Early dangerous stomachache recognition method based on machine learning

Abstract

The invention provides a dangerous stomachache early recognition method based on machine learning, and relates to the technical field of stomachache probability recognition, comprising the following steps: collecting stomachache data including symptom information, sign data, medical history information and examination reports of a patient; establishing an abdominal pain recognition model based on the abdominal pain data, the abdominal pain recognition model being used for processing symptom information, medical history information and an examination report of a patient through a natural language and processing sign data through numerical analysis to judge the probability of existence of a dangerous abdominal pain risk, training and evaluating whether the abdominal pain identification model needs to continue training optimization, if yes, continuing training, and if not, stopping training; and through the trained stomachache identification model, carrying out risk stomachache risk probability evaluation. The method has the advantage that the occurrence probability of dangerous stomachache can be comprehensively and rapidly evaluated.

Description

Technical Field

[0001] The present invention relates to the technical field of abdominal pain probability recognition, and more particularly, to an early recognition method for dangerous abdominal pain based on machine learning. Background Art

[0002] At present, abdominal pain is a common symptom in clinical practice, and its causes are complex and diverse, including gastrointestinal diseases, urinary system diseases, gynecological diseases, etc.

[0003] Dangerous abdominal pain refers to abdominal emergencies that may endanger life, such as acute pancreatitis, mesenteric artery embolism, etc., which require timely diagnosis and emergency treatment. Due to the rapid progression of these diseases, if not recognized and evaluated in time, the best treatment opportunity will be missed, leading to the aggravation of the condition and even endangering life. Existing technologies will assist medical staff in the early recognition of dangerous abdominal pain by means of intelligent data analysis, improving work efficiency and reducing the work burden of medical staff. However, existing technologies usually judge based on the patient's physical signs, and the types of collected data are limited, so the accuracy of evaluation is also limited.

[0004] Therefore, there is an urgent need for a systematic method that can utilize the multi-source heterogeneous data of patients to comprehensively and quickly evaluate the occurrence probability of dangerous abdominal pain, improving the efficiency and reliability of data processing. Summary of the Invention

[0005] The purpose of the present invention is to provide an early recognition method for dangerous abdominal pain based on machine learning, which can comprehensively and quickly evaluate the occurrence probability of dangerous abdominal pain.

[0006] The present invention is achieved through the following technical solutions: An early recognition method for dangerous abdominal pain based on machine learning, comprising the following steps: Collect abdominal pain data, where the abdominal pain data includes the patient's symptom information, physical sign data, medical history information, and examination reports; Based on the abdominal pain data, establish an abdominal pain recognition model, which is used to judge the probability of the risk of dangerous abdominal pain by processing the patient's symptom information, medical history information, and examination reports through natural language processing and processing the physical sign data through numerical analysis, train and evaluate whether the abdominal pain recognition model needs to be further trained and optimized. If so, continue training; if not, stop training; Evaluate the probability of the risk of dangerous abdominal pain through the trained abdominal pain recognition model.

[0007] Preferably, the method for collecting abdominal pain data includes: Collect the abdominal pain data through an electronic medical record system, a clinical research database, and wearable devices; Perform data cleaning and standardization processing on the collected abdominal pain data.

[0008] Preferably, the method for performing the data cleaning is to correct outliers and supplement missing values through a Kalman filter; The method for performing the normalization process is to perform Min-Max normalization or Z-score normalization.

[0009] Preferably, the abdominal pain recognition model includes: An input layer for inputting the abdominal pain data; A feature extraction layer for extracting features of the symptom information, the medical history information, and the examination report through a natural language model to obtain a first feature, and extracting features of the physical sign data through a non-linear function to obtain a second feature; A feature fusion layer for fusing the first feature and the second feature; An output layer for outputting whether there is a risk of dangerous abdominal pain according to the fused features.

[0010] Preferably, the method for extracting the first feature by extracting features of the symptom information, the medical history information, and the examination report through a natural language model is as follows: Establish a collection object library, which includes standard sentence vectors of standard description statements of multiple target feature data; Perform text cleaning on the sentences in the symptom information, the medical history information, and the examination report, and perform vectorization processing on each sentence through a BERT model to obtain multiple collection sentence vectors; Obtain the cosine distance between the i-th standard sentence vector and the j-th collection sentence vector , where i = 1, 2,..., N, j = 1, 2,..., M, and N and M are the total numbers of standard sentence vectors and collection sentence vectors respectively: ; Among them, and are the i-th standard sentence vector and the j-th collection sentence vector respectively and are both row vectors, represents taking the transpose; For each i-th standard sentence vector, put all collection sentence vectors whose cosine distance from it is less than a preset threshold into the i-th standard sentence set; Obtain the eigenvalue of the i-th standard sentence set :

[0011] Among them, represents the cosine similarity between the k-th collection sentence vector in the i-th standard sentence set and the i-th standard sentence vector, ​ is the total number of collected statement vectors in the i-th set of standard sentences; Obtain the first feature : .

[0012] Preferably, the method for extracting the features of the physical sign data through a non-linear function to obtain the second feature is: Perform the following processing on the k-th type of the physical sign data respectively: ; ; wherein, is the k-th type of the physical sign data is the characteristic value, and are the feature extraction weight and the feature extraction bias respectively, L is the total number of physical sign data, and tanh is the tanh activation function; Obtain the second feature : .

[0013] Preferably, the method for fusing the first feature and the second feature is: ; wherein, E is the fused feature.

[0014] Preferably, the method for outputting whether there is a risk of dangerous abdominal pain according to the fused feature is: Obtain the probability of having a risk of dangerous abdominal pain : ; wherein, and are the first training weight vector and the second training weight vector respectively and are row vectors, and are the first training bias and the second training bias respectively.

[0015] Preferably, the method for determining whether the recognition model needs to be further trained and optimized is: Obtain multiple groups of the abdominal pain data, output a judgment result through the recognition model, compare the judgment result with the actual result, if the accuracy reaches the preset threshold, it is determined that no further training and optimization are required, otherwise further training and optimization are still required.

[0016] Preferably, if further training and optimization are still required, include the abdominal pain data with incorrect judgment results in the training set for training.

[0017] The technical solution of the present invention has at least the following advantages and beneficial effects: The present invention integrates multi-source heterogeneous data such as symptom information, physical sign data, medical history information, and examination reports, realizes a comprehensive evaluation of multi-source data, avoids the limitations of data analysis brought by a single data source, and improves the recognition accuracy of the risk probability of dangerous abdominal pain by comprehensively analyzing various information; The present invention performs separate data processing and feature extraction on data with different structures, fully excavates the features of different modality information, can autonomously learn and optimize the influence of different data on dangerous abdominal pain, and improves the risk prediction ability of the model; The present invention performs semantic vectorization on text information such as medical history and symptoms, and then matches the patient's text with standard medical statements by calculating the cosine similarity, which can more accurately locate the feature data points to be extracted, and has a better normalization effect on the features extracted from different semantic information. At the same time, when processing numerical type data, the tanh non-linear transformation is used to enhance the fitting ability for complex relationships; The present invention uses regression logic to output the abdominal pain risk probability for the feature values from different modality data respectively, and then obtains the comprehensive abdominal pain risk probability through the average value, maintaining the integrity of the feature information extracted from multi-modal data, avoiding information imbalance, and at the same time facilitating the optimization of each modality of data according to its own characteristics, further improving the accuracy of their respective risk predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic flowchart of a method for early identification of dangerous abdominal pain based on machine learning provided by Embodiment 1 of the present invention; Figure 2 is a schematic structural diagram of an abdominal pain recognition model provided by Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and shown in the accompanying drawings here can be arranged and designed in various different configurations.

[0020] Embodiment 1 This embodiment provides a method for early identification of dangerous abdominal pain based on machine learning. Refer to Figure 1 , including the following steps: Collect abdominal pain data, where the abdominal pain data includes the patient's symptom information, physical sign data, medical history information, and examination reports; An abdominal pain recognition model is established based on the abdominal pain data. The abdominal pain recognition model is used to judge the probability of the risk of dangerous abdominal pain by processing the patient's symptom information, medical history information, and examination reports through natural language processing and processing the physical sign data through numerical analysis. Train and evaluate whether the abdominal pain recognition model needs to be further trained and optimized. If so, continue training; if not, stop training. Evaluate the probability of the risk of dangerous abdominal pain through the trained abdominal pain recognition model.

[0021] Among the specific data collected in this embodiment, as a preferred solution, the symptom information may include the location, nature, degree, duration, accompanying symptoms, etc. of the abdominal pain; the physical sign data may include body temperature, heart rate, blood pressure, abdominal palpation results, etc.; the medical history information may include past medical history, surgical history, allergy history, etc.; the examination report may include blood test, imaging examination results, etc.

[0022] Single data may cause inaccurate feature extraction. For example, relying only on physical sign data may miss some important medical history factors. For example, patients with a history of peptic ulcer may still have a high risk even if their physical signs are normal. Also, physical sign data may change due to factors such as tension, exercise, and medication, and single measurements may not be accurate. In this embodiment, the patient's symptom information, physical sign data, medical history information, and examination reports are collected, and the probability of the risk of dangerous abdominal pain is identified by extracting and fusing the features of multi-source heterogeneous data, which can be used as an intelligent auxiliary basis for medical staff to provide objective quantitative reference analysis.

[0023] Specifically, this embodiment comprehensively integrates various types of data. Among them, the symptom information can be derived from the patient's self-report. This data is more targeted and provides specific symptom characteristics. The disadvantage is that the lack of professionalism of the patient may lead to deviations in the expression; the physical sign data is an objective measurement, which helps to scientifically quantify the patient's physiological state. The disadvantage is that the influence of external factors on the patient's emotions, etc. may affect the physical signs and lead to inaccurate measurements; the medical history information and examination reports usually come from professional medical diagnoses and are highly professional. The disadvantage is that it cannot guarantee the comprehensive coverage of the key information required. Through the fusion of multi-source heterogeneous data in this embodiment, the multi-source heterogeneous data support and complement each other, and the comprehensive advantages make up for the defects of each other, avoiding the feature extraction deviation caused by information loss or data error and contradiction. Therefore, this embodiment can perform personalized analysis according to the characteristics of different patients and achieve a more targeted and flexible model output.

[0024] Finally, the trained abdominal pain recognition model can be applied. Input the current patient's abdominal pain data to be judged, and output the probability of the risk of dangerous abdominal pain through the abdominal pain recognition model.

[0025] In this embodiment, the method for collecting abdominal pain data includes: Collect the abdominal pain data through an electronic medical record system, a clinical research database, and wearable devices; Clean and standardize the collected abdominal pain data.

[0026] Specifically, the method for performing the data cleaning is to correct outliers and supplement missing values through a Kalman filter; The method for performing the standardization is to perform Min - Max normalization or Z - score standardization.

[0027] It should be noted that when collecting data, further broaden the data sources. In addition to the hospital's electronic medical record system, clinical research database, and wearable devices, case data from different regions and different hospitals can also be included, covering a more diverse patient population, including people of different ages, genders, and living environments, enabling the model to learn a wider range of abdominal pain - related feature patterns and enhancing the generalization ability of the model. For abdominal pain data, use domain knowledge for data augmentation. For example, perform synonym replacement on symptom descriptions, such as replacing "abdominal pain" with "abdominal ache"; for numerical data in inspection reports, make minor perturbations within a reasonable range to simulate measurement errors or differences between different detection devices. This increases the richness of the training data and improves the robustness of the model.

[0028] In the next step, refer to Figure 2 , the abdominal pain recognition model includes: An input layer for inputting the abdominal pain data; A feature extraction layer for extracting the features of the symptom information, the medical history information, and the inspection report through a natural language model to obtain a first feature, and extracting the features of the physical sign data through a non - linear function to obtain a second feature; A feature fusion layer for fusing the first feature and the second feature; An output layer for outputting whether there is a risk of dangerous abdominal pain based on the fused features.

[0029] Among the data collected in this embodiment, the symptom information, the medical history information, and the inspection report are text - format data described in natural language, and the physical sign data is numerical data. Therefore, the feature extraction layer will divide the multi - source heterogeneous data input by the input layer, that is, the abdominal pain data, into these two major categories and perform feature extraction processing on the data separately in different ways. Then, the features of the multi - source heterogeneous data are fused, and finally, the output layer outputs whether there is a risk of dangerous abdominal pain based on the fused features.

[0030] Furthermore, the method for extracting the features of the symptom information, the medical history information, and the inspection report through a natural language model to obtain the first feature is: Establish a collection object library, where the collection object library includes standard sentence vectors of multiple standard description statements of target feature data; Clean the text of the statements in the symptom information, the medical history information, and the inspection report, and perform vectorization processing on each sentence through a BERT model to obtain multiple collected sentence vectors; Obtain the cosine distance between the i-th standard sentence vector and the j-th collected sentence vector , i = 1, 2, …, N, j = 1, 2, …, M, where N and M are the total numbers of standard sentence vectors and collected sentence vectors respectively: ; wherein, and are respectively the i-th standard sentence vector and the j-th collected sentence vector and are both row vectors, represents taking the transpose; For each i-th standard sentence vector, put all the collected sentence vectors whose cosine distance from it is less than a preset threshold into the i-th standard sentence set; Respectively obtain the eigenvalues of the i-th standard sentence set :

[0031] wherein, represents the cosine similarity between the k-th collected sentence vector in the i-th standard sentence set and the i-th standard sentence vector, is the total number of collected sentence vectors in the i-th standard sentence set; Obtain the first feature : : .

[0032] In the above steps, a collection object library is set up. The collection object library includes standard description statements of multiple target feature data. That is to say, it is determined in advance which non-numerical indicators need to be referred to when medical staff generally conduct early assessments of dangerous abdominal pain. Then, a standard description statement is established for each of these non-numerical indicators. Usually, it is a description representing the health status. Using such a statement as a benchmark to construct features. For example, assume the following non-numerical indicators are required: whether there is pain in part A, whether there has been disease B, whether there is sweating, whether there are vomiting symptoms, whether there is an allergy to C, whether there are abnormalities such as intestinal obstruction or tumors, and whether liver function is normal. Then the established standard description statements can include: there is no pain in part A, there has been no disease B, there is no sweating, there are no vomiting symptoms, there is no allergy to C, there are no abnormalities such as intestinal obstruction or tumors, and liver function is normal. The non-numerical indicators to be referred to can be obtained from symptom information and / or medical history information and / or inspection reports respectively.

[0033] Therefore, the standard description statement of the target feature data is converted into a standard sentence vector, and the sentences from symptom information, medical history information, and inspection reports are converted into collected sentence vectors. Then, the similarity between the collected sentence vectors and the standard sentence vectors is compared one by one. If the j-th collected sentence vector is similar enough to the i-th standard sentence vector, it indicates that the non-numerical index types described by the two are the same. Therefore, the j-th collected sentence vector is classified into the i-th standard sentence set, and the i-th standard sentence set stores all the collected sentence vectors converted from the sentences that are judged to express the non-numerical index corresponding to the i-th standard sentence vector.

[0034] In this embodiment, the overall similarity between all the collected sentence vectors stored in the i-th standard sentence set and the i-th standard sentence vector is collected as the feature data of the non-numerical index corresponding to the i-th standard sentence vector to express the deviation between the actual situation and the standard situation of the patient. Comprehensive calculation of multi-source data can weaken the error caused by expression errors and also has a wider coverage, ensuring that all required non-numerical indexes can be covered. Finally, the feature data of all non-numerical indexes can be obtained to form the first feature.

[0035] On the other hand, the method for extracting the features of the physical sign data through a non-linear function to obtain the second feature is as follows: The following processing is performed on the k-th type of the physical sign data respectively: ; ; where is the k-th type of the physical sign data is the characteristic value, and are the feature extraction weight and the feature extraction bias respectively, L is the total number of physical sign data, and tanh is the tanh activation function; Obtain the second feature : .

[0036] The physical sign data here are the numerical indexes that medical staff generally need to refer to when conducting an early assessment of dangerous abdominal pain. For example, it can include parameters such as body temperature, heart rate, and blood pressure that can be measured and expressed numerically. Then, the features are extracted through the difference between the numerical value and the standard value, and finally, the features of all numerical indexes are integrated to form the second feature.

[0037] Furthermore, the method for fusing the first feature and the second feature is as follows: ; where E is the fused feature.

[0038] After obtaining the fused features, the method for outputting whether there is a risk of dangerous abdominal pain based on the fused features is as follows: Obtain the probability of having a risk of dangerous abdominal pain : ; wherein, and are the first training weight vector and the second training weight vector respectively and are row vectors, and are the first training bias and the second training bias respectively.

[0039] In this embodiment, the first feature and the second feature are directly concatenated for fusion. Text information belongs to high-dimensional sparse data, and the extracted features are usually relatively complex. On the contrary, numerical data is relatively compact. Problems such as poor data matching degree resulting in data feature loss or data conflict may occur when calculating the fused features. In this embodiment, the probabilities of the first feature and the second feature are calculated separately and then fused, so that the contributions of different data sources to the final decision are more reasonable. For example, if the degree of abdominal pain described by the patient is relatively mild, but the physical signs are extremely abnormal, direct concatenation may lead to blurred model features.

[0040] The above solution adopts the method of independent calculation and then averaging, which is equivalent to letting the features of different modalities "vote" separately, thereby reducing the influence of the errors of a single modality on the final result and improving the overall generalization ability. At the same time, the fault tolerance rate is also enhanced. For example, when an error occurs in any link of data processing of one modality, the other modality of the model can still make a relatively accurate decision, improving the stability of the prediction result. Since the final abdominal pain risk probability is obtained by synthesizing the prediction probabilities of different modalities, medical staff can view the risk probability of text analysis and the risk probability of physical sign data analysis separately, so as to more clearly understand the contributions of each data source to the final decision, which helps medical staff to make a comprehensive judgment in combination with experience when making a decision, avoid completely relying on the black box model, and enhance the interpretability of the model.

[0041] Finally, the method for identifying whether the recognition model needs to be further trained and optimized is as follows: Obtain multiple groups of the abdominal pain data, output a judgment result through the recognition model, compare the judgment result with the actual result. If the accuracy reaches the preset threshold, it is judged that there is no need to continue training and optimization, otherwise, it is still necessary to continue training and optimization.

[0042] Particularly, if it is still necessary to continue training and optimization, the abdominal pain data with incorrect judgment results will be incorporated into the training set for training.

[0043] In addition, during training, the learning rate can be adaptively updated through the RMSProp adaptive learning rate algorithm. The RMSProp adaptive learning rate algorithm can dynamically adjust the learning rate of each parameter based on the update history of the parameters. For models with different numbers of layers, they can automatically adapt to the complexity of the model without the need to manually adjust the learning rate frequently. The algorithm will automatically adjust according to the changes in the parameter gradients. The early abdominal pain recognition data may have a certain degree of sparsity, and some symptoms or examination indicators may be missing or rarely appear in some cases. The RMSProp algorithm can effectively handle this sparse data. Since it adjusts the learning rate based on the moving average of the squares of the gradients, for sparse features, their gradients are relatively small in most cases, and RMSProp will assign larger learning rates to these features, enabling the model to more fully learn these sparse but potentially important feature information, adapt to the non-stationarity of the data, avoid gradient vanishing and explosion, enhance the generalization ability of the model. Appropriate learning rate adjustment helps the model better learn the essential features of the data and avoid overfitting or underfitting. The RMSProp algorithm can enable the model to converge more stably to the optimal solution during training, thereby improving the generalization ability of the model to unknown data. In the early abdominal pain recognition task, this means that the model can more accurately identify and diagnose new abdominal pain cases, thus enabling the model to be trained more stably. In one implementation case, its update rule is as follows: Initialize parameters , set the learning rate and the decay coefficient , usually the decay coefficient is set to 0.9, set a small constant for numerical stability , usually set to ; In each iteration, calculate the gradient of the parameter ; Update the exponentially weighted moving average r of the cumulative squared gradient:

[0044] Calculate the parameter update amount : ; Finally, update the parameter .

[0045] When optimizing the model, the optimization methods include increasing the training data, modifying hyperparameters, etc.

[0046] Specifically, the possible defects are as follows: the model has insufficient generalization ability. That is, if the model performs well on the training set but the accuracy, recall rate and other indicators significantly decrease on the test set or new data, it indicates that the model has poor generalization ability. For example, in the early abdominal pain recognition model, the recognition accuracy is high on the training data of local hospitals, but the performance drops significantly when applied to the data of other regional hospitals. This may be because the case types and patient characteristics covered by the training data are not comprehensive enough; or the data distribution is unbalanced. That is, when the sample numbers of different categories (such as different causes of abdominal pain) in the training data vary greatly, the model may be biased towards the category with a large number of samples, resulting in poor recognition ability for the category with a small number of samples. For example, in abdominal pain recognition, the sample number of a certain rare cause of abdominal pain is extremely small, and it is difficult for the model to learn the characteristics of this category. Both of the above two defects can be solved by expanding the data source, cooperating with more hospitals to obtain case data, or collecting relevant information from public medical databases to increase the diversity and quantity of training data. For abdominal pain data, some data augmentation techniques can be used. For example, replace synonyms for symptom descriptions and change the sentence order; for examination index data, make small perturbations within a reasonable range to simulate measurement errors.

[0047] During the model training process, if the model converges too slowly, overfitting or underfitting occurs, it may be necessary to modify the hyperparameters. For example, too many training epochs may lead to overfitting, and an inappropriate learning rate setting will affect the model convergence speed and effect. The initial learning rate is set to 1e - 5, the batch size is 32, and the number of training epochs is 10. If the model converges slowly, the learning rate can be appropriately increased; if overfitting occurs, the number of training epochs can be reduced. Regarding the hyperparameters as variables, calculate the gradient of the model loss with respect to the hyperparameters through the backpropagation algorithm, and then adjust the values of the hyperparameters according to the gradient direction. The advantages are high computational efficiency and fast convergence speed.

[0048] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for early identification of dangerous abdominal pain based on machine learning, characterized in that: The following steps are involved: Collecting abdominal pain data, wherein the abdominal pain data includes the patient's symptom information, physical sign data, medical history information and examination report; An abdominal pain recognition model is established based on the abdominal pain data, wherein the abdominal pain recognition model is used to judge the probability of the existence of dangerous abdominal pain risk by natural language processing of the patient's symptom information, medical history information and examination report and by numerical analysis processing of the physical sign data, and to train and evaluate whether the abdominal pain recognition model needs to continue training and optimization, and if so, continue training, otherwise stop training; The probability of dangerous abdominal pain risk is assessed using the trained abdominal pain recognition model.

2. The method for early identification of dangerous abdominal pain based on machine learning according to claim 1, characterized in that: The method for collecting abdominal pain data includes: Collecting the abdominal pain data through an electronic medical record system, a clinical research database, and a wearable device; The collected abdominal pain data are cleaned and standardized.

3. The method for early identification of dangerous abdominal pain based on machine learning according to claim 2 is characterized in that: The method for performing the data cleaning is to correct outliers and supplement missing values ​​through a Kalman filter; The method for performing the standardization processing is to perform Min-Max normalization processing or Z-score standardization processing.

4. The method for early identification of dangerous abdominal pain based on machine learning according to claim 1, characterized in that: The abdominal pain recognition model includes: An input layer, used for inputting the abdominal pain data; A feature extraction layer, used for extracting the features of the symptom information, the medical history information and the examination report through a natural language model to obtain a first feature, and extracting the features of the vital sign data through a nonlinear function to obtain a second feature; A feature fusion layer, used to fuse the first feature and the second feature; The output layer is used to output whether there is a risk of dangerous abdominal pain based on the fused features.

5. The method for early identification of dangerous abdominal pain based on machine learning according to claim 4 is characterized in that: The method for obtaining the first feature by extracting the features of the symptom information, the medical history information and the examination report through a natural language model is: Establishing a collection object library, wherein the collection object library includes standard sentence vectors of standard description sentences of a plurality of target feature data; Performing text cleaning on the sentences in the symptom information, the medical history information, and the examination report, and performing vectorization processing on each sentence using a BERT model to obtain multiple collected sentence vectors; Get the cosine distance between the i-th standard sentence vector and the j-th collected sentence vector , i=1,2,…,N,j=1,2,…,M,N and M are the total number of standard sentence vectors and collected sentence vectors respectively: ； in, and They are the i-th standard sentence vector and the j-th collected sentence vector, and both are row vectors. represents transpose; For each i-th standard sentence vector, all vectors whose distance from their cosine is less than the preset threshold The collected sentence vector is put into the i-th standard sentence set; Get the feature values ​​of the i-th standard sentence set respectively : in, Represents the first sentence in the i-th standard sentence set The cosine similarity between the collected sentence vector and the i-th standard sentence vector, is the total number of collected sentence vectors in the i-th standard sentence set; Get the first feature : 。 6. The method for early identification of dangerous abdominal pain based on machine learning according to claim 5, characterized in that: The method of extracting the features of the vital sign data by a nonlinear function to obtain the second feature is: The following processing is performed on the k-th type of physical sign data respectively: ; ； in, is the kth type of physical sign data The characteristic value of and are feature extraction weight and feature extraction bias respectively, L is the total number of vital sign data, and tanh is the tanh activation function; Get the second feature : 。 7. The method for early identification of dangerous abdominal pain based on machine learning according to claim 6 is characterized in that: The method of fusing the first feature and the second feature is: ； Among them, E is the fused feature.

8. The method for early identification of dangerous abdominal pain based on machine learning according to claim 7, characterized in that: The method for outputting whether there is a risk of dangerous abdominal pain based on the fused features is: Get the probability of having dangerous abdominal pain risk : ； in, and are the first training weight vector and the second training weight vector respectively and are row vectors, and They are the first training bias and the second training bias respectively.

9. The method for early identification of dangerous abdominal pain based on machine learning according to claim 1, characterized in that: The method for identifying whether the model needs to continue training and optimization is: Acquire multiple groups of abdominal pain data, output judgment results through the recognition model, compare the judgment results with the actual results, and if the accuracy reaches a preset threshold, it is determined that there is no need to continue training optimization, otherwise training optimization still needs to be continued.

10. The method for early identification of dangerous abdominal pain based on machine learning according to claim 9, characterized in that: If further training optimization is needed, the abdominal pain data with erroneous judgment results will be included in the training set for training.

Images