Information pushing method and system based on subcutaneous injection home management system

Abstract

The invention provides an information pushing method and system based on a hypodermic injection home management system, and the method comprises the steps: obtaining a basic hypodermic injection data mining network, and carrying out the parameter optimization of a current label prediction network through the basic hypodermic injection data mining network, so as to determine a target hypodermic injection data mining network; then, a plurality of verification sample patient injection instances are further obtained, label prediction is conducted according to the current label prediction network, and an extended learning injection data sequence is determined; the data is then utilized for network parameter learning to determine a candidate hypodermic injection data mining network. And finally, testing a performance parameter value of the candidate hypodermic injection data mining network, if the performance parameter value is smaller than a preset value, determining the candidate hypodermic injection data mining network as a current label prediction network, otherwise, determining the candidate hypodermic injection data mining network as a target hypodermic injection data mining network. According to the method and the device, the patient injection instance data can be more accurately mined and pushed in the immunotherapy subcutaneous injection process, so that the accuracy and the efficiency of information pushing are improved.

Description

Technical Field

[0001] This application relates to the technical field of intelligent medical service optimization, and more particularly, to an information push method and system based on a subcutaneous injection home management system. Background Art

[0002] With the continuous progress of medical technology and the increasing demand for home health management, subcutaneous injection for immunotherapy, as a common treatment method, is increasingly widely used in the home environment. However, due to the lack of professional medical knowledge and operation experience of patients and their families, errors and improper operations often occur during the subcutaneous injection for immunotherapy, which not only affects the treatment effect but also may pose potential risks to the health of patients. Therefore, it is particularly important to develop a system and method that can effectively assist in the management of subcutaneous injection in the home environment.

[0003] In existing home management systems for subcutaneous injection of immunotherapy, personalized guidance and suggestions are usually provided by collecting and analyzing patients' injection data. However, these systems often face challenges in processing and analyzing a large amount of complex injection data. On the one hand, due to the diversity and complexity of patients' injection data, the system needs to have strong data processing and pattern recognition capabilities to extract useful information from it. On the other hand, with the continuous update and expansion of medical knowledge, the system also needs to have the ability of continuous learning and optimization to adapt to new injection knowledge and technologies. Summary of the Invention

[0004] To at least overcome the above deficiencies in the prior art, the purpose of this application is to provide an information push method and system based on a subcutaneous injection home management system.

[0005] In a first aspect, this application provides an information push method based on a subcutaneous injection home management system, which is applied to an intelligent medical service system. The method includes: Obtain a basic subcutaneous injection data mining network, which is generated by learning network parameters of a basic neural network based on a basic injection data sequence to be learned. The basic injection data sequence to be learned includes example patient injection instance data recorded by multiple subcutaneous injection home management systems. Each example patient injection instance data includes an example patient injection instance and a corresponding first injection knowledge attribute. The first injection knowledge attribute corresponding to the example patient injection instance is used to represent that the example patient injection instance is a patient injection instance of a target injection knowledge label; Using the basic subcutaneous injection data mining network as the current label prediction network, optimize the network parameters of the current label prediction network to determine the target subcutaneous injection data mining network, which is used to mine patient injection instances of target injection knowledge labels and push the mining results; In a possible implementation manner of the first aspect, the process of optimizing the network parameters includes: Step A: Obtain multiple verified example patient injection instances, and perform label prediction on the verified example patient injection instances according to the current label prediction network to determine the label prediction results corresponding to the verified example patient injection instances. The label prediction results include: patient injection instances that are target injection knowledge labels or patient injection instances that are not target injection knowledge labels; Step B: Determine an extended learning injection data sequence, which contains multiple extended learning injection data. Each extended learning injection data includes a verified example patient injection instance whose label is predicted as a target injection knowledge label by the current label prediction network and the second injection knowledge attribute corresponding to the verified example patient injection instance. The second injection knowledge attribute is used to characterize whether the verified example patient injection instance is actually a patient injection instance of the target injection knowledge label; Step C: Use the verified example patient injection instances of the extended learning injection data as inputs and the second injection knowledge attributes of the extended learning injection data as the expected label prediction results, and perform network parameter learning on the current label prediction network according to the extended learning injection data sequence to determine a candidate subcutaneous injection data mining network; Step D: Test the network performance parameter value of the candidate subcutaneous injection data mining network for mining patient injection instances of target injection knowledge labels. If the network performance parameter value is less than the preset value, determine the candidate subcutaneous injection data mining network as the current label prediction network and jump to A; If the network performance parameter value is not less than the preset value, determine that the candidate subcutaneous injection data mining network is the target subcutaneous injection data mining network.

[0006] In a possible implementation manner of the first aspect, the steps for obtaining the basic injection data sequence to be learned include: Obtain a reference training data sequence, which contains multiple example patient injection instances; Identify each example patient injection instance in the reference training data sequence according to the first identification network. The first identification network classifies the example patient injection instances into first example patient injection instances and non-first example patient injection instances. The first example patient injection instances include incomplete or incorrect example patient injection instances; Identifying each example patient injection instance in the reference training data sequence according to the second identification network, the second identification network classifying the example patient injection instances into second example patient injection instances and non-second example patient injection instances, the second example patient injection instances including example patient injection instances with non-standard operation procedures; Identifying each example patient injection instance in the reference training data sequence according to the third identification network, the third identification network classifying the example patient injection instances into third example patient injection instances and non-third example patient injection instances, the third example patient injection instances including example patient injection instances with abnormal reactions or complications; Identifying each example patient injection instance in the reference training data sequence according to the fourth identification network, the fourth identification network classifying the example patient injection instances into fourth example patient injection instances and non-fourth example patient injection instances, the fourth example patient injection instances including example patient injection instances with specific drugs or treatment regimens; Retaining the first example patient injection instances, second example patient injection instances, third example patient injection instances, and fourth example patient injection instances in the reference training data sequence, and determining the reference training data sequence after retaining the first example patient injection instances, second example patient injection instances, third example patient injection instances, and fourth example patient injection instances as the basic injection data sequence to be learned.

[0007] In a possible implementation manner of the first aspect, the obtaining of the basic subcutaneous injection data mining network includes: Inputting the example patient injection instances of the example patient injection instance data into the basic label prediction network to generate a label prediction result of the basic label prediction network for the example patient injection instances of the example patient injection instance data, the label prediction result including: a patient injection instance that is a target injection knowledge label or a patient injection instance that is not a target injection knowledge label; Comparing the label prediction result with the first injection knowledge attribute of the example patient injection instance data to generate a first learning error of the example patient injection instance data; Calculating a first network cost parameter of the basic label prediction network according to the first learning error of each example patient injection instance data in the injection data sequence to be learned; Updating the network parameter information of the basic label prediction network until the first network cost parameter of the basic label prediction network converges, and determining the basic label prediction network corresponding to the converged first network cost parameter as the basic subcutaneous injection data mining network.

[0008] In a possible implementation manner of the first aspect, in the step C, network parameter learning is performed on the current label prediction network according to the extended learning injection data sequence to determine a candidate subcutaneous injection data mining network, including: Loading the sample patient injection instances of the extended learning injection data into the current label prediction network, generating a label prediction result of the current label prediction network for the sample patient injection instances of the extended learning injection data, where the label prediction result includes: the sample patient injection instance of the extended learning injection data is a patient injection instance of the target injection knowledge label or the sample patient injection instance of the extended learning injection data is not a patient injection instance of the target injection knowledge label; Comparing the label prediction result with the second injection knowledge attribute of the extended learning injection data to generate a second learning error of the extended learning injection data; Calculating a second network cost parameter of the current label prediction network according to the second learning error of each extended learning injection data in the extended learning injection data sequence; Updating the network parameter information of the current label prediction network until the second network cost parameter of the current label prediction network converges, and determining the current label prediction network corresponding to the converged second network cost parameter as the candidate subcutaneous injection data mining network.

[0009] In a possible implementation manner of the first aspect, the determining step of the second injection knowledge attribute of the extended learning injection data includes: Performing feature slicing on the verification sample patient injection instance of the extended learning injection data based on a preset attention feature template to generate an attention patient injection instance corresponding to the verification sample patient injection instance, where the attention feature range corresponding to the preset attention feature template is smaller than the feature range of the verification sample patient injection instance; If the attention patient injection instance is a patient injection instance of the target injection knowledge label, configuring the second injection knowledge attribute to represent that the verification sample patient injection instance is actually a patient injection instance of the target injection knowledge label; If the attention patient injection instance is not a patient injection instance of the target injection knowledge label, configuring the second injection knowledge attribute to represent that the verification sample patient injection instance is actually not a patient injection instance of the target injection knowledge label.

[0010] In a possible implementation manner of the first aspect, determining whether the attention patient injection instance is a patient injection instance of the target injection knowledge label includes: Determining whether any patient injection instance segment in the attention patient injection instance is a patient injection instance of the target injection knowledge label; If any patient injection instance segment in the concerned patient injection instance is a patient injection instance of the target injection knowledge label, determine that the concerned patient injection instance is a patient injection instance of the target injection knowledge label; If no patient injection instance segment in the concerned patient injection instance is a patient injection instance of the target injection knowledge label, determine that the concerned patient injection instance is not a patient injection instance of the target injection knowledge label.

[0011] In a possible implementation manner of the first aspect, the loading the sample patient injection instances of the extended learning injection data into the current label prediction network to generate the label prediction result of the current label prediction network for the sample patient injection instances of the extended learning injection data includes: Performing feature masking processing on some patient injection instance segments of the sample patient injection instances of the extended learning injection data, so that only the patient injection instance segments of the concerned injection sample corresponding to the sample patient injection instances are not masked; Loading the sample patient injection instances of the extended learning injection data after the feature masking processing into the current label prediction network, so that the current label prediction network determines the label prediction result of the sample patient injection instances according to the unmasked patient injection instance segments in the input sample patient injection instances.

[0012] In a possible implementation manner of the first aspect, the comparing the label prediction result with the second injection knowledge attribute of the extended learning injection data to generate the second learning error of the extended learning injection data includes: If the label prediction result is a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verified sample patient injection instance is actually a patient injection instance of the target injection knowledge label, determine that the second learning error of the extended learning injection data is equal to 0; If the label prediction result is a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verified sample patient injection instance is actually not a patient injection instance of the target injection knowledge label, determine that the second learning error of the extended learning injection data is equal to 1; If the label prediction result is not a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verified sample patient injection instance is actually a patient injection instance of the target injection knowledge label, determine that the second learning error of the extended learning injection data is equal to 1; If the predicted label result is not a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verified example patient injection instance is actually not a patient injection instance of the target injection knowledge label, determine that the second learning error of the extended learning injection data is equal to 0; Wherein, calculating the second network cost parameter of the current label prediction network according to the second learning error of each extended learning injection data in the extended learning injection data sequence includes: Accumulate the second learning errors of all the extended learning injection data in the extended learning injection data sequence, and determine the accumulated learning error as the second network cost parameter of the current label prediction network.

[0013] In a possible implementation manner of the first aspect, the method further includes: Obtain a target patient injection instance; Input the target patient injection instance into the target subcutaneous injection data mining network, and generate a label prediction result of the target subcutaneous injection data mining network for the target patient injection instance; If the label prediction result of the target patient injection instance is a patient injection instance of the target injection knowledge label, determine the target patient injection instance as a patient injection instance of the target injection knowledge label, and push the information of all target patient injection instances of the target injection knowledge label to the target medical service terminal.

[0014] In a second aspect, an embodiment of the present application further provides a smart medical service system, the smart medical service system includes a processor and a machine-readable storage medium, and a computer program is stored in the machine-readable storage medium, and the computer program is loaded and executed by the processor to implement the information push method based on the subcutaneous injection home management system in the first aspect above.

[0015] Based on any of the above aspects, the beneficial effects of the present application are as follows: This application mainly realizes more accurate mining and pushing of patient injection instance data by learning and optimizing the network parameters of the basic neural network. Specifically, first, obtain the basic subcutaneous injection data mining network, and use it as the current label prediction network for parameter optimization to determine the target subcutaneous injection data mining network. Then, further obtain multiple verified example patient injection instances, perform label prediction according to the current label prediction network, and determine the extended learning injection data sequence. Next, use this data for network parameter learning to determine the candidate subcutaneous injection data mining network. Finally, test the performance parameter values of the candidate subcutaneous injection data mining network. If they are less than the preset values, determine it as the current label prediction network; otherwise, determine it as the target subcutaneous injection data mining network. Thus, the subcutaneous injection home management system based on the intelligent medical service system can more accurately mine and push patient injection instance data, thereby improving the accuracy and efficiency of information pushing and optimizing the home management experience of patients. At the same time, it also helps to improve the learning and prediction performance of the neural network, providing a new solution for various future medical and health management applications based on the intelligent medical service system.

[0016] That is to say, by using the basic subcutaneous injection data mining network for network parameter learning, this application can effectively extract feature information related to the target injection knowledge label from a large amount of example patient injection instance data recorded by the subcutaneous injection home management system, providing strong data support for subsequent network parameter optimization. Secondly, during the process of network parameter optimization, by continuously obtaining verified example patient injection instances, using the current label prediction network for label prediction, and then combining the extended learning injection data sequence for network parameter learning of the current label prediction network, the prediction accuracy and generalization ability of the network can be gradually improved. This iterative optimization method helps the network better adapt to various complex actual situations and improve the accuracy of mining patient injection instances with the target injection knowledge label. Finally, by testing the network performance parameter values of the candidate subcutaneous injection data mining network and comparing them with the preset values, it can be ensured that the finally determined target subcutaneous injection data mining network has good performance. This will help to more accurately mine patient injection instances with the target injection knowledge label and push the mining results for information, providing more timely and effective home management services for medical workers and patients. Thus, it can effectively improve the accuracy of mining patient injection instances with the target injection knowledge label and provide stronger support for medical services. Brief Description of the Drawings

[0017] Figure 1 It is a schematic flowchart of the information pushing method based on the subcutaneous injection home management system provided by the embodiment of this application; Figure 2A functional structure schematic block diagram of the intelligent medical service system for implementing the above information push method based on the subcutaneous injection home management system provided by the embodiments of the present application. Detailed implementation manners

[0018] Refer to Figure 1 As shown, the present application provides an information push method based on a subcutaneous injection home management system, including the following steps.

[0019] Step S110, obtain a basic subcutaneous injection data mining network, where the basic subcutaneous injection data mining network is generated by performing network parameter learning on a basic neural network according to a basic injection data sequence to be learned, and the basic injection data sequence to be learned includes example patient injection instance data recorded by a plurality of subcutaneous injection home management systems. Each example patient injection instance data includes an example patient injection instance and a corresponding first injection knowledge attribute, and the first injection knowledge attribute corresponding to the example patient injection instance is used to characterize that the example patient injection instance is a patient injection instance of a target injection knowledge label.

[0020] In this embodiment, during the subcutaneous injection process of immunotherapy, the intelligent medical service system first accesses a database or file system storing the basic subcutaneous injection data mining network. This basic subcutaneous injection data mining network is generated by performing network parameter learning on a basic neural network according to a basic injection data sequence to be learned.

[0021] Among them, each example patient injection instance data has a first injection knowledge attribute. This first injection knowledge attribute is used to identify whether an example patient injection instance belongs to a patient injection instance of a target injection knowledge label. In other words, if an example patient injection instance has a certain specific first injection knowledge attribute, then this example patient injection instance is considered to belong to the type of injection instance (i.e., the target injection knowledge label) concerned or studied in this embodiment.

[0022] Thus, using these patient injection instance data with special first injection knowledge attributes, the intelligent medical service system trains the basic neural network. The purpose of the training is to enable the basic subcutaneous injection data mining network to learn to identify those example patient injection instances with specific first injection knowledge attributes, that is, the example patient injection instances of the target injection knowledge label. After the training is completed, the intelligent medical service system obtains a basic subcutaneous injection data mining network that can initially identify the example patient injection instances belonging to the target injection knowledge label.

[0023] Specifically, the example patient injection instance usually contains information related to subcutaneous injection, which may include but is not limited to: Patient Information: Record the basic information of the patient receiving the injection, such as patient name, age, gender, unique identifier (such as medical record number), etc.

[0024] Injection Details: Describe in detail the specific circumstances of the injection, such as injection time, injection site, injection dose, name and concentration of the injected drug, injection speed, etc.

[0025] Device Information: If a smart syringe or other automated device is used, record the unique identifier, model number, serial number of the device, and possible status information (such as whether it is calibrated, battery status, etc.).

[0026] Operator Information: Record the information of the person performing the injection operation, such as name, role (such as medical staff, family member, or the patient himself / herself), etc.

[0027] Reactions and Side Effects: Record any reactions or side effects that occur to the patient during or after the injection, such as pain, redness, swelling, allergic reactions, etc.

[0028] Follow - up: Include follow - up information after the injection, such as patient status updates, whether further medical intervention is required, etc.

[0029] System Log: Record system activities related to the injection, such as data transmission, error messages, system status changes, etc., for troubleshooting and performance optimization.

[0030] Verification and Authorization Information: Ensure that only authorized personnel can access and modify injection data, and record relevant verification and authorization activities. Based on this, the first injection knowledge attribute is a specific attribute or tag directly associated with an example patient injection instance and is used to identify whether this example patient injection instance belongs to a specific, concerned injection category, that is, the target injection knowledge tag. That is to say, the first injection knowledge attribute can help the basic subcutaneous injection data mining network understand and distinguish different injection instances, especially those injection instances belonging to the categories that this embodiment particularly concerns. And the target injection knowledge tag is the category tag of the injection instances that this embodiment hopes the basic subcutaneous injection data mining network can particularly concern and accurately identify, representing a certain specific injection type or situation with clinical or research significance. For example, the target injection knowledge tag may correspond to a rare injection event, an injection under a specific treatment plan, or an injection behavior closely related to a certain disease state, which can be specifically designed by medical staff. The determination of this target injection knowledge tag is usually based on clinical needs, research objectives, or the purpose of data analysis. The task of the basic subcutaneous injection data mining network is to accurately identify those instances with the target injection knowledge tag by learning a large amount of injection instance data.

[0031] Step S120: Using the basic subcutaneous injection data mining network as the current label prediction network, optimize the network parameters of the current label prediction network to determine the target subcutaneous injection data mining network, which is used to mine patient injection instances of target injection knowledge labels and push the mining results for information.

[0032] Although the intelligent medical service system has constructed a basic subcutaneous injection data mining network capable of identifying target injection knowledge labels, this basic subcutaneous injection data mining network may not be precise or efficient enough. Therefore, the intelligent medical service system needs to further optimize this basic subcutaneous injection data mining network.

[0033] Specifically, the intelligent medical service system uses various advanced network optimization techniques and algorithms to fine-tune the parameters of the basic subcutaneous injection data mining network. The goal of optimization is to improve the accuracy and efficiency of the basic subcutaneous injection data mining network in identifying target injection knowledge labels.

[0034] After multiple rounds of optimization and verification, during the subcutaneous injection process of immunotherapy, the intelligent medical service system finally determines a network with the best performance, namely the target subcutaneous injection data mining network. This target subcutaneous injection data mining network can not only accurately identify patient injection instances belonging to target injection knowledge labels, but also remain efficient and stable when processing large-scale data. Once the target subcutaneous injection data mining network is ready, the intelligent medical service system can start using it to mine new patient injection instance data. When the intelligent medical service system discovers a new instance belonging to the target injection knowledge label, it pushes this information to relevant medical personnel or systems so that they can take necessary actions or make decisions in a timely manner.

[0035] In a possible implementation manner, the process of optimizing the network parameters includes: Step A: Obtain multiple verified sample patient injection instances, and perform label prediction on the verified sample patient injection instances according to the current label prediction network to determine the label prediction results corresponding to the verified sample patient injection instances. The label prediction results include: patient injection instances that are target injection knowledge labels or patient injection instances that are not target injection knowledge labels.

[0036] In this embodiment, for some hospitals, a large number of subcutaneous injection treatments are carried out every day. To improve the treatment quality and efficiency, the hospital can use the subcutaneous injection data mining network to automatically identify and manage patient injection instances. In particular, the hospital hopes that the current label prediction network can accurately identify a specific injection category - namely the target injection knowledge label, which may be related to the clinical trial of a new drug.

[0037] Thus, the intelligent medical service system can retrieve 500 subcutaneous injection instances within the most recent week from the hospital's electronic medical record system as verification sample patient injection instances. These verification sample patient injection instances include the basic information of the patients, details of the injected drugs, injection sites, injection times, etc. Then, the current basic subcutaneous injection data mining network can be used to predict these 500 verification sample patient injection instances. The basic subcutaneous injection data mining network outputs the probability of each verification sample patient injection instance belonging to the target injection knowledge label. According to a preset threshold (such as 0.5), the intelligent medical service system classifies these verification sample patient injection instances into two categories: instances that are target injection knowledge labels and instances that are not target injection knowledge labels.

[0038] Step B: Determine an extended learning injection data sequence. The extended learning injection data sequence includes multiple extended learning injection data. Each extended learning injection data includes a verification sample patient injection instance of a patient injection instance predicted by the current label prediction network label as a target injection knowledge label and the corresponding second injection knowledge attribute of the verification sample patient injection instance. The second injection knowledge attribute is used to characterize whether the verification sample patient injection instance is actually a patient injection instance of the target injection knowledge label.

[0039] In this embodiment, by comparing with the actual labels (manually marked by medical experts), the intelligent medical service system finds that 50 of the verification sample patient injection instances are misclassified. Among these 50 verification sample patient injection instances, 30 are actually target injection knowledge labels but are misjudged as non-target injection knowledge labels, and 20 are non-target injection knowledge labels but are misjudged as target injection knowledge labels.

[0040] In addition, the intelligent medical service system also identifies 40 verification sample patient injection instances. Although their classifications are correct, the current label prediction network has a low prediction confidence in them (i.e., the probability value is close to the threshold).

[0041] Thus, these 90 instances (50 misclassifications + 40 low confidences) can be combined with their actual labels to form an extended learning injection data sequence.

[0042] Step C: Using the verification sample patient injection instances of the extended learning injection data as the input and the second injection knowledge attribute of the extended learning injection data as the expected label prediction result, perform network parameter learning on the current label prediction network according to the extended learning injection data sequence to determine a candidate subcutaneous injection data mining network.

[0043] In this embodiment, necessary preprocessing can be performed on the extended learning injection data sequence, such as data cleaning, standardization, etc. Then, these extended learning injection data are used to retrain the current label prediction network. By adjusting the weights and parameters of the current label prediction network, the misclassifications and low-confidence predictions identified in step B are reduced.

[0044] During the training process, the intelligent medical service system monitors the performance changes of the current label prediction network in real time, such as the decrease in the loss function, the improvement in accuracy, etc.

[0045] Step D: Test the network performance parameter values of the candidate subcutaneous injection data mining network for mining patient injection instances of the target injection knowledge label. If the network performance parameter values are less than the preset values, determine the candidate subcutaneous injection data mining network as the current label prediction network and jump to A. If the network performance parameter values are not less than the preset values, determine that the candidate subcutaneous injection data mining network is the target subcutaneous injection data mining network.

[0046] In this embodiment, the intelligent medical service system can retrieve another 200 subcutaneous injection instances that did not participate in the previous training from the hospital's electronic medical record system as the test set. Use the candidate subcutaneous injection data mining network to predict these 200 instances and compare them with the actual labels. The evaluation metrics include accuracy, recall rate, F1 score, etc. If the performance of the candidate subcutaneous injection data mining network reaches or exceeds the preset standard (such as accuracy ≥ 95%), it is determined as the target subcutaneous injection data mining network and prepared to be deployed to the actual production environment; otherwise, analyze the deficiencies of the candidate subcutaneous injection data mining network based on the test results and repeat steps A to C for further optimization.

[0047] Based on the above steps, this application mainly realizes more accurate mining and pushing of patient injection instance data by performing network parameter learning and optimization on the basic neural network. Specifically, first, obtain the basic subcutaneous injection data mining network, and use it as the current label prediction network for parameter optimization to determine the target subcutaneous injection data mining network. Then, further obtain multiple verified example patient injection instances, perform label prediction according to the current label prediction network, and determine the extended learning injection data sequence. Next, use this data for network parameter learning to determine the candidate subcutaneous injection data mining network. Finally, test the performance parameter values of the candidate subcutaneous injection data mining network. If it is less than the preset value, it is determined as the current label prediction network; otherwise, it is determined as the target subcutaneous injection data mining network. Thus, the subcutaneous injection home management system based on the intelligent medical service system can more accurately mine and push patient injection instance data, thereby improving the accuracy and efficiency of information pushing and optimizing the home management experience of patients. At the same time, it also helps to improve the learning and prediction performance of the neural network, providing a new solution for various future medical and health management applications based on the intelligent medical service system.

[0048] That is to say, by using the basic subcutaneous injection data mining network for network parameter learning, this application can effectively extract feature information related to the target injection knowledge label from a large amount of example patient injection instance data recorded in the subcutaneous injection home management system, providing strong data support for subsequent network parameter optimization. Secondly, during the process of network parameter optimization, by continuously obtaining verified example patient injection instances, performing label prediction using the current label prediction network, and then combining the extended learning injection data sequence to perform network parameter learning on the current label prediction network, the prediction accuracy and generalization ability of the network can be gradually improved. This iterative optimization method helps the network better adapt to various complex actual situations and improve the accuracy of mining patient injection instances with the target injection knowledge label. Finally, by testing the network performance parameter values of the candidate subcutaneous injection data mining network and comparing them with the preset values, it can be ensured that the finally determined target subcutaneous injection data mining network has good performance. This will help to more accurately mine patient injection instances with the target injection knowledge label and push the mining results for information, providing more timely and effective home management services for medical workers and patients. Thus, the accuracy of mining patient injection instances with the target injection knowledge label can be effectively improved, providing stronger support for medical services.

[0049] In a possible implementation manner, the obtaining step of the basic injection data sequence to be learned includes: Step A110, obtain a reference training data sequence, where the reference training data sequence includes multiple example patient injection instances.

[0050] During the subcutaneous injection process of immunotherapy in this embodiment, the intelligent medical service system can send a request to obtain the reference training data sequence related to subcutaneous injection through the API interface of the electronic medical record system (EMR) of the medical institution. The EMR system responds to the request and returns a JSON-format file containing all subcutaneous injection records in the past year. Each record is a detailed JSON object, including fields such as patient ID, injection time, drug name, dose, injection site, and operator ID. After receiving the reference training data sequence, the intelligent medical service system parses it into a structured data sequence and stores it in memory for subsequent processing.

[0051] Step A120, identify each sample patient injection instance in the reference training data sequence according to the first recognition network. The first recognition network classifies the sample patient injection instances into first sample patient injection instances and non-first sample patient injection instances. The first sample patient injection instances include incomplete or incorrect sample patient injection instances.

[0052] The first recognition network is a deep learning model that has been pre-trained to identify incomplete or incorrect injection records. The intelligent medical service system inputs the reference training data sequence into this first recognition network and analyzes each sample patient injection instance one by one in a batch processing manner. The first recognition network identifies incomplete or incorrect records by checking the integrity of fields (such as whether there are null or invalid values), the consistency of data (such as whether the timestamp is reasonable, whether the dose is within the normal range), and the logical rationality (such as whether the injection time is earlier than the patient's admission time). For example, if the first recognition network finds that the drug dose field in a certain record is empty, or the injection time is later than the patient's discharge time, it will mark this record as a first sample patient injection instance.

[0053] Step A130, identify each sample patient injection instance in the reference training data sequence according to the second recognition network. The second recognition network classifies the sample patient injection instances into second sample patient injection instances and non-second sample patient injection instances. The second sample patient injection instances include sample patient injection instances with non-standard operation procedures.

[0054] The second recognition network is also a deep learning model, which focuses on identifying injection instances with non-standard operation processes. The intelligent medical service system inputs the reference training data sequence into this second recognition network again for batch processing. The second recognition network identifies non-standard operation processes by analyzing the operation details in the injection instances. This may include whether the selection of the injection site complies with the specifications, whether the drug ratio is accurate, whether the injection speed is too fast or too slow, etc. For example, if the second recognition network finds that the injection site in a certain record does not conform to the standard operation process (such as injecting in the thigh when it should be injected in the upper arm), or the injection speed is too fast (such as the standard operation process requires slow injection but the record shows rapid completion), it will mark this record as a second-sample patient injection instance. Similarly, the intelligent medical service system stores these marked instances separately in a list.

[0055] Step A140, identify each sample patient injection instance in the reference training data sequence according to the third recognition network. The third recognition network classifies the sample patient injection instances into third-sample patient injection instances and non-third-sample patient injection instances. The third-sample patient injection instances include sample patient injection instances with adverse reactions or complications.

[0056] The third recognition network is a deep learning model based on natural language processing, which can identify instances with adverse reactions or complications during the injection process. The intelligent medical service system extracts the patient reactions and subsequent medical record fields in the reference training data sequence and inputs them into this third recognition network. The third recognition network identifies the descriptions of abnormalities or complications by analyzing the text data. For example, it may search for keywords or phrases related to allergic reactions, local inflammation, infections, etc. If the third recognition network finds these keywords or phrases in the text, it will mark the corresponding record as a third-sample patient injection instance. The intelligent medical service system stores these marked instances separately from other types of instances.

[0057] Step A150, identify each sample patient injection instance in the reference training data sequence according to the fourth recognition network. The fourth recognition network classifies the sample patient injection instances into fourth-sample patient injection instances and non-fourth-sample patient injection instances. The fourth-sample patient injection instances include sample patient injection instances with specific drugs or treatment regimens.

[0058] The fourth recognition network is a deep learning model specifically used to recognize injection instances related to specific drugs or treatment regimens. The intelligent medical service system extracts fields such as drug names, dosages, and treatment regimens from the reference training data sequence and inputs them into this fourth recognition network. The fourth recognition network identifies relevant instances by matching a predefined list of specific drugs or treatment regimens. For example, it may look for records related to a new anti-cancer drug or a special immunotherapy regimen. If the fourth recognition network finds records that match the drugs or treatment regimens in the list, it marks these records as fourth sample patient injection instances. The intelligent medical service system stores these marked instances separately in a list for further in-depth analysis.

[0059] Step A160: Retain the first sample patient injection instance, the second sample patient injection instance, the third sample patient injection instance, and the fourth sample patient injection instance in the reference training data sequence, and determine the reference training data sequence after the retained first sample patient injection instance, second sample patient injection instance, third sample patient injection instance, and fourth sample patient injection instance as the basic injection data sequence to be learned.

[0060] After completing the recognition of the above four steps, the intelligent medical service system has four lists, which store the first, second, third, and fourth sample patient injection instances respectively. The intelligent medical service system combines these lists into a comprehensive list and removes duplicate instances (if any). Each instance in the comprehensive list contains all the information of the original record and the corresponding labels (such as the first sample, the second sample, etc.). This comprehensive list constitutes the basic injection data sequence to be learned, which will be used for subsequent data mining model training and optimization work.

[0061] In a possible implementation manner, the obtaining of the basic subcutaneous injection data mining network includes: Step B110: Input the sample patient injection instance of the sample patient injection instance data into the basic label prediction network to generate a label prediction result of the basic label prediction network for the sample patient injection instance of the sample patient injection instance data. The label prediction result includes: a patient injection instance that is a target injection knowledge label or a patient injection instance that is not a target injection knowledge label.

[0062] During the subcutaneous injection of immunotherapy in this embodiment, the intelligent medical service system first loads the pre-constructed basic label prediction network, which is a deep learning model dedicated to predicting the labels of patient injection instances. The intelligent medical service system batch extracts sample patient injection instance data from the previously constructed basic injection data sequence to be learned. These sample patient injection instance data contain various types of injection instances, including normal and abnormal ones. The intelligent medical service system inputs the extracted sample patient injection instance data into the basic label prediction network, thereby analyzing each sample patient injection instance data and attempting to predict whether it is a target injection knowledge label. The result output by the basic label prediction network is the probability that each of these sample patient injection instances is predicted as "a patient injection instance that is a target injection knowledge label" or "a patient injection instance that is not a target injection knowledge label".

[0063] Step B120: Compare the label prediction result with the first injection knowledge attribute of the sample patient injection instance data to generate the first learning error of the sample patient injection instance data.

[0064] The intelligent medical service system obtains the first injection knowledge attributes of the sample patient injection instance data. These first injection knowledge attributes are pre-annotated true labels used for comparison with the network prediction results. The intelligent medical service system compares the prediction result of the basic label prediction network with the first injection knowledge attribute. If the prediction result is consistent with the true label, it is considered a correct prediction; otherwise, it is considered an incorrect prediction. The intelligent medical service system calculates the first learning error of each sample patient injection instance data. The first learning error reflects the degree of difference between the network prediction result and the true label.

[0065] Step B130: Calculate the first network cost parameter of the basic label prediction network based on the first learning error of each sample patient injection instance data in the injection data sequence to be learned.

[0066] The intelligent medical service system calculates the first network cost parameter of the entire basic label prediction network according to the first learning error of each sample patient injection instance data. The first network cost parameter is an index measuring the overall performance of the network and is usually proportional to the learning error. The first network cost parameter may include mean square error, cross-entropy loss, etc., depending on the nature of the problem and the training objective of the network.

[0067] Step B140: Update the network parameter information of the basic label prediction network until the first network cost parameter of the basic label prediction network converges, and determine the basic label prediction network corresponding to the converged first network cost parameter as the basic subcutaneous injection data mining network.

[0068] The intelligent medical service system uses an optimization algorithm (such as gradient descent) to update the network parameter information of the basic label prediction network, aiming to reduce the first network cost parameter and improve the prediction accuracy of the network. The update process will continue until the first network cost parameter converges, that is, the performance of the network no longer improves significantly. Once the first network cost parameter converges, the intelligent medical service system will save the current network parameter information and determine this converged network as the basic subcutaneous injection data mining network. This network can be used for subsequent data mining tasks, such as identifying target knowledge labels in new injection instances.

[0069] In a possible implementation manner, in step C, learning network parameters of the current label prediction network according to the extended learning injection data sequence to determine a candidate subcutaneous injection data mining network includes: Step C110, loading the sample patient injection instances of the extended learning injection data into the current label prediction network, and generating a label prediction result of the current label prediction network for the sample patient injection instances of the extended learning injection data, where the label prediction result includes: the sample patient injection instance of the extended learning injection data is a patient injection instance of a target injection knowledge label or the sample patient injection instance of the extended learning injection data is not a patient injection instance of a target injection knowledge label.

[0070] For example, during the subcutaneous injection of immunotherapy, the intelligent medical service system reads the extended learning injection data sequence from the storage system, and these extended learning injection data contain more diverse and complex patient injection instances. The intelligent medical service system loads the sample patient injection instances in the extended learning injection data into the current label prediction network one by one. Each sample patient injection instance contains rich information, such as the basic information of the patient, injection operation details, drug usage, etc. The current label prediction network analyzes and processes each loaded sample patient injection instance, and uses the learned knowledge to try to predict whether these instances correspond to the target injection knowledge label. The result output by the current label prediction network is the probability or certainty score that each instance is predicted as "a patient injection instance of a target injection knowledge label" or "not a patient injection instance of a target injection knowledge label".

[0071] Step C120, comparing the label prediction result with the second injection knowledge attribute of the extended learning injection data to generate a second learning error of the extended learning injection data.

[0072] The intelligent medical service system obtains the second injection knowledge attributes of each sample patient's injection instance in the extended learning injection data. These second injection knowledge attributes are the true labels pre-annotated by experts and represent whether the instance truly matches the target injection knowledge. The intelligent medical service system compares the prediction result of the current label prediction network with the second injection knowledge attributes. If the prediction result is consistent with the true label, the prediction is considered correct; otherwise, it is considered incorrect. The intelligent medical service system calculates the second learning error of each extended learning injection data. This error reflects the degree of difference between the network prediction result and the true label and is an important indicator for evaluating the network performance.

[0073] Step C130, calculate the second network cost parameter of the current label prediction network according to the second learning error of each extended learning injection data in the extended learning injection data sequence.

[0074] The intelligent medical service system calculates the second network cost parameter of the entire current label prediction network according to the second learning error of each extended learning injection data. This parameter is a comprehensive indicator used to measure the overall performance of the network on the extended learning injection data.

[0075] The second network cost parameter may include mean square error, cross-entropy loss, etc., depending on the nature of the problem and the training objective of the network. The intelligent medical service system will evaluate the network performance according to this parameter and guide the subsequent network parameter update.

[0076] Step C140, update the network parameter information of the current label prediction network until the second network cost parameter of the current label prediction network converges, and determine the current label prediction network corresponding to the converged second network cost parameter as the candidate subcutaneous injection data mining network.

[0077] The intelligent medical service system uses optimization algorithms (such as stochastic gradient descent, Adam, etc.) to update the network parameter information of the current label prediction network. The goal of the update is to reduce the second network cost parameter and improve the prediction accuracy of the network on the extended learning injection data. The update process will continue for multiple iterations until the second network cost parameter converges or reaches the preset stop condition. In each iteration, the intelligent medical service system will adjust the network parameters according to the current network performance and learning error. Once the second network cost parameter converges or meets the stop condition, the intelligent medical service system will save the current network parameter information and determine this converged network as the candidate subcutaneous injection data mining network. This network has been trained with extended learning injection data and has stronger generalization ability and higher prediction accuracy.

[0078] The intelligent medical service system can further evaluate and test the candidate subcutaneous injection data mining network to ensure its performance and reliability in practical applications. If the evaluation results are satisfactory, this network can be deployed to actual data mining tasks for identifying and analyzing target injection knowledge tags in new patient injection instances.

[0079] In a possible implementation manner, the determining step of the second injection knowledge attribute of the extended learning injection data includes: Step D110, perform feature slicing on the verification example patient injection instance of the extended learning injection data based on a preset attention feature template to generate an attention patient injection instance corresponding to the verification example patient injection instance, where the attention feature range corresponding to the preset attention feature template is smaller than the feature range of the verification example patient injection instance.

[0080] In this embodiment, the intelligent medical service system first loads a preset attention feature template. This preset attention feature template defines which features are most relevant to the target injection knowledge tag and is predefined by experts based on experience and domain knowledge. The intelligent medical service system extracts verification example patient injection instances from the extended learning injection data. These verification example patient injection instances are key data for verifying data quality and label accuracy. For each verification example patient injection instance, the intelligent medical service system performs feature slicing using the preset attention feature template. This process is like cutting out the most critical and informative part from a complete dataset. The sliced data is called the attention patient injection instance, which only contains the features most relevant to the target injection knowledge tag.

[0081] Step D120, if the attention patient injection instance is a patient injection instance of the target injection knowledge tag, configure the second injection knowledge attribute to be used to represent that the verification example patient injection instance is actually a patient injection instance of the target injection knowledge tag.

[0082] Step D130, if the attention patient injection instance is not a patient injection instance of the target injection knowledge tag, configure the second injection knowledge attribute to be used to represent that the verification example patient injection instance is actually not a patient injection instance of the target injection knowledge tag.

[0083] In this embodiment, the intelligent medical service system analyzes and evaluates the injection instances of the patients of interest after slicing. Using the existing knowledge and algorithms, the intelligent medical service system attempts to determine whether these injection instances of the patients of interest correspond to the target injection knowledge tags. If an injection instance of a patient of interest is determined to match the target injection knowledge tag, the intelligent medical service system will configure its corresponding second injection knowledge attribute as "an injection instance of a patient that is a target injection knowledge tag". This means that the original verified example patient injection instance is also considered to match the target tag.

[0084] If an injection instance of a patient of interest is determined not to match the target injection knowledge tag, the intelligent medical service system will configure its corresponding second injection knowledge attribute as "an injection instance of a patient that is not a target injection knowledge tag". This also means that the original verified example patient injection instance is considered not to match the target tag.

[0085] In this way, the intelligent medical service system configures an accurate second injection knowledge attribute for each verified example patient injection instance in the extended learning injection data, providing a reliable data basis for subsequent network training and optimization.

[0086] Thus, the intelligent medical service system successfully determines the second injection knowledge attribute of the extended learning injection data, providing important label information for subsequent data mining network training. This process not only improves the utilization efficiency of the data and the accuracy of training, but also helps to enhance the reliability and practicality of the final mining results.

[0087] In a possible implementation manner, determining whether the injection instance of the patient of interest is an injection instance of a patient with a target injection knowledge tag includes: determining whether any one patient injection instance segment in the injection instance of the patient of interest is an injection instance of a patient with a target injection knowledge tag. If any one patient injection instance segment in the injection instance of the patient of interest is an injection instance of a patient with a target injection knowledge tag, it is determined that the injection instance of the patient of interest is an injection instance of a patient with a target injection knowledge tag. If no patient injection instance segment in the injection instance of the patient of interest is an injection instance of a patient with a target injection knowledge tag, it is determined that the injection instance of the patient of interest is not an injection instance of a patient with a target injection knowledge tag.

[0088] In this embodiment, the intelligent medical service system extracts a patient injection instance segment from the injection instance of the patient of interest. This patient injection instance segment contains key feature information related to the target injection knowledge.

[0089] The intelligent medical service system inputs the extracted patient injection instance segments into the label prediction network, which outputs a probability value or a certainty score indicating the likelihood that the patient injection instance segment contains the target injection knowledge label. Based on the output of the label prediction network, the intelligent medical service system determines whether the patient injection instance segment is a patient injection instance with the target injection knowledge label. If the probability value or score exceeds a preset threshold, it is considered an instance of the target label; otherwise, it is not.

[0090] The intelligent medical service system repeats the above judgment process for each patient injection instance segment in the concerned patient injection instances. If the target injection knowledge label is found in any patient injection instance segment, the intelligent medical service system immediately determines that the entire concerned patient injection instance is a patient injection instance with the target injection knowledge label. This is because as long as one segment contains the target label, it means that the entire instance may contain valuable information. If the target injection knowledge label is not found in all patient injection instance segments, the intelligent medical service system determines that the entire concerned patient injection instance is not a patient injection instance with the target injection knowledge label. This indicates that the instance may not contain valuable information relevant to the current task.

[0091] In this way, the intelligent medical service system can assign an accurate label to each concerned patient injection instance, and these labels will be used in the subsequent data mining network training and optimization processes.

[0092] In a possible implementation manner, step C110 may include: Step C111, performing feature masking processing on some patient injection instance segments of the sample patient injection instances of the extended learning injection data, so that only the patient injection instance segments corresponding to the concerned injection samples of the sample patient injection instances are not masked.

[0093] In this embodiment, the intelligent medical service system first extracts the sample patient injection instances from the extended learning injection data. These sample patient injection instances contain rich patient information and various features related to injections.

[0094] To improve the prediction efficiency and reduce the interference of irrelevant features, the intelligent medical service system performs feature masking processing on some patient injection instance segments using preset feature masking rules, so that the current label prediction network can only observe the part of information that is most relevant to the target injection knowledge label. After the feature masking processing, only the patient injection instance segments corresponding to the concerned injection samples in the sample patient injection instances are not masked. These segments are considered the most critical parts for the label prediction task by the intelligent medical service system.

[0095] Step C112: Load the sample patient injection instances of the extended learning injection data after feature masking processing into the current label prediction network, so that the current label prediction network determines the label prediction results of the sample patient injection instances based on the unmasked patient injection instance segments in the input sample patient injection instances.

[0096] The intelligent medical service system loads the sample patient injection instances of the extended learning injection data after feature masking processing into the current label prediction network. This current label prediction network has been preliminarily trained and has certain prediction capabilities. The current label prediction network receives these processed data as input and analyzes and processes them according to its internal complex algorithms and learned knowledge. In this process, the network mainly focuses on those unmasked patient injection instance segments because these segments contain the information most likely to affect the label prediction results. After a series of calculations and processes, the current label prediction network outputs the label prediction results of each sample patient injection instance. This result is a probability value or a certainty score, indicating the likelihood that the instance belongs to the target injection knowledge label. The intelligent medical service system saves these prediction results for use in subsequent steps. These results will be used for important tasks such as evaluating the performance of the network, calculating the learning error, and updating the network parameters.

[0097] Through the above steps, the intelligent medical service system has successfully processed the sample patient injection instances of the extended learning injection data using the current label prediction network and generated the corresponding label prediction results.

[0098] In a possible implementation manner, step C120 may include: Step C121: If the label prediction result is a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verified sample patient injection instance is actually a patient injection instance of the target injection knowledge label, determine that the second learning error of the extended learning injection data is equal to 0.

[0099] Step C122: If the label prediction result is a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verified sample patient injection instance is actually not a patient injection instance of the target injection knowledge label, determine that the second learning error of the extended learning injection data is equal to 1.

[0100] Step C123: If the label prediction result is not a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verified sample patient injection instance is actually a patient injection instance of the target injection knowledge label, determine that the second learning error of the extended learning injection data is equal to 1.

[0101] Step C124, if the label prediction result is not a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verified example patient injection instance is actually not a patient injection instance of the target injection knowledge label, determine that the second learning error of the extended learning injection data is equal to 0.

[0102] In this embodiment, the intelligent medical service system first obtains the label prediction results of the example patient injection instances of the extended learning injection data. These results are generated by the current label prediction network and represent the possibility of each instance belonging to the target injection knowledge label.

[0103] At the same time, the intelligent medical service system also obtains the corresponding second injection knowledge attributes of these example patient injection instances. These attributes are determined according to the previous data processing and analysis steps and are used to characterize whether each instance actually contains the target injection knowledge label.

[0104] The intelligent medical service system compares the label prediction result of each instance with its corresponding second injection knowledge attribute to check whether the prediction result of the network is consistent with the actual attribute.

[0105] If the label prediction result indicates that an instance is a patient injection instance of the target injection knowledge label, and the second injection knowledge attribute of this instance also indeed indicates that it is an instance of the target label (i.e., the two are consistent), then the intelligent medical service system determines that the second learning error of this instance is 0. This means that the network's prediction of this instance is completely accurate.

[0106] If the label prediction result indicates that an instance is a patient injection instance of the target injection knowledge label, but the second injection knowledge attribute of this instance indicates that it is actually not an instance of the target label (i.e., the two are inconsistent), then the intelligent medical service system determines that the second learning error of this instance is 1. This indicates that the network's prediction of this instance is incorrect.

[0107] Conversely, if the label prediction result indicates that an instance is not a patient injection instance of the target injection knowledge label, but the second injection knowledge attribute of this instance indicates that it is actually an instance of the target label (i.e., the two are inconsistent), then the intelligent medical service system also determines that the second learning error of this instance is 1. This is also a case of incorrect network prediction.

[0108] Finally, if the label prediction result indicates that an instance is not a patient injection instance of the target injection knowledge label, and the second injection knowledge attribute of this instance also indeed indicates that it is not an instance of the target label (i.e., the two are consistent), then the intelligent medical service system determines that the second learning error of this instance is 0. This means that the network's prediction of this instance is also accurate.

[0109] Through the above steps, the intelligent medical service system can calculate the second learning error of each sample patient's injection instance in the extended learning injection data, so as to evaluate the prediction accuracy of the entire data set. This error information will be used in the subsequent network optimization and training processes to improve the performance and prediction ability of the network.

[0110] Among them, step C130 may include: accumulating the second learning errors of all the extended learning injection data in the extended learning injection data sequence, and determining the accumulated learning error as the second network cost parameter of the current label prediction network.

[0111] In this embodiment, the intelligent medical service system performs an accumulation operation on these second learning errors. The accumulation process is like the intelligent medical service system summarizing all the errors to obtain an overall error index. This index can reflect the prediction accuracy of the current label prediction network on the entire extended learning injection data sequence. The intelligent medical service system determines the accumulated learning error as the second network cost parameter of the current label prediction network. This cost parameter is a quantitative index used to evaluate the performance of the network in the prediction task. The larger the value of the cost parameter, the greater the prediction error of the network and the worse the performance; on the contrary, the smaller the value of the cost parameter, the higher the prediction accuracy of the network and the better the performance.

[0112] Through this step, the intelligent medical service system successfully calculates the second network cost parameter of the current label prediction network. This parameter will provide an important reference basis for the subsequent network optimization and training, helping the intelligent medical service system continuously improve the prediction ability and performance of the network.

[0113] In a possible implementation manner, the method further includes: Step S130, obtaining a target patient's injection instance.

[0114] Step S140, inputting the target patient's injection instance into the target subcutaneous injection data mining network to generate a label prediction result of the target subcutaneous injection data mining network for the target patient's injection instance.

[0115] Step S150, if the label prediction result of the target patient's injection instance is a patient injection instance with a target injection knowledge label, determining the target patient's injection instance as a patient injection instance with a target injection knowledge label, and pushing the information of all target patient injection instances with target injection knowledge labels to the target medical service terminal.

[0116] In this embodiment, the intelligent medical service system receives a request from the medical service system, and the request contains injection instance data of the target patient. This data is recorded by medical staff when providing injection services to patients and contains rich patient information and various details related to injections. The intelligent medical service system saves this data as the injection instance of the target patient and prepares for subsequent analysis and processing.

[0117] Specifically, the intelligent medical service system inputs the obtained injection instance of the target patient into the trained target subcutaneous injection data mining network. This network is a deep learning model dedicated to analyzing patient injection data and has powerful data processing and prediction capabilities.

[0118] After receiving the input, the target subcutaneous injection data mining network starts to calculate and process according to its internal complex algorithms and learned knowledge. This process involves a large amount of data operations and pattern recognition, aiming to extract useful information from the input injection instance and predict the label it belongs to. After a series of calculations, the target subcutaneous injection data mining network outputs the label prediction result of the injection instance of the target patient. This result is a probability value or a certainty score, indicating the likelihood that the instance belongs to the target injection knowledge label.

[0119] After obtaining the label prediction result output by the target subcutaneous injection data mining network, the intelligent medical service system performs judgment processing. If the prediction result shows that the injection instance of the target patient belongs to the target injection knowledge label (i.e., the prediction result is consistent with the target injection knowledge label), then the intelligent medical service system determines this instance as the patient injection instance of the target injection knowledge label.

[0120] Once the patient injection instances of the target injection knowledge label are determined, the intelligent medical service system will organize the relevant information of these instances into data packets and prepare to push them to the target medical service terminal. This information includes but is not limited to the basic information of the patient, the specific details of the injection, as well as the label and probability predicted by the network, etc. Finally, the intelligent medical service system sends the data packet to the target medical service terminal through a secure network connection. After receiving this information, medical staff can make further diagnosis and treatment decisions according to the actual situation, thereby improving the quality and efficiency of medical services.

[0121] Figure 2 An intelligent medical service system 100 is provided in an embodiment of the present application, including a processor 1001, a memory 1003, and program code stored on the memory 1003. The processor 1001 executes the above program code to implement the steps of the information push method based on the subcutaneous injection home management system.

[0122] Figure 2The intelligent medical service system 100 shown includes: a processor 1001 and a memory 1003. Among them, the processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the intelligent medical service system 100 may further include a transceiver 1004, and the transceiver 1004 can be used for data interaction between the intelligent medical service system 100 and other intelligent medical service systems 100, such as data sending and / or data receiving, etc. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of the intelligent medical service system 100 does not constitute a limitation on the embodiments of the present application.

[0123] The processor 1001 can be a CPU (Central Processing Unit, central processor), a general-purpose processor, a DSP (Digital Signal Processor, data signal processor), an ASIC (Application Specific Integrated Circuit, application-specific integrated circuit), an FPGA (Field Programmable Gate Array, field programmable gate array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It can implement or execute various exemplary logic blocks, modules and circuits described in connection with the disclosure of the present application. The processor 1001 can also be a combination that implements computing functions, such as a combination including one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0124] The bus 1002 may include a path for transmitting information between the above components. The bus 1002 can be a PCI (Peripheral Component Interconnect, peripheral component interconnect standard) bus or an EISA (Extended Industry Standard Architecture, extended industry standard structure) bus, etc. The bus 1002 can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, Figure 2 only a thick line is used to represent it in the figure, but it does not mean that there is only one bus or one type of bus.

[0125] The memory 1003 can be a ROM (Read Only Memory), or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory), or other types of dynamic storage devices that can store information and instructions. It can also be an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, other magnetic storage devices, or any other medium that can be used to carry or store program code and can be read by a computer, which is not limited herein.

[0126] The memory 1003 is used to store the program code for implementing the embodiments of the present application and is controlled by the processor 1001 for execution. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0127] The embodiments of the present application provide a computer-readable storage medium, on which program code is stored. When the program code is executed by a processor, the steps and corresponding content of the foregoing method embodiments can be implemented.

[0128] The above are only optional implementation manners of some implementation scenarios of the present application. It should be noted that for those of ordinary skill in the art, without departing from the technical concept of the solution of the present application, based on other similar implementation means according to the technical idea of the present application, they also belong to the protection scope of the embodiments of the present application.

Claims

1. An information push method based on a subcutaneous injection home management system, characterized in that: Applied to a smart medical service system, the method includes: Obtain a basic subcutaneous injection data mining network, wherein the basic subcutaneous injection data mining network is generated by learning network parameters of a basic neural network based on a basic injection data sequence to be learned, wherein the basic injection data sequence to be learned includes a plurality of sample patient injection instance data recorded by a subcutaneous injection home management system, wherein each sample patient injection instance data includes a sample patient injection instance and a corresponding first injection knowledge attribute, wherein the first injection knowledge attribute corresponding to the sample patient injection instance is used to characterize that the sample patient injection instance is a patient injection instance of a target injection knowledge label; Taking the basic subcutaneous injection data mining network as the current label prediction network, optimizing the network parameters of the current label prediction network to determine a target subcutaneous injection data mining network, wherein the target subcutaneous injection data mining network is used to mine patient injection instances of target injection knowledge labels and push the mining results as information; The process of optimizing the network parameters includes: Step A, obtaining multiple verification sample patient injection instances, and performing label prediction on the verification sample patient injection instances according to the current label prediction network to determine the label prediction results corresponding to the verification sample patient injection instances, wherein the label prediction results include: patient injection instances with target injection knowledge labels or patient injection instances without target injection knowledge labels; Step B, determining an extended learning injection data sequence, wherein the extended learning injection data sequence comprises a plurality of extended learning injection data, each of which comprises a verification sample patient injection instance of a patient injection instance predicted by the current label prediction network label as a target injection knowledge label and a second injection knowledge attribute corresponding to the verification sample patient injection instance, wherein the second injection knowledge attribute is used to characterize whether the verification sample patient injection instance is actually a patient injection instance of the target injection knowledge label; Step C, taking the injection instance of the verification sample patient of the extended learning injection data as input, and the second injection knowledge attribute of the extended learning injection data as the expected label prediction result, and performing network parameter learning on the current label prediction network according to the extended learning injection data sequence to determine a candidate subcutaneous injection data mining network; Step D, testing the network performance parameter value of the candidate subcutaneous injection data mining network for mining the patient injection instance of the target injection knowledge label, if the network performance parameter value is less than a preset value, determining the candidate subcutaneous injection data mining network as the current label prediction network and jumping to the step A; If the network performance parameter value is not less than a preset value, it is determined that the candidate subcutaneous injection data mining network is a target subcutaneous injection data mining network.

2. The information push method based on the subcutaneous injection home management system according to claim 1 is characterized in that: The step of obtaining the basic injection data sequence to be learned comprises: Acquire a reference training data sequence, wherein the reference training data sequence includes multiple sample patient injection instances; Identifying each sample patient injection instance in the reference training data sequence according to a first recognition network, wherein the first recognition network divides the sample patient injection instance into a first sample patient injection instance and a non-first sample patient injection instance, wherein the first sample patient injection instance includes an incomplete or erroneous sample patient injection instance; Identifying each sample patient injection instance in the reference training data sequence according to a second recognition network, wherein the second recognition network divides the sample patient injection instance into a second sample patient injection instance and a non-second sample patient injection instance, wherein the second sample patient injection instance includes a sample patient injection instance of a non-standard operating procedure; Identifying each sample patient injection instance in the reference training data sequence according to a third recognition network, wherein the third recognition network divides the sample patient injection instances into third sample patient injection instances and non-third sample patient injection instances, wherein the third sample patient injection instances include sample patient injection instances with abnormal reactions or complications; identifying each sample patient injection instance in the reference training data sequence according to a fourth recognition network, wherein the fourth recognition network divides the sample patient injection instances into fourth sample patient injection instances and non-fourth sample patient injection instances, wherein the fourth sample patient injection instances include sample patient injection instances of a specific drug or treatment regimen; The first sample patient injection instance, the second sample patient injection instance, the third sample patient injection instance and the fourth sample patient injection instance in the reference training data sequence are retained, and the reference training data sequence after the retained first sample patient injection instance, the second sample patient injection instance, the third sample patient injection instance and the fourth sample patient injection instance is determined as the basic injection data sequence to be learned.

3. The information push method based on the subcutaneous injection home management system according to claim 1 is characterized in that: The method of obtaining a basic subcutaneous injection data mining network comprises: Inputting the sample patient injection instance of the sample patient injection instance data into a basic label prediction network, generating a label prediction result of the basic label prediction network for the sample patient injection instance of the sample patient injection instance data, wherein the label prediction result includes: a patient injection instance with a target injection knowledge label or a patient injection instance without a target injection knowledge label; Comparing the label prediction result with the first injection knowledge attribute of the sample patient injection instance data to generate a first learning error of the sample patient injection instance data; Calculating a first network cost parameter of the basic label prediction network according to a first learning error of each sample patient injection instance data in the injection data sequence to be learned; The network parameter information of the basic label prediction network is updated until the first network cost parameter of the basic label prediction network converges, and the basic label prediction network corresponding to the converged first network cost parameter is determined as the basic subcutaneous injection data mining network.

4. The information push method based on the subcutaneous injection home management system according to claim 1 is characterized in that: In the step C, performing network parameter learning on the current label prediction network according to the extended learning injection data sequence to determine a candidate subcutaneous injection data mining network includes: Loading the sample patient injection instance of the extended learning injection data into the current label prediction network, generating a label prediction result of the current label prediction network for the sample patient injection instance of the extended learning injection data, the label prediction result comprising: the sample patient injection instance of the extended learning injection data is a patient injection instance of a target injection knowledge label or the sample patient injection instance of the extended learning injection data is not a patient injection instance of a target injection knowledge label; Comparing the label prediction result with the second injection knowledge attribute of the extended learning injection data to generate a second learning error of the extended learning injection data; Calculating a second network cost parameter of the current label prediction network according to a second learning error of each extended learning injection data in the extended learning injection data sequence; The network parameter information of the current label prediction network is updated until the second network cost parameter of the current label prediction network converges, and the current label prediction network corresponding to the converged second network cost parameter is determined as a candidate subcutaneous injection data mining network.

5. The information push method based on the subcutaneous injection home management system according to claim 4 is characterized in that: The step of determining the second injection knowledge attribute of the extended learning injection data comprises: Performing feature slicing on the verification sample patient injection instance of the extended learning injection data based on a preset focus feature template to generate a focus patient injection instance corresponding to the verification sample patient injection instance, wherein the focus feature range corresponding to the preset focus feature template is smaller than the feature range of the verification sample patient injection instance; If the concerned patient injection instance is a patient injection instance of a target injection knowledge label, configuring the second injection knowledge attribute to be used to characterize that the verification sample patient injection instance is actually a patient injection instance of a target injection knowledge label; If the focused patient injection instance is not a patient injection instance of a target injection knowledge label, the second injection knowledge attribute is configured to characterize that the verification sample patient injection instance is not actually a patient injection instance of a target injection knowledge label.

6. The information push method based on the subcutaneous injection home management system according to claim 5 is characterized in that: Determining whether the concerned patient injection instance is a patient injection instance of a target injection knowledge tag includes: Determine whether any one of the patient injection instance segments in the focused patient injection instance is a patient injection instance of a target injection knowledge tag; If any one of the patient injection instance segments in the concerned patient injection instance is a patient injection instance of the target injection knowledge label, determining that the concerned patient injection instance is a patient injection instance of the target injection knowledge label; If no patient injection instance segment in the patient injection instance of interest is a patient injection instance of the target injection knowledge label, it is determined that the patient injection instance of interest is not a patient injection instance of the target injection knowledge label.

7. The information push method based on the subcutaneous injection home management system according to claim 5 is characterized in that: The step of loading the sample patient injection instance of the extended learning injection data into the current label prediction network to generate a label prediction result of the current label prediction network for the sample patient injection instance of the extended learning injection data comprises: Performing feature masking processing on some patient injection instance segments of the sample patient injection instance of the extended learning injection data, so that only the patient injection instance segments of the injection sample of interest corresponding to the sample patient injection instance are not masked; The sample patient injection instance of the extended learning injection data processed with feature masking is loaded into the current label prediction network, so that the current label prediction network determines the label prediction result of the sample patient injection instance based on the unmasked patient injection instance fragment in the input sample patient injection instance.

8. The information push method based on the subcutaneous injection home management system according to claim 5 is characterized in that: The step of comparing the label prediction result with the second injection knowledge attribute of the extended learning injection data to generate a second learning error of the extended learning injection data comprises: If the label prediction result is a patient injection instance of a target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verification sample patient injection instance is actually a patient injection instance of a target injection knowledge label, determining that the second learning error of the extended learning injection data is equal to 0; If the label prediction result is a patient injection instance of a target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verification sample patient injection instance is not actually a patient injection instance of a target injection knowledge label, determining that the second learning error of the extended learning injection data is equal to 1; If the label prediction result is not a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verification sample patient injection instance is actually a patient injection instance of the target injection knowledge label, determining that the second learning error of the extended learning injection data is equal to 1; If the label prediction result is not a patient injection instance of the target injection knowledge label and the second injection knowledge attribute of the extended learning injection data is used to characterize that the verification sample patient injection instance is actually not a patient injection instance of the target injection knowledge label, determining that the second learning error of the extended learning injection data is equal to 0; The step of calculating the second network cost parameter of the current label prediction network according to the second learning error of each extended learning injection data in the extended learning injection data sequence includes: The second learning errors of all the extended learning injection data in the extended learning injection data sequence are accumulated, and the accumulated learning errors are determined as the second network cost parameter of the current label prediction network.

9. The information push method based on the subcutaneous injection home management system according to any one of claims 1 to 8, characterized in that: The method further comprises: Obtain target patient injection instances; Inputting a target patient injection instance into the target subcutaneous injection data mining network to generate a label prediction result of the target subcutaneous injection data mining network for the target patient injection instance; If the label prediction result of the target patient injection instance is a patient injection instance of a target injection knowledge label, the target patient injection instance is determined as a patient injection instance of a target injection knowledge label, and information of all target patient injection instances of target injection knowledge labels is pushed to the target medical service terminal.

10. A smart medical service system, characterized in that: The smart medical service system includes a processor and a machine-readable storage medium, in which machine-executable instructions are stored. The machine-executable instructions are loaded and executed by the processor to implement the information push method based on the subcutaneous injection home management system as described in any one of claims 1-9.

Images