A Machine Learning-Based Method for Early Identification of Dangerous Abdominal Pain

Abstract

This invention provides a machine learning-based method for early identification of dangerous abdominal pain, relating to the field of abdominal pain probability recognition technology. The method includes the following steps: collecting abdominal pain data, including patient symptom information, physical signs data, medical history information, and examination reports; establishing an abdominal pain recognition model based on the abdominal pain data, wherein the model uses natural language processing of patient symptom information, medical history information, and examination reports, and numerical analysis of physical signs data to determine the probability of dangerous abdominal pain; training and evaluating whether the abdominal pain recognition model needs further training and optimization; if so, continuing training; otherwise, stopping training; and using the trained abdominal pain recognition model to assess the probability of dangerous abdominal pain. This invention has the advantage of achieving a comprehensive and rapid assessment of the probability of dangerous abdominal pain.

Description

Technical Field

[0001] This invention relates to the field of abdominal pain probability recognition technology, and more specifically, to a method for early identification of dangerous abdominal pain based on machine learning. Background Technology

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

[0003] Dangerous abdominal pain refers to potentially life-threatening abdominal emergencies, such as acute pancreatitis and mesenteric artery embolism, requiring timely diagnosis and emergency treatment. Because these conditions progress rapidly, failure to identify and assess them promptly can lead to missed opportunities for optimal treatment, resulting in worsening of the condition and even death. Existing technologies utilize intelligent data analysis to assist medical personnel in the early identification of dangerous abdominal pain, improving efficiency and reducing the workload of medical staff. However, current technologies typically rely on patient vital signs for assessment, and the limited types of data collected result in limited accuracy.

[0004] Therefore, there is an urgent need for a systematic approach that can utilize multi-source heterogeneous patient data to achieve a comprehensive and rapid assessment of the probability of dangerous abdominal pain, thereby improving the efficiency and reliability of data processing. Summary of the Invention

[0005] The purpose of this invention is to provide a method for early identification of dangerous abdominal pain based on machine learning, which can achieve a comprehensive and rapid assessment of the probability of occurrence of dangerous abdominal pain.

[0006] This invention is achieved through the following technical solution:

[0007] A machine learning-based method for early identification of dangerous abdominal pain includes the following steps:

[0008] Collect abdominal pain data, which includes the patient's symptom information, physical signs data, medical history information, and examination reports;

[0009] An abdominal pain recognition model is established based on the abdominal pain data. The abdominal pain recognition model is used to determine the probability of dangerous abdominal pain risk by processing the patient's symptom information, medical history information and examination report through natural language processing and processing vital sign data through numerical analysis. The model is trained and evaluated to determine whether it needs to be further trained and optimized. If so, training continues; otherwise, training is stopped.

[0010] The probability of dangerous abdominal pain is assessed by using a trained abdominal pain recognition model.

[0011] Preferably, the method for collecting abdominal pain data includes:

[0012] The abdominal pain data was collected through electronic medical record systems, clinical research databases, and wearable devices.

[0013] The collected abdominal pain data were cleaned and standardized.

[0014] Preferably, the data cleaning method involves using a Kalman filter to correct outliers and fill in missing values.

[0015] The standardization process is performed by performing Min-Max normalization or Z-score standardization.

[0016] Preferably, the abdominal pain recognition model includes:

[0017] An input layer is used to input the abdominal pain data;

[0018] The feature extraction layer is used to extract features from the symptom information, medical history information and examination report through a natural language model to obtain the first feature, and to extract features from the vital signs data through a nonlinear function to obtain the second feature.

[0019] The feature fusion layer is used to fuse the first feature and the second feature.

[0020] The output layer is used to output whether there is a risk of dangerous abdominal pain based on the fused features.

[0021] Preferably, the method for obtaining the first feature by extracting features from the symptom information, the medical history information, and the examination report using a natural language model is as follows:

[0022] Establish a collection object library, which includes standard sentence vectors of standard description statements for multiple target feature data;

[0023] The symptom information, medical history information and statements in the examination report are cleaned, and each sentence is vectorized using the BERT model to obtain multiple collected sentence vectors.

[0024] Obtain the cosine distance between the i-th standard sentence vector and the j-th collected sentence vector. Let i = 1, 2, ..., N, j = 1, 2, ..., M, and N and M be the total number of standard sentence vectors and the total number of collected sentence vectors, respectively.

[0025] ;

[0026] in, and These are the vectors of the i-th standard sentence and the j-th collected sentence, respectively, and both are row vectors. This represents the transpose;

[0027] For the i-th standard sentence vector, select all vectors whose distance from the cosine is less than a preset threshold. The collected sentence vectors are placed into the i-th standard sentence set;

[0028] Obtain the feature values ​​of the i-th set of standard sentences respectively. :

[0029]

[0030] in, Represents the i-th standard sentence set. The cosine similarity between the collected sentence vector and the i-th standard sentence vector. The total number of collected statement vectors in the i-th standard sentence set;

[0031] Obtain the first feature :

[0032] .

[0033] Preferably, the method for obtaining the second feature by extracting features from the vital sign data using a nonlinear function is as follows:

[0034] The following processing is performed on the k-th type of vital sign data respectively:

[0035] ;

[0036] ;

[0037] in, For the kth type of vital sign data The characteristic values, and These are the feature extraction weights and feature extraction biases, respectively; L is the total number of vital sign data; and tanh is the tanh activation function.

[0038] Obtaining the second feature :

[0039] .

[0040] Preferably, the method for fusing the first feature and the second feature is as follows:

[0041] ;

[0042] Where E represents the fused features.

[0043] Preferably, the method for determining whether the fused feature output carries a risk of dangerous abdominal pain is as follows:

[0044] Obtain the probability of having a dangerous abdominal pain risk. :

[0045] ;

[0046] in, and These are the first and second training weight vectors, respectively, and are row vectors. and These are the first training bias and the second training bias, respectively.

[0047] Preferably, the method for determining whether the recognition model needs further training and optimization is as follows:

[0048] Multiple sets of abdominal pain data are acquired, and the judgment result is output through the recognition model. The judgment result is compared with the actual result. If the accuracy reaches a preset threshold, it is determined that no further training and optimization is needed; otherwise, further training and optimization are needed.

[0049] Preferably, if further training and optimization are needed, the abdominal pain data with incorrect judgment results are included in the training set for training.

[0050] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0051] This invention integrates multi-source heterogeneous data such as symptom information, physical signs data, medical history information, and examination reports to achieve a comprehensive evaluation of multi-source data, avoiding the limitations of data analysis caused by a single data source. By comprehensively analyzing multiple types of information, it improves the accuracy of identifying the probability of dangerous abdominal pain.

[0052] This invention performs separate data processing and feature extraction on data with different structures, fully explores the features of different modal information, and can autonomously learn and optimize the impact of different data on dangerous abdominal pain, thereby improving the model's risk prediction ability.

[0053] This invention performs semantic vectorization on textual information such as medical history and symptoms, and then matches patient text with standard medical statements by calculating cosine similarity. This can more accurately locate the feature data points to be extracted, and also has a better normalization effect on features extracted from different semantic information. At the same time, when processing numerical data, tanh nonlinear transformation is used to enhance the fitting ability of complex relationships.

[0054] This invention uses regression logic to output the probability of abdominal pain risk for feature values ​​from different modal data, and then calculates the comprehensive probability of abdominal pain risk by averaging the values. This maintains the integrity of the feature information extracted from multimodal data, avoids information imbalance, and also makes it easier for each modality's data to be optimized according to its own characteristics, thereby further improving the accuracy of its respective risk prediction. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating a method for early identification of dangerous abdominal pain based on machine learning, as provided in Embodiment 1 of the present invention.

[0056] Figure 2 This is a schematic diagram of the abdominal pain recognition model provided in Embodiment 1 of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0058] Example 1

[0059] This embodiment provides a machine learning-based method for early identification of dangerous abdominal pain. (See attached document.) Figure 1 This includes the following steps:

[0060] Collect abdominal pain data, which includes the patient's symptom information, physical signs data, medical history information, and examination reports;

[0061] An abdominal pain recognition model is established based on the abdominal pain data. The abdominal pain recognition model is used to determine the probability of dangerous abdominal pain risk by processing the patient's symptom information, medical history information and examination report through natural language processing and processing vital sign data through numerical analysis. The model is trained and evaluated to determine whether it needs to be further trained and optimized. If so, training continues; otherwise, training is stopped.

[0062] The probability of dangerous abdominal pain is assessed by using a trained abdominal pain recognition model.

[0063] In this embodiment, the specific data collected may include, as a preferred option, symptom information such as the location, nature, degree, duration, and accompanying symptoms of abdominal pain; physical signs data such as body temperature, heart rate, blood pressure, and abdominal palpation results; medical history information such as past medical history, surgical history, and allergy history; and examination reports such as blood test results and imaging examination results.

[0064] Single data sets may lead to inaccurate feature extraction. For example, relying solely on physical signs may miss important medical history factors. Patients with a history of peptic ulcers may still have a higher risk even if their physical signs are normal. Furthermore, physical signs can change due to factors such as stress, exercise, and medication, making a single measurement potentially inaccurate. This embodiment collects patients' symptom information, physical signs data, medical history information, and examination reports. By extracting and fusing features from multi-source heterogeneous data, it identifies the probability of having dangerous abdominal pain, providing an intelligent auxiliary basis for medical staff and offering objective quantitative reference analysis.

[0065] Specifically, this embodiment comprehensively integrates multiple types of data. Symptom information can come from patient self-reports, which are more targeted and provide specific symptom characteristics. However, the patient's lack of expertise may lead to inaccurate descriptions. Physical signs data are objective measurements, helping to scientifically quantify the patient's physiological state. However, external factors such as the patient's emotions may affect the physical signs, leading to inaccurate measurements. Medical history information and examination reports usually come from professional medical diagnoses and are highly specialized, but they cannot guarantee comprehensive coverage of key information. This embodiment, through the fusion of multi-source heterogeneous data, allows these data to support and complement each other, combining their advantages and compensating for their shortcomings, avoiding feature extraction biases caused by missing information or contradictory data errors. Therefore, this embodiment can perform personalized analysis based on the characteristics of different patients, achieving more targeted and flexible model output.

[0066] Once the abdominal pain recognition model is trained, it can be applied by inputting the abdominal pain data of the patient to be judged, and outputting the probability of dangerous abdominal pain risk through the abdominal pain recognition model.

[0067] In this embodiment, the method for collecting abdominal pain data includes:

[0068] The abdominal pain data was collected through electronic medical record systems, clinical research databases, and wearable devices.

[0069] The collected abdominal pain data were cleaned and standardized.

[0070] Specifically, the data cleaning method involves using a Kalman filter to correct outliers and fill in missing values.

[0071] The standardization process is performed by performing Min-Max normalization or Z-score standardization.

[0072] It is worth noting that data sources were further broadened during data collection. In addition to hospital electronic medical record systems, clinical research databases, and wearable devices, case data from different regions and hospitals could be incorporated, covering a more diverse patient population, including people of different ages, genders, and living environments. This allows the model to learn a wider range of abdominal pain-related feature patterns, enhancing its generalization ability. For abdominal pain data, domain knowledge was used for data augmentation. For example, synonym replacement was performed on symptom descriptions, such as replacing "abdominal pain" with "abdominal ache"; and minor perturbations were made to numerical data in examination reports within a reasonable range to simulate measurement errors or differences between different testing devices. This increased the richness of the training data and improved the model's robustness.

[0073] In the following steps, please refer to Figure 2 The abdominal pain recognition model includes:

[0074] An input layer is used to input the abdominal pain data;

[0075] The feature extraction layer is used to extract features from the symptom information, medical history information and examination report through a natural language model to obtain the first feature, and to extract features from the vital signs data through a nonlinear function to obtain the second feature.

[0076] The feature fusion layer is used to fuse the first feature and the second feature.

[0077] The output layer is used to output whether there is a risk of dangerous abdominal pain based on the fused features.

[0078] In this embodiment, the symptom information, medical history information, and examination reports are text-based data described in natural language, while the vital signs data are numerical data. Therefore, the feature extraction layer divides the multi-source heterogeneous data (abdominal pain data) input from the input layer into these two categories and performs feature extraction processing on each data separately using different methods. 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.

[0079] Furthermore, the method for obtaining the first feature by extracting features from the symptom information, the medical history information, and the examination report using a natural language model is as follows:

[0080] Establish a collection object library, which includes standard sentence vectors of standard description statements for multiple target feature data;

[0081] The symptom information, medical history information and statements in the examination report are cleaned, and each sentence is vectorized using the BERT model to obtain multiple collected sentence vectors.

[0082] Obtain the cosine distance between the i-th standard sentence vector and the j-th collected sentence vector. Let i = 1, 2, ..., N, j = 1, 2, ..., M, and N and M be the total number of standard sentence vectors and the total number of collected sentence vectors, respectively.

[0083] ;

[0084] in, and These are the vectors of the i-th standard sentence and the j-th collected sentence, respectively, and both are row vectors. This represents the transpose;

[0085] For the i-th standard sentence vector, select all vectors whose distance from the cosine is less than a preset threshold. The collected sentence vectors are placed into the i-th standard sentence set;

[0086] Obtain the feature values ​​of the i-th set of standard sentences respectively. :

[0087]

[0088] in, Represents the i-th standard sentence set. The cosine similarity between the collected sentence vector and the i-th standard sentence vector. The total number of collected statement vectors in the i-th standard sentence set;

[0089] Obtain the first feature :

[0090] .

[0091] In the steps above, a data collection target library was established. This library includes standard descriptive statements for multiple target feature data. In other words, it pre-determines which non-numerical indicators medical personnel typically need to refer to when conducting early assessments of dangerous abdominal pain. Then, a standard descriptive statement is created for each of these non-numerical indicators, usually representing a description of health status. These statements serve as the basis for feature construction. For example, suppose the following non-numerical indicators are needed: pain in area A, history of disease B, cold sweats, vomiting, allergy to C, intestinal obstruction or tumor abnormalities, and normal liver function. Then the standard descriptive statements could include: no pain in area A, no history of disease B, no cold sweats, no vomiting, no allergy to C, no intestinal obstruction or tumor abnormalities, and normal liver function. The required non-numerical indicators can be obtained from symptom information and / or medical history information and / or examination reports.

[0092] Therefore, the standard descriptive statements of the target feature data are transformed into standard sentence vectors, and the statements from symptom information, medical history information, and examination reports are transformed 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 and the ith standard sentence vector are sufficiently similar, it means that the non-numerical indicator types they describe should be the same. Therefore, the j-th collected sentence vector is assigned to the ith standard sentence set. The ith standard sentence set stores all the collected sentence vectors transformed from statements that are judged to describe the non-numerical indicators corresponding to the ith standard sentence vector.

[0093] This embodiment uses the sum of similarity between the vectors of all collected sentences stored in the i-th standard sentence set and the vector of the i-th standard sentence as feature data of the non-numerical indicator corresponding to the i-th standard sentence vector. This represents the deviation between the patient's actual situation and the standard situation. Comprehensive calculation of data from multiple sources can reduce the error caused by expression errors and has a wider coverage, ensuring that all necessary non-numerical indicators are covered. Finally, feature data of all non-numerical indicators can be obtained to form the first feature.

[0094] On the other hand, the method for obtaining the second feature by extracting features from the vital sign data through a nonlinear function is as follows:

[0095] The following processing is performed on the k-th type of vital sign data respectively:

[0096] ;

[0097] ;

[0098] in, For the kth type of vital sign data The characteristic values, and These are the feature extraction weights and feature extraction biases, respectively; L is the total number of vital sign data; and tanh is the tanh activation function.

[0099] Obtaining the second feature :

[0100] .

[0101] The vital signs data here are the numerical indicators that medical personnel generally need to refer to when conducting early assessment of dangerous abdominal pain. For example, they may include parameters that can be measured and expressed numerically, such as body temperature, heart rate, and blood pressure. Then, features are extracted by the difference between the numerical values ​​and the standard values, and finally, the features of all numerical indicators are integrated to form a second feature.

[0102] Furthermore, the method for fusing the first feature and the second feature is as follows:

[0103] ;

[0104] Where E represents the fused features.

[0105] 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:

[0106] Obtain the probability of having a dangerous abdominal pain risk. :

[0107] ;

[0108] in, and These are the first and second training weight vectors, respectively, and are row vectors. and These are the first training bias and the second training bias, respectively.

[0109] This embodiment directly combines the first and second features for fusion. Textual information is high-dimensional sparse data, and the extracted features are usually quite complex, whereas numerical data is relatively compact. Calculating features after fusion may result in poor data matching, leading to feature loss or data conflicts. This embodiment calculates the probabilities of the first and second features separately before fusion, making the contributions of different data sources to the final decision more reasonable. For example, if a patient's reported abdominal pain is mild but their physical signs are severe, direct fusion might lead to blurred model features.

[0110] The above scheme employs an independent calculation followed by averaging method, essentially allowing features from different modalities to "vote" separately. This reduces the impact of errors from a single modality on the final result, improving overall generalization ability. It also enhances fault tolerance; for example, if an error occurs at any stage of data processing in one modality, another modality can still make a relatively accurate decision, improving the stability of the prediction results. Since the final probability of abdominal pain risk is derived from the combined prediction probabilities of different modalities, medical personnel can separately view the risk probabilities from text analysis and vital sign data analysis. This allows for a clearer understanding of the contribution of each data source to the final decision, helping medical personnel to make comprehensive judgments based on experience, avoiding complete reliance on a black-box model, and enhancing the model's interpretability.

[0111] Finally, the method for determining whether the recognition model needs further training and optimization is as follows:

[0112] Multiple sets of abdominal pain data are acquired, and the judgment result is output through the recognition model. The judgment result is compared with the actual result. If the accuracy reaches a preset threshold, it is determined that no further training and optimization is needed; otherwise, further training and optimization are needed.

[0113] Specifically, if further training and optimization are needed, the abdominal pain data with incorrect judgment results should be included in the training set for training.

[0114] Furthermore, during training, the RMSProp adaptive learning rate algorithm can be used to adaptively update the learning rate. The RMSProp algorithm dynamically adjusts the learning rate of each parameter based on its update history. For models with different numbers of layers, it can automatically adapt to the model's complexity without requiring frequent manual adjustments to the learning rate; the algorithm automatically adjusts based on changes in parameter gradients. Early abdominal pain identification data may exhibit sparsity, with certain symptoms or examination indicators possibly missing or rarely appearing in some cases. The RMSProp algorithm can effectively handle this sparse data. Because it adjusts the learning rate based on the moving average of the squared gradient, for sparse features, whose gradients are generally small, RMSProp assigns a larger learning rate to these features, enabling the model to learn these sparse but potentially important feature information more fully. This adapts to the non-stationarity of the data, avoids gradient vanishing and exploding, and enhances the model's generalization ability. Appropriate learning rate adjustment helps the model better learn the essential features of the data, avoiding overfitting or underfitting. The RMSProp algorithm enables the model to converge more stably to the optimal solution during training, thereby improving the model's generalization ability to unknown data. In early abdominal pain recognition tasks, this means the model can more accurately identify and diagnose new cases of abdominal pain, leading to more stable model training. In one implementation case, the update rule is:

[0115] Initialization parameters Set the learning rate and attenuation coefficient Attenuation coefficient Set it to 0.9 to set a small constant for numerical stability. , usually set as ;

[0116] In each iteration, the parameters are calculated. gradient ;

[0117] Update the exponentially weighted moving average r of the cumulative squared gradient:

[0118] Calculate parameter update amount : ;

[0119] Last updated parameters .

[0120] When optimizing a model, optimization methods include increasing training data and modifying hyperparameters.

[0121] Specifically, potential shortcomings include: insufficient model generalization ability. This means that while the model performs well on the training set, its accuracy and recall drop significantly on the test set or new data, indicating poor generalization. For example, early abdominal pain recognition models achieved high accuracy on training data from local hospitals, but their performance dropped drastically when applied to data from hospitals in other regions. This might be due to insufficient coverage of case types and patient characteristics in the training data; or unbalanced data distribution. When the number of samples for different categories (such as different types of abdominal pain causes) varies greatly in the training data, the model may favor the category with more samples, resulting in poor recognition of categories with fewer samples. For instance, in abdominal pain recognition, the number of samples for a rare cause of abdominal pain is extremely small, making it difficult for the model to learn the features of that category. Both of these shortcomings can be addressed by expanding data sources, collaborating with more hospitals to obtain case data, or collecting relevant information from publicly available medical databases to increase the diversity and quantity of training data. For abdominal pain data, data augmentation techniques can be employed. For example, synonyms can be replaced in symptom descriptions, and the order of sentences can be changed; minor perturbations can be made to the test indicator data within a reasonable range to simulate measurement errors.

[0122] During model training, if the model converges too slowly, or exhibits overfitting or underfitting, it may be necessary to modify the hyperparameters. For example, too many training epochs can lead to overfitting, and an improperly set learning rate can affect the model's convergence speed and performance. An initial learning rate of 1e - 5, a batch size of 32, and 10 training epochs are recommended. If the model converges slowly, the learning rate can be increased appropriately; if overfitting occurs, the number of training epochs can be reduced. Treating hyperparameters as variables and calculating the gradient of the model loss with respect to the hyperparameters using the backpropagation algorithm, then adjusting the hyperparameter values ​​based on the gradient direction, offers advantages such as high computational efficiency and fast convergence.

[0123] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early identification of dangerous abdominal pain based on machine learning, characterized in that, The method comprises the following steps: collecting abdominal pain data, which comprises symptom information, physical sign data, medical history information and examination reports of a patient; establishing an abdominal pain identification model based on the abdominal pain data, the abdominal pain identification model being used to determine the probability of dangerous abdominal pain risk by processing the symptom information, medical history information and examination reports of the patient through natural language processing and processing the physical sign data through numerical analysis, training the abdominal pain identification model and determining whether the abdominal pain identification model needs to be further trained and optimized, if yes, continuing to train, and if no, stopping training; evaluating the probability of dangerous abdominal pain risk through the trained abdominal pain identification model; the abdominal pain identification model comprises: an input layer for inputting the abdominal pain data; a feature extraction layer for extracting features of the symptom information, medical history information and examination reports through a natural language model to obtain first features and extracting features of the physical sign data through a nonlinear function to obtain second features; a feature fusion layer for fusing the first features and the second features; an output layer for outputting whether there is a dangerous abdominal pain risk according to the fused features; the method for extracting the features of the symptom information, medical history information and examination reports through a natural language model to obtain the first features comprises: establishing a collection object library comprising standard sentence vectors of standard description sentences of a plurality of target feature data; performing text cleaning on sentences in the symptom information, medical history information and examination reports, and performing vectorization processing on each sentence through a BERT model to obtain a plurality of collection sentence vectors; obtaining a cosine distance between the ith standard sentence vector and the jth 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: ； wherein, and are the ith standard sentence vector and the jth collected sentence vector, respectively, and are row vectors, denotes the transpose; respectively, put all the collection sentence vectors with the cosine distance less than a preset threshold value to the i-th standard sentence set; obtaining the feature values of the ith standard sentence set : wherein, represents the cosine similarity between the i-th standard sentence vector and the i-th collected sentence vector in the i-th standard sentence set, represents the cosine similarity between the i-th standard sentence vector and the i-th collected sentence vector in the i-th standard sentence set, represents the total number of collected sentence vectors in the i-th standard sentence set. acquiring a first feature : ； the method for extracting the features of the physical sign data through a nonlinear function to obtain the second features comprises: the following processing is performed on the kth physical sign data: ; ； wherein, is the kth feature value of the sign data , and are the feature extraction weight and the feature extraction bias, respectively, L is the total number of sign data, and tanh is a tanh activation function. acquiring a second feature : ； the method for fusing the first features and the second features comprises: ； wherein E is the fused features; the method for outputting whether there is a dangerous abdominal pain risk according to the fused features comprises: Obtaining a probability of risk of dangerous abdominal pain : ； wherein, and are a first training weight vector and a second training weight vector, respectively, and are row vectors, and are a first training bias and a second training bias, respectively.

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 the abdominal pain data comprises: collecting the abdominal pain data through an electronic medical record system, a clinical research database and a wearable device; performing data cleaning and standardization processing on the collected abdominal pain data.

3. The method of early identification of dangerous abdominal pain based on machine learning according to claim 2, characterized in that, The method for performing the data cleaning comprises performing abnormal value correction and missing value supplementation through a Kalman filter; The method for performing the standardization processing comprises performing Min-Max normalization processing or Z-score standardization processing. 4.The method of early identification of dangerous abdominal pain based on machine learning according to claim 1, wherein, the method for determining whether the identification model needs to be further trained and optimized comprises: obtaining a plurality of groups of the abdominal pain data, outputting a judgment result through the identification model, comparing the judgment result with an actual result, and determining that training and optimization do not need to be continued if the accuracy reaches a preset threshold, and otherwise, training and optimization need to be continued.

5. The method for early identification of dangerous abdominal pain based on machine learning according to claim 4, characterized in that, If training and optimization need to be continued, the abdominal pain data with incorrect judgment results are included in a training set for training.

Images