, obtain the input vector loss function value and the department probability loss function value, calculate the similarity between the input vector loss function value and the department probability loss function value, and obtain the prediction pair similarity value;

[0080] The determination unit 3042 is used to sort the target prediction pairs in descending order of the prediction pair similarity values ​​to obtain a prediction pair sequence, and determine the target prediction pair ranked first in the prediction pair sequence as the target result, which is used to indicate the target department information.

[0081] Optionally, before acquiring module 301, the device for processing physical examination data further includes:

[0082] The training module 305 is used to obtain a model training vector, and to train a preset initial deep language model and a preset initial twin network model based on the model training vector to obtain a target deep language model and a target twin network model.

[0083] Optionally, the training module 305 includes:

[0084] The second prediction unit 3051 is used to obtain a model training vector, call a preset deep language model, perform department probability prediction on the model training vector, and obtain multiple training prediction pairs, where the multiple training prediction pairs include the model training vector and multiple training prediction probabilities corresponding to the model training vector;

[0085] A first optimization unit 3052 is used to optimize a preset deep language model through a plurality of training prediction pairs to obtain a target deep language model;

[0086] The feature extraction unit 3053 is used to call the preset twin network model, perform feature extraction on multiple training prediction pairs, and obtain training department labels, where the training department labels are used to indicate the department information corresponding to the model training vector;

[0087] The second optimization unit 3054 is used to optimize the preset twin network model by training department labels to obtain a target twin network model.

[0088] Optionally, before the acquisition module 301, the physical examination data processing device further includes a construction module 306, including:

[0089] The desensitization unit 3061 is used to obtain historical data and perform desensitization processing on the historical data using a preset anonymization algorithm to obtain desensitized data, wherein the historical data includes historical physical examination data and department data corresponding to the historical physical examination data;

[0090] The loading unit 3062 is used to load the desensitized data into a preset database construction template to obtain a department information database.

[0091] In an embodiment of the present invention, by calling the target deep language model, probabilistic prediction is performed on the input vector to obtain an initial prediction pair, the initial prediction pair is sorted to obtain the sorting result, the target prediction pair is extracted according to the sorting result, and the target twin network model is called to calculate the similarity of the loss function and match the similarity of the department information for the target prediction pair to obtain the target result, thereby improving the accuracy of physical examination data analysis. This solution can be applied to the field of smart medical care, thereby promoting the construction of smart cities.

[0092] above Figure 3 and Figure 4 The physical examination data processing device in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The physical examination data processing device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0093] Figure 5: is a schematic diagram of the structure of a physical examination data processing device provided by an embodiment of the present invention. The physical examination data processing device 500 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 510 (for example, one or more processors) and a memory 520, and one or more storage media 530 (for example, one or more massive storage devices) storing application programs 533 or data 532. Among them, the memory 520 and the storage medium 530 can be short-term storage or permanent storage. The program stored in the storage medium 530 may include one or more modules (not shown in the figure), and each module may include a series of computer program operations in the physical examination data processing device 500. Furthermore, the processor 510 can be configured to communicate with the storage medium 530 to execute a series of computer program operations in the storage medium 530 on the physical examination data processing device 500.

[0094] The physical examination data processing device 500 may also include one or more power supplies 540, one or more wired or wireless network interfaces 550, one or more input and output interfaces 560, and / or one or more operating systems 531, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will appreciate that Figure 5 The structure of the physical examination data processing device shown does not constitute a limitation on the physical examination data processing device, and may include more or less components than shown in the figure, or combine certain components, or arrange the components differently.

[0095] The present invention also provides a physical examination data processing device, comprising: a memory and at least one processor, wherein a computer program is stored in the memory, and the memory and the at least one processor are interconnected via a line; the at least one processor calls the computer program in the memory so that the physical examination data processing device executes the steps in the above-mentioned physical examination data processing method.

[0096] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. A computer program is stored in the computer-readable storage medium. When the computer program runs on a computer, the computer executes the steps of the method for processing physical examination data.

[0097] Furthermore, the computer-readable storage medium may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function, etc.; the data storage area may store data created according to the use of the blockchain node, etc.

[0098] The blockchain referred to in this invention is a new application mode of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.

[0099] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0100] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0101] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for processing physical examination data, characterized in that: The method for processing the physical examination data comprises: Acquiring initial physical examination data, and preprocessing the initial physical examination data to obtain preprocessed physical examination data; Performing vector conversion on the preprocessed physical examination data to obtain an input vector, wherein the input vector includes a plurality of word vectors, a plurality of segment vectors, and a plurality of position vectors; Calling a preset target deep language model, performing department probability prediction on the input vector to obtain an initial prediction pair, sorting the initial prediction pair to obtain a sorting result, and extracting a target prediction pair according to the sorting result; Calling a preset target twin network model, performing similarity calculation of the loss function and similarity matching of the department information on the target prediction pair based on a preset department information database, and obtaining a target result; The vector conversion of the pre-processed physical examination data to obtain an input vector, wherein the input vector includes a plurality of word vectors, a plurality of segment vectors and a plurality of position vectors, comprises: based on a preset tag embedding layer, double-byte encoding the pre-processed physical examination data to obtain a plurality of word vectors; embedding the plurality of word vectors through a preset segment embedding layer to obtain a segment vector corresponding to each word vector; embedding the plurality of word vectors through a preset position embedding layer to obtain a position vector corresponding to each word vector; concatenating the plurality of word vectors, the segment vector corresponding to each word vector and the position vector corresponding to each word vector to obtain the input vector; The calling of a preset target deep language model, predicting the department probability of the input vector to obtain an initial prediction pair, sorting the initial prediction pair to obtain a sorting result, and extracting a target prediction pair according to the sorting result includes: calling a preset target deep language model, predicting the department probability corresponding to the input vector to obtain an initial prediction pair, the initial prediction pair including the input vector and a plurality of department prediction probabilities corresponding to the input vector; sorting the initial prediction pair in descending order of the plurality of department prediction probabilities to obtain a sorting result, extracting a preset number of prediction pairs from the sorting result to obtain a target prediction pair; The calling of the preset target twin network model, performing similarity calculation of the loss function and similarity matching of the department information on the target prediction pair based on the preset department information database, and obtaining the target result includes: calling the preset target twin network model, respectively calculating the loss function values ​​of the input vector in the target prediction pair and the multiple department prediction probabilities corresponding to the input vector, obtaining the input vector loss function value and the department probability loss function value, and calculating the similarity between the input vector loss function value and the department probability loss function value based on the preset department information database to obtain the prediction pair similarity value; sorting the target prediction pairs in descending order of the prediction pair similarity values ​​to obtain a prediction pair sequence, and determining the target prediction pair ranked first in the prediction pair sequence as the target result, and the target result is used to indicate the target department information.

2. The method for processing physical examination data according to claim 1, characterized in that: Before obtaining the initial physical examination data and preprocessing the initial physical examination data to obtain the preprocessed physical examination data, the method further includes: Obtain a model training vector, and train a preset initial deep language model and a preset initial twin network model based on the model training vector to obtain a target deep language model and a target twin network model.

3. The method for processing physical examination data according to claim 2, characterized in that: The acquiring model training vector, and training the preset initial deep language model and the preset initial twin network model based on the model training vector respectively to obtain the target deep language model and the target twin network model include: Obtain a model training vector, call a preset deep language model, perform department probability prediction on the model training vector, and obtain a plurality of training prediction pairs, wherein the plurality of training prediction pairs include the model training vector and a plurality of training prediction probabilities corresponding to the model training vector; Optimizing the preset deep language model through the multiple training prediction pairs to obtain a target deep language model; Calling a preset twin network model, performing feature extraction on the multiple training prediction pairs, and obtaining a training department label, wherein the training department label is used to indicate the department information corresponding to the model training vector; The preset twin network model is optimized through the training department label to obtain the target twin network model.

4. The method for processing physical examination data according to any one of claims 1 to 3, characterized in that: Before obtaining the initial physical examination data and preprocessing the initial physical examination data to obtain the preprocessed physical examination data, the method further includes: Acquire historical data, and perform desensitization processing on the historical data using a preset anonymization algorithm to obtain desensitized data, wherein the historical data includes historical physical examination data and department data corresponding to the historical physical examination data; The desensitized data is loaded into a preset database construction template to obtain a department information database.

5. A device for processing physical examination data, characterized in that: The physical examination data processing device comprises: An acquisition module, used for acquiring initial physical examination data, and preprocessing the initial physical examination data to obtain preprocessed physical examination data; A conversion module, used for performing vector conversion on the pre-processed physical examination data to obtain an input vector, wherein the input vector includes a plurality of word vectors, a plurality of segment vectors and a plurality of position vectors; A prediction module, used to call a preset target deep language model, perform department probability prediction on the input vector, obtain an initial prediction pair, sort the initial prediction pair to obtain a sorting result, and extract a target prediction pair according to the sorting result; A parsing module, used to call a preset target twin network model, perform similarity calculation of the loss function and similarity matching of the department information on the target prediction pair based on a preset department information database, and obtain a target result; The conversion module includes: an encoding unit, which is used to perform double-byte encoding on the pre-processed physical examination data based on a preset tag embedding layer to obtain a plurality of word vectors; a first embedding unit, which is used to perform embedding processing on the plurality of word vectors through a preset segment embedding layer to obtain a segment vector corresponding to each word vector; a second embedding unit, which is used to perform embedding processing on the plurality of word vectors through a preset position embedding layer to obtain a position vector corresponding to each word vector; a concatenation unit, which is used to concatenate the plurality of word vectors, the segment vector corresponding to each word vector, and the position vector corresponding to each word vector to obtain an input vector; The prediction module includes: a first prediction unit, used to call a preset target deep language model, predict the department probability corresponding to the input vector, and obtain an initial prediction pair, wherein the initial prediction pair includes the input vector and multiple department prediction probabilities corresponding to the input vector; a sorting unit, used to sort the initial prediction pair in descending order of the multiple department prediction probabilities, obtain a sorting result, extract a preset number of prediction pairs from the sorting result, and obtain a target prediction pair; The parsing module includes: a similarity calculation unit, which is used to call a preset target twin network model, and respectively calculate the loss function values ​​of the input vector in the target prediction pair and the multiple department prediction probabilities corresponding to the input vector to obtain the input vector loss function value and the department probability loss function value, and calculate the similarity between the input vector loss function value and the department probability loss function value based on a preset department information database to obtain a prediction pair similarity value; a determination unit, which is used to sort the target prediction pairs in order from large to small according to the prediction pair similarity values ​​to obtain a prediction pair sequence, and determine the target prediction pair ranked first in the prediction pair sequence as the target result, and the target result is used to indicate the target department information.

6. A device for processing physical examination data, characterized in that: The physical examination data processing device comprises: a memory and at least one processor, wherein the memory stores a computer program; The at least one processor calls the computer program in the memory to enable the physical examination data processing device to execute the physical examination data processing method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for processing physical examination data as described in any one of claims 1 to 4 is implemented.

Images