Electronic data analysis method for digestive system department

Abstract

The invention relates to the technical field of medical data analysis, in particular to a gastroenterology electronic data analysis method, which comprises the steps of generating a semantic feature vector, an image feature vector and an emotion feature vector, and performing feature fusion on the semantic feature vector, the image feature vector and the emotion feature vector to obtain a fusion feature vector; establishing a disease and emotion correlation analysis model, adopting a multi-layer perceptron model, taking the fusion feature vector as a prediction probability of an input layer and an output layer, and performing effect evaluation on a result output by the prediction probability through a physician; the emotion is used as an important analysis factor of a digestive tract disease, the emotional state of a patient is comprehensively mapped through two dimensions of stability and fluctuation amplitude, a more complex digestive tract disease and emotion association mode can be learned in combination with text data, image data and a model, and then the correlation between the digestive tract disease and the emotion is found. And prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and in particular to a method for analyzing electronic data of gastroenterology. Background Art

[0002] Medical data includes medical examination results, such as test reports, colonoscopy images, gastroscopy images, etc. Clinicians can refer to these medical data and images to improve work efficiency, which is an important foundation of modern medicine.

[0003] Electronic data from gastroenterology is one of the most common medical data. However, there are many items in electronic data from gastroenterology and the data is complex. In addition, with the deepening of modern medical knowledge, it is found that human emotions are also an important influencing factor of digestive diseases. Therefore, how to accurately process electronic data from gastroenterology using emotions as an influencing factor is a technical problem that needs to be solved at present. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for analyzing electronic data of gastroenterology.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A gastroenterology electronic data analysis method, comprising:

[0007] Step 1: Generate semantic feature vectors. Collect the patient's medical records and pathology reports, and use the language processing model to generate semantic feature vectors S for the text parts of the medical records and pathology reports.

[0008] Step 2: Generate an image feature vector. Generate an image feature vector V for the image part in the pathology report through an image processing model.

[0009] Step 3: Generate an emotion feature vector. In clinical practice, doctors extract emotion features from patients. Emotion feature extraction includes emotion stability score ESS and emotion fluctuation score EVI.

[0010] Emotional stability score ESS includes: ESS=

[0011] in, is the i-th emotion evaluation score, is the average emotion assessment score, n is the number of assessments, and ESS is used to measure the stability of emotions. The lower the score, the more stable the emotion;

[0012] Emotional volatility score EVI includes: EVI=

[0013] in, is the maximum value of the physician’s emotional assessment score for the patient. It is the minimum score of the physician's emotional assessment of the patient. EVI is used to measure the degree of emotional fluctuation of the patient. The higher the score, the greater the emotional fluctuation.

[0014] The emotional stability score ESS and the emotional volatility score EVI are combined to obtain the emotional feature vector E; the combination process includes: mapping the emotional stability score ESS and the emotional volatility score EVI to a multi-dimensional vector space, inputting a one-dimensional vector [ESS, EVI] composed of the emotional stability score ESS and the emotional volatility score EVI into the full neural network model, and transforming through multiple hidden layers:

[0015]

[0016]

[0017] E=

[0018] in, is the output of the first hidden layer of the neural network; is the output of the second hidden layer of the neural network; is the activation function, T represents the transposition operation; , , All are bias terms; is the weight matrix of the first hidden layer; is the weight matrix of the second hidden layer; is the weight matrix of the last fully connected layer;

[0019] Step 4: Feature fusion: fuse the semantic feature vector S, image feature vector V and emotion feature vector E to obtain the fused feature vector F; F = + + ;

[0020] in, is the attention weight of the semantic feature vector S; is the attention weight of the image feature vector V; is the attention weight of the emotion feature vector E;

[0021] calculate When a linear transformation is introduced, the semantic feature vector S is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ;

[0022] calculate When a linear transformation is introduced, the image feature vector V is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ;

[0023] calculate When a linear transformation is introduced, the emotion feature vector E is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ;

[0024] Step 5: Establish a disease and emotion correlation analysis model, using a multi-layer perceptron model, taking the fusion feature vector F as the input layer and the predicted probability P (Y=1│ );

[0025] P(Y=1│ )= ( + + + )

[0026] in, is the activation function, is the bias term of the output layer; is the weight coefficient of the semantic feature vector S; is the i-th element of the semantic feature vector S; m is the number of elements after the semantic feature vector S is expanded into a one-dimensional vector; is the weight coefficient of the image feature vector V; is the jth element of the image feature vector V; n is the number of elements after the image feature vector V is expanded into a one-dimensional vector; is the weight coefficient of the image feature vector V; is the kth element of the emotion feature vector E; p is the number of elements after the emotion feature vector E is expanded into a one-dimensional vector;

[0027] Step 6: Effect evaluation: The physician evaluates the predicted probability P (Y=1│ ) output results for effect evaluation.

[0028] Preferably, step six also includes: calculating the accuracy A, A= ;

[0029] Calculate the recall rate R, R= ;

[0030] Calculate the precision P, P= ;

[0031] Final evaluation value F1, F1= ;

[0032] Among them, TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example; TP, TN, FP, and FN are given by the predicted probability P (Y=1│ ) The output result value is generated;

[0033] Through the doctor's prediction probability P (Y = 1 | ) output result is evaluated to obtain the first evaluation value; the machine obtains the second evaluation value through accuracy A, recall rate R, precision P, and final evaluation value F1.

[0034] Preferably, if the result error between the first evaluation value and the second evaluation value is greater than a preset value, the physician's judgment shall prevail.

[0035] Preferably, the language processing model is a BERT model.

[0036] Preferably, the language processing model is a BERT model, and an adversarial training mechanism is introduced into the BERT model.

[0037] Preferably, the image processing model is a ResNet model.

[0038] Preferably, an attention module is introduced into the ResNet model.

[0039] Preferably, step five further includes: establishing a disease and emotion association analysis model, using multiple multi-layer perceptron models, and setting different initialization parameters for different multi-layer perceptron models.

[0040] Preferably, when normalizing in step 4, an IQR method is used to identify and correct or eliminate outliers.

[0041] Preferably, the patient's age, gender and occupation are extracted separately from the text portion of the medical record and pathology report.

[0042] The beneficial effect of the present invention is that: by taking emotion as an important analysis factor of digestive tract diseases, the emotion stability score ESS and the emotion fluctuation score EVI are calculated separately, ESS focuses on reflecting the stability of an individual's emotions over a period of time, and the stability means the change from one emotion to another, such as from joy to anger, and EVI focuses on reflecting the fluctuation amplitude of emotions, and the fluctuation amplitude means the fluctuation amplitude of an emotion, that is, a slight happiness changes to a very happy (a smile changes to a laugh), and the two indicators are calculated separately, and the patient's emotional state is comprehensively mapped from the two dimensions of stability and fluctuation amplitude. Combined with text data and image data, data normalization can unify the data scale, make subsequent analysis more accurate and improve the efficiency of analysis, so that different types of data can be effectively converted into feature representations for analysis, all are normalized and fused into feature F, and then F is used as the input layer, and a multi-layer perceptron (MLP) is used in combination with the fused features to perform disease and emotion association analysis, and the association pattern between emotion and digestive tract diseases is found through an algorithm or model, and the relationship between emotion and digestive tract diseases is found by using an algorithm and computing power, thereby improving prediction accuracy. Furthermore, calculating ESS and EVI separately allows the model to learn and analyze the relationship between these two emotional characteristics and diseases respectively. For example, patients with long-term stable emotions may have a relatively low probability of developing certain chronic diseases; while patients with large emotional fluctuations may develop some stress-induced diseases in the short term. When calculated separately, the model can learn more about the complex patterns and relationships between emotions and diseases, thereby improving the generalization ability of the model, enabling it to better adapt to different patient groups and various actual situations, and providing a more accurate reference for the doctor's final diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The figure is a flowchart of a method for electronic data analysis in gastroenterology according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to explain the technical content, achieved objectives and effects of the present invention in detail, the following is an explanation in combination with the implementation modes and the accompanying drawings.

[0045] Please refer to Figure 1 , a method for analyzing electronic data in gastroenterology, comprising:

[0046] Step 1: Generate semantic feature vectors. Collect the patient's medical records and pathology reports, and use the language processing model to generate semantic feature vectors S for the text parts of the medical records and pathology reports.

[0047] Step 2: Generate an image feature vector. Generate an image feature vector V for the image part in the pathology report through an image processing model.

[0048] Step 3: Generate an emotion feature vector. In clinical practice, doctors extract emotion features from patients. Emotion feature extraction includes emotion stability score ESS and emotion fluctuation score EVI.

[0049] Emotional stability score ESS includes: ESS=

[0050] in, is the i-th emotion evaluation score, is the average emotion assessment score, n is the number of assessments, and ESS is used to measure the stability of emotions. The lower the score, the more stable the emotion;

[0051] Emotional volatility score EVI includes: EVI=

[0052] in, is the maximum value of the physician’s emotional assessment score for the patient. It is the minimum score of the physician's emotional assessment of the patient. EVI is used to measure the degree of emotional fluctuation of the patient. The higher the score, the greater the emotional fluctuation.

[0053] The emotional stability score ESS and the emotional volatility score EVI are combined to obtain the emotional feature vector E; the combination process includes: mapping the emotional stability score ESS and the emotional volatility score EVI to a multi-dimensional vector space, inputting a one-dimensional vector [ESS, EVI] composed of the emotional stability score ESS and the emotional volatility score EVI into the full neural network model, and transforming through multiple hidden layers:

[0054]

[0055]

[0056] E=

[0057] in, is the output of the first hidden layer of the neural network; is the output of the second hidden layer of the neural network; is the activation function, T represents the transposition operation; , , All are bias terms; is the weight matrix of the first hidden layer; is the weight matrix of the second hidden layer; is the weight matrix of the last fully connected layer;

[0058] Step 4: Feature fusion: fuse the semantic feature vector S, image feature vector V and emotion feature vector E to obtain the fused feature vector F; F = + + ;

[0059] in, is the attention weight of the semantic feature vector S; is the attention weight of the image feature vector V; is the attention weight of the emotion feature vector E;

[0060] calculate When a linear transformation is introduced, the semantic feature vector S is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ;

[0061] calculate When a linear transformation is introduced, the image feature vector V is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ;

[0062] calculate When a linear transformation is introduced, the emotion feature vector E is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ;

[0063] Step 5: Establish a disease and emotion correlation analysis model, using a multi-layer perceptron model, taking the fusion feature vector F as the input layer and the predicted probability P (Y=1│ );

[0064] P(Y=1│ )= ( + + + )

[0065] Among them, P (Y = 1 | ) is given all the independent variables The probability that the dependent variable Y takes the value 1 under the condition that , that is, the probability of a specific disease state based on all disease- and emotion-related factors considered; is the activation function, is the bias term of the output layer; is the weight coefficient of the i-th element of the semantic feature vector S; is the i-th element of the semantic feature vector S; m is the number of elements after the semantic feature vector S is expanded into a one-dimensional vector; is the weight coefficient of the jth element of the image feature vector V; is the jth element of the image feature vector V; n is the number of elements after the image feature vector V is expanded into a one-dimensional vector; is the weight coefficient of the kth element of the image feature vector E; is the kth element of the emotion feature vector E; p is the number of elements after the emotion feature vector E is expanded into a one-dimensional vector;

[0066] Step 6: Effect evaluation: The physician evaluates the predicted probability P (Y=1│ ) output results for effect evaluation.

[0067] From the above description, it can be seen that by taking emotion as an important analysis factor of digestive tract diseases, the emotional stability score ESS and the emotional fluctuation score EVI are calculated separately. ESS focuses on reflecting the stability of an individual's emotions over a period of time. The stability means the change from one emotion to another, such as from joy to anger. EVI focuses on the fluctuation amplitude of emotions. The fluctuation amplitude means the fluctuation amplitude of an emotion, that is, a slight happiness changes to a very happy happiness (a smile changes to a laugh). The two indicators are calculated separately, and the patient's emotional state is fully mapped from the two dimensions of stability and fluctuation amplitude. Combined with text data and image data, data normalization can unify the data scale, make subsequent analysis more accurate and improve the efficiency of analysis, so that different types of data can be effectively converted into feature representations for analysis. They are all normalized and fused into feature F, and then F is used as the input layer. Multilayer perceptron (MLP) is used in combination with fusion features to perform disease and emotion association analysis. The association pattern between emotion and digestive tract diseases is found through algorithms or models. The multimodal fusion features are input into MLP. The model can learn more complex digestive tract disease and emotion association patterns, and then find the relationship between emotion and digestive tract diseases, thereby improving the prediction accuracy. Furthermore, calculating ESS and EVI separately allows the model to learn and analyze the relationship between these two emotional characteristics and diseases respectively. For example, patients with long-term stable emotions may have a relatively low probability of developing certain chronic diseases; while patients with large emotional fluctuations may develop some stress-induced diseases in the short term. When calculated separately, the model can learn more about the complex patterns and relationships between emotions and diseases, thereby improving the generalization ability of the model, enabling it to better adapt to different patient groups and various actual situations, and providing a more accurate reference for the doctor's final diagnosis.

[0068] Furthermore, step six also includes: calculating the accuracy A, A= ;

[0069] Calculate the recall rate R, R= ;

[0070] Calculate the precision P, P= ;

[0071] Final evaluation value F1, F1= ;

[0072] Among them, TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example; TP, TN, FP, and FN are given by the predicted probability P (Y=1│ ) The output result value is generated;

[0073] The physician predicts the probability P (Y = 1 | ) output result is evaluated to obtain the first evaluation value; the machine obtains the second evaluation value through accuracy A, recall rate R, precision P, and final evaluation value F1.

[0074] From the above description, it can be seen that the first manual evaluation value and the second evaluation value calculated by the machine can make up for the shortcomings of both sides, namely the uncertainty of manual labor and the lack of experience or different depth of understanding of the machine, and it is also convenient for doctors to refer to.

[0075] Furthermore, if the result error between the first evaluation value and the second evaluation value is greater than a preset value, the doctor's judgment shall prevail.

[0076] Furthermore, the language processing model is a BERT model.

[0077] Furthermore, the language processing model is a BERT model, and an adversarial training mechanism is introduced into the BERT model.

[0078] From the above description, we can see that through the adversarial training mechanism, adversarial perturbations are added during the training process to improve the robustness of the model to text data, so that it can better handle noisy text and adversarial samples.

[0079] Furthermore, the image processing model is a ResNet model.

[0080] Furthermore, the attention module is introduced into the ResNet model.

[0081] From the above description, it can be seen that by introducing the attention module, irrelevant features can be suppressed and the pertinence and effectiveness of image feature extraction can be improved.

[0082] Furthermore, step five further includes: establishing a disease and emotion association analysis model, using multiple multi-layer perceptron models, and setting different initialization parameters for different multi-layer perceptron models.

[0083] From the above description, it can be seen that the multilayer perceptron model with different initialization parameters can combine the prediction results of different models by voting or weighted averaging, thereby reducing the variance of the model.

[0084] Furthermore, when normalizing in step 4, the IQR method is used to identify and correct or eliminate outliers.

[0085] From the above description, we can see that, unlike the normal distribution, the electronic data of gastroenterology, such as the patient's age, symptom indicators, emotional scores, etc., often do not meet the normal distribution through the interquartile range (IQR). The IQR method is used to ensure that the normalized data is more representative and reduce the interference of outliers on subsequent analysis.

[0086] Furthermore, the patient's age, gender and occupation were extracted separately from the text portion of the medical records and pathology reports.

[0087] From the above description, we can see that by extracting the patient's age, gender and occupation separately, we can find patterns in subsequent statistics and thus find possible statistical significance.

[0088] Embodiment 1

[0089] A gastroenterology electronic data analysis method, comprising:

[0090] Step 1: Generate semantic feature vectors. Collect the patient's medical records and pathology reports, and use a language processing model to generate semantic feature vectors S for the text parts of the medical records and pathology reports. The language processing model is a BERT model, and an adversarial training mechanism is introduced into the BERT model. For the text parts of the medical records and pathology reports, extract the patient's age, gender, and occupation separately. If necessary, further expand them, such as weight, region, etc.

[0091] Step 2: Generate an image feature vector. Generate an image feature vector V for the image part in the pathology report through an image processing model; the image processing model is a ResNet model, and an attention module is introduced into the ResNet model.

[0092] Step 3: Generate an emotion feature vector. In clinical practice, doctors extract emotion features from patients. Emotion feature extraction includes emotion stability score ESS and emotion fluctuation score EVI.

[0093] Emotional stability score ESS includes: ESS=

[0094] in, is the i-th emotion evaluation score, is the average emotion assessment score, n is the number of assessments, and ESS is used to measure the stability of emotions. The lower the score, the more stable the emotion;

[0095] Emotional volatility score EVI includes: EVI=

[0096] in, is the maximum value of the physician’s emotional assessment score for the patient. It is the minimum score of the physician's emotional assessment of the patient. EVI is used to measure the degree of emotional fluctuation of the patient. The higher the score, the greater the emotional fluctuation.

[0097] The reason why doctors specifically extract the information from patients here is that if a professional emotion assessment scale is used, patients may not understand the meaning of the professional terms. Even if they do, everyone's subjective feelings are different. For example, fright and panic in the Beck Anxiety Inventory are relatively subjective concepts, and the degree is difficult to judge. Therefore, doctors need to make special judgments on this part.

[0098] The emotional stability score ESS and the emotional volatility score EVI are combined to obtain the emotional feature vector E; the combination process includes: mapping the emotional stability score ESS and the emotional volatility score EVI to a multi-dimensional vector space, inputting a one-dimensional vector [ESS, EVI] composed of the emotional stability score ESS and the emotional volatility score EVI into the full neural network model, and transforming through multiple hidden layers:

[0099]

[0100]

[0101] E=

[0102] in, is the output of the first hidden layer of the neural network; is the output of the second hidden layer of the neural network; is the activation function, T represents the transposition operation; , , All are bias terms; is the weight matrix of the first hidden layer; is the weight matrix of the second hidden layer; is the weight matrix of the last fully connected layer;

[0103] Step 4: Feature fusion: fuse the semantic feature vector S, image feature vector V and emotion feature vector E to obtain the fused feature vector F; F = + + ;

[0104] in, is the attention weight of the semantic feature vector S; is the attention weight of the image feature vector V; is the attention weight of the emotion feature vector E;

[0105] calculate When a linear transformation is introduced, the semantic feature vector S is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ;

[0106] calculate When a linear transformation is introduced, the image feature vector V is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ;

[0107] calculate When a linear transformation is introduced, the emotion feature vector E is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ;

[0108] During the above normalization, the IQR method is used to identify and correct or eliminate outliers, including:

[0109] Calculate the quartiles: First, sort the data from small to large. Then determine three quartiles, namely the first quartile (Q1), the second quartile (Q2, also known as the median), and the third quartile (Q3). Q1 represents the value at the 25% position of the sorted data, which means that 25% of the data is less than or equal to it; Q2 is the value at the middle position, dividing the data in half; Q3 represents the value at the 75% position, which means that 75% of the data is less than or equal to it.

[0110] Calculate the quartile range: The calculation formula for IQR is: IQR = Q3-Q1. It represents the range of the middle 50% of the data. Compared with the full range (maximum value minus minimum value), IQR is not sensitive to extreme values ​​and can more robustly reflect the degree of discreteness of the data.

[0111] Determine the range of outliers: Use IQR to determine the boundaries of outliers. Commonly used boundaries are:

[0112] Lower limit: Q1-1.5 IQR

[0113] Upper limit: Q3+1.5 IQR

[0114] Any data point that is less than the lower limit or greater than the upper limit is considered an outlier. This 1.5 times IQR is a commonly used empirical value, which can effectively identify outliers that are significantly different from most of the data in many cases. Of course, adaptive adjustments can also be made based on data characteristics and analysis purposes. This article only uses this as an example to illustrate this.

[0115] Outlier processing: Identified outliers can be corrected or eliminated.

[0116] Step 5: Establish a disease and emotion correlation analysis model. Establish a disease and emotion correlation analysis model, use multiple multi-layer perceptron models, and set different initialization parameters for different multi-layer perceptron models.

[0117] The fused feature vector F is used as the input layer, and the predicted probability P (Y=1│ );

[0118] P(Y=1│ )= ( + + + )

[0119] Among them, P (Y = 1 | ) is given all the independent variables The probability that the dependent variable Y takes the value 1 under the condition that , that is, the probability of a specific disease state based on all disease- and emotion-related factors considered; is the activation function, is the bias term of the output layer; is the weight coefficient of the i-th element of the semantic feature vector S; is the i-th element of the semantic feature vector S; m is the number of elements after the semantic feature vector S is expanded into a one-dimensional vector; is the weight coefficient of the jth element of the image feature vector V; is the jth element of the image feature vector V; n is the number of elements after the image feature vector V is expanded into a one-dimensional vector; is the weight coefficient of the kth element of the image feature vector E; is the kth element of the emotion feature vector E; p is the number of elements after the emotion feature vector E is expanded into a one-dimensional vector;

[0120] Step 6: Effect evaluation: The physician evaluates the predicted probability P (Y=1│ ) output results for effect evaluation.

[0121] Step 6 also includes: calculating the accuracy A, A= ;

[0122] Calculate the recall rate R, R= ;

[0123] Calculate the precision P, P= ;

[0124] Final evaluation value F1, F1= ;

[0125] Among them, TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example; TP, TN, FP, and FN are given by the predicted probability P (Y=1│ ) The output result value is generated; we use 0.5 as the boundary;

[0126] True Positive (TP): If the model predicts that the patient is ill, and the patient is actually ill, this is a true positive. For example, if the model calculates that the predicted probability of a patient suffering from a digestive tract disease is 0.7, and the patient actually suffers from a digestive tract disease, this is a true positive. TP reflects the model's ability to correctly identify patients with illness.

[0127] True negative example (TN): When the model predicts that the patient is not sick and the patient is not sick, it is a true negative example. For example, if the model predicts that the probability of a patient is 0.3 and the patient does not actually have a digestive tract disease, this is a true negative example. TN reflects the model's ability to correctly identify patients who are not sick.

[0128] False positive (FP): If the model predicts that the patient is sick, but the patient is not sick, this is a false positive. For example, if the model predicts that the patient is sick, the probability is 0.6, but the patient does not have a digestive tract disease, this is a false positive. FP means that the model mistakenly predicts that a patient who is not sick is sick, which will lead to misdiagnosis.

[0129] False negative (FN): When the model predicts a patient with no disease and the patient is not sick, but the patient is actually sick, this is a false negative. For example, if the model predicts a patient with a probability of 0.4, but the patient actually has a digestive tract disease, this is a false negative. FN means that the model fails to identify patients who are actually sick, resulting in missed diagnoses.

[0130] Through the doctor's prediction probability P (Y = 1 | ) output result is evaluated to obtain the first evaluation value; the machine obtains the second evaluation value through accuracy A, recall rate R, precision P, and final evaluation value F1.

[0131] If the error between the first evaluation value and the second evaluation value is greater than a preset value, the doctor's judgment shall prevail.

[0132] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's specification and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for analyzing electronic data in gastroenterology, characterized in that: include: Step 1: Generate semantic feature vectors. Collect the patient's medical records and pathology reports, and use the language processing model to generate semantic feature vectors S for the text parts of the medical records and pathology reports. Step 2: Generate an image feature vector. Generate an image feature vector V for the image part in the pathology report through an image processing model. Step 3: Generate an emotion feature vector. In clinical practice, doctors extract emotion features from patients. Emotion feature extraction includes emotion stability score ESS and emotion fluctuation score EVI. Emotional stability score ESS includes: ESS= ,in, is the i-th emotion evaluation score, is the average emotion assessment score, n is the number of assessments, and ESS is used to measure the stability of emotions. The lower the score, the more stable the emotion; Emotional volatility score EVI includes: EVI= ,in, is the maximum value of the physician’s emotional assessment score for the patient. It is the minimum score of the physician's emotional assessment of the patient. EVI is used to measure the degree of emotional fluctuation of the patient. The higher the score, the greater the emotional fluctuation. The emotional stability score ESS and the emotional volatility score EVI are combined to obtain the emotional feature vector E; the combination process includes: mapping the emotional stability score ESS and the emotional volatility score EVI to a multi-dimensional vector space, inputting a one-dimensional vector [ESS, EVI] composed of the emotional stability score ESS and the emotional volatility score EVI into the full neural network model, and transforming through multiple hidden layers: , 、E= ;in, is the output of the first hidden layer of the neural network; is the output of the second hidden layer of the neural network; is the activation function, T represents the transposition operation; , , All are bias terms; is the weight matrix of the first hidden layer; is the weight matrix of the second hidden layer; is the weight matrix of the last fully connected layer; Step 4: Feature fusion: fuse the semantic feature vector S, image feature vector V and emotion feature vector E to obtain the fused feature vector F; F = + + ; in, is the attention weight of the semantic feature vector S; is the attention weight of the image feature vector V; is the attention weight of the emotion feature vector E; calculate When a linear transformation is introduced, the semantic feature vector S is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ; calculate When a linear transformation is introduced, the image feature vector V is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ; calculate When a linear transformation is introduced, the emotion feature vector E is transformed through the weight matrix and the bias term Transform to get a new scalar value ; = ; Through the Softmax function Normalize to get the attention weight of semantic features ; = ; Step 5: Establish a disease and emotion correlation analysis model, using a multi-layer perceptron model, taking the fusion feature vector F as the input layer and the predicted probability P (Y=1│ ); P(Y=1│ )= ( + + + ) Among them, P (Y = 1 | ) is given all the independent variables The probability that the dependent variable Y takes the value 1 under the condition that , that is, the probability of a specific disease state based on all disease- and emotion-related factors considered; is the activation function, is the bias term of the output layer; is the weight coefficient of the i-th element of the semantic feature vector S; is the i-th element of the semantic feature vector S; m is the number of elements after the semantic feature vector S is expanded into a one-dimensional vector; is the weight coefficient of the jth element of the image feature vector V; is the jth element of the image feature vector V; n is the number of elements after the image feature vector V is expanded into a one-dimensional vector; is the weight coefficient of the kth element of the image feature vector E; is the kth element of the emotion feature vector E; p is the number of elements after the emotion feature vector E is expanded into a one-dimensional vector; Step 6: Effect evaluation: The physician evaluates the predicted probability P (Y=1│ ) output results for effect evaluation.

2. The method for analyzing electronic data of gastroenterology according to claim 1, characterized in that: Step 6 also includes: calculating the accuracy A, A= ; Calculate the recall rate R, R= ; Calculate the precision P, P = ; Final evaluation value F1, F1= ; Among them, TP is a true positive example, TN is a true negative example, FP is a false positive example, and FN is a false negative example; TP, TN, FP, and FN are given by the predicted probability P (Y=1│ ) The output result value is generated; Through the doctor's prediction probability P (Y = 1 | ) output result is evaluated to obtain the first evaluation value; the machine obtains the second evaluation value through accuracy A, recall rate R, precision P, and final evaluation value F1.

3. The method for analyzing electronic data of gastroenterology according to claim 2, characterized in that: If the error between the first evaluation value and the second evaluation value is greater than a preset value, the doctor's judgment shall prevail.

4. The method for analyzing electronic data of gastroenterology according to claim 1, characterized in that: The language processing model is a BERT model.

5. The method for analyzing electronic data of gastroenterology according to claim 4, characterized in that: The language processing model is a BERT model, and an adversarial training mechanism is introduced into the BERT model.

6. The method for analyzing electronic data of gastroenterology according to claim 1, characterized in that: The image processing model is a ResNet model.

7. The method for analyzing electronic data of gastroenterology according to claim 6, characterized in that: Introduce the attention module in the ResNet model.

8. The method for analyzing electronic data of gastroenterology according to claim 1, characterized in that: Step five further includes: establishing a disease and emotion association analysis model, using multiple multi-layer perceptron models, and setting different initialization parameters for different multi-layer perceptron models.

9. The method for analyzing electronic data of gastroenterology according to claim 1, characterized in that: When normalizing in step 4, the IQR method is used to identify and correct or eliminate outliers.

10. The gastroenterology electronic data analysis method according to claim 1, characterized in that: The patient's age, gender, and occupation were extracted separately from the text portion of the medical records and pathology reports.

Images