A method for processing medical reports of blood diseases

Abstract

The present application relates to the technical field of electronic digital data processing, and in particular to a method for processing medical reports of blood diseases. The method comprises: obtaining a medical report of a blood disease to be processed, which includes information related to the associated features of the blood disease; if the medical report of the blood disease to be processed is in text format, determining the medical report of the blood disease to be processed as the target text; if the medical report of the blood disease to be processed is in image format, performing OCR recognition on the medical report of the blood disease to be processed, and determining the text obtained after recognition as the target text; inputting the target text into a trained large model so that the trained large model outputs a label list X corresponding to the target text; traversing X, and converting x into j The jth position in the preset template text is written to obtain a structured text corresponding to the medical report of the blood disease to be processed. The present invention can accurately extract key information from the medical report of the blood disease.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a method for processing medical reports of blood diseases. Background Art

[0002] Medical reports on hematological diseases are a large number of text reports or picture reports generated in clinical practice, containing a wealth of key information; however, this key information mostly exists in the form of unstructured free text, which brings great challenges to information extraction, data analysis and other subsequent applications. How to accurately extract key information from medical reports on hematological diseases is an urgent problem to be solved. Summary of the invention

[0003] The purpose of the present invention is to provide a method for processing medical reports of blood diseases so as to accurately extract key information from the medical reports of blood diseases.

[0004] According to the present invention, a method for processing medical reports of blood diseases is provided, the method comprising the following steps:

[0005] S100, obtaining a medical report of a blood disease to be treated; the medical report of the blood disease to be treated includes information related to associated features of the blood disease.

[0006] S200, if the medical report of the blood disease to be processed is in text format, the medical report of the blood disease to be processed is determined as the target text; if the medical report of the blood disease to be processed is in image format, OCR recognition is performed on the medical report of the blood disease to be processed, and the text obtained after recognition is determined as the target text.

[0007] S300, input the target text into the trained large model, so that the trained large model outputs a label list X corresponding to the target text; X={x 1 ,x 2 ,…,x j ,…,x m}, x j is the label of the jth associated feature of blood disease corresponding to the target text. The value of j ranges from 1 to m, and m is the number of associated features of blood disease.

[0008] S400, traverse X, and convert x j Writing the jth position in the preset template text to obtain a structured text corresponding to the medical report of the blood disease to be processed; the preset template text includes the associated features of the blood disease, and the jth position in the preset template text has a corresponding relationship with the jth associated feature of the blood disease; the training process of the trained large model includes:

[0009] S040, use the initial training large model to obtain the structured text set D of the verification sample set, D={d 1 ,d 2 ,…,d k ,…,d z}, d k is the structured text of the kth validation sample obtained using the initial training large model, the value range of k is 1 to z, z is the number of validation samples; each validation sample is a medical report on blood disease; the validation sample set includes z validation samples.

[0010] S050, obtain the associated features of blood diseases and the Pearson correlation coefficient P of blood diseases, P={p 1 ,p 2 ,…,p j ,…,p m}, p j is the Pearson correlation coefficient between the j-th associated feature of blood disease and blood disease.

[0011] S060, traverse P, if |p j | is greater than or equal to the first preset value, then retain d k The content related to the jth associated feature of blood disease in d; otherwise, delete d k The content related to the j-th associated feature of blood disease; the first preset value is greater than 0.

[0012] S070, obtaining an updated structured text d obtained according to the first preset value k,new , and d k,new Performing standardization processing; the standardization processing includes: replacing the tags in the tag list in the text with preset characters.

[0013] S080, according to the structured text after standardization k,new and the trained first classification models to determine whether to determine the initial training large model as the trained large model; the number of the trained first classification models is greater than or equal to 2.

[0014] Furthermore, S080 includes:

[0015] S090, if the trained first classification model is d' k,new The output result of the trained first classification model is consistent with the preset output result corresponding to the k-th verification sample, then the k-th verification sample is determined as the first-class verification sample; if the trained first classification model is consistent with d' k,new If the output result of the trained first classification model is consistent with the preset output result corresponding to the k-th verification sample, the k-th verification sample is determined to be a second-class verification sample.

[0016] S0100, if the ratio of the number of second-category verification samples to the number of first-category verification samples in the verification sample set is less than or equal to a preset ratio threshold, the initially trained large model is determined to be the trained large model.

[0017] Furthermore, S0100 also includes: if the ratio of the number of second-category verification samples to the number of first-category verification samples in the verification sample set is greater than a preset ratio threshold, entering S0101.

[0018] S0101, traverse P, if |p j | is greater than or equal to the second preset value, then retain d k The content related to the jth associated feature of blood disease in d; otherwise, delete d k The content related to the j-th associated feature of blood disease; the second preset value is less than the first preset value, and the second preset value is greater than 0.

[0019] S0102, obtaining an updated structured text e obtained according to the second preset value k,new , and for e k,new Perform standardization.

[0020] S0103, obtain the structured text e' after standardization k,new , and e' k,new Input into the trained second classification model; the number of the trained second classification models is greater than or equal to 2.

[0021] S0104, if the trained second classification model is e' k,new The output result of the trained second classification model is consistent with the preset output result corresponding to the k-th verification sample, then the k-th verification sample is determined as the third-class verification sample; if the trained second classification model is consistent with e' k,new If the output result of the trained second classification model is consistent with the preset output result corresponding to the k-th verification sample, the k-th verification sample is determined to be a fourth type of verification sample.

[0022] S0105, if the ratio of the number of the fourth category verification samples to the third category verification samples in the verification sample set is less than or equal to a preset ratio threshold, the initially trained large model is determined to be the trained large model.

[0023] Furthermore, S0105 also includes: if the ratio of the number of fourth category verification samples to the third category verification samples in the verification sample set is greater than a preset ratio threshold, and the absolute value of the target difference is less than or equal to the preset difference threshold, the second training sample set is used to train the initial training large model to obtain a trained large model; the target difference is the difference between the ratio of the number of fourth category verification samples to the third category verification samples in the verification sample set and the ratio of the number of second category verification samples to the first category verification samples in the verification sample set.

[0024] Further, k The acquisition process includes:

[0025] S041, if the kth verification sample is in text format, the kth verification sample is determined as the kth input text; if the kth verification sample is in image format, OCR recognition is performed on the kth verification sample, and the text obtained after recognition is determined as the kth input text.

[0026] S042, input the kth input text into the initial training large model, so that the initial training large model outputs a label list Y corresponding to the kth input text k ; Y k ={y k,1 ,y k,2 ,…,y k,j ,…,y k,m},y k,j is the label of the jth associated feature of blood disease corresponding to the kth input text.

[0027] S043, traverse Y k , y k,j Write to the jth position in the preset template text and get d k .

[0028] Furthermore, labels in the label list include positive, high expression, +, increased, and abnormal, and the preset character is positive.

[0029] Furthermore, the first classification model includes support vector machine, logistic regression and gradient boosting tree.

[0030] Furthermore, before S040, the training process of the trained large model includes:

[0031] S010, obtain the first training sample set A, A={a 1 ,a 2 ,…,a i ,…,a n}, a i is the i-th training sample, the value range of i is 1 to n, n is the number of training samples in the first training sample set A, and each training sample is a medical report on blood disease.

[0032] S020, obtain the label B corresponding to the first training sample set, B={b 1 ,b 2 ,…,b i ,…,b n}, b i for a i The corresponding label, b i ={b i,1 ,b i,2 ,…,b i,j ,…,b i,m}, b i,j for a i The label of the jth associated feature corresponding to the blood disease.

[0033] S030, use A and B to train the pre-trained large model to obtain the initial training large model.

[0034] Compared with the prior art, the present invention has at least the following beneficial effects:

[0035] After obtaining the medical report of the hematological disease to be treated, the present invention first determines the format of the medical report of the hematological disease to be treated. If it is in text format, it is input into the trained large model without format conversion; if it is in picture format, the medical report of the hematological disease to be treated is converted into text format and then input into the trained large model; the trained large model has the ability to obtain the label of the associated feature of the hematological disease corresponding to the input text, so that after the target text is input into the trained large model, the label of the associated feature of the hematological disease corresponding to the target text can be obtained, that is, the label of the associated feature of the hematological disease corresponding to the medical report of the hematological disease to be treated; further, the present invention obtains the structured text corresponding to the medical report of the hematological disease to be treated by writing the label of each associated feature of the hematological disease into the corresponding position in the preset template text. The structured text corresponding to the medical report of the hematological disease to be treated obtained by the present invention includes the labels of all associated features of the hematological disease, realizes the extraction of key information in the medical report of the hematological disease to be treated, and based on the structured text corresponding to the medical report of the hematological disease to be treated obtained by the present invention, the key information in the medical report of the hematological disease to be treated can be quickly obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 A flowchart of a method for processing a medical report of a blood disease provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0039] According to this embodiment, Figure 1 As shown, a method for processing a medical report of a blood disease is provided, the method comprising the following steps:

[0040] S100, obtaining a medical report of a blood disease to be treated; the medical report of the blood disease to be treated includes information related to associated features of the blood disease.

[0041] In this embodiment, the medical report of the blood disease to be processed is a medical report related to the blood disease, such as a medical report obtained when the patient undergoes a diagnosis of the blood disease or an MRD assessment; the medical report of the blood disease includes information related to the associated characteristics of the blood disease, and the associated characteristics of the blood disease are known; as a specific embodiment, the medical report of the blood disease to be processed is an immune flow report, and the associated characteristics of the blood disease are cell type, percentage denominator, cell ratio qualitative, cell ratio quantitative (%) result value, phenotypic condition, expressed antigen, non-expressed antigen, highly expressed antigen, low expressed antigen, partially expressed antigen, CD4+ / CD8+ qualitative result, CD4+ / CD8+ quantitative result value, Kappa+ / Lambda+ qualitative result and Kappa+ / Lambda+ quantitative result value. As another specific embodiment, the medical report of the blood disease to be processed is a gene detection report, and the associated characteristics of the blood disease are gene name, detection method, qualitative result, quantitative result (%), international normalized value (%), examined gene site, mutation condition and mutation type.

[0042] S200, if the medical report of the blood disease to be processed is in text format, the medical report of the blood disease to be processed is determined as the target text; if the medical report of the blood disease to be processed is in image format, OCR recognition is performed on the medical report of the blood disease to be processed, and the text obtained after recognition is determined as the target text.

[0043] Those skilled in the art are aware that any OCR recognition method in the prior art falls within the scope of protection of the present invention. As a preferred specific implementation, for the text obtained after OCR recognition, the punctuation marks in the text are uniformly processed into English format, and if there are multiple consecutive spaces, only one space is retained; the text after punctuation unification and space deletion is used as the target text. Based on this preferred specific implementation, the readability of the target text is improved, and the accuracy of the label list obtained after the target text is input into the trained large model can be improved.

[0044] S300, input the target text into the trained large model, so that the trained large model outputs a label list X corresponding to the target text; X={x 1 ,x 2 ,…,x j ,…,x m}, x j is the label of the jth associated feature of blood disease corresponding to the target text, the value range of j is 1 to m, and m is the number of associated features of blood disease.

[0045] In this embodiment, the trained large model has the function of obtaining a label list corresponding to the input text, and the label list includes m labels, wherein the jth label is the label of the jth associated feature of the blood disease corresponding to the input text.

[0046] S400, traverse X, and convert x j Write the jth position in the preset template text to obtain a structured text corresponding to the medical report of the blood disease to be treated; the preset template text includes the associated features of the blood disease, and the jth position in the preset template text has a corresponding relationship with the jth associated feature of the blood disease.

[0047] In this embodiment, the structured text corresponding to the medical report of the blood disease to be processed includes relevant information of all the associated features of the blood disease. The preset template text includes all the associated features of the blood disease, but does not include the test results corresponding to each associated feature. The jth position in the preset template text has a corresponding relationship with the jth associated feature of the blood disease. j By writing into the jth position in the preset template text, the detection result corresponding to the jth associated feature of the blood disease can be obtained.

[0048] As a preferred specific implementation, the training process of the trained large model includes:

[0049] S010, obtain the first training sample set A, A={a 1 ,a 2 ,…,a i ,…,an}, a i is the i-th training sample, the value range of i is 1 to n, n is the number of training samples in the first training sample set A, and each training sample is a medical report on blood disease.

[0050] In this embodiment, the names of the medical reports corresponding to different training samples are the same. For example, different training samples are all immunocytometry reports of blood diseases. The difference between different training samples is that the corresponding patients are different, or the corresponding patients are the same but the corresponding report acquisition time is different.

[0051] S020, obtain the label B corresponding to the first training sample set, B={b 1 ,b 2 ,…,b i ,…,b n}, b i for a i The corresponding label, b i ={b i,1 ,b i,2 ,…,b i,j ,…,b i,m}, b i,j for a i The label of the jth associated feature corresponding to the blood disease.

[0052] In this embodiment, the associated features of the target label are known, and the label of the jth associated feature of the blood disease in the ith training sample is the detection result corresponding to the jth associated feature of the blood disease in the ith training sample. For example, if a certain associated feature is a quantitative result value of the cell proportion (%), then the label of the associated feature of the blood disease in the ith training sample is the detection result of the quantitative result value of the cell proportion (%) in the ith training sample; for another example, if a certain associated feature is a highly expressed antigen, then the label of the associated feature of the blood disease in the ith training sample is the detection result of the highly expressed antigen in the ith training sample.

[0053] S030, use A and B to train the pre-trained large model to obtain the initial training large model.

[0054] In this embodiment, the pre-trained large model refers to a general large model obtained by training based on a huge text corpus in the prior art. The pre-trained large model needs to be fine-tuned before it can be better applied to feature tasks.

[0055] In this embodiment, the initial training large model obtained after training the pre-trained large model using A and B has a certain reasoning ability, that is, the ability to obtain labels of associated features of blood diseases in medical reports of blood diseases. However, whether the reasoning ability of the initial training large model can meet the needs still needs to be verified in combination with subsequent steps.

[0056] Those skilled in the art are aware that any supervised training method in the prior art falls within the protection scope of the present invention.

[0057] S040, use the initial training large model to obtain the structured text set D of the verification sample set, D={d 1 ,d 2 ,…,d k ,…,d z}, d k is the structured text of the kth validation sample obtained using the initial training large model, the value range of k is 1 to z, z is the number of validation samples; each validation sample is a medical report on blood disease; the validation sample set includes z validation samples.

[0058] In this embodiment, the names of the medical reports corresponding to different verification samples are the same as the names of the medical reports corresponding to the training samples, and the difference between the different verification samples is that the corresponding patients are different, or the corresponding patients are the same but the corresponding report acquisition time is different. As a specific implementation method, d k The acquisition process includes:

[0059] S041, if the kth verification sample is in text format, the kth verification sample is determined as the kth input text; if the kth verification sample is in image format, OCR recognition is performed on the kth verification sample, and the text obtained after recognition is determined as the kth input text.

[0060] S042, input the kth input text into the initial training large model, so that the initial training large model outputs a label list Y corresponding to the kth input text k ; Y k ={y k,1 ,y k,2 ,…,y k,j ,…,y k,m},y k,j is the label of the jth associated feature of blood disease corresponding to the kth input text.

[0061] S043, traverse Y k , y k,j Write to the jth position in the preset template text and get d k .

[0062] The above-mentioned process of S041-S043 is similar to the process of S200-S400, and the main difference is that S200-S400 is structured text obtained based on the trained large model, and S041-S043 is structured text obtained based on the initially trained large model.

[0063] S050, obtain the associated features of blood diseases and the Pearson correlation coefficient P of blood diseases, P={p 1 ,p 2 ,…,p j ,…,p m}, p j is the Pearson correlation coefficient between the j-th associated feature of blood disease and blood disease.

[0064] Those skilled in the art know that any method for obtaining the Pearson correlation coefficient in the prior art falls within the protection scope of the present invention.

[0065] S060, traverse P, if |p j | is greater than or equal to the first preset value, then retain d k The content related to the jth associated feature of blood disease in d; otherwise, delete d k The content related to the j-th associated feature of blood disease; the first preset value is greater than 0.

[0066] As an optional specific implementation, the first preset value is an empirical value.

[0067] In this embodiment, d k The content related to the jth associated feature of blood disease refers to d k The j-th associated feature of blood disease itself and the label of the j-th associated feature of blood disease.

[0068] The updated structured text obtained based on S060 only retains the correlation features that are more closely related to blood diseases compared to the structured text before the update, which can reduce the complexity of reasoning of the subsequently trained first classification model and is conducive to improving the accuracy of the output results of the first classification model.

[0069] S070, obtaining an updated structured text d obtained according to the first preset value k,new , and d k,new Performing standardization processing; the standardization processing includes: replacing the tags in the tag list in the text with preset characters.

[0070] As a specific implementation, the labels in the label list include positive, high expression, +, increased and abnormal, and the preset character is positive.

[0071] The structured text after standardization obtained based on S070 reduces the complexity of expression compared to the structured text before standardization, can reduce the complexity of reasoning of the subsequently trained first classification model, and improve the accuracy of the output results of the first classification model.

[0072] S080, according to the structured text after standardizationk,new and the trained first classification models to determine whether to determine the initial training large model as the trained large model; the number of the trained first classification models is greater than or equal to 2.

[0073] As a specific embodiment, the training process of the trained first classification model includes: taking a number of structured texts updated based on a first preset value and standardized as training samples, labeling each training sample with a corresponding positive or negative label, and training the first classification model based on the labeled training samples; the trained first classification model has the function of labeling the input structured texts updated based on a first preset value and standardized with positive or negative labels. As a specific embodiment, the first classification model includes a support vector machine, a logistic regression, and a gradient boosting tree. Any process of training a support vector machine, a logistic regression, and a gradient boosting tree in the prior art falls within the protection scope of the present invention.

[0074] As a preferred specific implementation, S080 includes:

[0075] S090, if the trained first classification model is d' k,new The output result of the trained first classification model is consistent with the preset output result corresponding to the k-th verification sample, then the k-th verification sample is determined as the first-class verification sample; if the trained first classification model is consistent with d' k,new If the output result of the trained first classification model is consistent with the preset output result corresponding to the k-th verification sample, the k-th verification sample is determined to be a second-class verification sample.

[0076] In this embodiment, if the output result of the trained first classification model and the preset output result corresponding to the k-th verification sample are both positive labels, then the output result of the trained first classification model is determined to be the same as the preset output result corresponding to the k-th verification sample; if the output result of the trained first classification model and the preset output result corresponding to the k-th verification sample are both negative labels, then the output result of the trained first classification model is determined to be the same as the preset output result corresponding to the k-th verification sample; otherwise, it is determined that the output result of the trained first classification model is not the same as the preset output result corresponding to the k-th verification sample.

[0077] In this embodiment, if the trained first classification model is d' k,new The output result of the trained first classification model is consistent with the preset output result corresponding to the kth validation sample, and the output result of the trained first classification model is the same as the preset output result corresponding to the kth validation sample, and it is determined that the trained first classification model is k,new The accuracy is higher and the d' input to the trained first classification model k,newIt can be used to characterize the kth validation sample and further determine whether the initial training large model's reasoning ability for the kth validation sample meets the requirements; if the first classification model trained for d' k,new The output result of the trained first classification model is consistent with the preset output result corresponding to the kth validation sample, and the output result of the trained first classification model is different from the preset output result corresponding to the kth validation sample, and it is determined that the trained first classification model is k,new The accuracy is higher and the d' input to the trained first classification model k,new It cannot be used to characterize the kth validation sample.

[0078] In this embodiment, if the trained first classification model is d' k,new If the output results of are inconsistent, it is determined that different trained first classification models are not consistent with d' k,new There are differences in the reasoning results of the trained first classification model; if there are differences in the reasoning results of the standardized structured text corresponding to a large number of verification samples, the trained first classification model is retrained to improve the reasoning accuracy of the trained first classification model.

[0079] S0100, if the ratio of the number of second-category verification samples to the number of first-category verification samples in the verification sample set is less than or equal to a preset ratio threshold, the initially trained large model is determined to be the trained large model.

[0080] As an optional specific implementation manner, the preset ratio threshold is an empirical value. Optionally, the preset ratio threshold is 5%.

[0081] Based on S040-S0100, the accuracy of the initially trained large model can be judged, and the judgment process includes the process of updating and standardizing the structured text, so that the readability of the text input into the trained first classification model is improved and the complexity is reduced, which is beneficial to improving the accuracy of the output result of the trained first classification model; the output of the initially trained large model is a label list, and the label list includes multiple labels. When judging the accuracy of the initially trained large model, the number of comparisons corresponding to a single verification text is large, and the comparison process is complicated. Compared with judging the accuracy of the initially trained large model, the number of comparisons corresponding to a single verification text when judging the accuracy of the first classification model is only 1, and the comparison process is simple, which reduces the workload of the comparison link, and is beneficial to reducing the human resources occupied by the comparison link.

[0082] S0100 also includes: if the ratio of the number of second-category verification samples to the number of first-category verification samples in the verification sample set is greater than a preset ratio threshold, entering S0101.

[0083] S0101, traverse P, if |p j | is greater than or equal to the second preset value, then retain d kThe content related to the jth associated feature of blood disease in d; otherwise, delete d k The content related to the j-th associated feature of blood disease; the second preset value is less than the first preset value, and the second preset value is greater than 0.

[0084] In this embodiment, when the ratio of the number of second-category verification samples to the number of first-category verification samples in the verification sample set is greater than a preset ratio threshold, the second preset value is used as a reference for determining whether the associated features of blood diseases are closely related to blood diseases. Since the second preset value is smaller than the first preset value, the updated structure text obtained based on the second feature value can retain more relevant information about the associated features of blood diseases than the updated structure text obtained based on the first feature value.

[0085] S0102, obtaining an updated structured text e obtained according to the second preset value k,new , and for e k,new Perform standardization.

[0086] S0103, obtain the structured text e' after standardization k,new , and e' k,new Input to the trained second classification model.

[0087] S0104, if the trained second classification model is e' k,new The output result of the trained second classification model is consistent with the preset output result corresponding to the k-th verification sample, then the k-th verification sample is determined as the third-class verification sample; if the trained second classification model is consistent with e' k,new If the output result of the trained second classification model is consistent with the preset output result corresponding to the k-th verification sample, the k-th verification sample is determined to be a fourth type of verification sample.

[0088] As a specific implementation, the training process of the trained second classification model includes: taking a number of structured texts updated based on a second preset value and standardized as training samples, marking each training sample with a corresponding positive or negative label, and training the second classification model based on the labeled training samples; the trained second classification model has the function of marking the input structured text updated based on the second preset value and standardized with a positive label or a negative label.

[0089] S0105, if the ratio of the number of the fourth category verification samples to the third category verification samples in the verification sample set is less than or equal to a preset ratio threshold, the initially trained large model is determined to be the trained large model.

[0090] In this embodiment, the number of fourth-category verification samples and the number of third-category verification samples in the verification sample set are related to the second preset value. The number of associated features of blood diseases included in the structured text after standardized processing corresponding to the verification samples obtained based on the second preset value is relatively large. Therefore, the number of associated features of blood diseases referred to by the trained second classification model for reasoning is large. If the ratio of the number of fourth-category verification samples to the number of third-category verification samples in the verification sample set is less than or equal to the preset ratio threshold, it is determined that the initially trained large model can meet the reasoning requirements. The reason why the ratio of the number of second-category verification samples to the number of first-category verification samples is greater than the preset ratio threshold is that the first preset value selected when using the Pearson correlation coefficient to screen the associated features of blood diseases is large.

[0091] S0105 also includes: if the ratio of the number of fourth category verification samples to the number of third category verification samples in the verification sample set is greater than a preset ratio threshold, and the absolute value of the target difference is less than or equal to the preset difference threshold, then the second training sample set is used to train the initial training large model to obtain a trained large model; the target difference is the difference between the ratio of the number of fourth category verification samples to the number of third category verification samples in the verification sample set and the ratio of the number of second category verification samples to the number of first category verification samples in the verification sample set.

[0092] In this embodiment, if the absolute value of the target difference is less than or equal to the preset difference threshold, it is determined that the first preset value selected when using the Pearson correlation coefficient to screen the associated features of blood diseases is more appropriate, resulting in d' k,new The reason why it cannot be used to represent the kth validation sample is probably that the accuracy of the initial training model is low.

[0093] In this embodiment, the second training sample set is not equivalent to the first training sample set. The process of training the initial training large model using the second training sample set is similar to the process of training the pre-trained large model using the first training sample set, which will not be repeated here.

[0094] In this embodiment, if the ratio of the number of verification samples of the fourth category to the number of verification samples of the third category in the verification sample set is greater than the preset ratio threshold, and the difference between the ratio of the number of verification samples of the second category to the number of verification samples of the first category in the verification sample set and the ratio of the number of verification samples of the fourth category to the number of verification samples of the third category in the verification sample set is greater than the preset difference threshold, it is determined that the cause of d' k,newThe reason why it cannot be used to characterize the kth verification sample is most likely that the preset selected when using the Pearson correlation coefficient to screen the associated characteristics of blood diseases is inappropriate, then the third preset value is used instead of the second preset value, and the trained third classification model is used instead of the trained second classification model, and S0101-S0105 are repeated; the third preset value is less than the second preset value, and the third preset value is greater than 0. In this embodiment, the preset difference threshold is greater than 0. The training process of the trained third classification model includes: taking a number of structured texts updated and standardized based on the third preset value as training samples, labeling each training sample with a corresponding positive or negative label, and training the second classification model based on the labeled training samples; the trained second classification model has the function of labeling the input structured text updated and standardized based on the third preset value with a positive or negative label.

[0095] After obtaining the medical report of the hematological disease to be processed, the present embodiment first determines the format of the medical report of the hematological disease to be processed. If it is in text format, it is input into the trained large model without format conversion; if it is in image format, the medical report of the hematological disease to be processed is converted into text format and then input into the trained large model; the trained large model has the ability to obtain the label of the associated feature of the hematological disease corresponding to the input text, so that after inputting the target text into the trained large model, the label of the associated feature of the hematological disease corresponding to the target text can be obtained, that is, the label of the associated feature of the hematological disease corresponding to the medical report of the hematological disease to be processed; further, the present embodiment obtains the structured text corresponding to the medical report of the hematological disease to be processed by writing the label of each associated feature of the hematological disease into the corresponding position in the preset template text. The structured text corresponding to the medical report of the hematological disease to be processed obtained by the present embodiment includes the labels of all associated features of the hematological disease, which realizes the extraction of key information in the medical report of the hematological disease to be processed. Based on the structured text corresponding to the medical report of the hematological disease to be processed obtained by the present embodiment, the key information in the medical report of the hematological disease to be processed can be quickly obtained.

[0096] Although some specific embodiments of the present invention have been described in detail by way of example, it will be appreciated by those skilled in the art that the above examples are for illustration only and are not intended to limit the scope of the present invention. It will also be appreciated by those skilled in the art that various modifications may be made to the embodiments without departing from the scope and spirit of the present invention. The scope of the present invention is defined by the appended claims.

Claims

1. A method for processing medical reports of blood diseases, characterized in that: The method comprises the following steps: S100, obtaining a medical report of a blood disease to be treated; the medical report of the blood disease to be treated includes information related to associated features of the blood disease; S200, if the medical report of the blood disease to be processed is in text format, the medical report of the blood disease to be processed is determined as the target text; if the medical report of the blood disease to be processed is in image format, the medical report of the blood disease to be processed is subjected to OCR recognition, and the text obtained after recognition is determined as the target text; S300, input the target text into the trained large model, so that the trained large model outputs a label list X corresponding to the target text; X={x1,x2,…,x j ,…,x m }, x j is the label of the jth associated feature of blood disease corresponding to the target text, the value range of j is 1 to m, and m is the number of associated features of blood disease; S400, traverse X, and convert x j Writing the jth position in the preset template text to obtain a structured text corresponding to the medical report of the blood disease to be processed; the preset template text includes the associated features of the blood disease, and the jth position in the preset template text has a corresponding relationship with the jth associated feature of the blood disease; the training process of the trained large model includes: S040, use the initial training large model to obtain the structured text set D of the verification sample set, D = {d1, d2, ..., d k ,…,d z }, d k is the structured text of the kth validation sample obtained using the initial training large model, the value range of k is 1 to z, and z is the number of validation samples; each validation sample is a medical report on blood disease; the validation sample set includes z validation samples; S050, obtain the associated features of blood diseases and the Pearson correlation coefficient P of blood diseases, P={p1,p2,…,p j ,…,p m }, p j is the Pearson correlation coefficient between the jth associated feature of blood disease and blood disease; S060, traverse P, if |p j | is greater than or equal to the first preset value, then retain d k The content related to the jth associated feature of blood disease in d; otherwise, delete d k The content related to the jth associated feature of blood disease; the first preset value is greater than 0; S070, obtaining an updated structured text d obtained according to the first preset value k,new , and d k,new Performing standardization processing; the standardization processing includes: replacing the tags appearing in the tag list in the text with preset characters; S080, according to the structured text after standardization k,new and the trained first classification models to determine whether to determine the initial training large model as the trained large model; the number of the trained first classification models is greater than or equal to 2; S080 includes: S090, if the trained first classification model is d' k,new The output result of the trained first classification model is consistent with the preset output result corresponding to the k-th verification sample, then the k-th verification sample is determined as the first-class verification sample; if the trained first classification model is consistent with d' k,new If the output result of the trained first classification model is consistent with the preset output result corresponding to the k-th verification sample, the k-th verification sample is determined as the second-class verification sample; S0100, if the ratio of the number of second-category verification samples to the number of first-category verification samples in the verification sample set is less than or equal to a preset ratio threshold, the initially trained large model is determined to be the trained large model.

2. The method for processing medical reports of blood diseases according to claim 1, characterized in that: S0100 also includes: if the ratio of the number of the second type of verification samples to the number of the first type of verification samples in the verification sample set is greater than a preset ratio threshold, entering S0101; S0101, traverse P, if |p j | is greater than or equal to the second preset value, then retain d k The content related to the jth associated feature of blood disease in d; otherwise, delete d k The content related to the j-th associated feature of blood disease in the embodiment; the second preset value is less than the first preset value, and the second preset value is greater than 0; S0102, obtaining an updated structured text e obtained according to the second preset value k,new , and for e k,new Standardize the process; S0103, obtain the structured text e' after standardization k,new , and e' k,new Input into a trained second classification model; the number of the trained second classification models is greater than or equal to 2; S0104, if the trained second classification model is e' k,new The output result of the trained second classification model is consistent with the preset output result corresponding to the k-th verification sample, then the k-th verification sample is determined as the third-class verification sample; if the trained second classification model is consistent with e' k,new If the output result of the trained second classification model is consistent with the preset output result corresponding to the k-th verification sample, the k-th verification sample is determined to be a fourth-category verification sample; S0105, if the ratio of the number of the fourth category verification samples to the third category verification samples in the verification sample set is less than or equal to a preset ratio threshold, the initially trained large model is determined to be the trained large model.

3. The method for processing medical reports of blood diseases according to claim 2, characterized in that: S0105 also includes: if the ratio of the number of fourth category verification samples to the number of third category verification samples in the verification sample set is greater than a preset ratio threshold, and the absolute value of the target difference is less than or equal to the preset difference threshold, then the second training sample set is used to train the initial training large model to obtain a trained large model; the target difference is the difference between the ratio of the number of fourth category verification samples to the number of third category verification samples in the verification sample set and the ratio of the number of second category verification samples to the number of first category verification samples in the verification sample set.

4. The method for processing medical reports of blood diseases according to claim 1, characterized in that: d k The acquisition process includes: S041, if the kth verification sample is in text format, determining the kth verification sample as the kth input text; if the kth verification sample is in image format, performing OCR recognition on the kth verification sample, and determining the text obtained after recognition as the kth input text; S042, input the kth input text into the initial training large model, so that the initial training large model outputs a label list Y corresponding to the kth input text k ; Y k ={y k,1 ,y k,2 ,…,y k,j ,…,y k,m },y k,j is the label of the jth associated feature of blood disease corresponding to the kth input text; S043, traverse Y k , y k,j Write to the jth position in the preset template text and get d k .

5. The method for processing medical reports of blood diseases according to claim 1, characterized in that: The labels in the label list include positive, high expression, +, increased, and abnormal, and the default character is positive.

6. The method for processing medical reports of blood diseases according to claim 1, characterized in that: The first classification models included support vector machines, logistic regression, and gradient boosted trees.

7. The method for processing medical reports of blood diseases according to claim 1, characterized in that: Prior to S040, the training process for a trained large model included: S010, obtain the first training sample set A, A={a1,a2,…,a i ,…,a n }, a i is the i-th training sample, i ranges from 1 to n, n is the number of training samples in the first training sample set A, and each training sample is a medical report on a blood disease; S020, obtain the label B corresponding to the first training sample set, B={b1,b2,…,b i ,…,b n }, b i for a i The corresponding label, b i ={b i,1 ,b i,2 ,…,b i,j ,…,b i,m }, b i,j for a i The label of the jth associated feature of the corresponding blood disease; S030, use A and B to train the pre-trained large model to obtain the initial training large model.

Images