Medical record quality inspection method, device and storage medium

Abstract

A medical record quality inspection method, device, and storage medium, the method comprising: obtaining a medical record to be inspected, the medical record to be inspected including medical record content and diagnosis results, the diagnosis results including the name of at least one disease; extracting features from the medical record content to obtain a first feature; for each disease included in the diagnosis result, obtaining features of a typical medical record corresponding to the disease to obtain a second feature; for each disease included in the diagnosis result, performing similarity calculations on the second feature of the typical medical record corresponding to the disease and the first feature to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content. This application is based on the original diagnosis results in the medical record to be inspected, obtains typical medical records of the disease included in the original diagnosis results, compares them with the medical record content on the medical record to be inspected for similarity, and can automatically determine the occurrence of multiple diagnosis based on the comparison results.

Description

Technical Field

[0001] The present application relates to the field of intelligent medical technology, and more specifically to a medical record quality inspection method, device and storage medium. Background Art

[0002] DRGs (Diagnosis Related Groups) payment is a key component of medical insurance payment reform. The basic concept is to group illnesses into diagnosis-related groups based on the principal diagnosis, other diagnoses, major procedures, other procedures, and basic patient information on the first page of the inpatient medical record. The reimbursement amount is determined by multiplying the pre-determined rate by a preset weighting factor for each group.

[0003] The current mainstream approach to quality control of overwritten diagnoses relies on manual quality control. This is due to the inherent difficulty of this task, which requires a basic understanding of the symptoms, tests, medications, and procedures of various diseases. It also requires reviewing the entire inpatient medical record and verifying the preceding and following circumstances before a conclusion can be drawn. Therefore, current quality control methods typically rely on internal departmental inspections at the three levels of diagnosis and treatment: the attending physician, chief physician or deputy chief physician, and department director, who conduct checks at each level, or specialized inspections by quality control personnel from the hospital's quality inspection department.

[0004] The shortcomings of manual quality inspection are obvious. First, because the inpatient medical records contain a lot of content, it takes an average of 20 minutes to complete the quality inspection of the diagnosis of a medical record. Secondly, the quality inspectors need to have a basic understanding and knowledge of the diseases in each department. Because comorbidities and complications may be independent of the department where the patient is staying, the quality inspectors need to have a basic understanding of all diseases. Finally, manual quality inspection can easily lead to mutual cover-up. For example, if the hospital leads the quality inspection, it is very likely that doctors, departments, and quality inspectors will condone and cover up the behavior, which is not objective and true. Summary of the Invention

[0005] The present application is proposed to solve the above-mentioned problems. According to one aspect of the present application, a medical record quality inspection method is provided, the method comprising: obtaining a medical record to be inspected, the medical record to be inspected including medical record content and diagnosis results, the diagnosis results including the name of at least one disease; extracting features from the medical record content to obtain a first feature; for each disease included in the diagnosis result, obtaining features of a typical medical record corresponding to the disease to obtain a second feature; for each disease included in the diagnosis result, performing similarity calculation on the second feature of the typical medical record corresponding to the disease and the first feature to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content.

[0006] In one embodiment of the present application, obtaining the characteristics of the typical medical record corresponding to the disease to obtain the second feature includes: obtaining the characteristics of each of multiple typical medical records corresponding to the disease to obtain multiple second features; performing similarity calculation on the second feature of the typical medical record corresponding to the disease and the first feature to determine whether the disease is related to the medical record content, includes: performing similarity calculation on the first feature and each of the second features to obtain multiple similarity scores; calculating the average value of the multiple similarity scores and / or obtaining the maximum value of the multiple similarity scores; when the average value is greater than a first threshold value and / or the maximum value is greater than a second threshold value, determining that the disease is related to the medical record content.

[0007] In one embodiment of the present application, the method is performed based on a trained neural network, which is an end-to-end framework for joint training of self-supervised learning and supervised learning. The self-supervised learning includes feature learning of typical medical records, and the supervised learning includes feature similarity learning between typical medical records and atypical medical records.

[0008] In one embodiment of the present application, the neural network adopts a focal loss function to implement the joint training, wherein the focal loss function is designed based on the respective loss functions of the self-supervised learning and the supervised learning and the respective performance indicators of the two.

[0009] In one embodiment of the present application, the training set for self-supervised learning includes original samples, positive samples and negative samples, wherein: the original samples include typical medical record samples; the positive samples include replicated samples of the typical medical record samples; and the negative samples include randomly sampled samples of other typical medical record samples corresponding to different diseases from the typical medical record samples.

[0010] In one embodiment of the present application, the original sample and the positive sample constitute a positive correlation pair, and the original sample and the negative sample constitute a negative correlation pair; the loss function of the self-supervised learning is designed based on the following principle: making the representation distance of the positive correlation pair closer and closer, and the representation distance of the negative correlation pair farther and farther; the performance indicator of the self-supervised learning includes the result of subtracting the similarity score between the original sample and the positive sample from the similarity score between the original sample and the negative sample.

[0011] In one embodiment of the present application, the neural network is a neural network with a random dropout mechanism, and the typical medical record sample and the copy sample of the typical medical record sample are respectively input into the neural network, and based on the random dropout mechanism, the original sample and the positive sample are obtained.

[0012] In one embodiment of the present application, the atypical medical records correspond to multiple diseases, which serve as multiple labels. The supervised learning calculates feature similarity between the original samples of the typical medical records corresponding to each of the multiple diseases and the samples of the atypical medical records to obtain a prediction result of the disease type corresponding to the samples of the atypical medical records; the loss function of the supervised learning is designed based on the following principle: for each label in the multiple labels: when the prediction result is correct, a first loss function value is generated; when the prediction result is wrong, a second loss function value is generated, and all loss function values are weighted and summed to update the parameters of the neural network, wherein the first loss function value is less than the second loss function value.

[0013] In one embodiment of the present application, the neural network is a graph neural network, which constructs a graph for the medical record content based on the following method: preprocessing to obtain feature words of the medical record content, representing the feature words as nodes, representing the co-occurrence relationship between the feature words as edges, and calculating the edge weights through point mutual information to obtain a weighted graph of the medical record content.

[0014] According to another aspect of the present application, a medical record quality inspection method is provided, the method comprising: obtaining a medical record to be inspected, and inputting the medical record to be inspected into a trained neural network, wherein the medical record to be inspected includes medical record content and a diagnosis result, and the diagnosis result includes the name of at least one disease; the neural network is an end-to-end framework for joint training of self-supervised learning and supervised learning, the self-supervised learning includes feature learning of typical medical records, and the supervised learning includes feature similarity learning between typical medical records and atypical medical records; based on the neural network, features are extracted for the medical record content, and for each disease included in the diagnosis result, similarity is calculated between the features of the typical medical record corresponding to the disease and the features of the medical record content to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content.

[0015] In one embodiment of the present application, the neural network adopts a focal loss function to implement the joint training, wherein the focal loss function is designed based on the respective loss functions of the self-supervised learning and the supervised learning and the respective performance indicators of the two.

[0016] According to another aspect of the present application, a medical record quality inspection device is provided, which includes: a medical record acquisition module, used to acquire medical records to be inspected, wherein the medical records to be inspected include medical record content and diagnosis results, and the diagnosis results include the name of at least one disease; a feature extraction module, used to extract features for the medical record content to obtain a first feature; a similarity calculation module, used to obtain, for each disease included in the diagnosis result, features of typical medical records corresponding to the disease to obtain a second feature; and for each disease included in the diagnosis result, perform similarity calculation on the second feature of the typical medical record corresponding to the disease and the first feature to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content.

[0017] According to another aspect of the present application, a medical record quality inspection device is provided, which includes: an input module for obtaining medical records to be inspected and inputting the medical records to be inspected into a trained neural network, wherein the medical records to be inspected include medical record content and diagnosis results, and the diagnosis results include the name of at least one disease; the neural network is an end-to-end framework for joint training of self-supervised learning and supervised learning, the self-supervised learning includes feature learning of typical medical records, and the supervised learning includes feature similarity learning between typical medical records and atypical medical records; an output module for extracting features for the medical record content based on the neural network, and for each disease included in the diagnosis result, performing similarity calculation between the features of the typical medical record corresponding to the disease and the features of the medical record content, so as to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content.

[0018] According to another aspect of the present application, a medical record quality inspection device is provided, which includes a memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the processor executes the above-mentioned medical record quality inspection method.

[0019] According to another aspect of the present application, a storage medium is provided, on which a computer program is stored. When the computer program is run, the medical record quality inspection method described above is executed.

[0020] According to the medical record quality inspection method and device of the embodiment of the present application, the original diagnosis results in the medical records to be inspected are used as the basis, typical medical records of the diseases included in the original diagnosis results are obtained, and similarity comparison is performed with the medical record content on the medical records to be inspected. According to the comparison results, it can be determined whether the diseases included in the original diagnosis results are related to the medical record content, so that the diagnosis duplication situation can be automatically judged. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0022] Figure 1 A schematic flowchart of a medical record quality inspection method according to an embodiment of the present application is shown.

[0023] Figure 2 A schematic diagram of the training process of the neural network used in the medical record quality inspection method according to an embodiment of the present application is shown.

[0024] Figure 3 A schematic diagram showing the results obtained after quality inspection of medical records using the medical record quality inspection method according to an embodiment of the present application is shown.

[0025] Figure 4 A schematic flowchart of a medical record quality inspection method according to another embodiment of the present application is shown.

[0026] Figure 5 A schematic structural block diagram of a medical record quality inspection device according to an embodiment of the present application is shown.

[0027] Figure 6 A schematic structural block diagram of a medical record quality inspection device according to another embodiment of the present application is shown.

[0028] Figure 7 A schematic structural block diagram of a medical record quality inspection device according to another embodiment of the present application is shown. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present application more apparent, example embodiments according to the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments described herein. Based on the embodiments of the present application described in this application, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of this application.

[0030] First, refer to Figure 1 A medical record quality inspection method according to one embodiment of the present application is described. Figure 1 FIG. 1 shows a schematic flow chart of a medical record quality inspection method 100 according to an embodiment of the present application. Figure 1 As shown, the medical record quality inspection method 100 may include the following steps:

[0031] In step S110 , the medical records to be examined are obtained. The medical records to be examined include medical record content and diagnosis results. The diagnosis results include the name of at least one disease.

[0032] In step S120 , features are extracted from the medical record content to obtain a first feature.

[0033] In step S130 , for each disease included in the diagnosis result, the characteristics of the typical medical records corresponding to the disease are obtained to obtain a second characteristic.

[0034] In step S140, for each disease included in the diagnosis result, the second feature of the typical medical record corresponding to the disease is calculated with the first feature for similarity to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content.

[0035] In the embodiment of the present application, the medical records to be examined include medical record content and diagnosis results. Among them, the medical record content includes information such as patient gender, age, chief complaint information, current medical history, physical examination information, auxiliary examination information, etc. (such as later combined Figure 3 The diagnosis result is the original diagnosis made by the doctor based on the medical record content, including the diagnosis of a certain disease or multiple diseases by the medical record content (such as Figure 3The information of the original diagnosis part in the described example). For such medical records to be inspected, in an embodiment of the present application, quality inspection is carried out based on the diagnosis results in the medical records. Since the diagnosis result includes the name of at least one disease, the characteristics of the typical medical record of each disease can be obtained based on this (or the typical medical record of each disease can be obtained and then the characteristics are extracted), and the characteristics of the typical medical record of each disease are compared with the characteristics of the medical record content part in the current medical record to be inspected in turn for similarity, and then it can be determined based on the similarity comparison results whether the medical record content in the current medical record to be inspected is consistent with the typical medical record content corresponding to each disease in the diagnosis result. For example, for a disease in the diagnosis result, when the characteristics of the typical medical record of the disease are highly similar to the characteristics of the medical record content in the current medical record to be inspected, it can be considered that the medical record content in the current medical record to be inspected is indeed very likely to be caused by the disease, that is, the disease given in the diagnosis result is the correct diagnosis, not a miswritten or overwritten diagnosis. Conversely, for a disease in the diagnosis result, if the characteristics of the typical medical record of that disease are less similar to the characteristics of the medical record content in the current medical record to be tested, it can be considered that the medical record content in the current medical record to be tested is unlikely to be caused by that disease, that is, the disease given in the diagnosis result is a miswritten or overwritten diagnosis, which can be considered as an overwritten diagnosis. Therefore, after the above quality inspection process, it can be determined whether there is an overwritten diagnosis in the current medical record to be tested. If so, a quality inspection result can be given (for example, a certain disease in the original diagnosis is an overwritten diagnosis) and a diagnosis recommendation result can also be given (for example, removing the overwritten disease and retaining other diagnosis results).

[0036] Therefore, in general, the medical record quality inspection method according to the embodiment of the present application is based on the original diagnosis results in the medical records to be inspected, obtains typical medical records of the diseases included in the original diagnosis results, and compares them with the medical record contents on the medical records to be inspected for similarity. According to the comparison results, it can be determined whether the diseases included in the original diagnosis results are related to the medical record contents, so that the situation of multiple diagnosis can be automatically judged. It has the advantages of high rationality and fast inspection speed, can solve the current situation of insufficient quality inspection personnel and slow inspection speed, and carry out large-scale and efficient quality inspection. Whether it is for quality inspection leading departments such as the Medical Insurance Bureau or the Health Commission, or for quality inspection departments such as hospitals, there is a great demand for it, which can bring about a significant efficiency improvement in their work.

[0037] In an embodiment of the present application, the above-mentioned method 100 can be performed based on a trained neural network, wherein the neural network can be an end-to-end framework for joint training of self-supervised learning and supervised learning. Self-supervised learning includes feature learning of typical medical records, and supervised learning includes feature similarity learning between typical medical records and atypical medical records. That is to say, in an embodiment of the present application, an end-to-end framework is adopted to jointly train self-supervised learning and supervised learning. For the feature learning of typical medical records, self-supervised comparison technology is adopted to better grasp and characterize the characteristics of medical records; for the correlation between the medical records to be tested and the typical medical records, supervised multi-label training is adopted. The two training methods are jointly carried out and optimized, so that the model has better representation ability and performance. The self-supervised learning and supervised learning of the neural network are described below.

[0038] In an embodiment of the present application, the self-supervised learning task aims to obtain a better feature representation for typical medical records, so that the features of each typical medical record can well represent the disease corresponding to the typical medical record, so that the effect of similarity calculation between the medical record to be tested and the typical medical record can be more accurate, and the disease of the typical medical record with high similarity is used as the disease prediction result of the medical record to be tested, which will be more credible. In addition, the self-supervised training method can also improve the model's ability to model medical record text, so that the model as a whole can have a stronger performance in characterizing medical record text representation. Finally, the self-supervised training method can automatically generate positive and negative samples, without the need for a large amount of manual labeling, saving human resources. The following describes the sample construction, model design, and loss function construction in self-supervised learning.

[0039] First, a training set for self-supervised contrastive learning should be constructed based on typical medical records. Contrastive learning requires constructing positive and negative samples for a medical record to form a complete sample that can be used for contrastive learning. In an embodiment of the present application, the training set for self-supervised learning includes original samples, positive samples, and negative samples, where: original samples include typical medical record samples; positive samples include replicated samples of typical medical record samples; negative samples include randomly sampled samples of other typical medical record samples corresponding to different diseases than the typical medical record samples. This method can automatically generate positive and negative samples without the need for a large amount of manual labeling, saving human resources.

[0040] In an embodiment of the present application, the neural network to be trained can be a neural network with a random dropout mechanism; based on this, typical medical record samples and replicated samples of typical medical record samples can be input into the neural network respectively, and based on the random dropout mechanism of the neural network, the original samples and positive samples can be obtained.

[0041] That is, to construct a positive sample of a typical medical record, we can directly copy the original medical record and then pass it through a network model with a dropout mechanism. Dropout refers to the method of temporarily discarding a portion of neural network units from the network with a certain probability during the training process of a deep learning network. This is equivalent to finding a leaner network from the original network and training the network. It is mainly used to prevent network overfitting. Here, we feed a typical medical record and its copy into a network structure with a dropout mechanism. Since each dropout randomly discards a small number of neurons, after passing through the network, the vector representation will produce slightly inconsistent results due to the different discarded neurons. Such results can be used as one original sample and one as a positive sample of the original sample.

[0042] For the construction of negative samples of typical medical records, we can randomly select one typical medical record of other diseases with different diagnoses from the typical medical records according to the task orientation as a negative sample of the typical medical record, and then pass it through the model network (that is, the network model with dropout mechanism mentioned above), and use the obtained representation as the negative sample result.

[0043] Based on this, corresponding positive and negative samples can be automatically and quickly constructed for all existing typical medical records for comparative learning and training.

[0044] In an embodiment of the present application, the neural network (network model) used can be based on a graph neural network structure, and the graph neural network can construct a graph for a typical medical record in the following manner: pre-process the characteristic words of the typical medical record, represent the characteristic words as nodes (the embedding of the node is initialized with the word feature), represent the co-occurrence relationship between the characteristic words as edges (the co-occurrence relationship between words is determined by a sliding window), and calculate the edge weights through point mutual information (PMI), thereby obtaining a weighted graph of the typical medical record. Then, a graph neural network (such as a gated graph neural network) is used to learn the embedding of word nodes. A node can receive information from its adjacent neighbors and then merge it with its own representation for updating. When the layer runs on the first-order neighbors, t layers can be stacked t times to achieve high-order feature interaction, where one node can reach another node at a distance of t. Here we set t = 2, that is, each node can transmit information to its second-order neighbors at the farthest. For the self-supervised learning task of typical medical records, each sample, including the original sample, positive sample, and negative sample of the typical medical record, can be constructed as a graph respectively, and then graph convolution calculation (such as two-layer graph convolution calculation) is performed on each graph to obtain the feature representation of the original sample, positive sample, and negative sample of the typical medical record respectively.

[0045] In the embodiment of the present application, the original sample of a typical medical record and the positive sample constitute a positive correlation pair, and the original sample and the negative sample constitute a negative correlation pair; the design of the loss function needs to constrain the correlation of the medical record representation pairs, so that the representation distance of the positive correlation representation pairs is getting closer and closer, and the representation distance of the negative correlation representation pairs is getting farther and farther. Based on this, in one example, the following loss function is designed:

[0046]

[0047] in represents the similarity between the pair of sentence representations, for example, it can be calculated using the cosine similarity formula. г represents a hyperparameter, indicating the numerical scaling ratio. This loss function ensures that the more correlated the vector representation pair, the smaller the loss function; the less correlated the vector representation pair, the larger the loss function, which meets our model training objectives.

[0048] The above describes self-supervised learning. Based on the above training, the neural network can well extract the features of the medical record to be examined for similarity comparison with the features of typical medical records, as will be described below.

[0049] In an embodiment of the present application, for the prediction of the disease list of the medical records to be checked, the method we adopt is to compare the feature similarity of the medical records to be checked with all the typical medical records corresponding to the disease. In an example, multiple typical medical records can be obtained for each disease in the diagnosis result, and then for each disease in the diagnosis result, the average and / or maximum value of the similarity score between the features of the medical records to be checked and all the typical medical records of the disease are calculated. If the score is higher, it means that the medical record to be checked is more similar to the typical medical record corresponding to the disease, and the more inclined to the disease corresponding to the typical medical record group. By formulating a threshold in advance, the correlation can be flexibly controlled, and diseases above the average threshold (which can be called the first threshold, using the aforementioned similarity score average value for comparison) or the maximum threshold (which can be called the second threshold, using the aforementioned similarity score maximum value for comparison) are all regarded as possible diseases of the medical record to be checked, thereby forming a predicted disease list.

[0050] That is, for the step 130 of the method 100 described above, obtaining the characteristics of the typical medical record corresponding to the disease to obtain the second characteristic, it may include: obtaining the characteristics of each of the multiple typical medical records corresponding to the disease to obtain multiple second characteristics. Accordingly, for the step 140 of the method 100 described above, performing similarity calculation between the second characteristic of the typical medical record corresponding to the disease and the first characteristic to determine whether the disease is related to the medical record content, it may include: performing similarity calculation between the first characteristic and each second characteristic to obtain multiple similarity scores; calculating the average value of the multiple similarity scores and / or obtaining the maximum value of the multiple similarity scores; when the average value is greater than the first threshold value and / or the maximum value is greater than the second threshold value, determining that the disease is related to the medical record content.

[0051] As described above, in an embodiment of the present application, the neural network to be trained can be an end-to-end framework for joint training of self-supervised learning and supervised learning. Self-supervised learning includes feature learning of typical medical records, and supervised learning includes feature similarity learning between typical medical records and atypical medical records. The medical records to be tested are atypical medical records. During the training phase, the graph neural network described above can be used to construct the atypical medical records and then pass them through a graph convolutional network (such as a two-layer graph neural network) to obtain the feature representation of the atypical medical records. Afterwards, for example, the cosine similarity method can be used to calculate the similarity scores of the atypical medical record and all the typical medical records corresponding to its diagnosis list.

[0052] Among them, similarity calculation learning is performed based on supervised training. Specifically, the similarity between atypical medical records and typical medical records is calculated, and finally the diagnosis corresponding to the typical medical record with high similarity is used as the predicted diagnosis of the atypical medical record. Therefore, in essence, it is still to judge whether the disease list (diagnosis result) of the atypical medical record can be correctly predicted, so it can be regarded as a multi-label classification task (each disease in the disease list in the atypical medical record is a label, generally typical medical records correspond to one disease, and atypical medical records usually correspond to multiple diseases, of course, it does not rule out the case of corresponding to a single disease). When calculating single-label classification, the cross-entropy loss function can be used. Therefore, for multi-label classification, the cross-entropy loss function of single-label classification can be expanded, and the calculation method is as follows:

[0053]

[0054] Where m represents the total number of diseases, represents the true value of the i-th sample under the j-th disease, represents the output of the jth category of the i-th sample after being processed by the softmax function. This formula represents the calculation of the predicted result and all labels to generate the corresponding loss function value, which is then weighted. For each label, if the prediction is correct, the loss is small; if it is not, the loss is large. The results are then summed together. This formula constrains the model parameters as a whole, ensuring that the model trains towards all correct labels, continuously updating the parameters to achieve the optimal model results.

[0055] That is, in an embodiment of the present application, atypical medical records correspond to multiple diseases, which serve as multiple labels. Supervised learning calculates feature similarity between the original samples of typical medical records corresponding to each of the multiple diseases and the samples of atypical medical records, thereby obtaining a prediction result of the disease type corresponding to the sample of atypical medical records. The loss function of supervised learning is designed based on the following principle: for each of the multiple labels: when the prediction result is correct, a first loss function value is generated; when the prediction result is incorrect, a second loss function value is generated. All loss function values are weighted and summed to update the parameters of the neural network, where the first loss function value is less than the second loss function value.

[0056] We now describe the joint training of these two tasks, self-supervised learning and supervised learning.

[0057] Automatic multi-task training relies primarily on the design of loss functions. Traditional multi-task joint loss functions typically pre-set weights for each task's loss function. While this approach can address the issue of the loss functions of two tasks being on different orders of magnitude, it cannot resolve the issue of inconsistent optimization speeds across tasks. For example, if the loss functions of two tasks have very different value ranges—one between 0 and 1 and the other between 0 and 1000—pre-setting weights may superficially bring the two loss functions into line, but it fails to account for the varying optimization speeds of the different tasks. For example, suppose the loss of one task quickly decreases to between 0 and 0.2, while the loss of the other task requires many iterations to reduce to between 0.5 and 0.9. Despite the pre-set weights, the task with the larger loss function will continue to dominate model training. This means the model will continually update parameters for that task, while the other task remains untrained.

[0058] Based on this, in an embodiment of the present application, a focal loss function is used to implement joint training, wherein focal loss is designed based on the respective loss functions of self-supervised learning and supervised learning and the respective performance indicators of the two. For supervised learning tasks, the overlap rate of the predicted results of the disease types corresponding to the samples of atypical medical records and the disease labels in their diagnosis lists can be used as performance indicators. For self-supervised learning tasks, the average similarity between the original samples of typical medical records and positive samples minus the average similarity between the original samples of typical medical records and negative samples can be used as performance indicators, because if the original samples of typical medical records are more correlated with the positive samples, the more unrelated the original samples of typical medical records are with the negative samples, the greater the difference in the average values between the two will be, which is in line with our expectations. After obtaining the loss function values and performance indicator values of the two tasks respectively, the focal loss can be calculated, for example, by the following formula:

[0059] FL(k i )=-((1-k i ) gamma ×log(k i ))

[0060] loss = sum(FL(k i )×loss i )

[0061] where k i represents the performance value of task i, gamma represents the hyperparameter, for example, if it is set to 2, loss i Represents the loss function value of task i. This formula cleverly utilizes the relationship between loss function and performance, that is, if the performance of the task is good, the loss function will be smaller, and if the performance of the task is poor, the loss function will be larger. By dynamically evaluating the weight of a loss function based on the task performance, and then multiplying it by the corresponding loss function, the multi-task weighting is finally used as the final loss function to dynamically quantify the optimization direction of the two tasks. If the performance of task i is better, the proportion of its loss function value in the final loss will be smaller, so that the parameters are not optimized in this direction as much as possible; if the performance of task i is worse, the proportion of its loss function in the final loss function will be larger, so that the parameters are optimized in this direction.

[0062] Through focal loss, we achieve the goal of using a single model (neural network) to simultaneously support self-supervised comparative learning of typical medical records and supervised learning of typical and atypical medical records. By simultaneously estimating the performance and loss function of each task, we dynamically determine which task to prioritize in each iteration, enabling automatic and synchronous training of both tasks to find the globally optimal model parameters.

[0063] The above describes the training process of the neural network in this application, which can be combined with Figure 2 To better understand the above.

[0064] Based on the end-to-end framework of simultaneous training of the self-supervised contrastive learning and supervised multi-label classification model mentioned above, it is not necessary to fully execute these two goals when training for actual use. The best model results in the training phase can be saved, and the representations of typical medical records of each disease under the best model parameters can be extracted and saved. Then, for the medical records to be tested, it is only necessary to pass it through the graph network model, and then extract the representation (i.e., the first feature described in the above method 100) and compare it with the representation of the typical medical record (i.e., the second feature described in the above method 100). This method of exchanging space for time can make the actual application faster without losing any accuracy. For each disease in the diagnosis results of the medical records to be tested, after calculating the similarity results between the medical records to be tested and all typical medical records of the disease, the average value and the maximum value are calculated, and they are compared with the average value threshold and the maximum value threshold respectively. If one of them is greater than the threshold, it is considered that the disease corresponding to the typical medical record should be added to the disease list of the diagnosis results of the medical records to be tested; if both indicators are not greater than the corresponding thresholds, it is considered that the disease corresponding to this group of typical medical records should not be added to the disease list of the diagnosis results, which is a case of multiple diagnosis.

[0065] Figure 3 The following is a schematic diagram showing the results of quality inspection of medical records using the medical record quality inspection method according to an embodiment of the present application. Figure 3 As shown in the figure, an example of quality inspection for over-diagnosis is provided. We input the medical records to be inspected into the model and extract typical medical records for four diagnoses given by doctors (hemorrhagic anemia, gastric ulcer, gastrointestinal bleeding, and duodenitis). We then calculated the similarity between the medical records to be inspected and the typical medical records for these diseases. We found that the medical records to be inspected had high similarity with the typical medical records for hemorrhagic anemia, gastric ulcer, and gastrointestinal bleeding, but when compared with the typical medical records for duodenitis, the similarity was very low. Based on this, we believe that the diagnosis of duodenitis is an over-diagnosis case with quality issues (this result has been verified by doctors) and provide quality inspection results and diagnosis recommendations.

[0066] Based on the above description, the medical record quality inspection method according to the embodiment of the present application takes the original diagnosis results in the medical records to be inspected as the basis, obtains typical medical records of the diseases included in the original diagnosis results, and compares them with the medical record contents on the medical records to be inspected for similarity. According to the comparison results, it can be determined whether the diseases included in the original diagnosis results are related to the medical record contents, thereby automatically judging the situation of multiple diagnosis.

[0067] Furthermore, the medical record quality inspection method according to the embodiment of this application proposes an end-to-end computing framework for automatic quality inspection of multiple diagnostic write-ups. This framework utilizes a method of comparing the medical record to be inspected with the typical medical records corresponding to the original diagnosis, and through comprehensive evaluation, achieves automatic quality inspection of multiple diagnostic write-ups, which can effectively replace manual automatic quality inspection of multiple diagnostic write-ups. Because it is essentially based on the similarity calculation of multiple typical medical records, it will not produce outrageous errors overall. All prediction results are supported by typical medical records as evidence, which has a certain degree of explainability and performance assurance.

[0068] In addition, the medical record quality inspection method according to the embodiment of the present application proposes to jointly train self-supervised comparative learning of typical medical records with supervised learning of atypical and typical medical records. This enables the model to better capture the semantic features in medical records and better characterize the characteristics of medical records, thereby achieving better performance results. Compared with the two-stage implementation (non-joint training) method, it has the advantages of avoiding error transmission, better performance, and faster speed.

[0069] In addition, according to the medical record quality inspection method of the embodiment of the present application, it is proposed to use the dropout method to generate a representation of positive samples of typical medical records. This method uses the randomness of the dropout mechanism to cleverly generate positive samples that are highly similar to the original medical records. It is simple and effective, and very cleverly avoids the problems of inappropriateness, inappropriateness, and poor results faced in generating positive samples. The natural randomness of dropout and the adjustability of the proportion make the generated positive samples more in line with task requirements and bring better performance.

[0070] In addition, according to the medical record quality inspection method of the embodiment of the present application, it is proposed to use the focal-loss mechanism to jointly train self-supervised and supervised tasks, so that when the self-supervised and supervised tasks are trained, the training goals and directions between the tasks can be automatically balanced, and each time the task with worse performance is automatically optimized, and it is gradually iterated to finally obtain the optimal solution for both tasks instead of the optimal solution for a single task, and it has a good automatic correction capability. That is, the mechanism combines the performance and loss of each task, and guides the value of the loss function through performance, so that the task with worse performance automatically and dynamically obtains a larger loss function, and obtains more targeted training, avoiding the problems of inconsistent multi-task training speeds and inconsistent magnitudes of multi-task loss functions that weighted loss cannot cope with, and can dynamically adjust the direction of each training according to the performance of different tasks, supporting an end-to-end technical framework.

[0071] In addition, according to the medical record quality inspection method of the embodiment of the present application, during the supervised training process, the disease list in the diagnosis result given by the doctor is used as the real label for multi-label classification, which can avoid the problem of selecting the number of labels faced by multi-label classification; in the application quality inspection stage, the typical medical record content corresponding to the disease list in the diagnosis result of the medical record to be inspected is compared with the text content of the medical record to be inspected for similarity calculation, and there is no need to calculate the loss function or performance value, so it will not be troubled by the inability to determine the number of labels. In some existing multi-label classification technologies, for each medical record sample, the type of disease corresponding to the medical record is directly predicted without referring to the diagnosis result, that is, there is a problem of uncertain number of classifications, because some samples may have only one category label (disease), and some samples may have as many as a dozen category labels. How to automatically determine it is a difficulty; in addition, the mutual dependence between category labels is difficult to learn. Taking the relationship between diseases as an example, if hypertension develops seriously, it will cause complications in multiple parts such as the eyes. Therefore, how to solve the dependency between category labels is also a difficulty. However, the solution of the present application does not have such a problem, because the solution of the present application uses the disease list in the diagnosis results given by the doctor as the true labels of the multi-label classification, which can avoid the problem of selecting the number of labels faced by the multi-label classification; in addition, the solution of the present application uses the disease list in the diagnosis results given by the doctor as the true labels of the multi-label classification, and the labels have been determined, and there is no need to learn the dependencies between the labels.

[0072] The above is an exemplary description of a medical record quality inspection method according to an embodiment of the present application. Figure 4 A medical record quality inspection method according to another embodiment of the present application is described. Figure 4 FIG. 4 shows a schematic flow chart of a medical record quality inspection method 400 according to another embodiment of the present application. Figure 4 As shown, the medical record quality inspection method 400 may include the following steps:

[0073] In step S410, the medical records to be examined are obtained and input into the trained neural network, wherein the medical records to be examined include medical record content and diagnosis results, and the diagnosis results include the name of at least one disease; the neural network is an end-to-end framework for joint training of self-supervised learning and supervised learning, wherein self-supervised learning includes feature learning of typical medical records, and supervised learning includes feature similarity learning between typical medical records and atypical medical records.

[0074] In step S420, features are extracted from the medical record content based on a neural network, and for each disease included in the diagnosis result, similarity is calculated between the features of the typical medical record corresponding to the disease and the features of the medical record content to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content.

[0075] The medical record quality inspection method 400 according to an embodiment of the present application is generally similar to the medical record quality inspection method 100 according to the embodiment of the present application described above, except that the steps in the method are executed by the same neural network. The medical record quality inspection method 100 described above does not necessarily need to be the same. As long as the steps of method 100 can be implemented, it is not limited to whether a neural network is used or how many neural networks are used. Those skilled in the art can understand the specific operations and details of the medical record quality inspection method 400 in combination with the above description. For the sake of brevity, they will not be repeated here.

[0076] The following combination Figures 5 to 7 According to another aspect of the present application, medical record quality inspection devices 500, 600, and 700 are provided. Specifically, medical record quality inspection device 500 can be used to perform the medical record quality inspection method 100 described above; medical record quality inspection device 600 can be used to perform the medical record quality inspection method 400 described above; and medical record quality inspection device 700 can be used to perform either medical record quality inspection method 100 or 400 described above. Each of these devices will be briefly described below.

[0077] Figure 5 FIG. 5 shows a schematic structural block diagram of a medical record quality inspection device 500 according to an embodiment of the present application. Figure 5 As shown, the medical record quality inspection device 500 includes a medical record acquisition module 510, a feature extraction module 520, and a similarity calculation module 530. Among them, the medical record acquisition module 510 is used to obtain the medical record to be inspected, and the medical record to be inspected includes medical record content and diagnosis results, and the diagnosis results include the name of at least one disease; the feature extraction module 520 is used to extract features for the medical record content to obtain a first feature; the similarity calculation module 530 is used to obtain the features of the typical medical record corresponding to each disease included in the diagnosis result to obtain a second feature; and for each disease included in the diagnosis result, the second feature of the typical medical record corresponding to the disease is calculated with the first feature for similarity to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content. The medical record quality inspection device 500 according to the embodiment of the present application can be used to execute the medical record quality inspection method 100 described above. Those skilled in the art can understand its structure and operation in combination with the above description. For the sake of brevity, it will not be repeated here.

[0078] Figure 6 FIG. 6 is a schematic structural block diagram of a medical record quality inspection device 600 according to another embodiment of the present application. Figure 6As shown, the medical record quality inspection device 600 includes an input module 610 and an output module 620. The input module 610 is used to obtain the medical records to be inspected and input the medical records to be inspected into a trained neural network. The medical records to be inspected include medical record content and diagnosis results, and the diagnosis results include the name of at least one disease. The neural network is an end-to-end framework for joint training of self-supervised learning and supervised learning. Self-supervised learning includes feature learning of typical medical records, and supervised learning includes feature similarity learning between typical and atypical medical records. The output module 620 is used to extract features from the medical record content based on the neural network, and for each disease included in the diagnosis result, calculate the similarity between the features of the typical medical record corresponding to the disease and the features of the medical record content to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content. The medical record quality inspection device 600 according to the embodiment of the present application can be used to execute the medical record quality inspection method 400 described above. Those skilled in the art can understand its structure and operation in combination with the above description. For the sake of brevity, it will not be repeated here.

[0079] Figure 7 FIG. 1 shows a schematic structural block diagram of a medical record quality inspection device according to another embodiment of the present application. Figure 7 As shown, the medical record quality inspection device 700 may include a memory 710 and a processor 720. The memory 710 stores a computer program executed by the processor 720. When the computer program is executed by the processor 720, the processor 720 executes the medical record quality inspection method 100 or 400 according to the embodiment of the present application described above. Those skilled in the art can understand the specific operation of the medical record quality inspection device 700 according to the embodiment of the present application in combination with the above content. For the sake of brevity, the specific details are not repeated here.

[0080] In addition, according to an embodiment of the present application, a storage medium is further provided, on which program instructions are stored, and when the program instructions are run by a computer or a processor, the corresponding steps of the medical record quality inspection method of the embodiment of the present application are executed. The storage medium may include, for example, a memory card of a smart phone, a storage component of a tablet computer, a hard disk of a personal computer, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a portable compact disk read-only memory (CD-ROM), a USB memory, or any combination of the above storage media. The computer-readable storage medium may be any combination of one or more computer-readable storage media.

[0081] Based on the above description, the medical record quality inspection method and device according to the embodiment of the present application are based on the original diagnosis results in the medical record to be inspected, obtain typical medical records of the diseases included in the original diagnosis results, and compare them with the medical record contents on the medical record to be inspected for similarity. According to the comparison results, it can be determined whether the diseases included in the original diagnosis results are related to the medical record contents, so that the diagnosis duplication situation can be automatically judged.

[0082] Although example embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above example embodiments are merely illustrative and are not intended to limit the scope of the present application. Various changes and modifications may be made therein by those skilled in the art without departing from the scope and spirit of the present application. All such changes and modifications are intended to be included within the scope of the present application as required by the appended claims.

[0083] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0084] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units described is merely a logical function division. In actual implementation, other division methods may be used, such as combining or integrating multiple units or components into another device, or ignoring or not performing some features.

[0085] In the description provided herein, a large number of specific details are described. However, it is understood that the embodiments of the present application can be practiced without these specific details. In some instances, well-known methods, structures, and techniques are not shown in detail so as not to obscure the understanding of this description.

[0086] Similarly, it should be understood that in order to streamline the present application and aid in understanding one or more of the various inventive aspects, in the description of the exemplary embodiments of the present application, the various features of the present application are sometimes grouped together into a single embodiment, figure, or description thereof. However, this approach of the present application should not be interpreted as reflecting the intention that the application claimed for protection requires more features than those explicitly recited in each claim. More precisely, as reflected in the corresponding claims, the inventive point is that the corresponding technical problem can be solved with fewer features than all the features of a single disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into the detailed description, with each claim itself serving as a separate embodiment of the present application.

[0087] It will be understood by those skilled in the art that, except where mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or apparatus disclosed herein may be combined in any combination. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature providing the same, equivalent, or similar purpose.

[0088] Furthermore, those skilled in the art will appreciate that although some embodiments described herein include certain features included in other embodiments but not other features, combinations of features from different embodiments are intended to be within the scope of this application and to form different embodiments. For example, in the claims, any of the claimed embodiments may be used in any combination.

[0089] The various component embodiments of the present application can be implemented in hardware, or in a software module running on one or more processors, or in a combination thereof. Those skilled in the art will appreciate that a microprocessor or digital signal processor (DSP) can be used in practice to implement some or all of the functions of some modules according to the embodiments of the present application. The application can also be implemented as a part or all of a device program (e.g., a computer program and a computer program product) for performing the method described herein. Such a program implementing the present application can be stored on a computer-readable medium, or can have the form of one or more signals. Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0090] It should be noted that the above embodiments illustrate rather than limit the present application, and that those skilled in the art may devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbols placed between brackets should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present application may be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0091] The above description is merely a specific embodiment or illustration of a specific embodiment of the present application, and the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. The scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A medical record quality inspection method, characterized in that: The method combines self-supervised comparative learning of typical medical records with supervised learning of atypical and typical medical records, and adopts a focal loss function in the joint training process. For supervised learning tasks, the focal loss function uses the overlap rate between the predicted disease type corresponding to the sample of the atypical medical record and the disease label in the diagnosis list as a performance indicator. For self-supervised learning tasks, the average similarity between the original sample of the typical medical record and the positive sample minus the average similarity between the original sample of the typical medical record and the negative sample is used as a performance indicator. The training set of the self-supervised learning includes original samples, positive samples and negative samples, where: Inputting the typical medical record sample and the replicated sample of the typical medical record sample into a neural network with a random dropout mechanism respectively to obtain the original sample and the positive sample; Inputting randomly sampled samples of other typical medical record samples corresponding to different diseases from the typical medical record samples into the neural network with a random discard mechanism to obtain negative samples; The method further comprises: Obtaining a medical record to be examined, wherein the medical record to be examined includes medical record content and a diagnosis result, wherein the diagnosis result includes the name of at least one disease; Extracting features from the medical record content to obtain a first feature; For each disease included in the diagnosis result, obtaining a feature of a typical medical record corresponding to the disease to obtain a second feature; For each disease included in the diagnosis result, the similarity between the second feature of the typical medical record corresponding to the disease and the first feature is calculated to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content.

2. The method according to claim 1, characterized in that The step of obtaining the characteristic of the typical medical record corresponding to the disease to obtain the second characteristic includes: obtaining the characteristic of each of a plurality of typical medical records corresponding to the disease to obtain a plurality of second characteristics; The calculating the similarity between the second feature and the first feature of the typical medical record corresponding to the disease to determine whether the disease is related to the medical record content includes: calculating the similarity between the first feature and each of the second features to obtain multiple similarity scores; calculating the average of the multiple similarity scores and / or obtaining the maximum value of the multiple similarity scores; when the average value is greater than a first threshold and / or the maximum value is greater than a second threshold, determining that the disease is related to the medical record content.

3. The method according to claim 1, characterized in that The method is performed based on a trained neural network, which is an end-to-end framework for joint training of self-supervised learning and supervised learning. The self-supervised learning includes feature learning of typical medical records, and the supervised learning includes feature similarity learning between typical and atypical medical records.

4. The method according to claim 1, wherein The original sample and the positive sample constitute a positively correlated representation pair, and the original sample and the negative sample constitute a negatively correlated representation pair; the loss function of the self-supervised learning is designed based on the following principle: the representation distance of the positively correlated representation pair is increasingly closer, and the representation distance of the negatively correlated representation pair is increasingly farther; The performance indicator of the self-supervised learning includes a result of subtracting a similarity score between the original sample and the positive sample from a similarity score between the original sample and the negative sample.

5. The method according to claim 1, wherein The atypical medical records correspond to multiple diseases, which serve as multiple labels. The supervised learning calculates feature similarity between the original samples of the typical medical records corresponding to each of the multiple diseases and the samples of the atypical medical records to obtain a prediction result of the disease type corresponding to the samples of the atypical medical records; The loss function of the supervised learning is designed based on the following principle: for each of the multiple labels: when the prediction result is correct, a first loss function value is generated; when the prediction result is wrong, a second loss function value is generated, and all loss function values are weighted and summed to update the parameters of the neural network, wherein the first loss function value is less than the second loss function value.

6. The method according to claim 2, characterized in that The neural network is a graph neural network, which constructs a graph for the medical record content based on the following method: preprocessing to obtain feature words of the medical record content, representing the feature words as nodes, representing the co-occurrence relationship between the feature words as edges, and calculating the edge weights through point mutual information to obtain a weighted graph of the medical record content.

7. A medical record quality inspection method, characterized in that: The method comprises: Obtain a medical record to be examined and input the record into a trained neural network, wherein the medical record to be examined includes the medical record content and the diagnosis result, and the diagnosis result includes the name of at least one disease; the neural network is an end-to-end framework for joint training of self-supervised learning and supervised learning, wherein the self-supervised learning includes feature learning of typical medical records, and the supervised learning includes feature similarity learning between typical and atypical medical records; wherein the neural network is trained using the following method: The self-supervised comparative learning of typical medical records and the supervised learning of atypical and typical medical records are jointly trained, and a focal loss function is used in the joint training process; the focal loss function, for the supervised learning task, uses the overlap rate between the predicted disease type corresponding to the sample of the atypical medical record and the disease label in the diagnosis list as the performance indicator; for the self-supervised learning task, the average similarity between the original sample of the typical medical record and the positive sample minus the average similarity between the original sample of the typical medical record and the negative sample is used as the performance indicator; The training set of the self-supervised learning includes original samples, positive samples and negative samples, where: Inputting the typical medical record sample and the replicated sample of the typical medical record sample into a neural network with a random dropout mechanism respectively to obtain the original sample and the positive sample; Inputting randomly sampled samples of other typical medical record samples corresponding to different diseases from the typical medical record samples into the neural network with a random discard mechanism to obtain negative samples; Extracting features from the medical record content based on the neural network, and for each disease included in the diagnosis result, calculating similarity between features of a typical medical record corresponding to the disease and features of the medical record content to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is unrelated to the medical record content; Among them, the similarity calculation of the characteristics of the typical medical record corresponding to the disease and the characteristics of the medical record content to determine whether the disease is related to the medical record content includes: performing similarity calculation on the characteristics of the medical record content and the characteristics of multiple typical medical records respectively to obtain multiple similarity scores; calculating the average value of the multiple similarity scores and / or obtaining the maximum value of the multiple similarity scores; when the average value is greater than a first threshold and / or the maximum value is greater than a second threshold, determining that the disease is related to the medical record content.

8. The method according to claim 7, characterized in that The neural network uses a focal loss function to implement the joint training, wherein the focal loss function is designed based on the respective loss functions of the self-supervised learning and the supervised learning and the respective performance indicators of the two.

9. A medical record quality inspection device, characterized in that: The medical record quality inspection device is used to: jointly train self-supervised comparative learning of typical medical records and supervised learning of atypical and typical medical records, and adopt a focal loss function in the joint training process; the focal loss function, for supervised learning tasks, uses the overlap rate between the predicted disease type corresponding to the sample of the atypical medical record and the disease label in the diagnosis list as a performance indicator; for self-supervised learning tasks, the average similarity between the original sample of the typical medical record and the positive sample minus the average similarity between the original sample of the typical medical record and the negative sample is used as a performance indicator; The training set of the self-supervised learning includes original samples, positive samples and negative samples, where: Inputting the typical medical record sample and the replicated sample of the typical medical record sample into a neural network with a random dropout mechanism respectively to obtain the original sample and the positive sample; Inputting randomly sampled samples of other typical medical record samples corresponding to different diseases from the typical medical record samples into the neural network with a random discard mechanism to obtain negative samples; The medical record quality inspection device specifically includes: A medical record acquisition module is used to acquire medical records to be examined, wherein the medical records to be examined include medical record content and diagnosis results, and the diagnosis results include the name of at least one disease; A feature extraction module, configured to extract features from the medical record content to obtain a first feature; A similarity calculation module is used to obtain, for each disease included in the diagnosis result, the characteristics of the typical medical record corresponding to the disease to obtain a second feature; and for each disease included in the diagnosis result, perform similarity calculation on the second feature of the typical medical record corresponding to the disease with the first feature to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content.

10. The device according to claim 9, characterized in that When the feature of the typical medical record corresponding to the disease is obtained to obtain the second feature, the similarity calculation module is specifically used to: obtain the features of each of a plurality of typical medical records corresponding to the disease to obtain a plurality of second features; The calculating the similarity between the second feature and the first feature of the typical medical record corresponding to the disease to determine whether the disease is related to the medical record content includes: calculating the similarity between the first feature and each of the second features to obtain multiple similarity scores; calculating the average of the multiple similarity scores and / or obtaining the maximum value of the multiple similarity scores; when the average value is greater than a first threshold and / or the maximum value is greater than a second threshold, determining that the disease is related to the medical record content.

11. A medical record quality inspection device, characterized in that: The medical record quality inspection device comprises: An input module is used to obtain medical records to be examined and input the medical records to be examined into a trained neural network, wherein the medical records to be examined include medical record content and diagnosis results, and the diagnosis results include the name of at least one disease; the neural network is an end-to-end framework for joint training of self-supervised learning and supervised learning, wherein the self-supervised learning includes feature learning of typical medical records, and the supervised learning includes feature similarity learning between typical medical records and atypical medical records; wherein the neural network is trained by the following method: joint training of self-supervised comparative learning of typical medical records and supervised learning of atypical medical records and typical medical records, and a focal loss function is used in the joint training process; the focal loss function, for supervised learning tasks, uses the overlap rate of the predicted disease type corresponding to the sample of the atypical medical record and the disease label in the diagnosis list as a performance indicator; for self-supervised learning tasks, the original sample of the typical medical record The average similarity with the positive sample minus the average similarity between the original sample of the typical medical record and the negative sample is used as the performance indicator; the training set of the self-supervised learning includes the original sample, positive sample and negative sample, wherein: the typical medical record sample and the copy sample of the typical medical record sample are respectively input into the neural network with a random discard mechanism to obtain the original sample and the positive sample; the random sampling sample of other typical medical record samples corresponding to different diseases of the typical medical record sample is input into the neural network with a random discard mechanism to obtain the negative sample; the output module is used to extract features for the medical record content based on the neural network, and for each disease included in the diagnosis result, the similarity between the features of the typical medical record corresponding to the disease and the features of the medical record content is calculated to determine whether the disease is related to the medical record content, thereby determining whether the diagnosis result includes a disease name that is not related to the medical record content; Among them, the similarity calculation of the characteristics of the typical medical record corresponding to the disease and the characteristics of the medical record content to determine whether the disease is related to the medical record content includes: performing similarity calculation on the characteristics of the medical record content and the characteristics of multiple typical medical records respectively to obtain multiple similarity scores; calculating the average value of the multiple similarity scores and / or obtaining the maximum value of the multiple similarity scores; when the average value is greater than a first threshold and / or the maximum value is greater than a second threshold, determining that the disease is related to the medical record content.

12. A medical record quality inspection device, characterized in that: The device includes a memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the processor executes the medical record quality inspection method according to any one of claims 1 to 8.

13. A storage medium, characterized in that: The storage medium stores a computer program, which, when running, executes the medical record quality inspection method according to any one of claims 1 to 8.

Images