Gynecological clinical pathway prediction method based on neuron-level feature selection

Abstract

The application provides a gynecological clinical path prediction method based on neuron level feature selection, and belongs to the technical field of path prediction, and comprises the following steps: dividing gynecological electronic medical record data of a patient into time sequence features and baseline features, inputting the time sequence features into a trained GRU network for feature coding, obtaining dynamic change rules of various clinical indexes over time, and converting the baseline features into feature vectors through an embedding layer; calculating feature importance based on the patient's last gynecological visit record and the visit scene, projecting the dynamic change rules of various clinical indexes over time and the feature vectors into a new representation space based on the feature importance to obtain a projection matrix to generate a gating vector, and screening and fusing the features of a main task layer and an auxiliary task layer respectively to obtain a final feature representation and obtain a multidimensional prediction result. More personalized and accurate diagnosis and treatment suggestions are provided for doctors.

Description

Technical Field

[0001] The present invention relates to the technical field of path prediction, and in particular to a gynecology and obstetrics clinical path prediction method based on neuron-level feature selection. Background Art

[0002] With the development of medical informatization, obstetrics and gynecology hospitals have accumulated a large amount of electronic medical record data, including basic patient information, test reports, and medication disposition records. However, this data is often stored in a raw, discrete format, lacking appropriate organization and analysis methods, making it difficult to provide effective support for doctors' clinical decision-making. Although deep learning models have demonstrated excellent performance in clinical prediction tasks, they face challenges in practical applications, such as limited data and complex models. These models are prone to overfitting, which affects the reliability of predictions.

[0003] In clinical practice, experienced doctors often refer to previous similar cases to assist in diagnosis and treatment decisions. However, this experience requires long-term professional accumulation, and medical institutions generally face a shortage of expert resources. Furthermore, even patients with similar symptoms can differ significantly in age, physical condition, and other aspects, leaving doctors with a lack of quantitative methods to assess the impact of these differences on treatment outcomes. Furthermore, patients' health status changes dynamically, and simply categorizing patients into fixed categories cannot accurately reflect the progression of their condition. Currently, there is a lack of effective methods to capture and predict these dynamic changes.

[0004] Therefore, the present invention proposes a gynecology and obstetrics clinical pathway prediction method based on neuron-level feature selection. Summary of the Invention

[0005] This paper provides a method for predicting obstetrics and gynecology clinical pathways based on neuron-level feature selection. This method uses a separate auxiliary network to predict future clinical visits and extract long-term health status representations. Furthermore, through a neuron-level filtering gate mechanism, it adaptively selects feature units that are helpful for target prediction, thereby providing doctors with more personalized and accurate diagnosis and treatment recommendations. This method fully utilizes the hospital's existing clinical data, eliminating the need for additional labeling, and can provide doctors with more scientific and reliable decision support.

[0006] The present invention provides a method for predicting obstetrics and gynecology clinical pathways based on neuron-level feature selection, comprising:

[0007] Step 1: Divide the patient's obstetrics and gynecology electronic medical record data into time series features and baseline features, wherein the time series features include test data, vital signs, and medical records, and the baseline features include demographic characteristics and medical history information;

[0008] Step 2: Input the time series features into the trained GRU network for feature encoding to obtain the dynamic change patterns of various clinical indicators over time. At the same time, the baseline features are converted into feature vectors through the embedding layer;

[0009] Step 3: Calculate feature importance based on the patient's most recent obstetrics and gynecology visit record and the visit scenario. Project the dynamic changes of various clinical indicators over time based on feature importance and map the feature vectors to the new representation space to obtain the projection matrix to generate the gating vector. Then, filter and fuse the features of the main task layer and the auxiliary task layer to obtain the final feature representation. Feature importance calculation and feature screening constitute the neuron-level feature selection mechanism.

[0010] Step 4: Obtain a multi-dimensional prediction result of the clinical pathway based on the final feature representation.

[0011] The present invention provides a method for predicting obstetrics and gynecology clinical pathways based on neuron-level feature selection, which divides the obstetrics and gynecology electronic medical record data of patients into time series features and baseline features, including:

[0012] Collecting multimodal data from the obstetrics and gynecology electronic medical record, wherein the multimodal data includes: structured data and text data;

[0013] Using natural language processing technology to perform semantic analysis on the text data, extract key semantics, and combine the meaning of clinical indicators in the structured data to establish semantic associations between data items;

[0014] The multimodal data is divided into time series features and baseline features based on semantic association relationships.

[0015] The present invention provides a method for predicting a gynecological and obstetric clinical pathway based on neuron-level feature selection. In the process of obtaining a multi-dimensional prediction result of the clinical pathway according to the final feature representation, the method further includes:

[0016] Extracting the patient's disease type, treatment stage for each disease type, and disease severity from the medical history information, constructing a disease feature vector for each disease type, and performing cluster analysis on all disease feature vectors;

[0017] Determine, based on the disease mutual influence relationship table, the influence relationship between each cluster center and the current feature vector of the patient in the most recent obstetrics and gynecology visit record, and reversely infer the inspiration effect on each feature based on the influence relationship, wherein the inspiration effect includes: positive inspiration, negative inspiration, and no inspiration, wherein positive inspiration is assigned a value of 1, no inspiration is assigned a value of 0, and negative inspiration is assigned a value of -1;

[0018] Sort the values ​​of all the inspirations under each feature in descending order of influence. At the same time, sort the values ​​of all the inspirations under each feature in descending order of the value of the inspiration.

[0019] Performing linear fitting on the curve obtained by the first sorting and the curve obtained by the second sorting, respectively, to obtain a first fitting value and a second fitting value;

[0020] Sort the first fitting values ​​of all features from large to small to obtain the first serial number of each feature. At the same time, sort the second fitting values ​​of all features from large to small to obtain the second serial number of each feature. According to the first serial number and the second serial number, obtain the first rough attention coefficient;

[0021] Determine the absolute value of the difference between the first fitting value of the linear fitting result of the curve obtained by one sort and the original value of the corresponding point, determine the locked horizontal interval around the corresponding point, and obtain the original average value of the horizontal interval and the linear average , confirm consistency ;

[0022] Correcting the first rough attention coefficient according to the consistency to obtain a second rough attention coefficient, and using the second rough attention coefficient as the importance of the corresponding feature in the clinical pathway prediction process;

[0023] The present invention provides a method for predicting clinical pathways in obstetrics and gynecology based on neuron-level feature selection, which obtains a second rough attention coefficient, including:

[0024]

[0025]

[0026] in, represents the second coarse attention coefficient; represents the first coarse attention coefficient; represents the adjustment parameter; Indicates the corresponding absolute value of the difference; b1 is the set threshold.

[0027] The present invention provides a method for predicting clinical pathways in obstetrics and gynecology based on neuron-level feature selection, wherein each part of the temporal features corresponds to a trained GRU network;

[0028] Each component in the baseline features corresponds to a learnable embedding layer.

[0029] The present invention provides a method for predicting clinical pathways in obstetrics and gynecology based on neuron-level feature selection, which obtains the dynamic change patterns of various clinical indicators over time, including:

[0030] Obtain feature information of a set indicator for each part of the time series feature under the corresponding trained GRU network, extract the information structure under the same set indicator from all feature information, and obtain a first dynamic change law vector corresponding to the set indicator, wherein the first dynamic change law vector includes the local change value of the i1th set indicator at time t;

[0031] Based on the multi-scale feature extraction network, convolution kernels of different sizes are used to perform multi-scale convolution operations on the input time series features, extract feature information of different granularities, and capture the global dynamic change law vector of the time series features, wherein the global dynamic change law vector includes the global change value at time t; based on the global dynamic change law vector, the time series impact on each first dynamic change law vector is determined:

[0032]

[0033] in, is the global dynamic change law vector For the i-th first dynamic change law vector The timing impact of 、 They represent parameter vectors respectively;

[0034] Based on the intelligent layer, the first dynamic change law vector of all set indicators and the global dynamic change law vector are merged and processed, and the dynamic change law of various indicators over time is obtained by combining the temporal influence under different indicators. Among them, the output of the GRU network and the output of the multi-scale feature extraction network are used as the input of the intelligent layer.

[0035] The present invention provides a method for predicting clinical pathways in obstetrics and gynecology based on neuron-level feature selection. The method generates a gating vector by projecting the dynamic change patterns of various clinical indicators over time based on feature importance and mapping feature vectors to a new representation space to obtain a projection matrix, including:

[0036] Constructing a graph structure by mapping elements mapped to the new representation space and associated clinical indicator features, and performing in-depth analysis of the graph structure based on a graph neural network to obtain mapping logic;

[0037] Converting the mapping logic into a structured standard representation and inputting the result into a preset representation filling table to obtain a representation vector and a representation weight, wherein the preset representation filling table is a multi-layer perceptron model that takes the semantic and structural information of the standard representation as input and outputs the representation vector and the representation weight;

[0038] Based on all representation vectors and representation weights, construct the projection matrix , and use the regularization method to constrain the projection matrix, where , Bn represent the representation vectors corresponding to the first mapping element and the nth mapping element respectively; 、 Respectively represent the representation weights corresponding to the first mapping element and the nth mapping element;

[0039] The projection matrix is ​​quantized using an asymmetric quantization method, which maps the continuous values ​​in the matrix to a finite set of discrete values. The representation ability of the projection matrix is ​​enhanced by combining it with a generative adversarial network.

[0040] The enhanced projection matrix is ​​multiplied by the original eigenvector to obtain the projection result of each feature. The importance of the feature is evaluated based on the numerical value and distribution of the projection result combined with the representation weight;

[0041] Based on the feature importance scores, a gating vector is generated using the gating mechanism of the gated recurrent unit.

[0042] The present invention provides a method for predicting clinical pathways in obstetrics and gynecology based on neuron-level feature selection, which screens the features of the main task layer and the auxiliary task layer respectively, including:

[0043] Screening the features of the main task layer according to the value partitioning mechanism and using the gating vector;

[0044] Based on the auxiliary task layer, potential supplementary information of each gating vector is mined, and the features filtered by the main task layer are verified according to the potential supplementary information, and the features of the main task layer in progress are supplemented according to the verification association relationship.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] Through innovative neuron-level feature selection mechanism and multi-channel feature processing method, deep mining and efficient utilization of obstetrics and gynecology electronic medical record data are achieved, providing strong technical support for obstetric clinical decision-making.

[0047] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0048] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0050] Figure 1 This is a flowchart of a method for predicting obstetrics and gynecology clinical pathways based on neuron-level feature selection in an embodiment of the present invention;

[0051] Figure 2 This is a diagram of the main hierarchical structure in an embodiment of the present invention;

[0052] Figure 3 This is a feature extraction flow chart in an embodiment of the present invention;

[0053] Figure 4 This is a feature selection flow chart in an embodiment of the present invention;

[0054] Figure 5 Schematic diagram of an application scenario in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0056] The present invention provides a method for predicting clinical pathways in obstetrics and gynecology based on neuron-level feature selection. Figure 1 Shown, including:

[0057] Step 1: Divide the patient's obstetrics and gynecology electronic medical record data into time series features and baseline features, wherein the time series features include test data, vital signs, and medical records, and the baseline features include demographic characteristics and medical history information;

[0058] Step 2: Input the time series features into the trained GRU network for feature encoding to obtain the dynamic change patterns of various clinical indicators over time. At the same time, the baseline features are converted into feature vectors through the embedding layer;

[0059] Step 3: Calculate feature importance based on the patient's most recent obstetrics and gynecology visit record and the visit scenario. Project the dynamic changes of various clinical indicators over time based on feature importance and map the feature vectors to the new representation space to obtain the projection matrix to generate the gating vector. Then, filter and fuse the features of the main task layer and the auxiliary task layer to obtain the final feature representation. Feature importance calculation and feature screening constitute the neuron-level feature selection mechanism.

[0060] Step 4: Obtain a multi-dimensional prediction result of the clinical pathway based on the final feature representation.

[0061] In this example, the time series features, including test data (such as blood routine and biochemical test results), vital signs (such as blood pressure and heart rate monitoring data), and medical records (such as medication, surgery, and other treatment information), are fed into three trained GRU networks for feature encoding, thereby capturing the dynamic changes of various clinical indicators over time. Baseline features, including demographic characteristics (such as age, gravidity and parity) and medical history information (such as medical history and family history), are converted into dense feature vectors through two learnable embedding layers.

[0062] The features after the above processing are input into the feature fusion module, which integrates features from different sources and types to generate a unified feature representation. The key information is further refined through the feature selection step, and the fusion features are finally output for subsequent risk prediction and benefit assessment. This multi-channel feature processing solution can make full use of various types of information in the patient's electronic medical record and provide comprehensive data support for clinical decision-making. For example, the system can analyze the patient's test result change trend, treatment response characteristics, basic conditions and other dimensions, thereby providing a reliable basis for doctors to formulate personalized treatment plans, such as Figure 3 shown.

[0063] In the feature importance calculation stage, the system first calculates the feature importance based on the patient's most recent medical record, and then maps the features to a new representation space through importance projection. In order to improve the expressiveness of features, the system designs an innovative gating mechanism. By calculating the correlation of neuron activation values ​​and constructing a decorrelation loss function, different neurons are encouraged to learn complementary feature representations, such as Figure 4 shown.

[0064] During the feature selection phase, a separate gating strategy is employed to adaptively allocate features to the primary and auxiliary tasks. Specifically, the system generates gating vectors using a learnable projection matrix to filter features at both the primary and auxiliary task layers. For the primary task layer, the system prioritizes features directly related to target prediction; for the auxiliary task layer, it focuses on features that provide supplementary information. Finally, the system fuses these filtered features to produce the final feature representation. This feature selection mechanism not only effectively identifies and retains key information, but also reduces feature redundancy, improving the model's predictive performance.

[0065] Through this feature selection mechanism, we can adaptively identify key features for different patients, providing more reliable feature support for subsequent risk prediction and treatment plan development. We can dynamically adjust the importance of features based on specific medical scenarios and patient conditions, enabling personalized feature selection and prediction.

[0066] Time series features are dynamic medical data that change over time, including test data: routine blood tests (such as white blood cell count and hemoglobin), biochemical indicators (such as blood sugar and creatinine); vital signs: heart rate, blood pressure, body temperature, and blood oxygen saturation; and medical records: medication dosage, surgical operation records, and nursing records.

[0067] Baseline characteristics are relatively stable basic information about patients, including demographic characteristics: age, gestational age, BMI, and ethnicity; medical history information: history of chronic diseases (such as diabetes, hypertension), allergy history, and previous pregnancy complications. It should be noted that structured data is extracted from the hospital HIS system, and unstructured information is extracted from the electronic medical record text through NLP.

[0068] Among them, feature encoding is to convert the original data into a low-dimensional semantic vector, and the embedding layer is a neural network layer that maps discrete features to a continuous vector space.

[0069] Feature Importance The calculation of is as follows:

[0070] ,in, is the eigenvector of the jth feature, 、 、 、 is a trainable parameter.

[0071] 1) Input layer: responsible for receiving and preprocessing electronic medical record data

[0072] 1. Time series feature processing: processing time series data such as various inspection results

[0073] 2. Baseline feature embedding: processing static information such as demographic characteristics

[0074] 2) Processing layer: implement core algorithm flow

[0075] 1. Multi-channel feature extraction: encoding different types of features

[0076] 2. Main task network: complete the target prediction task

[0077] 3. Auxiliary prediction network: predicting future clinical indicators

[0078] 4. Neuron-level feature selection: screening important features

[0079] 3) Output layer: Generate multi-dimensional prediction results

[0080] 1. Risk prediction: assessing the patient's risk level

[0081] 2. Pathway analysis: analyzing disease development trends

[0082] 3. Similar cases: Match similar patient cases

[0083] In this embodiment, through the innovative neuron-level feature selection mechanism and multi-channel feature processing method, deep mining and efficient utilization of obstetrics and gynecology electronic medical record data are achieved, providing strong technical support for obstetric clinical decision-making.

[0084] like Figure 5 As shown, the present invention can be applied to three core scenarios: pregnancy management, delivery management, and postpartum follow-up. During the pregnancy management phase, the system conducts risk assessments starting with the first prenatal visit, develops personalized intervention plans, and conducts dynamic monitoring. During the delivery management phase, the system assesses delivery risks for hospitalized mothers, provides delivery method recommendations, and implements graded monitoring. During the postpartum follow-up phase, the system predicts the risk of postpartum complications, develops rehabilitation plans, and conducts long-term health management. Through a state transfer mechanism, the system achieves comprehensive and continuous health management for pregnant and postpartum women.

[0085] First, the technical solution of the present invention demonstrates significant advantages in addressing key obstetric issues such as pregnancy complication prediction. By designing an auxiliary prediction network to predict the changing trends of clinical indicators during pregnancy, a richer feature representation can be obtained without the need for additional labeled data. Furthermore, the innovative neuron-level filtering gate mechanism adaptively selects features relevant to pregnancy outcome prediction, such as key indicators like blood pressure and coagulation function, effectively avoiding the impact of feature redundancy on predictive performance. Experimental results demonstrate that this solution achieves excellent performance in tasks such as pregnancy complication warning and postpartum complication prediction.

[0086] Secondly, the technical solution of this invention provides a scientific basis for the tiered management of key obstetric populations. Through multi-channel feature extraction, the system comprehensively analyzes various maternal indicators, including pregnancy test results, changes in vital signs, and medical history. This design enables the model to capture long-term dependencies within the clinical pathway during pregnancy, promptly identifying high-risk women and providing decision support for obstetricians in developing personalized management plans.

[0087] From a clinical application perspective, the technical solution of this invention significantly improves the accuracy and interpretability of obstetric predictions. By explicitly modeling feature importance, the system helps doctors understand the key risk factors leading to pregnancy complications. In particular, for complications such as preeclampsia that threaten maternal and fetal safety, this solution provides early warning, providing a crucial basis for timely intervention. This not only optimizes the efficient allocation of obstetric medical resources, but also provides a strong guarantee for reducing risks to both mothers and babies.

[0088] Therefore, the technical solution of the present invention has important practical value in improving the performance of obstetric prediction models, promoting pregnancy and childbirth risk management, and promoting the development of smart medicine.

[0089] The present invention provides a method for predicting obstetrics and gynecology clinical pathways based on neuron-level feature selection, which divides the obstetrics and gynecology electronic medical record data of patients into time series features and baseline features, including:

[0090] Collecting multimodal data from the obstetrics and gynecology electronic medical record, wherein the multimodal data includes: structured data and text data;

[0091] Using natural language processing technology to perform semantic analysis on the text data, extract key semantics, and combine the meaning of clinical indicators in the structured data to establish semantic associations between data items;

[0092] The multimodal data is divided into time series features and baseline features based on semantic association relationships.

[0093] In this embodiment, multimodal data is comprehensive information including multiple data types.

[0094] Structured data is standardized data stored in a table format, such as test values ​​(white blood cell count 10.2×10 9 / L), vital signs (systolic blood pressure 120mmHg), demographic information (age 28 years, gestational age 32 weeks), and directly read database tables (such as test results table and patient basic information table).

[0095] Text data is unstructured free text, such as chief complaint ("recurrent dizziness during pregnancy with lower limb edema") and medical records ("Doctor's order: monitor blood pressure daily, take 30mg nifedipine controlled-release tablets orally"). The original text is extracted from medical record text fields (such as chief complaint and medical history).

[0096] In this embodiment, core information with medical significance in the text, such as the name of the disease ("gestational diabetes"), symptoms ("abdominal pain"), treatment measures ("cesarean section"), etc., can all be regarded as key semantics.

[0097] In this embodiment, the logical connection between the structured data and the text data, such as "blood sugar elevation" in the text corresponding to "fasting blood sugar value 8.2 mmol / L" in the structured test data, is regarded as a semantic association relationship.

[0098] In this example, for example, "blood pressure 110 / 70 mmHg at the first prenatal checkup" and "blood pressure 130 / 85 mmHg at the second prenatal checkup" are classified as time series features; data that is not related to time and describes the patient's inherent attributes (such as "age 30 years old" and "history of cesarean section") are classified as baseline features.

[0099] Combined with semantic association strength, noise is filtered out. For example, if a phrase like "occasional headaches mentioned by the patient" is not associated with structured examination results, it is not included as a baseline feature. Based on dynamic classification of disease types, such as patients with gestational hypertension, data related to blood pressure fluctuations are forcibly included in the time series features.

[0100] The beneficial effects of the above technical solution are: converting unstructured text into computable semantic information to improve data relevance, dividing time series and baselines based on semantic association relationships, ensuring the accuracy of data division, and through the deep integration of NLP and structured data, achieving a leap from "data stacking" to "semantic driving", providing high-quality feature input for subsequent clinical pathway prediction, disease diagnosis and other tasks.

[0101] The present invention provides a method for predicting a gynecological and obstetric clinical pathway based on neuron-level feature selection. In the process of obtaining a multi-dimensional prediction result of the clinical pathway according to the final feature representation, the method further includes:

[0102] Extracting the patient's disease type, treatment stage for each disease type, and disease severity from the medical history information, constructing a disease feature vector for each disease type, and performing cluster analysis on all disease feature vectors;

[0103] Determine, based on the disease mutual influence relationship table, the influence relationship between each cluster center and the current feature vector of the patient in the most recent obstetrics and gynecology visit record, and reversely infer the inspiration effect on each feature based on the influence relationship, wherein the inspiration effect includes: positive inspiration, negative inspiration, and no inspiration, wherein positive inspiration is assigned a value of 1, no inspiration is assigned a value of 0, and negative inspiration is assigned a value of -1;

[0104] Sort the values ​​of all the inspirations under each feature in descending order of influence. At the same time, sort the values ​​of all the inspirations under each feature in descending order of the value of the inspiration.

[0105] Performing linear fitting on the curve obtained by the first sorting and the curve obtained by the second sorting, respectively, to obtain a first fitting value and a second fitting value;

[0106] Sort the first fitting values ​​of all features from large to small to obtain the first serial number of each feature. At the same time, sort the second fitting values ​​of all features from large to small to obtain the second serial number of each feature. According to the first serial number and the second serial number, obtain the first rough attention coefficient;

[0107] Determine the absolute value of the difference between the first fitting value of the linear fitting result of the curve obtained by one sort and the original value of the corresponding point, determine the locked horizontal interval around the corresponding point, and obtain the original average value of the horizontal interval and the linear average , confirm consistency ;

[0108] Correcting the first rough attention coefficient according to the consistency to obtain a second rough attention coefficient, and using the second rough attention coefficient as the importance of the corresponding feature in the clinical pathway prediction process;

[0109] The present invention provides a method for predicting clinical pathways in obstetrics and gynecology based on neuron-level feature selection, which obtains a second rough attention coefficient, including:

[0110]

[0111]

[0112] in, represents the second coarse attention coefficient; represents the first coarse attention coefficient; represents the adjustment parameter; Indicates the corresponding absolute value of the difference; b1 is the set threshold.

[0113] In this example, natural language processing techniques are used to extract keywords such as disease type, treatment stage, and severity from the medical history text. These keywords are then converted into numerical values ​​using pre-set encoding rules to construct a disease feature vector. For example, from the text "The patient suffers from gestational diabetes, is currently receiving insulin treatment, and has a moderate condition," this information is extracted and encoded as [gestational diabetes code, 2, 2]. A clustering algorithm (such as K-means clustering) is then used to cluster all disease feature vectors, grouping patients with similar disease characteristics into the same category.

[0114] A disease interaction table is established. This table can be pre-set based on medical knowledge and clinical experience to record the interaction relationships between different diseases. Then, the similarity between each cluster center and the patient's current feature vector is calculated to determine the influence relationship. Based on the influence relationship table and the calculation results, the heuristic effect of each feature is determined and assigned a corresponding value. For example, if the gestational diabetes cluster center has a high correlation with the feature vector of the patient's current poor blood sugar control, and the disease interaction relationship table shows that gestational diabetes increases the risk of premature birth, then the heuristic effect of this feature is positive and is assigned a value of 1.

[0115] In this embodiment, the influence value of each feature is sorted according to the size of the influence relationship, which reflects the priority of the influence degree; and the second sorting is performed according to the size of the enlightenment value, which highlights the difference of the influence direction. The first fitting value and the second fitting value are obtained by linear fitting of the curves obtained by the two sortings, respectively. For example, for multiple features related to the risk of premature birth, the features are first sorted according to the influence relationship with the risk of premature birth, and then sorted according to the positive and negative and size of the enlightenment effect, and then the values reflecting the trend of the features are obtained by fitting.

[0116] The enlightenment effect represents the influence direction of the disease cluster center on the current features of the patient. Positive enlightenment indicates that the disease factor has a promoting or associated effect on the current disease development (for example, gestational hypertension may increase the risk of premature birth, and is assigned a value of 1); negative enlightenment indicates inhibition or reverse correlation (for example, good nutritional intervention during pregnancy may reduce the risk of pregnancy complications, and is assigned a value of -1); and no enlightenment indicates no obvious correlation (for example, the patient has no history of irrelevant diseases, and is assigned a value of 0).

[0117] The first fitting value and the second fitting value are the fitting values at the center position point after linear fitting of the curves obtained by the first sorting and the second sorting, respectively.

[0118] Based on the disease interaction relationship table and clinical experience, the actual correlation between diseases is considered, which is consistent with medical common sense and clinical practice. For example, the interaction relationship between common diseases such as gestational diabetes and gestational hypertension has been clearly studied in medicine, and the influence relationship and enlightenment effect are determined based on this, which has a solid medical basis. Through quantitative processing and data analysis of medical history information, complex text information is converted into calculable and comparable values, avoiding the limitations of subjective judgment. Using clustering analysis, sorting and fitting methods, valuable information is mined from a large amount of data to provide objective and accurate support for clinical path prediction. Considering the dynamic changes of the patient's condition, the importance of the features is re-evaluated by continuously updating the medical history information and the current feature vector, so that the scheme can adapt to the needs of different patients and different stages of the disease. For example, as the gestational age increases, the patient's disease condition and treatment stage may change, and the scheme can adjust the feature attention coefficient in time to provide real-time and effective reference for clinical decision-making.

[0119] In this embodiment, the determination of the horizontal interval is as follows:

[0120] Suppose the absolute value of the difference is d, the abscissa of the corresponding point is x0, and the expansion multiple is k. At this time, the horizontal interval is: [x0-kd, x0+kd], it should be noted that x0 is the center position point.

[0121] The expansion multiple k is matched from the difference absolute value-multiple control table, which contains the expansion multiple involved under different difference absolute values, and is directly used. Generally, the larger the difference absolute value, the worse the consistency, and the larger the corresponding expansion multiple. By examining the data in this interval, the overall situation of the data near the point can be more comprehensively analyzed, avoiding the excessive influence of a single point error or anomaly on the result, and further providing a data basis for subsequent calculation of the original average value and the linear average value.

[0122] If there are n0 points in the horizontal interval, the average value of the n0 points on the fitting straight line under the linear fitting of the n0 points is calculated , and the average value of the n0 points on the corresponding curve before linear fitting is .

[0123] In this embodiment, the consistency calculation formula is used to measure the consistency of the linear fitting result of the sorting curve with the original value. In clinical path prediction, by evaluating the consistency of the fitting value and the original value, it can be judged whether the feature importance trend obtained based on sorting fitting is reliable, and an effective reference is provided for subsequent coefficient correction.

[0124] In this embodiment, in clinical path prediction, different fitting consistency cases need different degrees of adjustment. According to the actual fitting difference, the adjustment parameter is reasonably determined, so as to more accurately reflect the importance of the feature when correcting the rough attention coefficient. In clinical path prediction, by the formula, the initially determined feature attention coefficient is optimized according to the fitting effect, so that the second rough attention coefficient obtained finally can more accurately reflect the importance of the feature in clinical path prediction, thereby assisting more accurate clinical decision-making.

[0125] The beneficial effects of the above technical solutions are: extracting key disease features from medical history information, analyzing the mutual relationship between them and the influence on the current condition, and quantifying the importance of each feature in clinical path prediction. Thus, it helps doctors more accurately grasp the patient's condition, develop personalized clinical paths, improve medical quality and efficiency, and reduce the risk of adverse pregnancy outcomes.

[0126] In obstetrics and gynecology clinical practice, doctors need to comprehensively consider the local changes in individual patient indicators and the overall development trend of the disease. For example, in pregnancy management, it is necessary to pay attention to the specific changes in various indicators of pregnant women and fetuses during each prenatal check-up, and to grasp the physiological development laws throughout the pregnancy. This solution meets the actual clinical needs by integrating local and global change laws, and helps to improve the accuracy and scientificity of medical decision-making. Therefore, the following technical solution is proposed. The present invention provides a method for predicting obstetrics and gynecology clinical pathways based on neuron-level feature selection, and each part of the time series feature corresponds to a trained GRU network;

[0127] Each component in the baseline features corresponds to a learnable embedding layer.

[0128] The present invention provides a method for predicting clinical pathways in obstetrics and gynecology based on neuron-level feature selection, which obtains the dynamic change patterns of various clinical indicators over time, including:

[0129] Obtain feature information of a set indicator for each part of the time series feature under the corresponding trained GRU network, extract the information structure under the same set indicator from all feature information, and obtain a first dynamic change law vector corresponding to the set indicator, wherein the first dynamic change law vector includes the local change value of the i1th set indicator at time t;

[0130] Based on the multi-scale feature extraction network, convolution kernels of different sizes are used to perform multi-scale convolution operations on the input time series features, extract feature information of different granularities, and capture the global dynamic change law vector of the time series features, wherein the global dynamic change law vector includes the global change value at time t; based on the global dynamic change law vector, the time series impact on each first dynamic change law vector is determined:

[0131]

[0132] in, is the global dynamic change law vector For the i-th first dynamic change law vector The timing impact of 、 They represent parameter vectors respectively;

[0133] Based on the intelligent layer, the first dynamic change law vector of all set indicators and the global dynamic change law vector are merged and processed, and the dynamic change law of various indicators over time is obtained by combining the temporal influence under different indicators. Among them, the output of the GRU network and the output of the multi-scale feature extraction network are used as the input of the intelligent layer.

[0134] In this embodiment, the predefined indicators are key medical indicators pre-determined for measurement and analysis in the time-series feature analysis of obstetrics and gynecology electronic medical records. For example, in obstetrics, these may include uterine contraction frequency, fetal heart rate, maternal blood pressure, and blood sugar levels; in gynecology, these may include hormone levels (such as estrogen and progesterone) and parameters related to pelvic ultrasound examinations (such as uterine size and endometrial thickness).

[0135] Feature information is obtained by processing time series features using a GRU network or a multi-scale feature extraction network, reflecting the inherent patterns and characteristics of the specified indicator. For example, after processing uterine contraction frequency data using a GRU network, information about the temporal trend and fluctuation of uterine contraction frequency is obtained.

[0136] Information structure is the organization and presentation of characteristic information across time for a given indicator. For example, the chronological arrangement of a woman's blood pressure measurements over a period of time, along with the changing relationships between these measurements, constitutes the information structure of the given indicator.

[0137] The first dynamic change pattern vector is a vector generated by the GRU network, reflecting the change in a specific indicator within a local time range. For example, for the fetal heart rate indicator, the first dynamic change pattern vector contains the change in the fetal heart rate every 5 minutes. For example, at time t = 30 minutes, the fetal heart rate increased by 10 beats / minute compared to the previous time point. This change value is an element in the vector.

[0138] The global dynamic change law vector includes the comprehensive change trends of the vital signs (such as blood pressure, heart rate, etc.) and test indicators (such as various blood routine indicators) of the entire obstetrics and gynecology patient group at a certain point in time. The small 3×1 convolution kernel captures the local feature changes during each uterine contraction, and the large 7×1 convolution kernel captures the overall trend of these features throughout the entire delivery process. Finally, the vector reflecting the global changes of the entire delivery process is obtained.

[0139] In this embodiment, the intelligent layer integrates these vectors based on the previously calculated temporal influences. Using methods such as weighted fusion and feature splicing, local and global variation patterns are organically combined to derive dynamic patterns of change over time for various indicators. For example, during labor, the intelligent layer integrates the local variation patterns of uterine contraction intensity with the global variation patterns of the entire labor process, yielding a comprehensive pattern that reflects how contraction intensity changes throughout labor as a function of the overall labor process.

[0140] In this embodiment, It is a value between 0 and 1, which measures the impact of global changes on local changes of specific indicators at time t. Refers to The purpose of concatenating vectors is to comprehensively consider the information of local changes and global changes in a unified vector space. 、 All of these can be learned and optimized through the model's training process. During training, by continuously adjusting these parameters, the model can find the most appropriate correlation between global and local changes based on actual data, making the calculated timing impact more consistent with actual conditions.

[0141] The beneficial effects of the above technical solution are: based on the GRU network, it can accurately capture the dynamic change details of each set indicator within the local time range, provide a basis for subsequent analysis, determine the global decoration law, grasp the overall change law of the time series characteristics from a macro perspective, avoid focusing only on local details and ignoring the overall trend, and clarify the time correlation and interaction between global changes and local changes based on the time series impact, providing key parameters for subsequent information merging. Based on the fusion of vectors at the intelligent layer, more comprehensive and accurate dynamic change laws of various indicators can be obtained, providing strong support for obstetrics and gynecology clinical pathway prediction, disease diagnosis and treatment decision-making.

[0142] Obstetrics and gynecology clinical indicators are numerous and complex. This solution aims to explore the intrinsic relationships between clinical indicators through a series of processing, quantify feature importance, and screen key features. These key features are used to more accurately analyze the dynamic changes of various clinical indicators over time, providing strong support for clinical pathway prediction, disease diagnosis, and treatment. Based on the projection of the dynamic changes of various clinical indicators over time based on feature importance and the mapping of feature vectors to a new representation space, the projection matrix is ​​generated to generate the gating vector, including:

[0143] Constructing a graph structure by mapping elements mapped to the new representation space and associated clinical indicator features, and performing in-depth analysis of the graph structure based on a graph neural network to obtain mapping logic;

[0144] Converting the mapping logic into a structured standard representation and inputting the result into a preset representation filling table to obtain a representation vector and a representation weight, wherein the preset representation filling table is a multi-layer perceptron model that takes the semantic and structural information of the standard representation as input and outputs the representation vector and the representation weight;

[0145] Based on all representation vectors and representation weights, construct the projection matrix , and use the regularization method to constrain the projection matrix, where , Bn represent the representation vectors corresponding to the first mapping element and the nth mapping element respectively; 、 Respectively represent the representation weights corresponding to the first mapping element and the nth mapping element;

[0146] The projection matrix is ​​quantized using an asymmetric quantization method, which maps the continuous values ​​in the matrix to a finite set of discrete values. The representation ability of the projection matrix is ​​enhanced by combining it with a generative adversarial network.

[0147] The enhanced projection matrix is ​​multiplied by the original eigenvector to obtain the projection result of each feature. The importance of the feature is evaluated based on the numerical value and distribution of the projection result combined with the representation weight;

[0148] Based on the feature importance scores, a gating vector is generated using the gating mechanism of the gated recurrent unit.

[0149] In this embodiment, a mapped element refers to an element obtained by mapping a clinical indicator feature into a new representation space. For example, in obstetrics and gynecology medical records, the corresponding element obtained after the clinical indicator feature "blood sugar level of a pregnant woman" is mapped into the new space through a transformation (such as a linear transformation or a nonlinear transformation) is a mapped element.

[0150] In this embodiment, the graph structure is a data structure consisting of nodes and edges. In this scenario, nodes can be mapping elements and clinical indicator features, and edges represent the associations between them. For example, the mapping elements corresponding to the two clinical indicator features "pregnant woman's blood sugar level" and "insulin dosage" can be nodes, and the possible causal relationship or correlation between them can be represented by edges.

[0151] In this example, the graph neural network learns the relationships and feature propagation between nodes in the graph. Mapping logic refers to the inherent connections and mechanisms between mapped elements and clinical indicator features. For example, a graph neural network analysis could reveal a logical relationship such as "changes in a pregnant woman's diet can lead to changes in blood sugar levels, which in turn affect insulin dosage adjustments."

[0152] In this embodiment, the structured standard representation is a representation format in which the mapping logic is organized according to certain specifications and formats. For example, the above mapping logic can be organized into a structured format such as "[Cause (change in pregnant woman's diet), Result 1 (change in blood sugar level), Result 2 (adjustment of insulin dosage)]".

[0153] In this embodiment, a representation filling table (multi-layer perceptron model) is preset. For example, structured information such as "the relationship between maternal age and risk of premature birth" is input, and a vector representing the relationship and the important weight of the relationship in the overall feature are output.

[0154] In this embodiment, the representation vector is a vector form used to express the mapping logic. For example, the mapping logic for "exercise during pregnancy and weight management" is represented by a vector (such as [0.3, 0.5, 0.2]). The dimension and value of the vector are determined by the specific model settings and learning.

[0155] In this embodiment, the asymmetric quantization method converts continuous values ​​in a matrix into a finite number of discrete values, and the positive and negative values ​​may be treated differently during the conversion process. For example, the continuous weight values ​​[0.1, 0.3, -0.2] in the projection matrix are converted into discrete values ​​[1, 2, -1].

[0156] The gating vector is a binary or weighted vector, such as [1,0,1], which indicates that some features are open (value 1) and some features are closed (value 0).

[0157] The enhanced projection matrix is ​​multiplied by the original eigenvector to obtain the projection result for each feature. For example, the original eigenvector represents various indicators of a pregnant woman (age, gestational age, medical history, etc.). After being transformed by the projection matrix, the new vector is obtained. The importance of each feature is evaluated based on the numerical value and distribution of the projection result, combined with the representation weight. For example, a feature with a larger numerical value and a higher corresponding weight indicates that it is more important in the overall feature set.

[0158] The beneficial effects of the above technical solution are: graph structure and graph neural network can well handle the complex situation of obstetrics and gynecology clinical data with multiple indicators and the relationship between indicators, and mine hidden information; multi-layer perceptron can quantify complex logic; operations such as projection matrix can effectively evaluate and screen features. The entire solution is adapted to the characteristics of clinical data. By generating gated vectors to screen features, it can help doctors focus on the key points from a large amount of clinical data and improve decision-making efficiency and accuracy. The solution integrates multiple advanced technologies such as graph neural networks, multi-layer perceptrons, generative adversarial networks, etc. Each technology plays a unique role in the corresponding link, and processes and analyzes clinical data from different angles. The technology integration is scientific and reasonable, and can effectively achieve the design goals.

[0159] In tasks such as obstetrics and gynecology clinical pathway prediction and disease diagnosis, the features of the main task layer are the core focus, but there may be incomplete information. Although the features of the auxiliary task layer do not directly participate in the main task prediction, they may contain valuable supplementary information for the main task. This scheme aims to improve processing efficiency by filtering the main task layer features through gating vectors, while using the auxiliary task layer to mine potential supplementary information, verify and supplement the main task layer, so as to analyze the condition more comprehensively and accurately, and provide a more reliable basis for clinical decision-making. Therefore, it is proposed to filter the features of the main task layer and the auxiliary task layer separately, including:

[0160] Screening the features of the main task layer according to the value partitioning mechanism and using the gating vector;

[0161] Based on the auxiliary task layer, potential supplementary information of each gating vector is mined, and the features filtered by the main task layer are verified according to the potential supplementary information, and the features of the main task layer in progress are supplemented according to the verification association relationship.

[0162] In this embodiment, it is set that the features corresponding to the elements with a value of 1 in the gating vector are retained, and the features corresponding to the elements with a value of 0 are eliminated, which is a value partitioning mechanism.

[0163] In this embodiment, the main task layer is a set of features directly related to the main clinical prediction or diagnosis task. For example, in the task of predicting whether a pregnant woman will have a premature birth, the features of the main task layer may include the pregnant woman's gestational age, uterine contraction frequency, cervical length and other indicators directly related to premature birth. The auxiliary task layer is a set of features that provide supplementary information. These features are not directly involved in the prediction of the main task, but are helpful for understanding the overall condition or assisting the judgment of the main task. For example, in the prediction of premature birth, the features of the auxiliary task layer may include the psychological state of the pregnant woman, the details of nutritional intake during pregnancy, etc. Although this information does not directly determine whether premature birth will occur, it may have an indirect impact on the risk of premature birth.

[0164] In this embodiment, potential supplementary information is information hidden within the auxiliary task layer features that can supplement or verify the features filtered by the primary task layer. For example, when analyzing the risk of premature birth in a pregnant woman, the auxiliary task layer feature "whether the pregnant woman has recently experienced a stressful event" may contain potential supplementary information that affects the pregnant woman's physical stress response and, therefore, indirectly affects the risk of premature birth.

[0165] In this embodiment, the verification association relationship is a logical connection and mutual verification relationship between the potential supplementary information in the auxiliary task layer and the features selected by the main task layer. For example, if the main task layer selects "abnormal uterine contraction frequency in pregnant women" as a key feature, the potential supplementary information in the auxiliary task layer, "pregnant women have recently experienced high mental stress," could be used to verify and supplement the features of the main task layer through the association relationship that "high mental stress can lead to abnormal uterine contraction frequency."

[0166] In this embodiment, the verification association relationship is a logical connection and mutual verification relationship between the potential supplementary information in the auxiliary task layer and the features selected by the main task layer. For example, if the main task layer selects "abnormal uterine contraction frequency in pregnant women" as a key feature, the potential supplementary information in the auxiliary task layer, "pregnant women have recently experienced high mental stress," could be used to verify and supplement the features of the main task layer through the association relationship that "high mental stress can lead to abnormal uterine contraction frequency."

[0167] The beneficial effects of the above technical solution are: through the screening of the gating vector, it is possible to remove features in the main task layer that have low relevance to the current main clinical task, reduce redundant information, improve the computational efficiency and prediction accuracy of the model, explore the potential supplementary information of the auxiliary task layer and verify and supplement the features of the main task layer, enrich the feature dimension, enable the model to take into account more factors that may affect the clinical results, and improve the comprehensiveness and reliability of the model.

[0168] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for predicting clinical pathways in obstetrics and gynecology based on neuron-level feature selection, characterized in that: include: Step 1: Divide the patient's obstetrics and gynecology electronic medical record data into time series features and baseline features, wherein the time series features include test data, vital signs, and medical records, and the baseline features include demographic characteristics and medical history information; Step 2: Input the time series features into the trained GRU network for feature encoding to obtain the dynamic change patterns of various clinical indicators over time. At the same time, the baseline features are converted into feature vectors through the embedding layer; Step 3: Calculate feature importance based on the patient's most recent obstetrics and gynecology visit record and the visit scenario. Project the dynamic changes of various clinical indicators over time based on feature importance and map the feature vectors to the new representation space to obtain the projection matrix to generate the gating vector. Then, filter and fuse the features of the main task layer and the auxiliary task layer to obtain the final feature representation. Feature importance calculation and feature screening constitute the neuron-level feature selection mechanism. Step 4: Obtain a multi-dimensional prediction result of the clinical pathway based on the final feature representation; The baseline features include demographic characteristics and medical history information, which are converted into dense feature vectors through two learnable embedding layers; Wherein, step 3 includes: Constructing a graph structure by mapping elements mapped to the new representation space and associated clinical indicator features, and performing in-depth analysis of the graph structure based on a graph neural network to obtain mapping logic; Converting the mapping logic into a structured standard representation and inputting the result into a preset representation filling table to obtain a representation vector and a representation weight, wherein the preset representation filling table is a multi-layer perceptron model that takes the semantic and structural information of the standard representation as input and outputs the representation vector and the representation weight; Based on all representation vectors and representation weights, construct the projection matrix , and use the regularization method to constrain the projection matrix, where , Bn represent the representation vectors corresponding to the first mapping element and the nth mapping element respectively; 、 Respectively represent the representation weights corresponding to the first mapping element and the nth mapping element; The projection matrix is ​​quantized using an asymmetric quantization method, which maps the continuous values ​​in the matrix to a finite set of discrete values. The representation ability of the projection matrix is ​​enhanced by combining it with a generative adversarial network. The enhanced projection matrix is ​​multiplied by the original eigenvector to obtain the projection result of each feature. The importance of the feature is evaluated based on the numerical value and distribution of the projection result combined with the representation weight; Based on the feature importance scores, a gating vector is generated using the gating mechanism of the gated recurrent unit.

2. The obstetrics and gynecology clinical pathway prediction method based on neuron-level feature selection according to claim 1, characterized in that: The patient's obstetrics and gynecology electronic medical record data is divided into time series features and baseline characteristics, including: Collecting multimodal data from the obstetrics and gynecology electronic medical record, wherein the multimodal data includes: structured data and text data; Using natural language processing technology to perform semantic analysis on the text data, extract key semantics, and combine the meaning of clinical indicators in the structured data to establish semantic associations between data items; The multimodal data is divided into time series features and baseline features based on semantic association relationships.

3. The obstetrics and gynecology clinical pathway prediction method based on neuron-level feature selection according to claim 1, characterized in that: The process of obtaining the multi-dimensional prediction result of the clinical pathway according to the final feature representation also includes: Extracting the patient's disease type, treatment stage for each disease type, and disease severity from the medical history information, constructing a disease feature vector for each disease type, and performing cluster analysis on all disease feature vectors; Determine, based on the disease mutual influence relationship table, the influence relationship between each cluster center and the current feature vector of the patient in the most recent obstetrics and gynecology visit record, and reversely infer the inspiration effect on each feature based on the influence relationship, wherein the inspiration effect includes: positive inspiration, negative inspiration, and no inspiration, wherein positive inspiration is assigned a value of 1, no inspiration is assigned a value of 0, and negative inspiration is assigned a value of -1; Sort the values ​​of all the inspirations under each feature in descending order of influence. At the same time, sort the values ​​of all the inspirations under each feature in descending order of the value of the inspiration. Performing linear fitting on the curve obtained by the first sorting and the curve obtained by the second sorting, respectively, to obtain a first fitting value and a second fitting value; Sort the first fitting values ​​of all features from large to small to obtain the first serial number of each feature. At the same time, sort the second fitting values ​​of all features from large to small to obtain the second serial number of each feature. According to the first serial number and the second serial number, obtain the first rough attention coefficient; Determine the absolute value of the difference between the first fitting value of the linear fitting result of the curve obtained by one sort and the original value of the corresponding point, determine the locked horizontal interval around the corresponding point, and obtain the original average value of the horizontal interval and the linear average , confirm consistency ; The first rough attention coefficient is modified according to the consistency to obtain a second rough attention coefficient, which is used as the importance of attention of the corresponding feature in the clinical pathway prediction process.

4. The obstetrics and gynecology clinical pathway prediction method based on neuron-level feature selection according to claim 1, characterized in that: The second rough attention coefficient is obtained, including: in, represents the second coarse attention coefficient; represents the first coarse attention coefficient; represents the adjustment parameter; Indicates the corresponding absolute value of the difference; b1 is the set threshold.

5. The obstetrics and gynecology clinical pathway prediction method based on neuron-level feature selection according to claim 1, characterized in that: Each part of the temporal features corresponds to a trained GRU network; Each component in the baseline features corresponds to a learnable embedding layer.

6. The obstetrics and gynecology clinical pathway prediction method based on neuron-level feature selection according to claim 5, characterized in that: Obtain the dynamic changes of various clinical indicators over time, including: Obtain feature information of a set indicator for each part of the time series feature under the corresponding trained GRU network, extract the information structure under the same set indicator from all feature information, and obtain a first dynamic change law vector corresponding to the set indicator, wherein the first dynamic change law vector includes the local change value of the i1th set indicator at time t; Based on the multi-scale feature extraction network, convolution kernels of different sizes are used to perform multi-scale convolution operations on the input time series features, extract feature information of different granularities, and capture the global dynamic change law vector of the time series features, wherein the global dynamic change law vector includes the global change value at time t; based on the global dynamic change law vector, the time series impact on each first dynamic change law vector is determined: in, is the global dynamic change law vector For the i-th first dynamic change law vector The timing impact of 、 They represent parameter vectors respectively; Based on the intelligent layer, the first dynamic change law vector of all set indicators and the global dynamic change law vector are merged and processed, and the dynamic change law of various indicators over time is obtained by combining the temporal influence under different indicators. Among them, the output of the GRU network and the output of the multi-scale feature extraction network are used as the input of the intelligent layer.

7. The obstetrics and gynecology clinical pathway prediction method based on neuron-level feature selection according to claim 1, characterized in that: The features of the main task layer and the auxiliary task layer are screened separately, including: Screening the features of the main task layer according to the value partitioning mechanism and using the gating vector; Based on the auxiliary task layer, potential supplementary information of each gating vector is mined, and the features filtered by the main task layer are verified according to the potential supplementary information, and the features of the main task layer in progress are supplemented according to the verification association relationship.

Images