Intelligent monitoring intervention method and system for rehabilitation correction of drug addicts

Abstract

According to the intelligent monitoring intervention method and system for rehabilitation correction of the drug abstinence person, involved information between abstinence symptom marks is taken into consideration, meanwhile, differences of relevance between abstinence reactions under different abstinence symptom classifications are taken into consideration, and by means of a graph convolutional network under different abstinence symptom classification scenes, the rehabilitation correction of the drug abstinence person can be achieved. And detecting to obtain involved information of withdrawal symptom marks in different withdrawal symptom classification scenes, and reasoning confidence degree distribution of the target object monitoring sequence belonging to each preset withdrawal symptom classification according to the overall representation vector of the target object monitoring sequence. And iterating the withdrawal symptom characterization vector of the alternative withdrawal symptom marker based on confidence distribution and the involved information of the withdrawal symptom marker of each withdrawal symptom classification scene, so as to determine a target withdrawal symptom marker corresponding to the target object monitoring sequence according to the iterated withdrawal symptom characterization vector, thereby improving the detection precision of the withdrawal symptom marker, and improving the detection accuracy of the withdrawal symptom marker. And reasonable intervention on patients abstained from drugs in the correction process is helped.

Description

Technical Field

[0001] This application relates to the field of data processing, and more particularly, to an intelligent monitoring and intervention method and system for the rehabilitation and correction of drug addicts. Background Art

[0002] In the field of drug rehabilitation and correction, accurately identifying the withdrawal symptoms of drug addicts is crucial for formulating effective intervention strategies. Traditionally, the identification of withdrawal symptoms mainly relies on doctors' clinical experience and subjective judgment. This method is not only time-consuming and laborious but also vulnerable to human factors, resulting in inconsistencies and inaccuracies in the identification results. With the rapid development of big data technology, Internet of Things technology, and medical information technology, especially the application of machine learning and deep learning technologies, new possibilities have been provided for the automatic identification of withdrawal symptoms. When dealing with withdrawal symptoms, existing automatic identification methods often ignore the complex correlation relationships between withdrawal symptom markers and the differences in these correlation relationships under different withdrawal symptom classification scenarios. Withdrawal symptoms do not exist in isolation; there are often mutual influences and involvements among them, and this relationship is particularly important in the rehabilitation process of drug addicts. At the same time, under different withdrawal symptom classifications (such as mood disorders, sleep disorders, physical pain, etc.), the correlation between withdrawal symptoms may show significant differences, and this difference has an important impact on the accurate identification of withdrawal symptoms. When dealing with the task of identifying withdrawal symptoms, existing machine learning models usually regard each withdrawal symptom marker as an independent entity, ignoring the correlation information between them, resulting in poor performance of the models in complex and changing drug rehabilitation scenarios. In addition, although some methods attempt to introduce graph structures to represent the relationships between symptom markers, they often lack differential processing of the correlations under different withdrawal symptom classification scenarios and cannot make full use of the feature information under specific classifications to improve the identification accuracy. Summary of the Invention

[0003] The object of the present invention is to provide an intelligent monitoring and intervention method and system for the rehabilitation and correction of drug addicts. This application is implemented as follows: In a first aspect, the present application provides an intelligent monitoring and intervention method for the rehabilitation and correction of drug addicts, including: obtaining a target object monitoring sequence and a plurality of alternative withdrawal symptom markers, and respectively extracting characterization information from the target object monitoring sequence and each alternative withdrawal symptom marker to obtain an overall characterization vector of the target object monitoring sequence and a withdrawal symptom characterization vector of each alternative withdrawal symptom marker; inferring the confidence distribution of the target object monitoring sequence corresponding to each preset withdrawal symptom classification based on the overall characterization vector; obtaining a graph convolutional network corresponding to each preset withdrawal symptom classification, where the graph convolutional network is used to indicate the involvement information between each alternative withdrawal symptom marker under the preset withdrawal symptom classification; performing a graph context embedding operation on the withdrawal symptom characterization vectors of each alternative withdrawal symptom marker according to the confidence distribution and the graph convolutional network corresponding to each preset withdrawal symptom classification to obtain an operation-after withdrawal symptom characterization vector corresponding to each alternative withdrawal symptom marker; determining the inference confidence of each alternative withdrawal symptom marker belonging to the target object monitoring sequence according to the operation-after withdrawal symptom characterization vector corresponding to each alternative withdrawal symptom marker; determining the target withdrawal symptom marker corresponding to the target object monitoring sequence among each alternative withdrawal symptom marker based on the inference confidence; and analyzing whether to execute a correction and intervention strategy based on the target withdrawal symptom marker.

[0004] In a second aspect, the present application provides a computer system, including: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method described above is implemented.

[0005] The beneficial effects of the present application at least include: The present application not only takes into account the involvement information between withdrawal symptom markers, but also takes into account the differences in the correlations between withdrawal reactions under different withdrawal symptom classifications. It can use the graph convolutional network in different withdrawal symptom classification scenarios to detect the involvement information of withdrawal symptom markers in different withdrawal symptom classification scenarios, and infer the confidence distribution of the target object monitoring sequence belonging to each preset withdrawal symptom classification based on the overall characterization vector of the target object monitoring sequence. Based on the confidence distribution and the involvement information of withdrawal symptom markers in each withdrawal symptom classification scenario, the withdrawal symptom characterization vectors of alternative withdrawal symptom markers are iteratively updated to determine the target withdrawal symptom marker corresponding to the target object monitoring sequence according to the iterated withdrawal symptom characterization vectors. In this way, the feature information of different withdrawal symptom classifications is integrated for withdrawal symptom detection, and the correlation between alternative withdrawal symptom markers is measured under specific withdrawal symptom classifications, which can improve the detection accuracy of withdrawal symptom markers (i.e., specific withdrawal symptoms) and help with reasonable intervention in the correction process of drug addicts. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 It is a flowchart of an intelligent monitoring and intervention method for the rehabilitation and correction of drug addicts provided by an embodiment of the present application.

[0007] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present application. Detailed implementation manners

[0008] In the embodiment of the present application, the execution subject of the intelligent monitoring and intervention method for the rehabilitation and correction of drug addicts is a computer system, including but not limited to servers, personal computers, laptops, tablets, smartphones, etc. As Figure 1 shown, the method includes: Step S100: Obtain a monitoring sequence of a target object and multiple alternative withdrawal symptom markers, and respectively extract characterization information from the monitoring sequence of the target object and each alternative withdrawal symptom marker to obtain an overall characterization vector of the monitoring sequence of the target object and a withdrawal symptom characterization vector of each alternative withdrawal symptom marker.

[0009] In the embodiment of the present application, the computer system can access the monitoring device network of a drug rehabilitation center and obtain the monitoring sequence of the target object (i.e., the drug addict) in real time or regularly. These monitoring sequences contain rich physiological, psychological and behavioral data, such as heart rate, blood pressure, sleep quality, emotional state, daily activity level, etc. These data are recorded in the form of a time series, forming a multi-dimensional data set. For example, a typical monitoring sequence may include a heart rate change graph for several consecutive days, the length and quality assessment of each night's sleep, and the emotional stability score obtained through a questionnaire survey. It can be understood that these data can be monitored by wearable devices and corresponding monitoring instruments, and the specific method is not limited.

[0010] At the same time, the computer system also maintains a library of alternative withdrawal symptom markers. This library is constructed based on professional knowledge and past cases in the field of drug rehabilitation, and contains various possible withdrawal symptoms, such as anxiety, depression, irritability, insomnia, body pain, appetite change, etc. Each symptom marker represents a challenge or problem that may be encountered during the drug rehabilitation process.

[0011] Next, the computer system extracts the characterization information from these monitoring sequences and alternative withdrawal symptom markers, which is the process of feature extraction. The purpose of this step is to convert the original monitoring data and symptom markers into a form that is easy for the computer to process and understand, namely vectors or feature sets. For the monitoring sequences of the target object, the computer system may adopt various feature extraction techniques, such as time series analysis, signal processing, and wavelet transform, etc., to capture the key information in the data. For example, the system can evaluate the activity of the autonomic nervous system of drug addicts by calculating heart rate variability (HRV), which reflects their ability to cope with stress; at the same time, sleep cycle analysis is used to identify the presence of insomnia symptoms. These features are integrated into an overall characterization vector, which synthesizes the global features in the monitoring sequences and provides a basis for subsequent withdrawal symptom classification.

[0012] For the alternative withdrawal symptom markers, the system may adopt methods such as text embedding, word vector representation, or feature encoding based on domain knowledge. For example, for the symptom marker "anxiety", the system can map it into a high-dimensional vector space, and this vector captures the semantic similarities and differences between anxiety and other symptoms (such as depression, insomnia). In addition, the system can also assign specific feature values to each symptom marker according to professional manuals or doctor guidelines, such as the severity, duration, and scope of influence of the symptom. Finally, through this step, the computer system obtains the overall characterization vector of the monitoring sequences of the target object and the withdrawal symptom characterization vectors of each alternative withdrawal symptom marker. These vectors not only retain the key information of the original data but also exist in an efficient and computable form, laying a solid foundation for subsequent confidence inference, graph context embedding, and other operations.

[0013] In practical applications, the execution of step S100 depends on the support of advanced data processing technologies and machine learning algorithms, such as recurrent neural networks (RNNs) for time series analysis, Word2Vec or BERT for text embedding, etc. Through the comprehensive application of these technologies, the computer system can achieve a deep understanding and accurate characterization of the complex monitoring data of drug addicts.

[0014] Step S200: Infer the confidence distribution corresponding to each preset withdrawal symptom classification for the monitoring sequences of the target object based on the overall characterization vector.

[0015] Specifically, the computer system first receives the overall characterization vector from step S100. This is a high-dimensional vector that synthesizes the global features of the target object in multiple monitoring dimensions, such as the change trend of physiological indicators, the fluctuation of mental state, and the abnormal performance of behavior patterns, etc. These features are encoded numerically in the vector, and each dimension represents an aspect of the monitoring data.

[0016] Next, the computer system utilizes a pre-trained machine learning model (such as Support Vector Machine (SVM), Random Forest (RF), or Deep Neural Network (DNN), etc.) to perform inference tasks. These models have learned the mapping relationships between different withdrawal symptom classifications and the characteristics of monitoring data during the training phase, and can predict the possible withdrawal symptoms and their probabilities of the target object based on the input overall representation vector.

[0017] Taking the deep neural network as an example, the model may contain multiple hidden layers, and each layer transforms the input features into a higher-level abstract representation through non-linear transformation. During the inference phase, the overall representation vector is passed to the model as input. After being processed layer by layer, a vector is finally output, and each element of this vector corresponds to the confidence score of a preset withdrawal symptom classification. These scores reflect the probabilities of the target object's monitoring sequence belonging to each classification, thus constituting a confidence distribution.

[0018] For example, assume that the preset withdrawal symptom classifications include four types: anxiety, depression, insomnia, and physical discomfort. After the computer system executes step S200, it can output a confidence distribution vector in the form of [0.2, 0.5, 0.1, 0.2], where 0.5 corresponds to the depression classification, indicating that the system believes the probability of the target object currently having depressive symptoms is the highest, while the confidence levels of the other three classifications are relatively low.

[0019] This inference result is crucial for drug rehabilitation and correction. It helps the medical team quickly identify the main withdrawal challenges that drug addicts may face, so that personalized intervention strategies can be formulated accordingly. For example, if the system infers a relatively high probability of depressive symptoms, then the medical team can consider increasing the frequency of psychological counseling, adjusting the drug treatment plan, or introducing other non-drug intervention measures to relieve depressive symptoms and promote the comprehensive rehabilitation of drug addicts.

[0020] Step S300: Obtain the graph convolutional network corresponding to each preset withdrawal symptom classification. The graph convolutional network is used to indicate the involvement information between each alternative withdrawal symptom label under the preset withdrawal symptom classification.

[0021] In the embodiments of the present application, step S300 involves the application of a Graph Convolutional Network (GCN) to capture and represent the complex correlation relationships among various alternative withdrawal symptom markers under different withdrawal symptom classifications. First, the computer system pre-constructs and trains multiple graph convolutional networks, with each network corresponding to a preset withdrawal symptom classification. These graph convolutional networks are deep learning models designed based on graph-structured data. They can process graphs composed of nodes (i.e., alternative withdrawal symptom markers) and edges (i.e., the involvement information between symptom markers), and update the feature representation of each node by iteratively aggregating the feature information of neighboring nodes. In the context of drug rehabilitation, each alternative withdrawal symptom marker is regarded as a node in the graph, and the involvement information between them constitutes the edges in the graph. This involvement information reflects the frequency or conditional probability of co-occurrence of different symptom markers in the same monitoring data, and is an important clue for understanding the mutual influence between withdrawal symptoms. For example, anxiety and depression often co-occur in the monitoring data of many drug addicts, indicating a strong correlation between these two symptoms; while insomnia may appear as a result of anxiety or depression, reflecting a conditional involvement relationship.

[0022] To construct these graph convolutional networks, the computer system first needs to collect a large number of monitoring data samples and label the withdrawal symptom markers that appear in each sample. Then, statistical analysis or machine learning methods (such as matrix factorization, probabilistic graphical models, etc.) are used to quantify the involvement information between symptom markers and generate graph-structured data. These data are then used to train the graph convolutional networks so that the networks can learn the internal correlation patterns between symptom markers.

[0023] In step S300, the computer system selects the corresponding graph convolutional network according to the preset withdrawal symptom classification being currently processed. This network has learned how to update the feature representation of each symptom marker based on the involvement information between symptom markers under this classification. Specifically, the network will iteratively aggregate the feature information of the neighboring nodes of each symptom marker and incorporate this information into the feature vector of the current node. After multiple iterations, the feature vector of each symptom marker will contain rich context information, reflecting the complex correlation relationships between it and other symptom markers.

[0024] Taking the classification of anxiety symptoms as an example, it is assumed that the corresponding graph convolutional network has been trained. In step S300, when the computer system processes the monitoring data related to anxiety, it will activate this network and input the initial feature vectors of the alternative withdrawal symptom labels (these vectors may come from the result of the characterization information extraction in step S100). The network will then update these feature vectors according to the predetermined iterative rules and finally output the new feature representations of each symptom label under the anxiety classification. These representations not only contain the information of the symptom labels themselves but also incorporate the feature information of other symptom labels associated with them, providing a more comprehensive and in-depth perspective for subsequent reasoning and intervention.

[0025] Step S400: According to the confidence distribution and the graph convolutional networks corresponding to each preset withdrawal symptom classification, perform graph context embedding operations on the withdrawal symptom representation vectors of each alternative withdrawal symptom label to obtain the post-operation withdrawal symptom representation vectors corresponding to each alternative withdrawal symptom label.

[0026] In the embodiment of the present application, step S400 aims to combine the confidence distribution obtained in step S200 with the graph convolutional networks corresponding to each preset withdrawal symptom classification in step S300 to perform graph context embedding operations on the withdrawal symptom representation vectors of the alternative withdrawal symptom labels, so as to further enrich and refine the feature representations of these vectors. Specifically, the computer system first reviews the result of step S200, that is, the confidence distribution of the target object monitoring sequence corresponding to each preset withdrawal symptom classification. This distribution reveals the possible withdrawal symptoms and their probabilities of the target object under different classifications. Then, the system selects the corresponding graph convolutional network according to the current classification being processed. These networks have learned the involvement information between the symptom labels through training and can reflect their association relationships under specific classifications. Now, the system enters the core part of the graph context embedding operation. For each alternative withdrawal symptom label, the system first obtains its withdrawal symptom representation vector obtained in step S100. This vector is the result of the original monitoring data after feature extraction and contains the initial feature information of the symptom label. Then, the system uses the selected graph convolutional network to perform graph feature encoding on these representation vectors according to the involvement information between the symptom labels. In the process of graph feature encoding, the graph convolutional network will iteratively aggregate the feature information of the neighbor nodes of each symptom label and update the feature vector of the current node. This process not only considers the features of the symptom label itself but also incorporates the feature information of other symptom labels associated with it. Through multiple iterations, the representation vectors of each alternative withdrawal symptom label will be given rich context information, enabling them to more accurately reflect the comprehensive state of the symptom label under specific classifications.

[0027] For example, assume that the monitoring data of the target object at a certain time point shows a relatively high confidence level of anxiety and depression symptoms. When dealing with anxiety classification, the system selects the corresponding graph convolutional network and performs graph context embedding operations on the representation vectors of the two symptom labels, "anxiety" and "depression". Since anxiety and depression often coexist among drug addicts undergoing detoxification, the graph convolutional network will capture the strong correlation between them and incorporate this relationship into the updated feature vectors. Therefore, the representation vector of the "anxiety" symptom after graph feature encoding will not only contain the feature information of the anxiety symptom itself but also incorporate the context information related to the depression symptom, thus more comprehensively reflecting the actual state of the target object under anxiety classification.

[0028] Finally, the output of step S400 is the post-operation withdrawal symptom representation vectors corresponding to each alternative withdrawal symptom label. These vectors not only contain the initial feature information of the symptom labels but also incorporate rich context information, providing a more accurate and comprehensive input for determining the inference confidence in the subsequent steps.

[0029] Step S500: Determine the inference confidence of each alternative withdrawal symptom label belonging to the monitoring sequence of the target object based on the post-operation withdrawal symptom representation vectors corresponding to each alternative withdrawal symptom label.

[0030] In the embodiment of this application, step S500 is responsible for further analyzing and determining the inference confidence of these symptom labels belonging to the monitoring sequence of the target object based on the post-operation withdrawal symptom representation vectors corresponding to each alternative withdrawal symptom label obtained in step S400. This process not only relies on the feature information extracted and encoded in the previous steps but also involves the comprehensive evaluation and probability inference of the features.

[0031] When the computer system executes step S500, it examines the post-operation withdrawal symptom representation vectors of each alternative withdrawal symptom label. These vectors have incorporated rich context information through graph context embedding operations and can more comprehensively reflect the comprehensive features of the symptom labels under specific classifications. Then, the system uses one or more machine learning algorithms (such as logistic regression, support vector machines, random forests, etc.) to analyze these feature vectors to determine the association strength between each symptom label and the monitoring sequence of the target object, that is, the inference confidence. The calculation of the inference confidence is usually based on the similarity or matching degree between the feature vectors and the target monitoring sequence. For example, the system can calculate the cosine similarity between the post-operation representation vector of each symptom label and the overall representation vector of the target monitoring sequence, or use a distance metric (such as the Euclidean distance) to evaluate their proximity. These similarity or distance values are then converted into a probability form, that is, the inference confidence, to represent the likelihood of each symptom label appearing in the target monitoring sequence.

[0032] In practical applications, the calculation of inference confidence may involve comprehensive consideration of multiple feature dimensions. The system can assign different weights to different features based on prior knowledge or the suggestions of domain experts to emphasize the importance of certain features in symptom recognition. These weights can be trained and optimized as part of the machine learning algorithm to ensure the accuracy and reliability of the inference confidence.

[0033] Taking a specific scenario as an example, assume that the monitoring data of the target object within a certain period shows abnormal symptoms such as increased heart rate, decreased sleep quality, and mood swings. In step S500, the computer system calculates the inference confidence between each of them and the target monitoring sequence according to the operation posterior representation vectors marked with alternative withdrawal symptoms such as "anxiety" and "depression". If the representation vector of the "anxiety" symptom marker shows a high similarity with the target monitoring sequence in terms of features such as heart rate variability, abnormal sleep cycle, and unstable emotional state, then the system can assign a high inference confidence to "anxiety", indicating that this symptom marker is highly relevant to the current state of the target object. Finally, the output of step S500 is a set of inference confidence values, each corresponding to an alternative withdrawal symptom marker. These confidence values not only reflect the association strength between the symptom marker and the target monitoring sequence, but also provide an important basis for symptom marker selection and corrective intervention strategy formulation in subsequent steps.

[0034] Step S600: Determine the target withdrawal symptom marker corresponding to the target object's monitoring sequence among each alternative withdrawal symptom marker according to the inference confidence.

[0035] In the embodiment of the present application, step S600 determines the target withdrawal symptom marker corresponding to the target object's monitoring sequence based on the inference confidence of each alternative withdrawal symptom marker calculated in step S500. This process is the final judgment made by the system on the current withdrawal symptom state of the target object after comprehensive analysis of the monitoring data. When the computer system executes step S600, it reviews the inference confidence values output in step S500. These values form a confidence distribution, reflecting the association strength and likelihood between each alternative withdrawal symptom marker and the target object's monitoring sequence. Subsequently, the system adopts a decision-making strategy to select the most likely target withdrawal symptom marker from the alternative markers according to these confidence values.

[0036] The choice of the decision-making strategy can be determined according to the specific application scenario and requirements. For example, it includes the following common methods: Threshold method: The system sets a confidence threshold. Only when the inference confidence of an alternative marker exceeds this threshold is it considered a target withdrawal symptom marker. If multiple markers exceed the threshold simultaneously, one or more with the highest confidence may be selected as the target.

[0037] Sorting and Selection Method: The system simply sorts the alternative markers in descending order of inference confidence, and then selects the marker with the highest confidence as the target withdrawal symptom marker. If the coexistence of multiple symptoms needs to be considered, multiple markers with the top few confidences can be selected.

[0038] Probability Model Method: In some complex scenarios, the system may adopt more advanced probability models (such as Bayesian networks, hidden Markov models, etc.) to comprehensively consider the mutual influence and conditional probabilities among multiple alternative markers, so as to make more accurate decisions.

[0039] Taking a specific scenario as an example, assume that the monitoring data of the target object over a period of time shows various physiological and psychological abnormalities. The inference confidences calculated in step S500 indicate that the two alternative withdrawal symptom markers of "anxiety" and "depression" have relatively high confidence values, 0.7 and 0.6 respectively, while the confidences of other markers are relatively low. In step S600, the computer system may adopt the sorting and selection method and directly select the "anxiety" with the highest confidence as the target withdrawal symptom marker. If the system sets a threshold (such as 0.5), both "anxiety" and "depression" will be considered possible target symptoms, but "anxiety" is given priority due to its higher confidence.

[0040] Finally, the output of step S600 is the target withdrawal symptom marker corresponding to the monitoring sequence of the target object, which provides a direct basis for formulating subsequent correction and intervention strategies. The computer system can adjust the type and intensity of the intervention measures according to different target symptoms to ensure that drug addicts can receive targeted help and support.

[0041] Step S700: Analyze whether to execute the correction and intervention strategy based on the target withdrawal symptom marker.

[0042] In the embodiment of this application, step S700 analyzes and decides whether to execute the corresponding correction and intervention strategy based on the target withdrawal symptom marker determined in step S600. This process reflects the transformation of the results of the computer system's in-depth analysis of the monitoring data into actual intervention measures, aiming to promote the rehabilitation process of drug addicts through precise intervention. When the computer system executes step S700, it will first review the target withdrawal symptom marker output in step S600. This marker is the conclusion drawn by the system after a comprehensive analysis of the monitoring sequence of the target object and represents the most likely withdrawal symptom currently present in the target object. Then, the system matches the correction measures corresponding to the target symptom according to the preset intervention strategy library and expert rule set.

[0043] The intervention strategy library may contain multiple intervention programs for different withdrawal symptoms. These programs are developed by rehabilitation experts based on clinical experience and research findings, aiming to help drug addicts relieve withdrawal symptoms, enhance self-control ability and rebuild a healthy lifestyle through various means such as psychological support, drug treatment, physical therapy, social skills training, etc. After matching the appropriate intervention strategy, the computer system further analyzes whether to execute the strategy according to the specific situation of the target object and the availability of resources. This analysis may involve considerations of various aspects such as the current physiological state, psychological state, social environment, and family support of the target object. For example, if the target object is determined to have severe anxiety symptoms, the system can recommend a series of psychological counseling and relaxation training courses and evaluate whether these courses are suitable for the target object's schedule and financial ability.

[0044] If the analysis result shows that it is necessary and feasible to execute the intervention strategy, the computer system sends the relevant suggestions to the rehabilitation team or medical professionals so that they can take measures in a timely manner. These measures may include arranging psychological counseling sessions, adjusting the drug treatment plan, providing emergency medical assistance, etc. At the same time, the system will also record the implementation situation and effect feedback of the intervention strategy for subsequent continuous optimization and improvement of the intervention strategy.

[0045] Taking a specific scenario as an example, assume that the target object is determined to have moderate depressive symptoms. In step S700, the computer system matches a series of intervention strategies for depressive symptoms, such as cognitive behavioral therapy, antidepressant drug treatment, and social skills training, etc. After analyzing the personal situation and resources of the target object, the system can suggest trying cognitive behavioral therapy first because it usually has lower risks and higher acceptance. Subsequently, the system sends specific suggestions for implementing this therapy to the rehabilitation team and tracks and records its effect for subsequent evaluation.

[0046] As an implementation manner, in step S100, the representation information extraction is respectively performed on the target object monitoring sequence and each alternative withdrawal symptom marker to obtain the overall representation vector of the target object monitoring sequence and the withdrawal symptom representation vectors of each alternative withdrawal symptom marker. Specifically, it may include: Step S110: Perform a vector embedding operation on the object monitoring data of the target object monitoring sequence under not less than one type to obtain the monitoring data embedding vectors of the target object monitoring sequence under not less than one type; Step S120: Perform a vector embedding operation on each alternative withdrawal symptom marker respectively to obtain the withdrawal symptom embedding vectors of each alternative withdrawal symptom marker; Step S130: Generate a set of representation vectors according to the withdrawal symptom embedding vectors of each alternative withdrawal symptom marker and the prior feature vectors; Step S140: Embed the monitoring data into the vector and the set of characterization vectors for weight focusing operation to obtain the overall characterization vector of the monitoring sequence of the target object and the withdrawal symptom characterization vectors of each alternative withdrawal symptom label.

[0047] In the above step S110, the computer system first performs vector embedding operations on various types of monitoring data in the monitoring sequence of the target object. These monitoring data may include, but are not limited to, time series data of physiological indicators such as heart rate, blood pressure, sleep quality, and daily activity level. Each type of monitoring data is regarded as an independent feature dimension, and the system uses vector embedding techniques (such as Word2Vec, TF-IDF weighting, or deep learning embedding layers) to convert these raw data into vector representations in a high-dimensional space, that is, monitoring data embedding vectors.

[0048] For example, for heart rate data, the system can convert the heart rate value at each time point into a vector of a fixed length, where each dimension of the vector captures a certain characteristic of the heart rate change (such as periodicity, variability, etc.). Similarly, sleep quality data may also be converted into a vector containing multiple features such as sleep duration, proportion of deep sleep, and number of REM sleep cycles.

[0049] Next, the system performs vector embedding operations on each symptom label in the set of alternative withdrawal symptom labels. These symptom labels, such as anxiety, depression, irritability, etc., represent various psychological and behavioral problems that may occur during the drug withdrawal process. The system uses text embedding techniques (such as Word2Vec, BERT) or feature encoding methods based on domain knowledge to convert each symptom label into a high-dimensional vector, that is, a withdrawal symptom embedding vector. These vectors capture the semantic similarities and differences between the symptom labels and provide a basis for subsequent analysis.

[0050] After obtaining the withdrawal symptom embedding vectors of all alternative withdrawal symptom labels, the system generates a set of characterization vectors based on these vectors and possible prior feature vectors (such as statistical features based on historical data, expert knowledge, etc.). This set is a sequence data that contains all the feature vectors related to each alternative withdrawal symptom label. These feature vectors not only represent the characteristics of the symptom labels themselves but may also incorporate additional context information or domain knowledge.

[0051] Finally, the system performs a weight focusing operation to combine the monitoring data embedding vectors with the set of characterization vectors to generate the overall characterization vector of the monitoring sequence of the target object and the withdrawal symptom characterization vectors of each alternative withdrawal symptom label. This step is usually based on an attention mechanism (such as the self-attention layer in the Transformer model) or other weighted aggregation methods, and determines their respective weights by calculating the similarity or correlation between different monitoring data vectors and characterization vectors.

[0052] Specifically, the system can first calculate the similarity scores between each monitoring data embedding vector and all vectors in the set of characterization vectors, and then perform a weighted sum on the characterization vectors according to these scores to obtain the preliminary characterization vectors of each alternative withdrawal symptom label. Next, the system can apply the attention mechanism again to interact the preliminary characterization vectors with the monitoring data embedding vectors to further adjust the weights and generate the final withdrawal symptom characterization vectors. For the overall characterization vector of the monitoring sequence of the target object, the system can adopt a similar method, but process all the monitoring data embedding vectors as a whole to capture the global features of the entire monitoring sequence.

[0053] As an implementation manner, in step S140, a weight focusing operation is performed on the monitoring data embedding vectors and the set of characterization vectors to obtain the overall characterization vector of the monitoring sequence of the target object and the withdrawal symptom characterization vectors of each alternative withdrawal symptom label. Specifically, it may include: Step S141: Determine the monitoring data embedding vectors of the currently to-be-processed target type among the monitoring data embedding vectors of not less than one type.

[0054] In the embodiment of the present application, step S141 is the starting step of the weight focusing operation (a part of step S140), which involves selecting the currently to-be-processed target type from multiple types of monitoring data embedding vectors.

[0055] The computer system faces a set of embedding vectors of monitoring data of multiple types. These monitoring data may cover multiple aspects such as physiology, psychology, and behavior. Each type is collected through specific sensors or evaluation tools and is converted into a vector representation in a high-dimensional space through vectorization processing. For example, physiological monitoring data may include heart rate, blood pressure, respiratory rate, etc., and they are respectively converted into corresponding heart rate embedding vectors, blood pressure embedding vectors, etc.; while psychological monitoring data may involve aspects such as emotional state and cognitive function, obtained through questionnaires or psychological tests, and converted into corresponding emotional state embedding vectors, cognitive function embedding vectors, etc.

[0056] In step S141, the computer system determines the monitoring data embedding vectors of the currently to-be-processed target type from this set according to a certain rule or strategy. This selection process can be based on multiple factors, such as the importance, timeliness, relevance to the target withdrawal symptoms of the monitoring data, or the preset processing order, etc.

[0057] In step S141, the system can first select the heart rate embedding vector as the target type to be processed currently. This is because the heart rate is an important indicator that can quickly reflect an individual's physiological stress state and is very useful for preliminarily evaluating the withdrawal reaction of drug addicts undergoing rehabilitation. Of course, this selection is not fixed, and the system can adjust the strategy according to specific circumstances. For example, if the system detects that the person's sleep quality has significantly declined recently, then it can preferentially select the sleep quality embedding vector for processing.

[0058] Once the monitoring data embedding vector under the currently proposed target type is determined, the system will enter step S142, and the weight focusing operation related to the monitoring data of this type will be carried out accordingly. Through this series of iterative processes, the system can gradually integrate various types of monitoring data and finally form a comprehensive representation of the monitoring sequence of the target object, providing strong data support for subsequent withdrawal symptom inference and the formulation of rehabilitation intervention strategies.

[0059] Step S142: Perform a weight focusing operation on the monitoring data embedding vector and the set of representation vectors under the target type to obtain the focused representation vector of the monitoring sequence of the target object and the focused representation vectors of each alternative withdrawal symptom label.

[0060] In the embodiment of this application, step S142 is a key step in the weight focusing operation (the core link in step S140), which involves deeply fusing the monitoring data embedding vector under the current target type with the set of representation vectors to generate a more targeted focused representation vector.

[0061] When the computer system executes step S142, it has already determined the monitoring data embedding vector under the currently proposed target type (such as the heart rate embedding vector), and has a set of representation vectors including the withdrawal symptom embedding vectors of all alternative withdrawal symptom labels and possible prior feature vectors. These vectors together constitute the basic data framework for subsequent analysis.

[0062] Next, the system uses advanced machine learning models or algorithms (such as the attention mechanism model) to perform the weight focusing operation. The core of this step is to evaluate the influence degree of the monitoring data of the target type on each alternative withdrawal symptom label and assign corresponding weights to them. By this method, the system can capture the complex correlation relationships between different monitoring data and the target withdrawal symptoms.

[0063] In a specific implementation, the system can construct an attention layer or a similar mechanism that takes as input the monitoring data embedding vectors and the set of characterization vectors under the target type. The attention layer calculates the similarity or correlation scores between the monitoring data embedding vectors and each vector in the set of characterization vectors, and these scores reflect the "attention" degree of the monitoring data to each withdrawal symptom marker. Subsequently, the system performs a weighted sum of the characterization vectors based on these scores to generate a focused characterization vector.

[0064] Taking the heart rate embedding vector as an example, assume that the system is processing the heart rate monitoring data of a drug rehabilitation patient. Heart rate data can usually reflect an individual's physiological stress state and is closely related to withdrawal symptoms such as anxiety and depression. In step S142, the system interacts the heart rate embedding vector with each withdrawal symptom embedding vector in the set of characterization vectors and calculates the similarity scores between them. If the heart rate data shows an obvious accelerating trend, the system can assign a higher weight to the withdrawal symptom embedding vector related to anxiety because heart rate acceleration is a common physiological response in an anxious state.

[0065] After the weight focusing operation, the system not only obtains the focused characterization vector of the monitoring sequence of the target object under the current target type monitoring data (i.e., an attention feature that synthesizes the influence of various factors), but also obtains the focused characterization vectors of each alternative withdrawal symptom marker. These focused characterization vectors not only retain the information of the original characterization vectors, but also incorporate the context features closely related to the target monitoring data, providing more accurate data support for subsequent symptom reasoning and intervention strategy formulation.

[0066] Through such a process, the computer system can effectively integrate and utilize various types of monitoring data, bringing a more personalized and precise solution to the monitoring work of drug rehabilitation treatment.

[0067] Step S143: Determine a new set of characterization vectors based on the focused characterization vector of the monitoring sequence of the target object and the focused characterization vectors of each alternative withdrawal symptom marker.

[0068] Step S143 is responsible for updating or determining a new set of characterization vectors based on the focused characterization vector of the monitoring sequence of the target object and the focused characterization vectors of each alternative withdrawal symptom marker generated in the current iteration process. This process not only consolidates the influence of the current monitoring data on the withdrawal symptom characteristics, but also provides a more refined data basis for subsequent iterations or comprehensive analysis. The following is a detailed explanation of step S143, illustrated with specific application scenarios and examples: When the computer system executes step S143, it has obtained the set of focused representation vectors under the current target type monitoring data through step S142. This set includes the overall focused representation vector of the target object monitoring sequence (reflecting the comprehensive characteristics of the current monitoring data) and the focused representation vectors of each alternative withdrawal symptom marker (reflecting the association strength between the monitoring data and specific symptoms). Next, the computer system determines a new set of representation vectors based on these focused representation vectors. However, it should be noted that the "new" here does not mean completely replacing the original set of representation vectors, but updating or expanding on the original basis. Specifically, the system can adopt one of the following methods or a combination of multiple methods to process: Direct update: If the focus of the system design is to capture the most direct impact of the current monitoring data on the characteristics of withdrawal symptoms, then it can directly use the set of focused representation vectors as the new set of representation vectors. In this case, the subsequent analysis of the system will be mainly based on these focused representation vectors.

[0069] Weighted average: To comprehensively consider the combined impact of multiple monitoring data on the characteristics of withdrawal symptoms, the system can adopt the method of weighted average. It will assign a weight to each focused representation vector (this weight can be determined based on the importance, timeliness, or other relevant factors of the monitoring data), and then sum them with weights to obtain the new set of representation vectors. This method can balance the contribution degrees of different monitoring data to the inference of withdrawal symptoms.

[0070] Feature fusion: In some cases, the system may want to retain some feature information in the original set of representation vectors while integrating the new information of the focused representation vectors. At this time, the system can adopt feature fusion techniques (such as feature concatenation, feature mapping, etc.) to combine the original representation vectors with the focused representation vectors to form a new set of representation vectors. This method can enrich the dimension and depth of the representation vectors and improve the accuracy and robustness of subsequent analysis.

[0071] Taking an actual scenario as an example, assume that the computer system is processing the heart rate and sleep quality monitoring data of a drug addict. In step S142, the system has generated the set of focused representation vectors under the heart rate and sleep quality. In step S143, if the system processes these vectors in the way of weighted average, it can assign different weights to them according to the importance of the heart rate and sleep quality data (such as obtained through expert evaluation or historical data analysis), and fuse or replace the weighted focused representation vectors with other feature vectors in the original set of representation vectors (such as emotional state, social activities, etc.), and finally determine a new set of representation vectors. This set will be used as the input data for subsequent steps (such as the iterative processing in step S144 or the symptom inference in step S150).

[0072] Through such a processing process, the computer system can more accurately capture the impact of the current monitoring data on the characteristics of withdrawal symptoms on the basis of retaining the characteristics of the original monitoring data, providing more powerful data support for subsequent drug rehabilitation and correction work.

[0073] Step S144: Jump to the step of determining the monitoring data embedding vector of the currently to-be-processed target type under the monitoring data embedding vectors of not less than one type for iterative execution, and stop when the monitoring data embedding vectors of all types are processed. According to the current new set of representation vectors, obtain the overall representation vector of the target object monitoring sequence and the withdrawal symptom representation vectors of each alternative withdrawal symptom label.

[0074] In the embodiment of the present application, as the iterative execution and summary link of step S140, step S144 plays a crucial role. It ensures that the computer system can comprehensively and systematically process the monitoring data embedding vectors of all types and finally integrate them into the overall representation vector of the target object monitoring sequence and the withdrawal symptom representation vectors of each alternative withdrawal symptom label.

[0075] After a complete iteration of steps S141 to S143, the computer system has generated corresponding focused representation vectors for the monitoring data embedding vectors of the currently to-be-processed target type and may have updated the set of representation vectors based on these focused representation vectors. However, since the monitoring data often covers multiple types (such as heart rate, blood pressure, sleep quality, emotional state, etc.), the computer system repeats this process to ensure that each type of monitoring data can be fully utilized and analyzed. Therefore, in step S144, the computer system automatically jumps to step S141 to re-determine the monitoring data embedding vector of the next to-be-processed target type. This process is iterative, and each iteration will execute the operations of steps S141 to S143 for a new type of monitoring data. As the iteration progresses, the system gradually accumulates more feature information about the target object in different monitoring dimensions.

[0076] The number of iterative executions depends on the total number of monitoring data types. When the monitoring data embedding vectors of all types are processed, the iterative process stops. At this time, the computer system has constructed a set of representation vectors containing rich context information. This set not only reflects the comprehensive state of the target object under multiple monitoring data but also emphasizes the features most relevant to the target withdrawal symptoms through the weight focusing operation.

[0077] After the iteration is completed, the system further processes the current latest set of representation vectors to obtain the overall representation vector of the target object monitoring sequence and the withdrawal symptom representation vectors of each alternative withdrawal symptom label. This process may involve weighted averaging, feature fusion, or other advanced data processing techniques on the vectors in the set of representation vectors. The finally generated overall representation vector is a comprehensive feature representation that comprehensively summarizes the overall state of the target object during the monitoring period; while the withdrawal symptom representation vectors of each alternative withdrawal symptom label respectively reflect the degree of association between these symptoms and the current state of the target object.

[0078] As an implementation, the graph convolutional network includes the label vertices corresponding to each alternative withdrawal symptom label and the connecting lines between the label vertices, and the connecting lines are used to indicate the involvement information between the two connected label vertices; based on this, in step S400, according to the confidence distribution and the graph convolutional networks corresponding to each preset withdrawal symptom classification, perform graph context embedding operations on the withdrawal symptom representation vectors of each alternative withdrawal symptom label to obtain the post-operation withdrawal symptom representation vectors corresponding to each alternative withdrawal symptom label, which may specifically include: Step S410: For each alternative withdrawal symptom label, determine the target label vertex corresponding to the alternative withdrawal symptom label in the graph convolutional networks corresponding to each preset withdrawal symptom classification.

[0079] In the embodiments of the present application, when the computer system executes step S410, it traverses the set of all alternative withdrawal symptom labels. These labels represent various psychological and behavioral problems that may occur during the drug rehabilitation process, such as anxiety, depression, irritability, insomnia, etc. Each label is a potential node in the graph convolutional network and is used to represent a specific withdrawal symptom. Next, for each alternative withdrawal symptom label in the set, the system queries multiple predefined preset withdrawal symptom classifications and their corresponding graph convolutional networks. These graph convolutional networks are constructed based on a large amount of monitoring data and expert knowledge, and they describe the association relationships between symptom labels under different withdrawal symptom classifications in the form of graphs. Each network contains a series of label vertices, and these vertices correspond one by one to the alternative withdrawal symptom labels.

[0080] After determining the currently processed alternative withdrawal symptom label, the system traverses the graph convolutional networks corresponding to all preset withdrawal symptom classifications to find the corresponding vertex of the label in each network. Since different graph convolutional networks may focus on different symptom classifications and association patterns, the same alternative withdrawal symptom label may correspond to different vertices or positions in different networks.

[0081] For example, assume that the set of alternative withdrawal symptom markers contains the marker "anxiety". When processing "anxiety", the system first looks for a graph convolutional network directly related to the anxiety classification. In this network, "anxiety" will have a clearly corresponding target marker vertex, and this vertex is connected by connection lines to the vertices of other symptom markers closely related to anxiety, such as "depression" and "panic". At the same time, the system also checks the graph convolutional networks corresponding to other preset withdrawal symptom classifications. Although "anxiety" may not be the main symptom of concern in these networks, it may still exist as a secondary or related symptom marker and have a corresponding vertex position. Through this mapping process, the computer system finds the corresponding target marker vertices for each alternative withdrawal symptom marker in all relevant preset withdrawal symptom classification graph convolutional networks. These vertices will become the starting points for performing graph context embedding operations in subsequent steps, allowing the system to utilize the powerful capabilities of the graph convolutional network to capture and represent the complex correlation relationships between symptom markers.

[0082] Step S420: For each graph convolutional network corresponding to a preset withdrawal symptom classification, determine the adjacent representation vector of the target marker vertex under the preset withdrawal symptom classification according to the involvement information between the target marker vertex and the remaining marker vertices in the graph convolutional network.

[0083] In the embodiment of the present application, step S420 uses the involvement information between the target marker vertex and the remaining marker vertices in the graph convolutional network to determine the adjacent representation vector of the target marker vertex under a specific preset withdrawal symptom classification. When the computer system executes step S420, it has completed the determination of the corresponding target marker vertices for each alternative withdrawal symptom marker in the graph convolutional network (step S410). Next, for each graph convolutional network corresponding to a preset withdrawal symptom classification, the system further analyzes the connection relationship between the target marker vertex and its adjacent vertices. These connection relationships are represented by edges (connection lines) in the graph convolutional network and contain rich involvement information. The involvement information can, for example, reflect the co-occurrence frequency, conditional probability, or other statistical correlations of two symptom markers in the monitoring data, and they are learned based on a large amount of historical data when constructing the graph convolutional network. In the context of drug rehabilitation, this involvement information may reveal a strong correlation between anxiety and depression, or insomnia as a common accompanying phenomenon of anxiety symptoms.

[0084] For each target marker vertex, the system first identifies all adjacent vertices directly connected to it. These adjacent vertices represent other symptom markers that have a significant association with the target symptom marker under a specific classification. Then, the system extracts the representation vectors of each adjacent vertex according to the structure and parameters of the graph convolutional network. These representation vectors are obtained through feature extraction and embedding operations in the previous steps, and they encode the feature information of the adjacent symptom markers in the form of high-dimensional vectors.

[0085] It should be noted that the extraction of adjacent representation vectors may not be limited to directly connected vertices, but may also involve indirectly adjacent vertices that can be reached through multiple paths. This depends on the specific implementation and parameter configuration of the graph convolutional network. Some advanced graph convolutional models (such as Graph Attention Networks) can consider a wider range of neighborhood information and dynamically adjust the weights of different adjacent vertices through an attention mechanism.

[0086] Through step S420, the computer system determines a set of adjacent representation vectors for each target labeled vertex under a specific preset withdrawal symptom classification. These vectors not only contain the feature information of adjacent symptom labels, but also reflect the complex association relationship between them and the target symptom label through the structure of the graph convolutional network. These adjacent representation vectors will be used in subsequent steps to update the feature representation of the target labeled vertex, so as to generate a more comprehensive and accurate post-operation withdrawal symptom representation vector.

[0087] Step S430: According to the confidence distribution, integrate the adjacent representation vectors of the target labeled vertex under each preset withdrawal symptom classification to obtain the integrated adjacent representation vectors of the alternative withdrawal symptom labels.

[0088] When the computer system executes step S430, it has obtained the adjacent representation vectors of the target labeled vertex under each preset withdrawal symptom classification through step S420. These vectors respectively represent the feature information of other symptom labels associated with the target symptom label from different classification perspectives. At the same time, the confidence distribution generated by step S200 provides the system with a probability assessment of the target object monitoring sequence belonging to each preset withdrawal symptom classification. In the process of integrating the adjacent representation vectors, the computer system considers the influence of the confidence distribution. Each probability value in the confidence distribution reflects the matching degree between the target object monitoring sequence and the corresponding preset withdrawal symptom classification. Therefore, the system weights the adjacent representation vectors under each classification according to these probability values, so that the classification more relevant to the target object contributes a greater weight in the integration process.

[0089] The specific integration operation may involve various mathematical and machine learning techniques, such as weighted average, feature splicing, attention mechanism, etc. Taking weighted average as an example, the system can calculate the weighted sum of the adjacent representation vectors under each classification, where the weight is determined by the confidence of this classification. In this way, integrating the adjacent representation vectors can comprehensively reflect the influence degree of adjacent symptom labels under different classifications on the target symptom label.

[0090] In addition, some advanced graph convolutional network models (such as Graph Attention Networks) can directly implement the attention mechanism on graph-structured data, automatically learn the importance weights of different adjacent vertices (i.e., symptom markers), and perform feature fusion accordingly. This method can capture the complex correlation relationships between symptom markers more flexibly and generate more accurate and meaningful integrated adjacent representation vectors.

[0091] Step S440: Perform iterative operations on the withdrawal symptom representation vectors of the alternative withdrawal symptom markers according to the integrated adjacent representation vectors to obtain the post-operation withdrawal symptom representation vectors corresponding to the alternative withdrawal symptom markers.

[0092] When the computer system executes step S440, it has obtained the integrated adjacent representation vectors of the alternative withdrawal symptom markers through step S430. This vector is the result of comprehensively considering the feature information of other symptom markers associated with the target symptom marker under multiple preset withdrawal symptom classifications and weighted integration based on the confidence distribution. At the same time, the system also retains the original withdrawal symptom representation vectors of the alternative withdrawal symptom markers, which are generated through previous feature extraction and embedding operations and directly reflect the characteristic attributes of the symptom markers themselves.

[0093] Next, the system uses a graph convolutional network (GCN) or other machine learning models suitable for processing graph-structured data to perform iterative operations on the integrated adjacent representation vectors and the original withdrawal symptom representation vectors. This operation usually involves multiple steps, aiming to continuously update the feature representations of each symptom marker through the message passing mechanism so that it can incorporate more context information.

[0094] During the iterative process, the system can adopt various strategies to integrate these two types of representation vectors. A common method is feature concatenation, that is, directly concatenating the integrated adjacent representation vector and the original withdrawal symptom representation vector into a longer vector. However, this method can increase the dimension and sparsity of the vector, which is not conducive to subsequent processing. Therefore, a more commonly used method is to dynamically adjust the contribution degrees of the two representation vectors through weighted summation or the attention mechanism. Taking the graph convolutional network as an example, the system can define one or more graph convolutional layers, and each layer contains a series of learnable parameters (such as weight matrices and bias terms). At the beginning of the iteration, the system takes the integrated adjacent representation vector and the original withdrawal symptom representation vector as inputs, and through the calculation of the graph convolutional layer, generates new feature representations. These new feature representations not only retain the information of the original vectors but also incorporate the features of adjacent symptom markers and the correlation relationships between them. Then, the system takes these new feature representations as inputs for the next round of iteration and repeats the above process until a predetermined number of iterations is reached or the convergence condition is satisfied.

[0095] After multiple iterations, the system finally generates the post - operation withdrawal symptom representation vectors corresponding to the alternative withdrawal symptom markers. This vector not only contains the characteristic information of the symptom markers themselves but also incorporates broader context information by integrating adjacent representation vectors and iterative operations. Such post - operation representation vectors more comprehensively and accurately reflect the state of the symptom markers in the context of specific monitoring data, providing strong support for subsequent symptom reasoning and the formulation of corrective intervention strategies.

[0096] As an implementation, the confidence distribution includes the confidence levels corresponding to each preset withdrawal symptom classification for the monitoring sequence of the target object. Based on this, in step S430, according to the confidence distribution, an integration operation is performed on the adjacent representation vectors of the target marker vertices under each preset withdrawal symptom classification to obtain the integrated adjacent representation vectors of the alternative withdrawal symptom markers, which may specifically include: Step S431: For each preset withdrawal symptom classification, an integration operation is performed on the confidence level of the monitoring sequence of the target object belonging to the preset withdrawal symptom classification and the adjacent representation vectors of the target marker vertices under the preset withdrawal symptom classification to obtain the initial integrated adjacent representation vectors of the alternative withdrawal symptom markers under the preset withdrawal symptom classification. Step S432: Determine the integrated adjacent representation vectors of the alternative withdrawal symptom markers according to the initial integrated adjacent representation vectors of the alternative withdrawal symptom markers under each preset withdrawal symptom classification.

[0097] In step S431, the computer system performs an integration operation for each preset withdrawal symptom classification to generate the initial integrated adjacent representation vectors of the alternative withdrawal symptom markers under that classification. The core of the integration operation is to combine the confidence level of the monitoring sequence of the target object belonging to that classification with the adjacent representation vectors of the target marker vertices under the classification. Specifically, the system first extracts the confidence levels corresponding to each preset withdrawal symptom classification of the monitoring sequence of the target object. These confidence levels reflect the matching degree between the monitoring sequence and different classifications and are obtained through reasoning by a machine learning model in previous steps (such as step S200). Then, for each preset withdrawal symptom classification, the system performs a weighted sum or other form of integration on the confidence level under that classification and the adjacent representation vectors of the target marker vertices (i.e., the corresponding vertices of the alternative withdrawal symptom markers in the graph convolutional network). The weighting coefficient here is the confidence level of that classification, which ensures that the classifications more relevant to the target object contribute greater weights in the integration process.

[0098] For example, assume that the alternative withdrawal symptom is labeled as "anxiety", and the confidence levels of the target object monitoring sequence under the two classifications of "mood disorder" and "sleep disorder" are 0.7 and 0.3 respectively. Under the "mood disorder" classification, the adjacent representation vectors of the "anxiety" vertex may include the characteristic information of labels such as "depression" and "irritability". The system multiplies these adjacent representation vectors by the confidence level of 0.7 (or performs other forms of weighting processing) to obtain the initial integrated adjacent representation vector of "anxiety" under the "mood disorder" classification. Similarly, the system will also perform a similar process on the adjacent representation vectors under the "sleep disorder" classification.

[0099] After generating the initial integrated adjacent representation vectors under all preset withdrawal symptom classifications, the system enters step S432 to determine the final integrated adjacent representation vector of the alternative withdrawal symptom label based on these initial vectors. The implementation method of this step may vary depending on the specific application scenario and system design, but the core goal is to synthesize the initial vectors under different classifications to form an integrated vector that can comprehensively reflect the context information of the alternative withdrawal symptom label. A common method is to splice or perform weighted averaging on all the initial vectors. The splicing operation will directly connect all the initial vectors into a longer vector, but it can increase the dimension and sparsity of the vector; while the weighted averaging operation will assign weights to different initial vectors according to the confidence level of the classification or other relevant factors, and calculate their weighted sum as the final vector.

[0100] Taking weighted averaging as an example, the system can assign weights of 0.7 and 0.3 respectively to the initial integrated adjacent representation vectors under the "mood disorder" and "sleep disorder" classifications (corresponding to their confidence levels), and then add these two vectors according to the weights to obtain the final integrated adjacent representation vector. This vector not only contains the adjacent symptom characteristic information of the "anxiety" symptom under different classifications, but also reflects the importance difference of this information through confidence level weighting.

[0101] As an implementation method, in step S500, according to the post-operation withdrawal symptom representation vectors corresponding to each alternative withdrawal symptom label, determine the inference confidence level of each alternative withdrawal symptom label belonging to the target object monitoring sequence, which may specifically include: Step S510: For each alternative withdrawal symptom label, perform an integration operation on the withdrawal symptom representation vector and the post-operation withdrawal symptom representation vector of the alternative withdrawal symptom label to obtain the target withdrawal symptom representation vector of the alternative withdrawal symptom label; Step S520: According to the target withdrawal symptom representation vectors of each alternative withdrawal symptom label, determine the inference confidence level of each alternative withdrawal symptom label belonging to the target object monitoring sequence.

[0102] In step S510, the computer system performs an integration operation for each alternative withdrawal symptom marker to generate its target withdrawal symptom representation vector. This vector is a comprehensive representation of the withdrawal symptom representation vector of the alternative withdrawal symptom marker (generated in step S100, reflecting the initial characteristic information of the symptom marker) and the post-operation withdrawal symptom representation vector (generated in step S400, incorporating the context information in the graph convolutional network).

[0103] The specific implementation of the integration operation may vary depending on the system design. However, common methods include weighted summation, feature concatenation, or feature fusion using deep learning models. Weighted summation sums the elements of the corresponding dimensions in the two representation vectors after weighting them according to a certain weight assignment strategy; feature concatenation directly concatenates the two vectors end to end to form a longer vector; and feature fusion may further learn the non-linear combination method of the two vectors through neural network layers (such as fully connected layers, convolutional layers, etc.).

[0104] Taking weighted summation as an example, the system can assign certain weights to the withdrawal symptom representation vector and the post-operation withdrawal symptom representation vector respectively (these weights can be preset or obtained through learning), and then sum the weighted elements of their corresponding dimensions to obtain the target withdrawal symptom representation vector. This vector not only retains the initial characteristics of the symptom marker but also incorporates rich context information, providing comprehensive data support for the calculation of the subsequent inference confidence.

[0105] After obtaining the target withdrawal symptom representation vectors of each alternative withdrawal symptom marker, the system enters step S520 to determine the inference confidence of each symptom marker belonging to the target object monitoring sequence based on these vectors. The inference confidence reflects the matching degree between the symptom marker and the current withdrawal state of the target object, and is an important basis for subsequent symptom recognition and the formulation of correction and intervention strategies.

[0106] The specific method for determining the inference confidence may involve machine learning classifiers, regression models, or other types of decision algorithms. For example, the system can train a support vector machine (SVM) classifier, use the target withdrawal symptom representation vector as input, and output the probability value of each symptom marker belonging to the target monitoring sequence as the inference confidence. Alternatively, the system can also adopt a logistic regression model to obtain a confidence score between 0 and 1 by calculating the dot product between the target vector and the model parameters and activating through the sigmoid function.

[0107] Regardless of the algorithm used, its core goal is to evaluate the correlation between the alternative withdrawal symptom marker and the target object monitoring sequence based on the feature information in the target withdrawal symptom representation vector. The higher the inference confidence of a symptom marker, the more likely it is the main withdrawal symptom currently present in the target object, and thus more attention and targeted correction and intervention measures are required.

[0108] As another implementation, in step S100, the characterization information is extracted from the target object monitoring sequence and each alternative withdrawal symptom marker respectively to obtain the overall characterization vector of the target object monitoring sequence and the withdrawal symptom characterization vectors of each alternative withdrawal symptom marker, which can be specifically implemented through a monitoring data recognition network. That is, step S100 may include: Step S100': Based on the monitoring data recognition network, the characterization information is extracted from the target object monitoring sequence and each alternative withdrawal symptom marker respectively to obtain the overall characterization vector of the target object monitoring sequence and the withdrawal symptom characterization vectors of each alternative withdrawal symptom marker.

[0109] When the computer system executes step S100', it first relies on a pre-trained monitoring data recognition network. This network is a complex machine learning model, usually containing multiple hierarchical structures such as convolutional layers, pooling layers, fully connected layers, etc., aiming to extract high-level feature representations from the original monitoring data. These feature representations not only capture the essential attributes of the data but also facilitate subsequent analysis and processing.

[0110] In the embodiments of the present application, the monitoring data recognition network can process various types of monitoring data, including but not limited to physiological indicators (such as heart rate, blood pressure), psychological assessment results (such as emotional state scores), behavior records (such as daily activity levels), etc. These data form the basis of the target object monitoring sequence and reflect the comprehensive state of drug addicts during the rehabilitation process. For the target object monitoring sequence, the monitoring data recognition network will receive the raw data of the entire sequence as input and gradually extract the key features in the sequence through a series of forward propagation operations. These features may include periodic patterns, trend changes, outliers, etc. in the time series, and they jointly constitute the overall characterization vector of the target object monitoring sequence. This vector is a point in a high-dimensional space and can comprehensively and accurately represent the global features of the monitoring sequence.

[0111] Meanwhile, for each alternative withdrawal symptom marker, the monitoring data recognition network will also perform similar characterization information extraction operations. However, since the symptom markers themselves are discrete text or symbol representations, the network needs to first convert them into continuous representations in a high-dimensional vector space through an embedding layer (such as Word2Vec, GloVe, etc.). Then, the network will further process these embedding vectors, extract semantic features, context information, etc. related to the symptom markers, and generate the corresponding withdrawal symptom characterization vectors.

[0112] In practical applications, the monitoring data recognition network may be a deep learning model, such as a recurrent neural network (RNN) and its variants (such as LSTM, GRU) for processing time series data, or a convolutional neural network (CNN) combined with an attention mechanism for processing multi-dimensional feature data. These models learn the mapping relationship from the original monitoring data to the high-level feature representation through a large amount of training data, enabling them to quickly and accurately perform the task of extracting characterization information on new monitoring sequences and symptom markers.

[0113] Through the implementation of step S100’, the computer system can efficiently obtain the overall characterization vector of the target object monitoring sequence and the withdrawal symptom characterization vectors of each alternative withdrawal symptom marker. These vectors not only provide rich feature information for subsequent steps (such as the confidence inference in step S200, the graph context embedding operation in step S400, etc.), but also significantly improve the intelligence level and processing efficiency of the entire monitoring system.

[0114] As an implementation manner, before step S100’, based on the monitoring data recognition network, respectively extracting the characterization information from the target object monitoring sequence and each alternative withdrawal symptom marker to obtain the overall characterization vector of the target object monitoring sequence and the withdrawal symptom characterization vectors of each alternative withdrawal symptom marker, it may also include the training process of the monitoring data recognition network, specifically including: Step S101: Obtain a debugging example set and the graph convolutional networks corresponding to each preset withdrawal symptom classification. The debugging example set includes multiple object monitoring sequence examples and the actual withdrawal symptom markers of the object monitoring sequence examples; Step S102: Based on the monitoring data recognition network, respectively extract the characterization information from the object monitoring sequence examples and each alternative withdrawal symptom marker to obtain the example overall characterization vector of the object monitoring sequence examples and the example withdrawal symptom characterization vectors of each alternative withdrawal symptom marker; Step S103: According to the example overall characterization vector, infer the confidence distribution of the object monitoring sequence examples corresponding to each preset withdrawal symptom classification; Step S104: According to the confidence distribution and the graph convolutional networks corresponding to each preset withdrawal symptom classification, perform a graph context embedding operation on the example withdrawal symptom characterization vectors of each alternative withdrawal symptom marker to obtain the processed example withdrawal symptom characterization vectors corresponding to each alternative withdrawal symptom marker; Step S105: According to the processed example withdrawal symptom characterization vectors corresponding to each alternative withdrawal symptom marker, determine the inference confidence of each alternative withdrawal symptom marker belonging to the object monitoring sequence examples; Step S106: According to the inference confidence and the actual withdrawal symptom markers of the object monitoring sequence examples, correct the internal configuration parameters in the monitoring data recognition network to obtain a debugged monitoring data recognition network.

[0115] In the embodiments of the present application, the implementation of step S100' not only depends on the efficient operation of the monitoring data recognition network, but also requires the network to be carefully designed and fully trained in the early stage. This process ensures that the network can accurately extract valuable feature information from the original monitoring data, laying a solid foundation for subsequent analysis and reasoning. The following is a detailed explanation of the training process of the monitoring data recognition network in step S100', illustrated with specific application scenarios and examples.

[0116] Before starting to train the monitoring data recognition network, the computer system first needs to prepare a set of debugging example sets. This set of example sets is carefully selected and contains monitoring sequence examples of multiple objects (i.e., drug addicts undergoing rehabilitation) and the corresponding actual withdrawal symptom markers. Each monitoring sequence example is a series of monitoring data points arranged in chronological order, which may cover multiple dimensions such as physiological indicators (e.g., heart rate, blood pressure), psychological assessment results (e.g., emotional state scores), and behavior records (e.g., daily activity levels). The actual withdrawal symptom markers are comprehensively judged by professional doctors or rehabilitation teams based on monitoring results and clinical manifestations, and they accurately indicate the main withdrawal symptoms of the object during the monitoring period.

[0117] In addition to the debugging example sets, the system also needs to obtain graph convolutional networks corresponding to each preset withdrawal symptom classification. These graph convolutional networks are trained with a large amount of data in the early stage, and they represent the association relationships between symptom markers under different withdrawal symptom classifications in the form of graphs. Each network contains a series of labeled vertices (representing different withdrawal symptom markers) and connecting lines (representing the implicated information between symptom markers), and this information is crucial for subsequent graph context embedding operations.

[0118] After obtaining the debugging example sets and graph convolutional networks, the system starts to extract characterization information from the data in the example sets using the monitoring data recognition network. This step is one of the core links in the training process, which involves converting the original monitoring data into high-level feature representations for subsequent analysis and processing.

[0119] For each object monitoring sequence example, the monitoring data recognition network receives the original data of the entire sequence as input and gradually extracts the key features in the sequence through its internal multi-layer structure (such as convolutional layers, pooling layers, fully connected layers, etc.). These features may include periodic patterns, trend changes, outliers, etc. in the time series, and they jointly constitute the overall characterization vector of the monitoring sequence example. This vector is a point in a high-dimensional space that can comprehensively and accurately represent the global features of the monitoring sequence.

[0120] Meanwhile, for each alternative withdrawal symptom marker, the monitoring data recognition network also performs a similar operation of extracting characterization information. However, since the symptom markers themselves are discrete text or symbol representations, the network needs to first convert them into continuous representations (i.e., embedding vectors) in a high-dimensional vector space through an embedding layer. These embedding vectors capture the semantic features and context information of the symptom markers, providing a basis for subsequent graph context embedding operations. Then, the network further processes these embedding vectors to generate corresponding withdrawal symptom characterization vectors.

[0121] After obtaining the overall characterization vectors of the sample of the object monitoring sequence, the computer system uses these vectors to infer the confidence distribution of the sample corresponding to each preset withdrawal symptom classification. This step usually involves a classifier or regression model that can evaluate the probability or score of the sample belonging to different withdrawal symptom classifications based on the feature information in the overall characterization vector of the sample.

[0122] Specifically, the system can use a multi-layer perceptron (MLP) or a support vector machine (SVM) as a classifier. These classifiers have learned the mapping relationship from feature representations to classification labels through a large amount of training data. During the inference process, the classifier receives the overall characterization vector of the sample as input and calculates the confidence or probability values of the sample belonging to each preset withdrawal symptom classification through its internal parameters. These confidence values form a confidence distribution, reflecting the matching degree between the sample and different classifications.

[0123] After obtaining the confidence distribution, the system starts to perform graph context embedding operations on the sample withdrawal symptom characterization vectors of the alternative withdrawal symptom markers using the graph convolutional networks corresponding to each preset withdrawal symptom classification. This step aims to further enrich and refine the feature representation of the withdrawal symptom characterization vectors, enabling them to more accurately reflect the association relationships and context information between symptom markers.

[0124] Specifically, for the sample withdrawal symptom characterization vector of each alternative withdrawal symptom marker, the system determines the target marker vertex of the marker in the graph convolutional networks corresponding to each preset withdrawal symptom classification. Then, the system calculates the involvement information (i.e., the information represented by the connection lines) between the target marker vertex and adjacent vertices according to the structure and parameters of the graph convolutional network. These information reflect the association strength and mutual influence between adjacent symptom markers and the target symptom marker.

[0125] Next, the system assigns weights to the implicated information under different classifications according to the confidence distribution. Classifications with higher confidence will have greater influence when assigning weights, thus ensuring that symptom markers more relevant to the target object contribute more information during the embedding process. Then, the system uses this weighted implicated information to perform iterative update operations on the sample withdrawal symptom representation vectors, generating processed sample withdrawal symptom representation vectors. This vector not only contains the characteristic information of the symptom markers themselves but also incorporates rich context information and association relationships.

[0126] After obtaining the processed sample withdrawal symptom representation vectors of each alternative withdrawal symptom marker, the computer system uses these vectors to determine the inference confidence of each symptom marker belonging to the sample of the object monitoring sequence. This step usually involves another classifier or regression model that can evaluate the correlation or matching degree between the symptom marker and the target monitoring sequence sample based on the characteristic information in the processed representation vectors.

[0127] Specifically, a logistic regression model or a deep neural network can be used as the classifier. These classifiers have learned the mapping relationship from feature representations to classification labels through a large amount of training data. During the inference process, the classifier receives the processed sample withdrawal symptom representation vectors as input and calculates the confidence or probability value of the symptom marker belonging to the target monitoring sequence sample through its internal parameters. These confidence values reflect the matching strength and likelihood between the symptom marker and the target monitoring sequence.

[0128] Finally, after obtaining the inference confidence of all alternative withdrawal symptom markers, the computer system compares these confidences with the actual withdrawal symptom markers of the object monitoring sequence sample to evaluate the performance of the monitoring data recognition network. This step is the feedback link of the training process, which adjusts the internal configuration parameters (such as weights, biases, etc.) of the network by comparing the differences between the predicted results and the actual results, thereby improving the accuracy and generalization ability of the network.

[0129] Specifically, a loss function can be used to calculate the difference between the predicted confidence and the actual marker. The loss function is a metric for measuring the degree of prediction error, and it usually includes multiple terms to separately consider errors in different aspects (such as classification errors, confidence biases, etc.). Then, the system uses the backpropagation algorithm to optimize the loss function, adjusting the internal configuration parameters of the network to minimize the loss value. This optimization process is an iterative process that will be iterated multiple times on the entire debugging sample set until the convergence condition is met or the preset number of iterations is reached.

[0130] During the optimization process, the system can adopt various optimization algorithms (such as gradient descent method, Adam optimizer, etc.) to accelerate convergence and avoid falling into local optimal solutions. At the same time, the system will also record the loss value and performance metrics (such as accuracy, recall rate, F1 score, etc.) after each iteration to monitor the training progress and evaluate the network performance.

[0131] After multiple iterations of optimization, the system will finally obtain a well-debugged monitoring data recognition network. This network can not only perform well on the debug sample set but also has good generalization ability, enabling accurate characterization information extraction and inference confidence calculation on new monitoring sequences and symptom markers. It provides strong support for subsequent steps (such as confidence inference in step S200, graph context embedding operation in step S400, etc.).

[0132] In summary, the training process of the monitoring data recognition network in step S100' is a complex and delicate process, which involves multiple aspects such as the preparation of the debug sample set, characterization information extraction, confidence inference, graph context embedding operation, inference confidence calculation, and network parameter optimization. Through the implementation of this process, the computer system can build an efficient and accurate monitoring data recognition network, providing a solid technical foundation for the intelligent application of drug rehabilitation monitoring.

[0133] As an implementation method, in step S106, according to the inference confidence and the actual withdrawal symptom markers of the object monitoring sequence samples, the internal configuration parameters in the monitoring data recognition network are corrected to obtain a well-debugged monitoring data recognition network, which can specifically include: Step S1061: Determine the debugging error of the monitoring data recognition network according to the inference confidence and the actual withdrawal symptom markers of the object monitoring sequence samples; Step S1062: If the debugging error does not reach the preset debugging cut-off requirement, correct the internal configuration parameters in the monitoring data recognition network according to the debugging error; Step S1063: Divide each object monitoring sequence sample according to the confidence distribution corresponding to each preset withdrawal symptom classification of the object monitoring sequence sample to obtain the set of object monitoring sequence samples corresponding to each preset withdrawal symptom classification; Step S1064: For each preset withdrawal symptom classification, iterate the involvement information of each alternative withdrawal symptom marker in the graph convolutional network corresponding to the preset withdrawal symptom classification according to the actual withdrawal symptom markers of the object monitoring sequence samples in the set of object monitoring sequence samples corresponding to the preset withdrawal symptom classification to obtain the iterated graph convolutional network corresponding to the preset withdrawal symptom classification; Step S1065: Identify the network for monitoring data and the iterated graph convolutional networks corresponding to each preset withdrawal symptom classification. Jump to the step of extracting characterization information from the object monitoring sequence sample and each alternative withdrawal symptom label based on the network for monitoring data, and iterate. Stop the iteration when the debugging error of the network for monitoring data reaches the preset debugging cut-off requirement to obtain a well-debugged network for monitoring data.

[0134] The computer system first determines the debugging error of the network for monitoring data based on the inference confidence and the actual withdrawal symptom label of the object monitoring sequence sample. The debugging error is an important indicator for measuring the difference between the network prediction result and the actual result, which reflects the accuracy and reliability of the network in the current training state.

[0135] Specifically, the system traverses each object monitoring sequence sample in the debugging sample set. For each sample, it compares its inference confidence with the actual withdrawal symptom label. The inference confidence is calculated in step S105, which represents the probability or score that the alternative withdrawal symptom label belongs to this sample; while the actual withdrawal symptom label is comprehensively judged by professional doctors or the rehabilitation team based on clinical manifestations and monitoring data, representing the true state of the sample.

[0136] To quantify this difference, the system can use the cross-entropy loss function, mean squared error loss function, or other suitable loss functions for classification or regression tasks. Taking the cross-entropy loss function as an example, it calculates the similarity between the inference confidence distribution and the actual withdrawal symptom label distribution. The lower the similarity (i.e., the greater the difference), the higher the loss value. The system aggregates or averages the loss values of all samples to obtain the final debugging error.

[0137] This debugging error not only reflects the prediction accuracy of the network on a single sample but also synthesizes the overall performance of all samples. It is the basis for network parameter correction and optimization in subsequent steps.

[0138] If the debugging error does not reach the preset debugging cut-off requirement (i.e., the current network performance does not meet the expected standard), the computer system corrects the internal configuration parameters in the network for monitoring data according to the debugging error. This process is usually achieved through the backpropagation algorithm and the optimizer.

[0139] The backpropagation algorithm is a method for updating network parameters by calculating the gradient of the loss function with respect to the network parameters. In step S1062, the system first calculates the gradient of the debugging error with respect to the internal configuration parameters of the network (such as weights, biases, etc.), and these gradients indicate how the parameters should be adjusted to reduce the debugging error.

[0140] Then, the system uses an optimizer (such as gradient descent, Adam optimizer, etc.) to update the network parameters according to the gradient. The role of the optimizer is to accelerate the convergence process and avoid falling into the local optimal solution. It will calculate the specific values ​​of the parameter update based on the size and direction of the gradient, learning rate, momentum and other parameters, and apply these values ​​to the network parameters.

[0141] By repeatedly executing steps S1061 and S1062 (i.e., calculating the debugging error and correcting the internal configuration parameters), the computer system can gradually reduce the debugging error and improve the accuracy and generalization ability of the monitoring data recognition network.

[0142] While correcting the network parameters, the system also needs to divide the sample set according to the confidence distribution of the object monitoring sequence samples corresponding to each preset withdrawal symptom classification. The purpose of this step is to provide basic data support for the subsequent iterative optimization of the graph convolutional network.

[0143] Specifically, the system traverses each object monitoring sequence sample in the debugging sample set, and for each sample, checks the confidence distribution corresponding to each preset withdrawal symptom classification. The confidence distribution is calculated in step S103, which represents the probability or score of the sample belonging to different classifications.

[0144] Then, the system determines the main category to which the sample belongs based on the maximum value in the confidence distribution or the probability value exceeding a certain threshold. For example, if the confidence of a sample in the "mood disorder" category is the highest and exceeds 0.5 (assuming the threshold is 0.5), the sample is classified into the object monitoring sequence sample set corresponding to the "mood disorder" category.

[0145] Through this division process, the system can obtain a set of object monitoring sequence samples corresponding to multiple preset withdrawal symptom classifications. The samples in each set belong to the same classification or have similar classification features, which facilitates the subsequent optimization of graph convolutional networks for specific classifications.

[0146] After obtaining the sample set of object monitoring sequences corresponding to each preset withdrawal symptom classification, the computer system iteratively optimizes the graph convolutional network for each classification. This step aims to update the involved information (i.e., the information represented by the connecting lines) of each candidate withdrawal symptom marker in the graph convolutional network according to the actual withdrawal symptom marker, thereby improving the performance of the network under a specific classification.

[0147] Specifically, for each preset withdrawal symptom classification and its corresponding object monitoring sequence sample set, the system performs the following operations: Count actual withdrawal symptom markers: First, count the actual withdrawal symptom markers of all samples in the collection to understand which symptom markers appear more frequently under this classification and the co-occurrence relationship between them.

[0148] Evaluate the involvement information: Then evaluate whether the involvement information among the alternative withdrawal symptom markers in the current graph convolutional network accurately reflects the co-occurrence relationship in the actual data. This can be achieved by calculating the similarity or correlation between the involvement information in the network and the co-occurrence frequency of symptom markers in the actual data.

[0149] Iteratively update the involvement information: If the evaluation result shows a large difference between the involvement information in the network and the actual data, then these information need to be iteratively updated. The update process may involve operations such as adjusting the weights of connection lines, adding or deleting connection lines, etc., to more accurately reflect the association relationship between symptom markers.

[0150] Verify the update effect: After each update, the system will use the validation set (or a part of the debugging sample set) to verify the update effect. The validation set is a data set independent of the training set, used to evaluate the performance of the network on unseen data. If the performance of the updated graph convolutional network on the validation set is improved, it indicates that the update is effective; otherwise, other update strategies or adjustment of update parameters may need to be tried.

[0151] By iteratively executing the above operations multiple times (i.e., counting the actual withdrawal symptom markers, evaluating the involvement information, iteratively updating the involvement information, verifying the update effect), the computer system can gradually optimize the graph convolutional network corresponding to each preset withdrawal symptom classification, making it more accurately reflect the association relationship between symptom markers in the actual data.

[0152] After correcting the internal configuration parameters of the monitoring data recognition network and iteratively optimizing the graph convolutional network corresponding to each preset withdrawal symptom classification, the computer system returns to the starting point of step S100', and re-performs operations such as extracting representation information and calculating inference confidence based on the corrected network and the iterated graph convolutional network. This process is an iterative training process, which will be iterated multiple times on the entire debugging sample set until the preset debugging cut-off requirement is met or the preset number of iterations is reached.

[0153] Specifically, in each iteration, the system performs the following operations: Extract representation information: Based on the corrected monitoring data recognition network, perform representation information extraction operations on the object monitoring sequence samples and alternative withdrawal symptom markers in the debugging sample set (step S102). This step will generate key data such as the overall sample representation vector and the sample withdrawal symptom representation vector.

[0154] Confidence inference: Infer the confidence distribution of the object monitoring sequence samples corresponding to each preset withdrawal symptom classification based on the overall sample representation vector (step S103). This step will generate confidence distribution data for subsequent graph context embedding operations.

[0155] Graph context embedding operation: Perform a graph context embedding operation on the sample withdrawal symptom representation vector according to the confidence distribution and the iterated graph convolutional network (step S104). This step generates a processed sample withdrawal symptom representation vector for subsequent inference confidence calculation.

[0156] Inference confidence calculation: Determine the inference confidence of each alternative withdrawal symptom label belonging to the object monitoring sequence sample according to the processed sample withdrawal symptom representation vector (step S105). This step generates inference confidence data for subsequent performance evaluation and error calculation.

[0157] Performance evaluation and error calculation: Calculate the debugging error based on the inference confidence and the actual withdrawal symptom labels of the object monitoring sequence sample (step S1061). Then, determine whether the debugging error reaches the preset debugging cut-off requirement; if not, continue to perform operations such as parameter correction and graph convolutional network optimization (steps S1062 to S1064); if so, stop the iterative training process and obtain a well-debugged monitoring data recognition network.

[0158] Through this iterative training process, the computer system can continuously optimize the performance of the monitoring data recognition network, enabling it to more accurately extract valuable feature information from the original monitoring data and perform effective inference analysis in practical applications. This not only improves the intelligent level of drug rehabilitation monitoring but also provides more reliable data support for the formulation of subsequent rehabilitation intervention strategies.

[0159] As an implementation manner, in step S1063, according to the confidence distribution of each object monitoring sequence sample corresponding to each preset withdrawal symptom classification, divide each object monitoring sequence sample to obtain a set of object monitoring sequence samples corresponding to each preset withdrawal symptom classification, which may specifically include: Step S10631: According to the confidence distribution of each object monitoring sequence sample corresponding to each preset withdrawal symptom classification, determine the target withdrawal symptom type to which the object monitoring sequence sample belongs in each preset withdrawal symptom classification; Step S10632: Classify each object monitoring sequence sample according to the target withdrawal symptom type to which each object monitoring sequence sample belongs, and obtain a set of object monitoring sequence samples corresponding to each preset withdrawal symptom classification.

[0160] In the embodiment of the present application, the computer system determines the target withdrawal symptom type to which each sample belongs according to the confidence distribution of each object monitoring sequence sample corresponding to each preset withdrawal symptom classification. This process is the basis for subsequent classification, ensuring that the samples can be accurately assigned to the classification that best matches their actual status.

[0161] The confidence distribution is calculated in step S103, which represents the probability or score that an object monitoring sequence example belongs to different preset withdrawal symptom classifications. This distribution is usually a vector, and each element of the vector corresponds to the confidence value of a classification. For example, assume there are three preset withdrawal symptom classifications: "emotional disorder", "sleep disorder", and "physical pain". Then, the confidence distribution of an object monitoring sequence example may be [0.7, 0.2, 0.1], indicating that the confidence of this example belonging to the "emotional disorder" classification is 0.7, and the confidences belonging to the "sleep disorder" and "physical pain" classifications are 0.2 and 0.1 respectively.

[0162] Methods for determining the target withdrawal symptom type: When determining the target withdrawal symptom type, the computer system usually adopts one of the following methods: Maximum confidence method: Select the classification with the highest confidence value in the confidence distribution as the target withdrawal symptom type. This is the simplest and most commonly used method, which assumes that the classification with the highest confidence is most likely to represent the actual state of the example. Continuing with the above example, the target withdrawal symptom type of the example will be determined as "emotional disorder".

[0163] Threshold method: Set one or more thresholds. Only when the confidence value of a certain classification exceeds these thresholds, is this classification regarded as a candidate target type. Then, select the one with the highest confidence from these candidate types as the target type. This method can increase the strictness of classification and avoid assigning examples to classifications with lower confidence. For example, the threshold can be set to 0.5, and only the confidence of the "emotional disorder" classification exceeds this threshold to become a candidate target type.

[0164] Comprehensive evaluation method: Consider the overall characteristics of the confidence distribution (such as entropy, variance, etc.), as well as other relevant information of the example (such as anomalies in monitoring data, preliminary diagnosis by doctors, etc.), and determine the target withdrawal symptom type through comprehensive evaluation. This method is more complex but may also be more accurate.

[0165] After determining the target withdrawal symptom type of each object monitoring sequence example, the computer system then needs to classify these examples according to the target type to obtain the set of object monitoring sequence examples corresponding to each preset withdrawal symptom classification. This step is the basis for the subsequent optimization of the graph convolutional network, which ensures that the optimization process can be carried out according to the characteristics and requirements of each classification.

[0166] Classification process: The classification process is relatively simple and intuitive. The computer system only needs to traverse all object monitoring sequence examples and assign them to the corresponding sets according to their target withdrawal symptom types. Each set contains all examples belonging to the same classification, and these examples will be treated as a whole in the subsequent graph convolutional network optimization process.

[0167] Representation of sets: Each set of object monitoring sequence examples can be represented in various ways, depending on the requirements of subsequent processing. A common representation is to use a list or an array, where each element is a data structure of an object monitoring sequence example (which may contain information such as raw monitoring data, processed feature vectors, actual withdrawal symptom labels, etc.). Another representation is to use a dictionary or a hash table, where the key is the classification name or classification identifier, and the value is a list or an array of object monitoring sequence examples under the corresponding classification.

[0168] Example illustration: Suppose there are 100 object monitoring sequence examples in the debugging example set. After the processing in step S10631, the target withdrawal symptom types of each example are determined. Now, the computer system classifies these examples according to the types. Suppose there are three classifications: "Emotional disorder", "Sleep disorder", and "Physical pain". Then the classification results may be as follows (taking the dictionary representation as an example): classification_results = { "Emotional disorder": [example 1, example 5,..., example 80], # List of examples belonging to the "Emotional disorder" classification; "Sleep disorder": [example 3, example 12,..., example 95], # List of examples belonging to the "Sleep disorder" classification; "Physical pain": [example 2, example 7,..., example 45]# List of examples belonging to the "Physical pain" classification;} In this example, classification_results is a dictionary that contains three keys (i.e., classification names) and the corresponding values (i.e., lists of examples belonging to that classification). In this way, the computer system can easily obtain all the examples under the corresponding classification according to the classification name, and then perform targeted optimization on the graph convolutional network. Through the implementation of step S1063, the computer system has successfully classified the object monitoring sequence examples according to their target withdrawal symptom types, and obtained the sets of object monitoring sequence examples corresponding to each preset withdrawal symptom classification. These sets provide basic data support for the subsequent optimization of the graph convolutional network, ensuring that the optimization process can be carried out according to the characteristics and requirements of each classification, thereby improving the accuracy and reliability of the entire monitoring data recognition network.

[0169] As an implementation manner, in step S1064, according to the actual withdrawal symptom markers of the object monitoring sequence examples in the object monitoring sequence example set corresponding to the preset withdrawal symptom classification, the involvement information of each alternative withdrawal symptom marker in the graph convolutional network corresponding to the preset withdrawal symptom classification is iterated to obtain the iterated graph convolutional network corresponding to the preset withdrawal symptom classification. Specifically, it may include: Step S10641: According to the actual withdrawal symptom markers of the object monitoring sequence examples in the object monitoring sequence example set corresponding to the preset withdrawal symptom classification, in the object monitoring sequence example set corresponding to the preset withdrawal symptom classification, determine the object monitoring sequence example subsets of each alternative withdrawal symptom marker. Each object monitoring sequence example subset of an alternative withdrawal symptom marker includes the object monitoring sequence examples corresponding to the corresponding alternative withdrawal symptom marker; Step S10642: For each alternative withdrawal symptom marker, perform a quantitative evaluation on the object monitoring sequence examples in the object monitoring sequence example subset of the alternative withdrawal symptom marker, and determine the target confidence that the object monitoring sequence example includes the remaining alternative withdrawal symptom markers when the object monitoring sequence example includes the alternative withdrawal symptom marker under the preset withdrawal symptom classification. The target confidence is used to indicate the involvement information between the alternative withdrawal symptom marker and the remaining alternative withdrawal symptom markers; Step S10643: According to the target confidence, iterate the involvement information of each alternative withdrawal symptom marker in the graph convolutional network corresponding to the preset withdrawal symptom classification to obtain the iterated graph convolutional network corresponding to the preset withdrawal symptom classification.

[0170] In step S10641, the computer system needs to determine the object monitoring sequence example subsets of each alternative withdrawal symptom marker according to the actual withdrawal symptom markers of the examples in the object monitoring sequence example set corresponding to the preset withdrawal symptom classification. The purpose of this step is to subdivide the original example set according to the alternative withdrawal symptom markers so as to perform independent quantitative evaluation on each marker subsequently.

[0171] Understanding of the actual withdrawal symptom markers: The actual withdrawal symptom markers are comprehensively judged by professional doctors or rehabilitation teams based on the monitoring data and clinical manifestations in the object monitoring sequence examples, and they accurately point out the main withdrawal symptoms existing in the examples during the monitoring period. These markers are discrete text or symbol representations, such as "anxiety", "depression", "insomnia", etc.

[0172] Method for determining example subsets: For each preset withdrawal symptom classification, the computer system traverses each example in the set of example object monitoring sequences corresponding to this classification to check whether a specific alternative withdrawal symptom marker is included in its actual withdrawal symptom markers. If it is included, this example is added to the subset of the example object monitoring sequences of this alternative withdrawal symptom marker. In this way, for each alternative withdrawal symptom marker, a subset containing all relevant examples can be obtained.

[0173] After determining the subsets of the example object monitoring sequences of each alternative withdrawal symptom marker, the computer system then needs to quantitatively evaluate the examples in each subset to determine the target confidence level that when an example of the object monitoring sequence contains a certain alternative withdrawal symptom marker, it also contains the remaining alternative withdrawal symptom markers under the preset withdrawal symptom classification. This step is the key process for extracting and quantifying the involved information.

[0174] Methods for quantitative evaluation: Quantitative evaluation usually involves methods such as statistical analysis, probability calculation, or machine learning model prediction. In practical applications, the computer system can select a suitable method according to the characteristics and requirements of the example subset. The following are some possible methods for quantitative evaluation: Co-occurrence frequency statistics: Calculate the co-occurrence frequency of each alternative withdrawal symptom marker and the remaining markers in the example subset. The higher the co-occurrence frequency, the stronger the involved information between the two markers. For example, the proportion of examples containing the "depression" marker in the "anxiety" example subset can be statistically calculated as the target confidence level between "anxiety" and "depression".

[0175] Conditional probability calculation: Use the conditional probability formula to calculate the probability of the remaining markers appearing under the condition of a given alternative withdrawal symptom marker. This probability value can be used as the target confidence level to indicate the involved information between the markers. For example, P(depression|anxiety) can be calculated, that is, the conditional probability of "depression" appearing in the "anxiety" example subset.

[0176] Machine learning model prediction: Train a machine learning classifier or regression model to predict the occurrence probability or confidence level of the remaining markers under the condition of a given alternative withdrawal symptom marker. This method can capture the complex non-linear relationships between the markers and improve the accuracy of quantitative evaluation.

[0177] Understanding of the target confidence level: The target confidence level is the result of quantitative evaluation, which indicates the possibility or strength that when an example of the object monitoring sequence contains a certain alternative withdrawal symptom marker, it also contains the remaining alternative withdrawal symptom markers under the preset withdrawal symptom classification. This confidence level value is used to indicate the involved information between the markers and is the basis for subsequent iterative updating of the graph convolutional network.

[0178] After obtaining the target confidence levels among various alternative withdrawal symptom markers, the computer system then needs to iteratively update the involvement information of each alternative withdrawal symptom marker in the graph convolutional network corresponding to the preset withdrawal symptom classification based on these confidence levels. This step is a crucial link in optimizing the network structure and improving the network performance.

[0179] Understanding of the graph convolutional network: The graph convolutional network is a neural network model specifically designed for processing graph-structured data. In the graph convolutional network, each node represents an entity (such as an alternative withdrawal symptom marker), and each edge represents the relationship between entities (such as involvement information). The network updates the feature representation of each node by iteratively aggregating the feature information of neighboring nodes, thereby capturing the complex relationship patterns in the graph structure.

[0180] Iterative update method for involvement information: In a computer system, the iterative update of involvement information usually involves the following steps: Determine the update strategy: Determine the update strategy based on the target confidence levels and the characteristics of the graph convolutional network. The update strategy may include adjusting the weights of edges, adding or deleting edges, updating node features, etc.

[0181] Execute the update operation: Iteratively update the involvement information in the graph convolutional network according to the update strategy. For example, if the target confidence level indicates a strong involvement information between two markers, the weight of the edge between them can be increased; if the target confidence level is low, the edge can be considered for deletion or the weight can be reduced.

[0182] Verify the update effect: After each update, use the validation set (or a part of the debugging example set) to verify the update effect. The validation set is a data set independent of the training set and is used to evaluate the performance of the network on unseen data. If the performance of the updated graph convolutional network on the validation set has improved, it indicates that the update is effective; otherwise, the update strategy or parameters may need to be adjusted.

[0183] By iteratively executing steps S10641 to S10643 (i.e., determining the subset of examples, quantitatively evaluating the involvement information, and iteratively updating the graph convolutional network) multiple times, the computer system can gradually optimize the graph convolutional network structure corresponding to the preset withdrawal symptom classification, making it more accurately reflect the symptom marker association relationships in the actual data. This not only improves the performance of the network under a specific classification but also provides a more reliable basis for subsequent monitoring data identification and analysis.

[0184] As an implementation, in step S10642, for each alternative withdrawal symptom marker, a quantitative evaluation is performed on the object monitoring sequence sample subset of the alternative withdrawal symptom marker to determine the target confidence that when the object monitoring sequence sample contains the alternative withdrawal symptom marker under the preset withdrawal symptom classification, the object monitoring sequence sample also includes the remaining alternative withdrawal symptom markers, including: Step S106421: For each alternative withdrawal symptom marker, determine the first number of the object monitoring sequence sample in the object monitoring sequence sample subset of the alternative withdrawal symptom marker; Step S106422: In the object monitoring sequence sample subset of the alternative withdrawal symptom marker, determine the second number of the object monitoring sequence sample that also includes the remaining alternative withdrawal symptom markers; Step S106423: According to the first number and the second number, determine the target confidence that when the object monitoring sequence sample contains the alternative withdrawal symptom marker under the preset withdrawal symptom classification, the object monitoring sequence sample also includes the remaining alternative withdrawal symptom markers.

[0185] In step S106421, the computer system traverses the object monitoring sequence sample subset of each alternative withdrawal symptom marker to determine the total number of object monitoring sequence samples in the subset.

[0186] After determining the total number of samples in the object monitoring sequence sample subset, the computer system then needs to count the number of samples that also include the remaining alternative withdrawal symptom markers in the subset. This process requires traversing each sample in the subset, checking whether its actual withdrawal symptom markers also include other markers except the current alternative withdrawal symptom marker, and counting the samples that meet the conditions.

[0187] After obtaining the total number of samples in the object monitoring sequence sample subset and the number of samples that also include the remaining alternative withdrawal symptom markers, the computer system can calculate the target confidence based on these two numbers. The target confidence represents the possibility or intensity that when the object monitoring sequence sample contains a certain alternative withdrawal symptom marker under the preset withdrawal symptom classification, it also includes the remaining alternative withdrawal symptom markers.

[0188] Calculation method: The calculation of the target confidence usually takes the form of a ratio or conditional probability. In this scenario, an intuitive and simple method is to use the ratio of the number of co-occurring samples to the total number of samples as the target confidence. That is, for any two alternative withdrawal symptom markers A and B, under the preset withdrawal symptom classification C, the target confidence between A and B can be calculated as: ; Example illustration: Take the co-occurrence relationship between "anxiety" and "depression" as an example. Under the classification of "emotional disorder", the subset of object monitoring sequence examples of "anxiety" contains 50 examples (the first number), and among them, 25 examples also contain the "depression" label (the second number). Therefore, the target confidence between "anxiety" and "depression" can be calculated as: ; This means that under the classification of "emotional disorder", when the object monitoring sequence example contains the "anxiety" label, there is a 50% probability that it also contains the "depression" label.

[0189] Similarly, the computer system can also calculate the target confidence between "anxiety" and all other alternative withdrawal symptom labels, so as to obtain a complete confidence matrix or table, which is used to represent the association relationship between alternative withdrawal symptom labels under the classification of "emotional disorder".

[0190] By quantitatively evaluating the alternative withdrawal symptom labels and calculating the target confidence, the computer system can more accurately capture and represent the implicated information between symptom labels. This information is crucial for the subsequent iterative optimization of the graph convolutional network. Specifically, they can be used as the basis for updating the edge weights in the network structure, so that the network can pay more attention to those label pairs with higher confidence when propagating feature information, thereby generating more accurate and meaningful feature representations.

[0191] In addition, the results of quantitative evaluation can also provide valuable references for drug rehabilitation and correction work. For example, doctors or rehabilitation teams can formulate more personalized and targeted correction and intervention strategies based on the association relationship between labels in the confidence matrix. If there is a strong association between the main withdrawal symptom label of a certain patient and multiple other labels, then special attention needs to be paid to the occurrence and development of these associated symptoms during the treatment process, so as to take corresponding measures for intervention in a timely manner.

[0192] As an implementation manner, step S101, obtaining the debugging example set and the graph convolutional network corresponding to each preset withdrawal symptom classification, may specifically include: Step S1011: Obtain the debugging example set, and determine the object monitoring sequence example sets of each alternative withdrawal symptom label in the debugging example set according to the actual withdrawal symptom labels of the object monitoring sequence examples in the debugging example set. The object monitoring sequence example set of each alternative withdrawal symptom label includes the object monitoring sequence examples of the corresponding alternative withdrawal symptom label in the debugging example set; Step S1012: For each alternative withdrawal symptom marker, perform a quantitative evaluation on the object monitoring sequence examples in the set of object monitoring sequence examples of the alternative withdrawal symptom marker to determine the confidence that when the object monitoring sequence example contains the alternative withdrawal symptom marker, the object monitoring sequence example also includes the remaining alternative withdrawal symptom markers. The confidence represents the implicated information between the alternative withdrawal symptom marker and the remaining alternative withdrawal symptom markers; Step S1013: According to the confidence, set the implicated information between the marker vertices corresponding to the alternative withdrawal symptom markers in the graph convolutional network of each preset withdrawal symptom classification, and obtain the graph convolutional network corresponding to each preset withdrawal symptom classification after setting.

[0193] In step S1011, the computer system first obtains a set of debugging example sets. This set of example sets is carefully selected and contains the monitoring sequence examples of multiple objects (i.e., drug addicts undergoing rehabilitation) and the actual withdrawal symptom markers corresponding to these examples. Each monitoring sequence example is a series of monitoring data points arranged in chronological order, and these data points may cover multiple dimensions such as physiological indicators (such as heart rate, blood pressure), psychological assessment results (such as emotional state scores), and behavior records (such as daily activity levels). The actual withdrawal symptom markers are comprehensively judged by professional doctors or rehabilitation teams based on the monitoring results and clinical manifestations, and they accurately indicate the main withdrawal symptoms present in the object during the monitoring period. After obtaining the debugging example set, the computer system then needs to determine the set of object monitoring sequence examples of each alternative withdrawal symptom marker according to the actual withdrawal symptom markers in the examples. The purpose of this step is to subdivide the original example set according to the alternative withdrawal symptom markers for subsequent independent quantitative evaluation of each marker.

[0194] Through such a subdivision operation, the computer system prepares a dedicated set of object monitoring sequence examples for each alternative withdrawal symptom marker. These sets provide the basic data support for subsequent quantitative evaluation and graph convolutional network initialization.

[0195] In step S1012, the computer system performs a quantitative evaluation on the examples in the set of object monitoring sequence examples of each alternative withdrawal symptom marker to determine the confidence that when the object monitoring sequence example contains a certain alternative withdrawal symptom marker, it also includes the remaining alternative withdrawal symptom markers. This confidence represents the implicated information between the alternative withdrawal symptom markers, that is, the frequency or intensity of their co-occurrence in the same monitoring sequence example.

[0196] The process of quantitative evaluation usually involves methods such as statistical analysis, probability calculation, or machine learning model prediction. In practical applications, the computer system can select appropriate methods according to the characteristics and requirements of the example set. The following are some possible quantitative evaluation methods: Co-occurrence frequency statistics: Calculate the co-occurrence frequency of each alternative withdrawal symptom label with the remaining labels in the sample set. The higher the co-occurrence frequency, the stronger the involvement information between the two labels. For example, the proportion of samples containing the "depression" label in the "anxiety" sample set can be counted as the co-occurrence confidence between "anxiety" and "depression".

[0197] Conditional probability calculation: Use the conditional probability formula to calculate the probability of the remaining labels appearing given a certain alternative withdrawal symptom label. This probability value can be used as the co-occurrence confidence to indicate the involvement information between the labels. For example, P(depression|anxiety) can be calculated, that is, the conditional probability of "depression" appearing in the "anxiety" sample set.

[0198] Machine learning model prediction: Train a machine learning classifier or regression model to predict the probability or confidence of the remaining labels appearing given a certain alternative withdrawal symptom label. This method can capture the complex non-linear relationships between the labels and improve the accuracy of quantitative evaluation.

[0199] Regardless of which quantitative evaluation method is adopted, the goal of the computer system is to calculate the confidence that when the object monitoring sequence sample contains a certain alternative withdrawal symptom label, it also contains the remaining alternative withdrawal symptom labels. This confidence value will be used as the basis for setting the involvement information between the label vertices in the subsequent graph convolutional network.

[0200] In step S1013, the computer system sets the involvement information between the label vertices corresponding to the alternative withdrawal symptom labels in the graph convolutional network of each preset withdrawal symptom classification according to the confidence value calculated in step S1012. This step is a key link in the initialization of the graph convolutional network, which ensures that the network can accurately reflect the symptom label association relationship in the actual data.

[0201] The graph convolutional network is a neural network model specifically used to process graph-structured data. In the graph convolutional network, each node represents an entity (such as an alternative withdrawal symptom label), and each edge represents the relationship between entities (such as involvement information). The weight of the edge is usually used to represent the strength or confidence of the relationship. Therefore, when initializing the graph convolutional network, the weights of the edges need to be set according to the association relationship in the actual data.

[0202] The same operation will also be applied to all other pairs of alternative withdrawal symptom labels. The computer system traverses all label pairs and sets the weights of the corresponding edges in the graph convolutional network according to their co-occurrence confidence. Finally, the graph convolutional network corresponding to each preset withdrawal symptom classification will be initialized into a graph structure model containing rich involvement information, providing a solid foundation for subsequent training and optimization.

[0203] Through the implementation of step S101, the computer system successfully obtained the debugging example set and initialized the graph convolutional network corresponding to each preset withdrawal symptom classification. These preparatory works provided the necessary data support and network structure foundation for the subsequent training process, ensuring that the entire monitoring data recognition system could operate accurately and efficiently. At the same time, by quantifying and evaluating the involvement information between symptom markers and integrating it into the graph convolutional network, the computer system further improved the intelligence level and robustness of the system, providing more reliable technical support for the monitoring work of drug rehabilitation and correction.

[0204] As an implementation manner, in step S1012, for each alternative withdrawal symptom marker, the object monitoring sequence examples in the set of object monitoring sequence examples of the alternative withdrawal symptom marker are quantitatively evaluated to determine the confidence that when the object monitoring sequence example contains the alternative withdrawal symptom marker, the object monitoring sequence example also includes the remaining alternative withdrawal symptom markers. Specifically, it may include: Step S10121: For each alternative withdrawal symptom marker, determine the first number of the object monitoring sequence examples in the set of object monitoring sequence examples of the alternative withdrawal symptom marker; Step S10122: In the set of object monitoring sequence examples of the alternative withdrawal symptom marker, determine the second number of the object monitoring sequence examples that also include the remaining alternative withdrawal symptom markers; Step S10123: According to the first number and the second number, determine the confidence that when the object monitoring sequence example contains the alternative withdrawal symptom marker, the object monitoring sequence example also includes the remaining alternative withdrawal symptom markers.

[0205] In step S10121, the computer system first needs to traverse the set of object monitoring sequence examples of each alternative withdrawal symptom marker to determine the total number of object monitoring sequence examples in this set. This total number represents the overall scale of the examples containing a specific alternative withdrawal symptom marker in this debugging example set and is the basis for calculating the co-occurrence confidence later.

[0206] In step S10122, in the set of object monitoring sequence examples of the determined alternative withdrawal symptom marker, the computer system further screens out the examples that also contain other alternative withdrawal symptom markers and counts the number of these examples. This process is the core of the quantitative evaluation and directly determines the calculation result of the subsequent confidence.

[0207] The key to this step lies in traversing and screening. The computer system efficiently traverses each example in the example set and accurately identifies the examples that contain multiple markers at the same time. This requires the system to have strong data processing capabilities and efficient algorithm support.

[0208] In step S10123, the computer system will calculate the confidence that when an object monitoring sequence example contains a specific alternative withdrawal symptom marker, it also contains other alternative withdrawal symptom markers, based on the two numbers obtained in steps S10121 and S10122 (i.e., the total number of examples and the number of examples that simultaneously contain the remaining alternative withdrawal symptom markers). This confidence reflects the association strength and co-occurrence frequency between the markers, and is an important basis for setting the edge weights during the initialization of the graph convolutional network.

[0209] Calculation method: The calculation of co-occurrence confidence usually takes the form of a ratio. Specifically, it is to divide the number of examples that simultaneously contain two markers by the total number of examples that contain a specific marker. That is: ; where A represents the alternative withdrawal symptom marker currently being processed, and B represents any other alternative withdrawal symptom marker.

[0210] Example illustration: Take the co-occurrence relationship between "anxiety" and "depression" as an example. In the set of object monitoring sequence examples of "anxiety", the total number of examples is 50 (the first number), and the number of examples that simultaneously contain "anxiety" and "depression" is 25 (the second number). Therefore, the co-occurrence confidence between "anxiety" and "depression" can be calculated as: ; This means that in the monitoring sequence examples containing the "anxiety" marker, there is a 50% possibility of also containing the "depression" marker.

[0211] The same calculation method also applies to the calculation of the co-occurrence confidence between "anxiety" and all other alternative withdrawal symptom markers. By traversing all possible combinations of marker pairs and calculating their co-occurrence confidence, the computer system can obtain a complete confidence matrix or table to represent the association relationship between alternative withdrawal symptom markers under a specific classification.

[0212] Quantitatively evaluating the association relationship between alternative withdrawal symptom markers not only provides important data support for the initialization of the graph convolutional network, but also lays a solid foundation for subsequent monitoring data identification and analysis work. Specifically, the significance and application of quantitative evaluation are reflected in the following aspects: Improve network accuracy: By using the confidence obtained from quantitative evaluation as the basis for edge weights in the graph convolutional network, the network can pay more attention to those marker pairs with higher association strength when propagating feature information. In this way, the feature representations generated by the network will be more accurate and meaningful, thereby improving the accuracy of the entire monitoring data identification system.

[0213] Enhancing network robustness: The quantitative evaluation takes into account the co-occurrence frequency and association strength between different markers, enabling the network to operate more robustly in the face of complex and changing monitoring data. Even if some monitoring data is missing or contains outliers, the network can still make effective inferences and compensations through the information of other associated markers.

[0214] Guiding rehabilitation intervention: The results of the quantitative evaluation can also provide valuable references for drug rehabilitation and correction work. Doctors or rehabilitation teams can formulate more personalized and targeted rehabilitation intervention strategies based on the association relationships between markers in the confidence matrix. For example, if there are strong association relationships between the main withdrawal symptom markers of a patient and multiple other markers, then special attention needs to be paid to the occurrence and development of these associated symptoms during the treatment process in order to take corresponding intervention measures in a timely manner.

[0215] Optimizing resource allocation: The quantitative evaluation can also help drug rehabilitation institutions optimize resource allocation. By understanding the association relationships and co-occurrence frequencies between different symptom markers, the institution can arrange resources such as medical staff, treatment equipment, and rehabilitation facilities more reasonably to ensure that patients can receive timely and effective treatment and support.

[0216] As an implementation method, in step S1061, according to the inference confidence and the actual withdrawal symptom markers of the object monitoring sequence example, determine the debugging error of the monitoring data recognition network, which can specifically include: Step S10611: Determine the marker debugging error according to the inference confidence and the actual withdrawal symptom markers of the object monitoring sequence example; Step S10612: Determine the classification confusion degree of the object monitoring sequence example according to the confidence distribution of each preset withdrawal symptom classification corresponding to the object monitoring sequence example; Step S10613: Determine the target withdrawal symptom type to which the object monitoring sequence example belongs according to the confidence distribution of each preset withdrawal symptom classification corresponding to the object monitoring sequence example, and determine the number of examples corresponding to each preset withdrawal symptom classification according to the target withdrawal symptom types to which each object monitoring sequence example belongs; Step S10614: Determine the classification debugging error according to the classification confusion degree of each object monitoring sequence example and the number of examples corresponding to each preset withdrawal symptom classification; Step S10615: Determine the debugging error of the monitoring data recognition network according to the marker debugging error and the classification debugging error.

[0217] In step S10611, the computer system determines the labeling debugging error based on the inference confidence of each object monitoring sequence example and the actual withdrawal symptom label. The labeling debugging error is an important indicator to measure the prediction accuracy of the network on a single example, which reflects the deviation degree between the network prediction result and the actual result.

[0218] For each example, the computer system compares its inference confidence distribution with the actual withdrawal symptom label. The inference confidence distribution is calculated in step S105, which represents the probability or score of the example belonging to different preset withdrawal symptom classifications; while the actual withdrawal symptom label is comprehensively judged by professional doctors or rehabilitation teams based on clinical manifestations and monitoring data, representing the true state of the example.

[0219] To quantify this difference, the computer system usually adopts the cross-entropy loss function or other suitable classification loss functions to calculate the labeling debugging error. The cross-entropy loss function can measure the difference between two probability distributions. When the inference confidence distribution is closer to the actual label distribution, the cross-entropy loss is smaller, indicating that the network's prediction on this example is more accurate.

[0220] In step S10612, the computer system evaluates the classification performance of the object monitoring sequence example set as a whole, especially the distribution of examples among different preset withdrawal symptom classifications. The classification confusion degree (also called information entropy) is an indicator used to measure the complexity and confusion degree of this distribution.

[0221] Specific implementation method: First, the computer system determines the target withdrawal symptom type to which each example belongs according to the confidence distribution of each example corresponding to each preset withdrawal symptom classification. This is usually achieved by selecting the classification with the highest confidence (such as using the maximum confidence method). Then, the system calculates the number of examples included in each classification to understand the distribution of the example set in each category.

[0222] Next, the system uses the concept of information entropy to calculate the classification confusion degree. Information entropy is a basic concept in information theory, used to describe the uncertainty or amount of information of a random variable. In a classification problem, the category information entropy reflects the complexity and confusion degree of the category distribution in the example set. The larger the information entropy, the more uniform the distribution of the example set in each category, and the lower the discrimination degree between categories; conversely, the smaller the information entropy, the more concentrated the example set in one or some categories, and the higher the discrimination degree between categories.

[0223] Although the content of step S10613 overlaps with that of step S10612 to a certain extent, here we focus more on explaining how to determine the target withdrawal symptom type of each example according to the confidence distribution and count the number of examples in each classification.

[0224] Specific implementation method: For each object monitoring sequence example, the computer system determines the target withdrawal symptom type according to the confidence distribution corresponding to each preset withdrawal symptom classification. This is usually achieved by selecting the classification with the highest confidence. Then, the system conducts classification statistics on all examples to obtain the number of examples included in each classification.

[0225] In step S10614, the computer system determines the classification debugging error according to the classification confusion degree (information entropy) and the number of examples corresponding to each preset withdrawal symptom classification. The classification debugging error is an error index that measures the overall classification performance of the network in the example set. It comprehensively considers the classification confusion degree of the example set and the example distribution under each category. The calculation of the classification debugging error usually involves weighted summation or other forms of combination of the classification confusion degree and the number of examples. Since the classification confusion degree has been quantitatively represented by the information entropy, the computer system can directly use this index to reflect the complexity and confusion degree of the example set classification. At the same time, the number of examples, as a weight factor, can affect the contribution degree of different classifications in the calculation of the classification debugging error. Specifically, the classification with a larger number of examples may have a greater impact on the classification debugging error because they represent the main part of the example set. It should be noted that in practical applications, the specific calculation method of the classification debugging error may vary depending on system design and requirements. Some systems can adopt more complex weighting strategies or introduce other performance indicators (such as accuracy, recall rate, F1 score, etc.) to comprehensively evaluate the classification performance.

[0226] Suppose the number of examples in the three classifications of "emotional disorder", "sleep disorder", and "physical pain" in the debugging example set are 30, 40, and 30 respectively, and the category information entropy is relatively high (indicating a large classification confusion degree). In this case, the computer system can calculate the classification debugging error according to the information entropy and the number of examples. Specifically, the system can assign a higher weight to the information entropy to reflect the complexity and confusion degree of the classification; at the same time, since the number of examples in the "sleep disorder" classification is the largest, the contribution degree of this classification in the calculation of the classification debugging error may also increase accordingly.

[0227] Finally, in step S10615, the computer system determines the debugging error of the monitoring data recognition network according to the labeled debugging error and the classification debugging error. The debugging error is an important index that measures the overall performance of the network. It comprehensively considers the prediction accuracy of the network on a single example and the performance on the overall classification of the example set.

[0228] The labeled debugging error and the classified debugging error are usually assigned different weights to reflect their importance in the calculation of the debugging error. For example, the labeled debugging error may focus more on measuring the prediction accuracy of the network on a single sample; while the classified debugging error focuses more on measuring the performance of the network in the overall classification performance of the sample set. The computer system calculates the sum or weighted sum of the debugging errors based on these weights as the final debugging error value. Suppose the labeled debugging error is 0.2 (indicating the average prediction error of the network on a single sample), and the classified debugging error is 0.3 (indicating the error of the network in the overall classification performance of the sample set). In this case, the computer system can calculate the sum or weighted sum of the debugging errors according to the preset weights (such as the weight of the labeled debugging error is 0.6 and the weight of the classified debugging error is 0.4). For example, if the weighted sum method is used to calculate the debugging error, the final debugging error value may be 0.2×0.6 + 0.3×0.4 = 0.24. This value reflects the overall performance of the monitoring data recognition network in the current debugging state, providing a basis for subsequent network parameter correction and optimization.

[0229] As an implementation manner, in step S102, based on the monitoring data recognition network, the characterization information extraction is respectively performed on the object monitoring sequence sample and each alternative withdrawal symptom label to obtain the overall sample characterization vector of the object monitoring sequence sample and the sample withdrawal symptom characterization vector of each alternative withdrawal symptom label, which may specifically include: Step S1021: Based on the monitoring data recognition network, perform a vector embedding operation on the object monitoring sequence sample to obtain the sample monitoring data embedding vector of the object monitoring sequence sample; Step S1022: Based on the monitoring data recognition network, perform a vector embedding operation on each alternative withdrawal symptom label respectively to obtain the sample withdrawal symptom embedding vector of each alternative withdrawal symptom label; and generate a characterization vector set according to the sample withdrawal symptom embedding vector of each alternative withdrawal symptom label and the adjustable prior feature vector; Step S1023: Perform a weight focusing operation on the sample monitoring data embedding vector and the characterization vector set to obtain the overall sample characterization vector of the object monitoring sequence sample and the sample withdrawal symptom characterization vector of each alternative withdrawal symptom label; Based on this, in step S106, the internal configuration parameters in the monitoring data recognition network are corrected according to the inference confidence and the actual withdrawal symptom markers of the object monitoring sequence examples, and a well - debugged monitoring data recognition network is obtained, including: correcting the internal configuration parameters and the prior feature vectors in the monitoring data recognition network according to the inference confidence and the actual withdrawal symptom markers of the object monitoring sequence examples, and obtaining a well - debugged monitoring data recognition network. In the intelligent process of drug rehabilitation monitoring, steps S102 and S106, as the core links of the training and debugging of the monitoring data recognition network, undertake the important task of extracting valuable feature information from the original monitoring data and continuously optimizing the network performance. The following is a detailed explanation of each step in the implementation methods of these two steps, illustrated with specific application scenarios and examples.

[0230] In step S1021, the computer system performs a vector embedding operation on the object monitoring sequence examples using the monitoring data recognition network. Vector embedding is a method of converting discrete data (such as text, symbols, etc.) into points in a continuous vector space. It can capture the similarities and differences between data, facilitating subsequent calculations and processing. For each object monitoring sequence example, the computer system arranges the monitoring data therein (such as heart rate, blood pressure, emotional state score, etc.) into a sequence according to the time order or specific rules. Then, the system uses the embedding layer in the monitoring data recognition network (such as Word2Vec, GloVe, or a custom embedding matrix) to convert each monitoring data point in the sequence into a vector of a fixed dimension. These vectors not only retain the feature information of the original monitoring data but also enable the similarities and differences between data to be quantitatively measured through the vectorized representation.

[0231] Suppose an object monitoring sequence example contains heart rate monitoring data for three consecutive days, and the heart rate values for each day are recorded as specific values at a series of time points. In step S1021, the computer system converts these heart rate values into vector representations. For example, the heart rate data for the first day may be converted into a 100 - dimensional vector, where each dimension represents the distribution or change trend of the heart rate value in different frequency bands. Through such a vector embedding operation, the system obtains an embedded vector of the example monitoring data that can comprehensively reflect the heart rate state of the object.

[0232] In step S1022, the computer system performs vector embedding operations on each alternative withdrawal symptom marker and generates a set of characterization vectors based on the embedded vectors and adjustable prior feature vectors. This step aims to convert discrete symptom markers into representations in a continuous vector space and enrich the information content of the characterization vectors by introducing prior knowledge. For each alternative withdrawal symptom marker (such as "anxiety", "depression", "insomnia", etc.), the computer system first uses an embedding layer to convert it into a vector of a fixed dimension (i.e., an example withdrawal symptom embedded vector). Then, the system constructs a set of adjustable prior feature vectors based on prior knowledge or the advice of domain experts. These prior feature vectors may represent certain inherent properties of the symptom markers (such as severity, duration, scope of influence, etc.) or the association relationships between other symptom markers. Finally, the system combines or concatenates the example withdrawal symptom embedded vectors with the prior feature vectors to generate a set of characterization vectors containing richer information.

[0233] Suppose the set of alternative withdrawal symptom markers contains two markers, "anxiety" and "depression". In step S1022, the computer system first converts them into vector representations respectively. Then, the system constructs a set of prior feature vectors based on prior knowledge, and these vectors may include dimensions describing symptom severity (such as mild, moderate, severe), dimensions describing symptom duration (such as short-term, medium-term, long-term), etc. Finally, the system concatenates the example withdrawal symptom embedded vectors of "anxiety" and "depression" with the corresponding prior feature vectors to generate two characterization vectors containing richer information. These characterization vectors not only contain the characteristic information of the symptom markers themselves but also incorporate relevant information from prior knowledge.

[0234] In step S1023, the computer system performs a weight focusing operation on the example monitoring data embedded vector and the set of characterization vectors to generate an example overall characterization vector of the object monitoring sequence example and an example withdrawal symptom characterization vector for each alternative withdrawal symptom marker. The weight focusing operation is a method of emphasizing important features and suppressing unimportant features by calculating weights, which can improve the information density and distinguishability of the characterization vectors.

[0235] The system uses an attention mechanism or other weighting strategies to calculate the weights between the example monitoring data embedded vector and each characterization vector in the set of characterization vectors. These weights reflect the association strength and importance between the monitoring data points and the symptom markers. Then, the system performs weighted summation or weighted concatenation operations on the characterization vectors according to these weights to generate an example overall characterization vector and an example withdrawal symptom characterization vector. The example overall characterization vector synthesizes all the monitoring data point information in the monitoring sequence example and emphasizes the features associated with the symptom markers; while the example withdrawal symptom characterization vector performs specialized feature extraction and weighting processing for each alternative withdrawal symptom marker.

[0236] Continuing with the previous example of the object monitoring sequence. In step S1023, the computer system first calculated the weights between the example monitoring data embedding vector and the representation vectors of "anxiety" and "depression". These weights may be derived based on the co-occurrence frequency, conditional probability, or other statistical metrics between the monitoring data points and the symptom markers. Then, the system performed a weighted sum operation on the representation vectors according to these weights to generate example withdrawal symptom representation vectors for "anxiety" and "depression". At the same time, the system also performed a similar weighting process on the example monitoring data embedding vector (possibly also considering the temporal relationship or dependency between different monitoring data points) to generate an example overall representation vector. These two vectors respectively represent the feature representations of the object monitoring sequence example under the overall and specific symptom markers.

[0237] Step S106 is the feedback and adjustment link in the training process of the monitoring data recognition network. It corrects the internal configuration parameters (such as weights, biases, etc.) and prior feature vectors of the network according to the difference between the inference confidence and the actual withdrawal symptom markers to improve the accuracy and robustness of the network.

[0238] In step S106, the computer system determines the debugging error of the network based on the inference confidence and the actual withdrawal symptom markers (such as by calculating through the cross-entropy loss function or other suitable classification loss functions). Then, the system backpropagates the gradient information according to the debugging error and uses an optimization algorithm (such as the gradient descent method, Adam optimizer, etc.) to update the internal configuration parameters and prior feature vectors of the network. This process is an iterative process that will be iterated multiple times on the entire debugging example set until the preset convergence condition is met or the preset number of iterations is reached.

[0239] For each object monitoring sequence example, the computer system compares the difference between its inference confidence distribution and the actual withdrawal symptom markers and calculates the corresponding marker debugging error (such as by calculating through the cross-entropy loss function). At the same time, the system also calculates the classification debugging error according to the overall classification performance of the example set (such as by calculating the weighted sum of the category information entropy and the number of examples). Finally, the system performs a weighted sum of the marker debugging error and the classification debugging error or adopts other combination methods to obtain the final debugging error value, and updates the network parameters and prior feature vectors according to this value.

[0240] Suppose that in a certain iteration cycle, the debugging error value calculated by the computer system is relatively high, indicating that there is still a large room for improvement in the prediction accuracy and classification performance of the current network. In this case, the system starts the backpropagation process to calculate the gradient information and uses an optimization algorithm to update the parameters such as the weights and biases of the network. At the same time, the system also adjusts the prior feature vectors according to the debugging error to make them more conform to the distribution and features of the actual data. By repeatedly iterating and executing such a parameter update process, the system can gradually reduce the debugging error value and improve the accuracy and robustness of the network. Finally, when the debugging error value reaches the preset convergence condition or the number of iterations reaches the preset value, the iteration process stops, and the debugged monitoring data recognition network is output for subsequent use.

[0241] An embodiment of the present application provides a computer system, as Figure 2 shown, the computer system 100 includes: a processor 101 and a memory 103. Among them, the processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may further include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation to the embodiments of the present application.

[0242] An embodiment of the present application provides a computer system. The computer system in the embodiment of the present application includes: one or more processors; a memory; one or more computer programs, where one or more computer programs are stored in the memory and are configured to be executed by one or more processors. When the one or more programs are executed by the processor, the above method is implemented.

Claims

1. An intelligent monitoring and intervention method for the rehabilitation and correction of drug addicts, characterized in that, Including: Obtaining a monitoring sequence of a target object and a plurality of alternative withdrawal symptom markers, and respectively extracting characterization information from the monitoring sequence of the target object and each alternative withdrawal symptom marker to obtain an overall characterization vector of the monitoring sequence of the target object and a withdrawal symptom characterization vector of each alternative withdrawal symptom marker; Inferring a confidence distribution of each preset withdrawal symptom classification corresponding to the monitoring sequence of the target object according to the overall characterization vector; Obtaining a graph convolutional network corresponding to each preset withdrawal symptom classification, where the graph convolutional network is used to indicate the involvement information between each alternative withdrawal symptom marker under the preset withdrawal symptom classification; Performing a graph context embedding operation on the withdrawal symptom characterization vectors of each alternative withdrawal symptom marker according to the confidence distribution and the graph convolutional network corresponding to each preset withdrawal symptom classification to obtain an operation-after withdrawal symptom characterization vector corresponding to each alternative withdrawal symptom marker; Determining an inference confidence of each alternative withdrawal symptom marker belonging to the monitoring sequence of the target object according to the operation-after withdrawal symptom characterization vector corresponding to each alternative withdrawal symptom marker; Determining a target withdrawal symptom marker corresponding to the monitoring sequence of the target object among each alternative withdrawal symptom marker according to the inference confidence; Analyzing whether to execute a correction intervention strategy based on the target withdrawal symptom marker.

2. The method according to claim 1, wherein The respectively extracting characterization information from the monitoring sequence of the target object and each alternative withdrawal symptom marker to obtain an overall characterization vector of the monitoring sequence of the target object and a withdrawal symptom characterization vector of each alternative withdrawal symptom marker includes: Performing a vector embedding operation on the object monitoring data of the monitoring sequence of the target object under not less than one type to obtain a monitoring data embedding vector of the monitoring sequence of the target object under not less than one type; Performing a vector embedding operation on each alternative withdrawal symptom marker respectively to obtain a withdrawal symptom embedding vector of each alternative withdrawal symptom marker; Generating a set of characterization vectors according to the withdrawal symptom embedding vectors of each alternative withdrawal symptom marker and a prior feature vector; Performing a weight focusing operation on the monitoring data embedding vector and the set of characterization vectors to obtain an overall characterization vector of the monitoring sequence of the target object and a withdrawal symptom characterization vector of each alternative withdrawal symptom marker.

3. The method according to claim 2, characterized in that, The performing a weight focusing operation on the monitoring data embedding vector and the set of characterization vectors to obtain an overall characterization vector of the monitoring sequence of the target object and a withdrawal symptom characterization vector of each alternative withdrawal symptom marker includes: Determining a monitoring data embedding vector of a target type to be currently processed among the monitoring data embedding vectors under the not less than one type; Performing a weight focusing operation on the monitoring data embedding vector of the target type and the set of characterization vectors to obtain a focused characterization vector of the monitoring sequence of the target object and a focused characterization vector of each alternative withdrawal symptom marker; Determining a new set of characterization vectors according to the focused characterization vector of the monitoring sequence of the target object and the focused characterization vectors of each alternative withdrawal symptom marker; Jump to the step of determining the monitoring data embedding vector of the current target type to be processed by the monitoring data embedding vector under the at least one type, and perform iterative execution. Stop when the monitoring data embedding vectors under all types are processed. According to the current new set of feature vectors, obtain the overall feature vector of the target object monitoring sequence and the withdrawal symptom feature vectors of each alternative withdrawal symptom label.

4. The method according to claim 1, wherein The graph convolutional network includes label vertices corresponding to each alternative withdrawal symptom label, and connection lines between the label vertices. The connection lines are used to indicate the involvement information between the two connected label vertices. The graph context embedding operation is performed on the withdrawal symptom feature vectors of each alternative withdrawal symptom label according to the confidence distribution and the graph convolutional network corresponding to each preset withdrawal symptom classification, to obtain the post-operation withdrawal symptom feature vectors corresponding to each alternative withdrawal symptom label, including: For each alternative withdrawal symptom label, determine the target label vertex corresponding to the alternative withdrawal symptom label in the graph convolutional network corresponding to each preset withdrawal symptom classification. For the graph convolutional network corresponding to each preset withdrawal symptom classification, determine the adjacent feature vector of the target label vertex under the preset withdrawal symptom classification according to the involvement information between the target label vertex and the remaining label vertices in the graph convolutional network. According to the confidence distribution, perform an integration operation on the adjacent feature vectors of the target label vertex under each preset withdrawal symptom classification to obtain the integrated adjacent feature vector of the alternative withdrawal symptom label. According to the integrated adjacent feature vector, perform an iterative operation on the withdrawal symptom feature vector of the alternative withdrawal symptom label to obtain the post-operation withdrawal symptom feature vector corresponding to the alternative withdrawal symptom label.

5. The method according to claim 4, wherein The confidence distribution includes the confidence of the target object monitoring sequence corresponding to each preset withdrawal symptom classification. The integration operation of the adjacent feature vectors of the target label vertex under each preset withdrawal symptom classification according to the confidence distribution to obtain the integrated adjacent feature vector of the alternative withdrawal symptom label includes: For each preset withdrawal symptom classification, perform an integration operation on the confidence that the target object monitoring sequence belongs to the preset withdrawal symptom classification and the adjacent feature vector of the target label vertex under the preset withdrawal symptom classification to obtain the initial integrated adjacent feature vector of the alternative withdrawal symptom label under the preset withdrawal symptom classification. Determine the integrated adjacent feature vector of the alternative withdrawal symptom label according to the initial integrated adjacent feature vectors of the alternative withdrawal symptom label under each preset withdrawal symptom classification.

6. The method according to claim 1, wherein Determining the inference confidence that each alternative withdrawal symptom label belongs to the target object monitoring sequence according to the post-operation withdrawal symptom feature vectors corresponding to each alternative withdrawal symptom label includes: For each alternative withdrawal symptom label, perform an integration operation on the withdrawal symptom feature vector and the post-operation withdrawal symptom feature vector of the alternative withdrawal symptom label to obtain the target withdrawal symptom feature vector of the alternative withdrawal symptom label. Determine the inference confidence of each alternative withdrawal symptom marker belonging to the target object monitoring sequence according to the target withdrawal symptom representation vector marked by each alternative withdrawal symptom; The step of respectively extracting the representation information of the target object monitoring sequence and each alternative withdrawal symptom marker to obtain the overall representation vector of the target object monitoring sequence and the withdrawal symptom representation vector of each alternative withdrawal symptom marker includes: Based on the monitoring data recognition network, respectively extract the representation information of the target object monitoring sequence and each alternative withdrawal symptom marker to obtain the overall representation vector of the target object monitoring sequence and the withdrawal symptom representation vector of each alternative withdrawal symptom marker.

7. The method according to claim 6, wherein Before the step of, based on the monitoring data recognition network, respectively extract the representation information of the target object monitoring sequence and each alternative withdrawal symptom marker to obtain the overall representation vector of the target object monitoring sequence and the withdrawal symptom representation vector of each alternative withdrawal symptom marker, further includes: Obtain a debugging example set and a graph convolutional network corresponding to each preset withdrawal symptom classification, where the debugging example set includes multiple object monitoring sequence examples and the actual withdrawal symptom markers of the object monitoring sequence examples; Based on the monitoring data recognition network, respectively extract the representation information of the object monitoring sequence examples and each alternative withdrawal symptom marker to obtain the overall example representation vector of the object monitoring sequence examples and the example withdrawal symptom representation vector of each alternative withdrawal symptom marker; According to the overall example representation vector, infer the confidence distribution of each object monitoring sequence example corresponding to each preset withdrawal symptom classification; According to the confidence distribution and the graph convolutional network corresponding to each preset withdrawal symptom classification, perform a graph context embedding operation on the example withdrawal symptom representation vector of each alternative withdrawal symptom marker to obtain the processed example withdrawal symptom representation vector corresponding to each alternative withdrawal symptom marker; According to the processed example withdrawal symptom representation vector corresponding to each alternative withdrawal symptom marker, determine the inference confidence of each alternative withdrawal symptom marker belonging to the object monitoring sequence example; According to the inference confidence and the actual withdrawal symptom marker of the object monitoring sequence example, correct the internal configuration parameters in the monitoring data recognition network to obtain a debugged monitoring data recognition network.

8. The method according to claim 7, wherein The step of, according to the inference confidence and the actual withdrawal symptom marker of the object monitoring sequence example, correct the internal configuration parameters in the monitoring data recognition network to obtain a debugged monitoring data recognition network, includes: According to the inference confidence and the actual withdrawal symptom marker of the object monitoring sequence example, determine the debugging error of the monitoring data recognition network; If the debugging error does not reach the preset debugging cut-off requirement, correct the internal configuration parameters in the monitoring data recognition network according to the debugging error; According to the confidence distribution of each object monitoring sequence example corresponding to each preset withdrawal symptom classification, divide each object monitoring sequence example to obtain an object monitoring sequence example set corresponding to each preset withdrawal symptom classification; For each preset withdrawal symptom classification, according to the actual withdrawal symptom labels of the object monitoring sequence examples in the set of object monitoring sequence examples corresponding to the preset withdrawal symptom classification, iterate the involvement information of each alternative withdrawal symptom label in the graph convolutional network corresponding to the preset withdrawal symptom classification to obtain the iterated graph convolutional network corresponding to the preset withdrawal symptom classification; According to the corrected monitoring data recognition network and the iterated graph convolutional networks corresponding to each preset withdrawal symptom classification, jump to the step of iteratively extracting the characterization information of the object monitoring sequence examples and each alternative withdrawal symptom label in the monitoring data recognition network based on the monitoring data recognition network, and stop the iteration when the debugging error of the monitoring data recognition network reaches the preset debugging cut-off requirement to obtain a well-debugged monitoring data recognition network.

9. The method according to claim 8, wherein The dividing of each object monitoring sequence example according to the confidence distribution of each object monitoring sequence example corresponding to each preset withdrawal symptom classification to obtain the set of object monitoring sequence examples corresponding to each preset withdrawal symptom classification includes: Determine the target withdrawal symptom type to which the object monitoring sequence example belongs in each preset withdrawal symptom classification according to the confidence distribution of the object monitoring sequence example corresponding to each preset withdrawal symptom classification; Classify each object monitoring sequence example according to the target withdrawal symptom type to which each object monitoring sequence example belongs to obtain the set of object monitoring sequence examples corresponding to each preset withdrawal symptom classification; The iterating of the involvement information of each alternative withdrawal symptom label in the graph convolutional network corresponding to the preset withdrawal symptom classification according to the actual withdrawal symptom labels of the object monitoring sequence examples in the set of object monitoring sequence examples corresponding to the preset withdrawal symptom classification to obtain the iterated graph convolutional network corresponding to the preset withdrawal symptom classification includes: According to the actual withdrawal symptom labels of the object monitoring sequence examples in the set of object monitoring sequence examples corresponding to the preset withdrawal symptom classification, determine, in the set of object monitoring sequence examples corresponding to the preset withdrawal symptom classification, the subset of object monitoring sequence examples of each alternative withdrawal symptom label, and the subset of object monitoring sequence examples of each alternative withdrawal symptom label includes the object monitoring sequence examples corresponding to the corresponding alternative withdrawal symptom label; For each alternative withdrawal symptom label, quantitatively evaluate the object monitoring sequence examples in the subset of object monitoring sequence examples of the alternative withdrawal symptom label to determine the target confidence level that when the object monitoring sequence example includes the alternative withdrawal symptom label under the preset withdrawal symptom classification, the object monitoring sequence example also includes the remaining alternative withdrawal symptom labels, and the target confidence level is used to indicate the involvement information between the alternative withdrawal symptom label and the remaining alternative withdrawal symptom labels; According to the target confidence level, iterate the involvement information of each alternative withdrawal symptom label in the graph convolutional network corresponding to the preset withdrawal symptom classification to obtain the iterated graph convolutional network corresponding to the preset withdrawal symptom classification.

10. A computer system, characterized in that, Including: One or more processors; A memory; One or more computer programs; Wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processor, the method according to any one of claims 1 to 9 is implemented.

Images