An information monitoring and circulation platform supporting a three-level management mode

Abstract

The application belongs to the field of information management, and particularly relates to an information monitoring and circulation platform supporting a three-level management mode, which is constructed into a three-level mode chain by a first subsystem, a second subsystem and a third subsystem based on a block chain algorithm; the first subsystem extracts target representation information of a user through dialect recognition and entity extraction, and after rule screening, the target representation information is classified and stored and a strategy is pushed, while a dynamic adjustment cycle collection is simultaneously performed; the second subsystem matches a disposal strategy according to a classified disposal node network and a quick index, performs hierarchical evaluation and disposal, and when a threshold is not reached, a data packet is constructed and submitted to the third subsystem; the third subsystem performs advanced disposal and evaluation, performs root cause analysis by fusing multi-source logs and abnormal information, and performs real-time adjustment on resources and strategies of the lower-level subsystems until an evaluation index is satisfied; the application realizes multi-level collaborative management, closed-loop disposal and system self-adaptive optimization, and improves the accuracy and reliability of information processing.

Description

Technical Field

[0001] This invention belongs to the field of information management, and in particular relates to an information monitoring and circulation platform that supports a three-level management model. Background Technology

[0002] Under the current tiered healthcare system, primary healthcare services generally suffer from inconsistent screening standards and traditional, inefficient data collection methods, resulting in low accuracy in identifying high-risk cases and difficulty in accurately delivering health education content based on individual characteristics. At the regional healthcare level, there are technical bottlenecks in integrating multi-source, heterogeneous medical data; the treatment process lacks information-based closed-loop management and reliable evidence storage mechanisms; resource allocation still relies heavily on manual judgment, and both response speed and allocation rationality need improvement. Meanwhile, tertiary hospitals, as the supporting central hubs, have not yet established a systematic empowerment and dynamic control mechanism for lower-level institutions, and lack the ability to continuously optimize based on full-chain data, leading to low overall efficiency in the utilization of medical resources and insufficient collaborative service capabilities. These problems hinder the improvement of the quality of primary healthcare services, such as pain management, and make it difficult to adapt to the growing demand for precise, tiered, and continuous healthcare services. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes an information monitoring and flow platform supporting a three-tiered management model. This platform comprises a first subsystem, a second subsystem, and a third subsystem, constructing a three-tiered model chain based on blockchain algorithms. The first subsystem extracts user target representation information through dialect recognition and entity extraction, filters it according to rules, classifies and stores it, and pushes strategies, while dynamically adjusting cyclical data collection. The second subsystem performs hierarchical evaluation and processing based on a classified processing node network and a fast index matching processing strategy. When thresholds are not met, a data packet is constructed and submitted to the third subsystem. The third subsystem performs escalated processing and evaluation, integrates multi-source logs and anomaly information for root cause analysis, and adjusts the resources and strategies of lower-level subsystems in real time until the evaluation indicators are met. This invention achieves multi-level collaborative management, closed-loop processing, and system adaptive optimization, improving the accuracy and reliability of information processing.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] An information monitoring and flow platform supporting a three-tier management model includes:

[0006] The three-level model chain is constructed by combining a first subsystem, a second subsystem, and a third subsystem with a blockchain algorithm. The first subsystem is used to collect different types of user target representation information and to filter and classify the user target representation information according to preset data filtering rules. At the same time, based on the filtered user target representation information and preset key filtering fields, a preset push strategy library is called to perform information push and fixed-point cyclic collection operations.

[0007] The second subsystem performs hierarchical classification cyclical evaluation and disposal based on the user target representation information stored in the classification system, combined with the preset hierarchical classification evaluation criteria and disposal strategy library; when the disposal result within the preset time does not reach the preset evaluation threshold, the corresponding user information and disposal strategy information are recorded.

[0008] The third subsystem executes a preset escalation assessment process based on the recorded user information and handling strategy information combined with the deep Q network; at the same time, it performs anomaly analysis based on the escalation assessment results and the log records collected in real time by the first and second subsystems; and adjusts the resources of the first and second subsystems according to the anomaly analysis results until the preset assessment indicator thresholds of each subsystem are met.

[0009] Specifically, the first subsystem includes a chief complaint identification module and an initial screening module;

[0010] The chief complaint recognition module is used to extract and correct user speech text information according to a preset dialect recognition model, and to identify and extract user target representation information based on the corrected user speech text information combined with an entity extraction algorithm to obtain a set of user target representation information keywords; the data filtering rules are constructed by combining an initial filtering score threshold with a filtering keyword library and a matching algorithm.

[0011] The initial screening module is used to evaluate user target representation information and screen keywords based on the acquired user target representation information keyword set and a comprehensive evaluation algorithm. When the initial evaluation score of the corresponding user target representation is greater than the preset initial screening score threshold or a screening keyword is matched, a first user data package is constructed based on the user basic information, user target representation information keyword set and initial evaluation score that meet the conditions. The first user data package is then hashed using a hash algorithm and uploaded to the distributed database corresponding to the second subsystem through the smart contract in the three-level model chain for classified storage.

[0012] Specifically, the first subsystem also includes an information push module, a cyclic collection and adjustment module, and a first analysis module;

[0013] The information push module is used to push first strategy information based on the user target representation information keyword set, the user status information after evaluation and processing corresponding to the second subsystem and the third subsystem, the final evaluation score and the evaluation score change trend, and in combination with the preset push strategy library, through a matching algorithm.

[0014] The cyclic collection and adjustment module is used to adjust the time interval and frequency of cyclic collection of target users in real time according to the round of evaluation and treatment of the corresponding user after evaluation and treatment, the evaluation score of each target characterization, the trend of evaluation score change and the screening keywords. It also obtains the cyclic screening evaluation result based on the cyclic collection information. If the cyclic screening evaluation result meets the data screening rules again, the corresponding user information is uploaded and saved again. If it does not meet the rules, the corresponding user information is deleted from the distributed database.

[0015] The first analysis module is used to record in real time the data collection, filtering, strategy push and cyclic collection process information corresponding to the first subsystem, and to perform horizontal anomaly analysis and location on the preset first subsystem chain in combination with the preset node anomaly analysis model to obtain the first anomaly feature information, and to feed back the first anomaly feature information to the preset third subsystem chain in the third subsystem through the three-level mode chain.

[0016] Specifically, the second subsystem includes a first processing module;

[0017] The first processing module is used to match the corresponding type processing strategy based on the user type and user target representation information keywords in the labeled first user data packet, combined with a preset classification processing node network and a preset fast index, and to process the corresponding user using the matching processing strategy. It also collects, records and preprocesses the user target representation information keywords within a preset time length after the current user is processed, and records the average processing rate of the current processing strategy within the preset time length. Each node in the classification processing node network corresponds to a processing strategy.

[0018] The preset fast index is constructed by combining the user type, skill level, index frequency, evaluation score decrease of real-time user target representation information after the corresponding handling strategy, the number of rounds and probability of re-evaluation after the current handling strategy, and the evaluation satisfaction rate after handling with a hash algorithm.

[0019] Specifically, the second subsystem also includes a first evaluation module, an upgrade discrimination module, and a second analysis module;

[0020] The first evaluation module is used to perform real-time evaluation based on the keywords of the user target representation information after treatment, combined with preset evaluation indicators and evaluation algorithms, to obtain the first evaluation score and the trend of the first evaluation score change of the real-time user target representation information, and at the same time, based on the trend of the first evaluation score change, to obtain the relief rate of the target user within a preset time period.

[0021] The escalation judgment module is used to perform escalation judgment based on the target user's remission rate within a preset time period and preset escalation judgment conditions. Specifically:

[0022] When the mitigation rate of the target user within a preset time period meets the preset escalation judgment condition, the first user data packet corresponding to the target user and the second user data packet constructed by combining the information of each round of handling strategy with the hash algorithm are fed back to the preset third subsystem chain in the third subsystem through the three-level mode chain.

[0023] If the relief rate of the target user within the preset time period does not meet the preset escalation criteria, the first satisfaction level of the corresponding target user with the current treatment strategy information is evaluated and fed back to the third subsystem and the first subsystem through the three-level mode chain.

[0024] Specifically, the second subsystem also includes a second analysis module; the second analysis module is used to perform anomaly analysis and localization on the handling strategies of corresponding users who meet the escalation judgment conditions within the corresponding preset time length in the classification and handling node network, obtain second anomaly feature information, and feed the second anomaly feature information back to the third subsystem through the three-level pattern chain.

[0025] Specifically, the third subsystem includes an upgraded handling matching module and an upgraded assessment module;

[0026] The escalation handling matching module is used to obtain the corresponding escalation handling strategy and matching accuracy by combining the second user data packet with the escalation handling strategy library configured by the third subsystem through a matching algorithm, and to collect the second user status information after the corresponding user is handled according to the escalation handling strategy in real time.

[0027] The upgrade evaluation module is used to obtain a second evaluation score after upgrade processing based on the second user status information and the second evaluation algorithm. When the second evaluation score is less than the preset processing completion evaluation threshold, the processing of the corresponding user ends. At the same time, based on the evaluation information of the corresponding user on the matched upgrade processing strategy during the upgrade processing, the second satisfaction rate of the current user is obtained, and the second satisfaction rate of the corresponding upgrade processing strategy is fed back to the upgrade processing strategy library to adjust the index information of the current upgrade processing strategy in real time.

[0028] When the second evaluation score is greater than or equal to the preset disposal completion evaluation threshold, the third subsystem presets the abnormal decision node in the first subsystem chain to perform cyclic matching evaluation of the escalation disposal strategy until the second evaluation score is less than the preset disposal completion evaluation threshold, and records the corresponding cycle number and the corresponding escalation disposal strategy information that does not meet the condition.

[0029] Specifically, the third subsystem also includes an anomaly analysis and decision-making module and an anomaly adjustment module;

[0030] The anomaly analysis and decision-making module is used to obtain anomaly feature association information and difference information by combining the first anomaly feature information and the second anomaly feature information with an association algorithm. Based on the anomaly feature association information and difference information, the average handling rate of the current handling strategy within a preset time length, the first satisfaction rate, and the relief rate of the target user within a preset time length, and combined with the anomaly decision-making nodes in the preset third subsystem chain of the third subsystem, an associated anomaly adjustment strategy is obtained.

[0031] The anomaly adjustment module is used to adjust the anomaly node information located in the first subsystem and the second subsystem in real time according to the associated anomaly adjustment strategy and the anomaly decision node, reinforcement adjustment node and historical anomaly adjustment index channel information in the first subsystem chain, with the goal of minimizing delay and maximizing anomaly adjustment rate and adjustment speed, until the preset evaluation index threshold of the corresponding anomaly node is met.

[0032] Specifically, the first subsystem executes the first-level management mode, the second subsystem executes the second-level management mode, and the third subsystem executes the third-level management mode;

[0033] The first-level management mode generates a first user data packet with hash label through the initial screening module and the chief complaint identification module, and automatically uploads it to the distributed database managed by the second-level management mode in conjunction with the blockchain smart contract;

[0034] The second-level management mode uses the first disposal module to match the disposal strategy for the first user data packet based on the classification disposal node network and fast index. After the first evaluation module and the escalation judgment module make judgments, the cases to be escalated are constructed into second user data packets and fed back to the third-level management mode through the same blockchain on-chain channel.

[0035] The third-level management mode uses the upgraded handling matching module to match and handle the second user data packets, and uses the anomaly analysis and decision module to integrate the real-time logs and anomaly feature information of the first and second subsystems to generate associated anomaly adjustment strategies. The anomaly adjustment module then uses historical index channels and enhanced adjustment nodes to perform real-time parameterized adjustments to the anomaly nodes in the first or second subsystems based on historical index channels and enhanced adjustment nodes, until they meet the preset evaluation index thresholds.

[0036] Specifically, the construction process of the classified disposal node network includes:

[0037] Based on the historical case database, an initial set of nodes is obtained by extracting all independent handling strategies and assigning a unique node identifier to each strategy. Each node stores the complete content, applicable conditions and execution parameters of the corresponding strategy.

[0038] Based on the logical correlation between strategies and the frequency of historical joint use, a directed graph model is used to construct the connection edges between nodes. Each edge is assigned an initial weight according to the historical treatment evaluation score, treatment success rate, conversion time length, and the number of rounds of reprocessing after treatment, thus obtaining a treatment strategy network topology with weighted relationships.

[0039] Based on a multi-dimensional feature index system, a fast retrieval entry point is built for each node through a hybrid index structure. The index key includes user type, target character keyword weight, urgency level of handling, required resource type, historical handling success rate and average handling rate within a preset time period, to obtain a strategy index that supports multi-condition combined queries.

[0040] Based on real-time feedback data after real-time processing, the node index weight and edge relationship weight are adjusted through simulation algorithm, and the failure mode is located by mining association rules to obtain the classified processing node network.

[0041] Based on the blockchain's evidence storage requirements, the structure hash and decision log of the node network are synchronized to the chain through smart contracts to obtain the handling strategy record. The handling strategy record is then uploaded to the abnormal decision node for real-time abnormal simulation decision adjustment until the corresponding abnormal adjustment is completed.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This invention addresses the shortcomings of existing technologies by constructing a three-tiered collaborative management model based on blockchain technology, achieving a breakthrough improvement in the entire process of medical data management. The first subsystem employs dialect recognition and multimodal processing technologies to effectively solve the problem of distorted chief complaint information caused by dialect differences in primary healthcare, ensuring the accuracy and completeness of information from the data source. The second subsystem, through the design of a classified treatment node network and a rapid indexing mechanism, achieves intelligent matching and dynamic optimization of treatment strategies, significantly improving the response speed and processing efficiency of primary healthcare services. The third subsystem utilizes advanced anomaly analysis algorithms and resource adjustment mechanisms to construct a cross-level closed-loop control system, enabling the system to have self-optimization and continuous improvement capabilities. The system corresponding to this application relies on blockchain technology to construct an immutable and fully traceable data trust system, and achieves automated execution of business processes through smart contracts. Ultimately, it has achieved significant results in optimizing the allocation of medical resources, improving the standardization of service quality, and enhancing the efficiency of multi-level collaboration, providing a complete technical solution for the digital transformation of the hierarchical medical system. Attached Figure Description

[0044] Figure 1 This is a module architecture diagram of an information monitoring and circulation platform supporting a three-level management mode according to the present invention;

[0045] Figure 2 This is a simplified diagram of the three-level pattern chain of the present invention;

[0046] Figure 3 This is a flowchart of the fixed-length hash value segmentation and implantation process of the present invention. Detailed Implementation

[0047] The current three-tiered healthcare system (village level, township / community level, and tertiary hospitals) faces the following core problems in pain patient management: dialect differences lead to large errors in extracting patient complaints and distorted pain representations; cross-level data transmission has low reliability and is at risk of tampering; primary care strategies lack precise matching and dynamic optimization, resulting in delayed responses at higher levels; and anomalies throughout the entire process are difficult to locate and adjust in a closed loop. Therefore, there is an urgent need to build a system adapted to this model to achieve accurate extraction of pain information in dialect environments, reliable data flow across systems, collaborative optimization between primary and higher levels, and full-process anomaly identification and closed-loop management.

[0048] Please see Figure 1 and Figure 2This invention provides an embodiment of an information monitoring and circulation platform supporting a three-tier management model, comprising: a three-tier model chain, which is constructed by combining a first subsystem, a second subsystem, and a third subsystem with a blockchain algorithm; the first subsystem executes a first-tier management model, the second subsystem executes a second-tier management model, and the third subsystem executes a third-tier management model; it should be further noted that the first-tier management model in this embodiment is a village-level model, where users screen target users, and the target representation information corresponds to publicity management and subsequent random visits; the second-tier management model is a township / community model, used for screening... The selected data is used for initial processing and process evaluation; the third-level management model is a tertiary hospital model, used to upgrade the treatment of target users who cannot be treated in townships / communities. At the same time, it controls and adjusts the management based on the anomalies or resource allocation problems of the first-level and second-level management models, provides training and guidance for the medical staff corresponding to the first-level and second-level management models, and conducts performance evaluations for the medical staff corresponding to the first-level and second-level management models. Based on the corresponding performance evaluations and the established reward mechanism, incentives are provided to ensure that the management models corresponding to the first-level and second-level management models are in an optimal state in real time.

[0049] The first subsystem is used to collect different types of user target representation information, and to filter and classify the user target representation information according to preset data filtering rules; at the same time, based on the filtered user target representation information and preset key filtering fields, it calls a preset push strategy library to perform information push and fixed-point cyclic collection operations.

[0050] It should be further explained that the different types of user classification in this embodiment are based on the established disease type database. It should also be explained that the user target representation information in this embodiment is the pain representation information corresponding to different types of diseases. Furthermore, it should be explained that the data filtering rules in this embodiment are constructed by combining an initial filtering score threshold with a filtering keyword database and a matching algorithm. Finally, it should be explained that the initial filtering score and the following evaluation score in this embodiment are both obtained through specific evaluation using the VAS (Visual Analogue Scale) index score corresponding to the pain level.

[0051] The second subsystem performs hierarchical classification cyclical evaluation and disposal based on the user target representation information stored in the classification system, combined with the preset hierarchical classification evaluation criteria and disposal strategy library; when the disposal result within the preset time does not reach the preset evaluation threshold, the corresponding user information and disposal strategy information are recorded.

[0052] The third subsystem executes a preset escalation assessment process based on recorded user information and handling strategy information, combined with a deep Q-network. Simultaneously, it performs anomaly analysis based on the escalation assessment results and real-time log data collected by the first and second subsystems. Based on the anomaly analysis results, it adjusts resources for the first and second subsystems until the preset assessment thresholds for each subsystem are met. In this embodiment, the construction and operation of the deep Q-network in the escalation assessment process are based on the following state, action, and reward settings to achieve adaptive decision-making: State Space S t The multidimensional feature vector extracted from the second user data packet at time t is defined as follows: It includes user type encoding (e.g., One-Hot encoding for postoperative pain and tumor pain), a word frequency-inverse document frequency weighted vector of the target representation information keyword set, an index sequence of historical k-round treatment strategies (converted to a dense vector through the embedding layer), the current real-time second assessment score (i.e., the normalized VAS score and its trend slope within a preset time window), the target user's relief rate within the preset time window (i.e., the percentage decrease in assessment score), and resource status variables from the real-time log information collected by the third subsystem, including the current workload rate of the corresponding department's doctors (i.e., current patient volume / maximum patient volume), the percentage of remaining inventory of drugs required for the corresponding strategy, and the idle rate of available nursing equipment; Action Space A t Defined as the set of identifiers for all candidate strategies in the escalation treatment strategy library, where each strategy identifier is associated with a specific expert team configuration, treatment plan combination, and expected resource consumption coefficient; reward function R t The design employs a multidimensional weighted sum, where the core reward is the decrease in the second assessment score after the escalation (i.e., the normalized difference between the VAS scores before and after the escalation) multiplied by a weight α; the user experience item is the second user satisfaction rate collected after this escalation (calculated in real-time through a satisfaction questionnaire) multiplied by a weight β; and the resource consumption item is the negative sum of the doctor's working hours, drug costs, and equipment occupancy time consumed in this strategy execution multiplied by a weight γ. Using a Q-learning framework, with the goal of minimizing the Bellman error, it utilizes the (S) stored in the experience replay pool... t A t R t S t+1 The quadruple is used to train the deep neural network and output the Q-value estimate for each action. When making a decision, the ε-greedy strategy is used to select the strategy identifier with the largest Q-value as the current escalation treatment plan, thereby realizing adaptive optimization matching of escalation strategies for complex pain cases.

[0053] It should be further explained that the first subsystem in this embodiment includes a chief complaint identification module, an initial screening module, an information push module, a cyclic collection and adjustment module, and a first analysis module;

[0054] The chief complaint recognition module is used to extract and correct user voice text information according to a preset dialect recognition model, and to identify and extract user target representation information based on the corrected user voice text information combined with an entity extraction algorithm to obtain a set of user target representation information keywords. It should be further noted that in this embodiment, the set of user target representation information keywords includes different disease types and pain description keywords under the corresponding disease types.

[0055] It should be further explained that the construction and training process of the dialect recognition model in this embodiment includes:

[0056] S1. Based on the original user speech sample library covering multiple dialect regions and the corresponding standard text annotation library, the original speech data is denoised by spectral subtraction through the speech signal preprocessing module, and frame-by-frame processing is performed by Hamming window weighting. The static and dynamic spectral features of the speech signal are extracted frame by frame using the Mel frequency cepstral coefficient algorithm to obtain a 39-dimensional MFCC feature vector.

[0057] S2. Based on the speech signal after frame processing, the autocorrelation algorithm is used to calculate the fundamental frequency value of each frame of speech, and the fundamental frequency trajectory curve is generated by linear interpolation to obtain the F0 feature sequence that reflects the tone pattern and tone value changes of different dialects.

[0058] S3. Based on the same speech sample, an energy-based endpoint detection algorithm is used to detect syllable boundaries and silent segments, and the number of syllables and the number and duration of pauses are counted per unit time to obtain a speech rate feature set containing syllable rate and pause interval.

[0059] S4. Based on the MFCC feature vector, F0 feature sequence and speech rate feature set, multi-dimensional alignment and splicing are performed through feature layer fusion algorithm to obtain a comprehensive multi-dimensional dialect feature set that integrates spectrum, tone and rhythm information.

[0060] S5. Based on the comprehensive multidimensional dialect feature set, the local spectral patterns are extracted by performing multi-layer convolution and pooling operations through a ResNet architecture convolutional neural network to obtain high-dimensional abstract spectral features.

[0061] S6. Based on the same comprehensive multidimensional dialect feature set, the temporal structure is recursively processed by a bidirectional LSTM network step by step to obtain the long-term dependency features of connected speech and sound change phenomena in the modeled dialect.

[0062] S7. Based on the high-dimensional abstract spectral features and long-term dependency features, the feature weight distribution is calculated through the Bahdanau attention mechanism and the key frame features are weighted and converged to obtain a context-aware feature representation that enhances dialect discriminative features.

[0063] S8. Based on labeled speech samples and their standard text, the initial dialect recognition model is iteratively trained using the stochastic gradient descent algorithm with the goal of minimizing the cross-entropy loss function, to obtain a preliminarily optimized dialect recognition model. It should be further noted that in this embodiment, during model training, dialect variant feature parameters collected from each village-level node are aggregated using a federated learning framework to dynamically update the acoustic model parameters of the dialect recognition model. The specific process includes:

[0064] S81. Based on the edge computing devices deployed at each village-level node, the collected dialect speech samples and corresponding dialect variant labels are stored locally. The dialect variant labels include at least dialect area classification, sub-dialect classification and special pronunciation feature annotation.

[0065] S82. Dialect variant feature extraction network trained locally extracts dialect variant feature parameters from dialect speech samples. The feature extraction network contains a combination structure of multi-layer convolutional neural network and attention mechanism, and the output is a fixed-dimensional feature vector.

[0066] S83. Each village-level node uses dialect variant feature parameters to fine-tune the acoustic model parameters of the initial dialect recognition model locally to obtain locally updated acoustic model parameters.

[0067] S84. Gaussian noise is added to the locally updated acoustic model parameters using differential privacy technology to generate the updated model parameters with noise.

[0068] S85. Each village-level node transmits the updated model parameters after adding noise to the federated learning aggregation server through a secure encrypted channel.

[0069] S86, the federated learning aggregation server uses a weighted average algorithm to aggregate the received model parameter updates from each node. The weights are dynamically calculated based on the number of dialect variant samples and data quality indicators of each node.

[0070] S87. Apply the aggregated model parameter update to the acoustic model parameters of the global dialect recognition model to complete one round of federated learning update.

[0071] S88. Distribute the updated global acoustic model parameters to each village-level node to replace the acoustic model parameters of the local model.

[0072] S89. Repeat steps S2 to S8 to continuously optimize the dialect recognition model's ability to recognize dialect variants.

[0073] The dialect variant feature parameters refer to the feature representations extracted from dialect speech samples that can characterize the acoustic properties of a specific dialect variant; the acoustic model parameters refer to the neural network parameters responsible for acoustic feature extraction and pattern recognition in the dialect recognition model.

[0074] S9. Based on the recognition results of the preliminary optimization model, a dual-array Trie tree is used to automatically search for and replace dialect words in a dialect-to-Mandarin dictionary, obtaining the first corrected text. It should be further explained that the process of using a dual-array Trie tree in this embodiment to automatically search for and replace dialect words in a dialect-to-Mandarin dictionary includes:

[0075] A dual-array Trie tree is constructed based on a dialect-to-Mandarin dictionary. This dictionary contains a dialect vocabulary set and its one-to-one corresponding Mandarin vocabulary set. The dialect vocabulary set includes, but is not limited to, words unique to each dialect region, words with pronunciation variations, and words with semantic shifts. By decomposing all words in the dialect vocabulary set into character units according to their character sequences, and using the ASCII code value of each character unit as the index key, a base array and a check array are constructed for the dual-array Trie tree. The base array stores the transition index value of each node, and the check array stores the parent node index of the corresponding node to verify the prefix validity, thus realizing the prefix tree structure storage of dialect vocabulary. Based on the recognition results text of the preliminary optimized model, character pointers are used to access the text... Starting from the initial position, consecutive character units are extracted as the sequence to be matched. The ASCII code value of the first character unit is used as the initial index to query the base array to obtain the index of the next hop node. The validity of the prefix of the current sequence is verified by checking the array. The character pointer is advanced by looping and iterating until no further matching is possible or the end of the text is reached to obtain the longest matching dialect word. The corresponding Mandarin word of the dialect word is retrieved from the reference dictionary through the mapping pointer stored in the terminal node of the double array Trie tree. The matching dialect word in the recognition result text is replaced with the corresponding Mandarin word. If there are multiple nested matching dialect words, the longest matching priority principle is used to perform the replacement to obtain the first corrected text.

[0076] S10. Based on the initial and final variation patterns defined in the preset pronunciation variation rule library, the systematic pronunciation errors in the first corrected text are corrected by a finite state transcription machine to obtain the second corrected text.

[0077] S11. Based on the second corrected text, identify semantically ambiguous field fragments by using the BiLSTM-CRF entity extraction algorithm based on BIO annotation to obtain a set of semantic units to be calibrated.

[0078] It should be further explained that the construction process of the preset pronunciation variation rule base in this embodiment includes a phonetic comparison study based on the seven major Chinese dialect areas and sub-dialects, including a set of initial consonant variation patterns and a set of final vowel variation patterns. The set of initial consonant variation patterns covers the substitution patterns of aspirated and unaspirated initial consonants (such as the confusion between "p" and "b" in Wu dialect), the conversion patterns between alveolar and retroflex consonants (such as the alternation of "z" and "zh" in Cantonese), and the variation patterns of labiodental and velar consonants (such as the difference between "f" and "h" in Min dialect). The replacement of each pattern is associated with a specific dialect area identifier and syllable initial position condition. The set of vowel variation patterns includes monophthong and diphthong merging patterns (such as the confusion of "ai" and "ei" in Southwestern Mandarin), nasalized final dropout patterns (such as the transformation of "an" to "a" in Xiang dialect), and medial omission patterns (such as the absence of the "u" medial in Hakka dialect). Each pattern is associated with a dialect area identifier and final vowel type condition. All patterns are marked with the probability of variation and the corresponding standard Mandarin phonetic mapping relationship.

[0079] It should be further explained that the correction process implemented by the finite-state transcription machine in this embodiment includes: constructing a state transition system for the transcription machine based on a pronunciation variation rule base, defining an initial state, intermediate state, and termination state, and the transition function between states includes input symbols (variant phonetic symbols in the first corrected text), output symbols (corresponding standard Mandarin phonetic symbols), and triggering conditions (dialect area identifier matching, syllable position matching); splitting the first corrected text into initial consonant units and final vowel units according to syllables, and inputting them sequentially into the finite-state transcription machine; receiving the first initial consonant unit in the initial state, matching the corresponding initial consonant variation pattern according to the transition function, and if the triggering condition is met, transitioning to the intermediate state and outputting the standard initial consonant; receiving the final vowel unit in the intermediate state, matching the corresponding final vowel variation pattern according to the transition function, and if the triggering condition is met, transitioning to the termination state and outputting the standard final vowel; directly retaining and transferring phonetic units that do not match any variation pattern to the termination state; aggregating all output standard initial consonants and standard final vowels, recombining them into a corrected syllable sequence to obtain the second corrected text.

[0080] S12. Based on the semantic units to be calibrated and their corresponding contextual acoustic and lexical features in the multidimensional dialect feature set, the probability distribution of their belonging to each candidate standard semantic is calculated by the GIS algorithm in the maximum entropy model to obtain the semantic disambiguation results.

[0081] It should be further explained that the process of obtaining the probability distribution of each candidate standard semantic in this embodiment includes:

[0082] Based on the semantic unit to be calibrated and its corresponding contextual acoustic and lexical features in the multidimensional dialect feature set, a set of feature functions is constructed to transform the spectral peaks and fundamental frequency contours in the acoustic features and the dialect terms and contextual word sequences in the lexical features into binary feature indicator values. By statistically analyzing the joint occurrence frequency of the feature functions with respect to each candidate standard semantic, the empirical expected value of the feature functions with respect to the semantic class is obtained. The weight parameters of each feature function are iteratively adjusted using a generalized iterative scaling algorithm, so that the expected value of the feature functions calculated by the model continuously approaches the empirical expected value, and finally a feature weight set that satisfies the maximum entropy constraint is obtained. Based on this feature weight set, by calculating the weighted sum of the output values ​​of all feature functions corresponding to the semantic unit to be calibrated and applying softmax normalization, the probability distribution of the unit belonging to each candidate standard semantic is obtained, thereby achieving semantic disambiguation.

[0083] S13. Based on the semantic disambiguation results, the original ambiguous field is replaced by selecting the candidate standard semantic with the highest probability using the Viterbi algorithm, generating the final standardized text that conforms to the user's true expression intent. It should be further noted that in the text processing of S9 to S13 of this embodiment, when a semantic conflict is detected in pain description keywords, the secondary correction process of the finite-state transcription machine is automatically triggered, and the correction result is verified by combining the F0 feature sequence of the real-time speech segment to ensure the accuracy of the pain representation information. The specific process includes:

[0084] S131. Construct the text segment to be detected based on the semantic disambiguation results, and perform semantic conflict detection through a preset pain keyword conflict rule base; the conflict rule base includes a mutual exclusion relationship table of pain description keywords, a combination constraint table of degree adverbs and pain types, and a symptom description logical contradiction pattern library.

[0085] S132. When a semantic conflict is detected, the secondary correction module of the finite state transcription machine is automatically activated. The transcription machine includes an acoustic feature verification layer and a semantic coordination layer.

[0086] S133. Extract the F0 feature sequence of the corresponding conflicting text segment in the original speech segment through the acoustic feature verification layer, and calculate the acoustic matching degree between the feature sequence and the candidate correction result.

[0087] S134. The semantic coordination layer receives each candidate standard semantic and its probability distribution in the semantic disambiguation results, and performs weighted sorting based on the acoustic matching degree.

[0088] S135. The finite-state transcription machine generates new state transition paths based on the reordering results, where each state corresponds to a candidate standard semantic selection, and the state transition weight is determined by both acoustic matching degree and semantic probability.

[0089] S136. Use the Viterbi algorithm to find the optimal state path and select the candidate standard semantic sequence with the highest comprehensive score as the final correction result.

[0090] S137. Compare the correction result with the original recognized text. When the correction confidence exceeds the preset threshold, automatically replace the original conflicting field.

[0091] S138. Record the type, triggering reason, and correction result of this correction operation, and update the abnormal pattern record table of the dialect variant feature library.

[0092] The semantic conflict refers to the occurrence of contradictory combinations of pain description terms in the same text segment; the acoustic matching degree refers to the similarity measure of the acoustic features of the candidate corrected text and the acoustic features of the original speech segment; and the comprehensive score is the weighted sum of the acoustic matching degree and the semantic probability.

[0093] The process begins with speech feature extraction. Spectral subtraction denoising and Hamming window frame segmentation are used to extract MFCC feature vectors. This is combined with an autocorrelation algorithm to generate F0 feature sequences reflecting dialect tone changes, and energy endpoint detection to obtain speech rate feature sets. This multi-dimensional deconstruction of the speech signal covers spectrum, tone, and rhythm information, avoiding the omission of key acoustic differences in dialects by single features, thus providing a rich and accurate feature base for the model. A feature layer fusion algorithm then aligns and stitches multiple features to ensure that the comprehensive multi-dimensional dialect feature set can fully characterize the acoustic properties of different dialects. Next, in the model training phase, a ResNet architecture is used to extract high-dimensional abstract spectral features to capture local dialect spectral patterns. A bidirectional LSTM network models the long-term dependency features of connected speech and sound change phenomena, and the Bahdanau attention mechanism strengthens the discriminative features of keyframes, significantly improving the model's ability to represent complex dialect pronunciation phenomena. In particular, a federated learning framework aggregates dialect data from various village-level nodes, combined with differential latent... The technology fine-tunes acoustic model parameters locally and encrypts and transmits updates, avoiding the privacy risks of centralized storage of dialect data. It can also dynamically adapt to dialect variations at different village levels (such as sub-dialects and special pronunciations), continuously optimizing the model's ability to recognize niche dialects and solving the problem of insufficient generalization in speech recognition in multi-dialect regions. Finally, in the text processing and correction stage, a dual-array Trie tree enables rapid standardization and replacement of dialect words, a finite-state transcription machine corrects systematic pronunciation errors based on a pronunciation variation rule library, and the BiLSTM-CRF algorithm and the maximum entropy GIS algorithm work together to complete entity extraction and semantic disambiguation. Furthermore, it triggers a secondary correction process for semantic conflicts of pain description keywords, that is, it combines real-time speech F0 feature sequences to verify the correction results, and selects the optimal correction scheme by weighted sorting of acoustic matching degree and semantic probability, completely eliminating the bias in pain representation information caused by dialect ambiguity, and ensuring that the final standardized text can accurately reflect the user's real pain state. The overall process forms a closed loop from feature extraction and model training to text correction. It is not only suitable for speech scenarios in multi-dialect regions (especially at the village level), but also significantly improves the recognition accuracy of dialect speech to pain representation keywords. This provides high-quality data support for the first subsystem to accurately screen pain patients and for the subsequent subsystem to formulate treatment strategies, avoiding deviations in the pain management process caused by dialect recognition errors.

[0094] The initial screening module is used to evaluate user target representation information and screen keywords based on the acquired user target representation information keyword set and a comprehensive evaluation algorithm. When the initial evaluation score of the corresponding user target representation is greater than the preset initial screening score threshold or a screening keyword is matched, a first user data package is constructed based on the user basic information, user target representation information keyword set and initial evaluation score that meet the conditions. The first user data package is then hashed using a hash algorithm and uploaded to the distributed database corresponding to the second subsystem through the smart contract in the three-level model chain for classified storage.

[0095] It should be further explained that the keywords selected in this embodiment include, but are not limited to, pain, ache, tumor, cancer, etc. It should also be explained that when the chief complaint identification module identifies the corresponding user as a pain patient, a screening assessment is performed. When the initial VAS assessment score of the corresponding user is greater than or equal to 4 or when the screening keywords exist, the corresponding user information is screened and uploaded to a pain monitoring list constructed by a distributed database, which is classified by disease type and initial VAS assessment score, for distributed storage.

[0096] Further explanation is needed regarding the implementation process of hash labeling using a hash algorithm in this embodiment. This includes: based on the complete data domain of user basic information, user target representation information keyword set, and initial evaluation score in the first user data packet, determining the concatenation order of each data item through a preset field sorting rule. Specifically, the user basic information is extracted in the field order of user unique identifier, gender, age, and region; the user target representation information keyword set is sorted in descending order of keyword frequency and converted into a string sequence; and the initial evaluation score is converted into a character-type numerical value with two decimal places. By concatenating the sorted data items sequentially into a continuous data string, SHA-256 hashing is applied. The algorithm performs a one-way hash operation on the continuous data string to generate a fixed-length hash value. This hash value is then associated with the metadata field of the first user data packet. The metadata field includes the data packet generation timestamp, the first subsystem node identifier, and the data version number. A field mapping mechanism ensures that the hash value uniquely corresponds to the complete content of the first user data packet. The hash value is then hexadecimal encoded to obtain a storable hash label string. This string is embedded into the header verification field of the first user data packet, forming a hash-labeled first user data packet. This enables the identification of the data packet's content integrity and uniqueness through the hash value, providing a basis for data consistency verification for subsequent smart contract verification and on-chain evidence storage.

[0097] Further explanation is needed; please refer to [link / reference]. Figure 3The implementation process of dividing a fixed-length hash value and embedding it into a three-level model chain in this embodiment includes: a 256-bit hash value generated based on the SHA-256 algorithm is divided into three sub-hash values ​​according to a preset segmentation rule. The first sub-hash value contains the first 85 bits of binary value, corresponding to the first subsystem verification segment; the second sub-hash value contains the middle 85 bits of binary value, corresponding to the second subsystem verification segment; and the third sub-hash value contains the last 86 bits of binary value, corresponding to the third subsystem verification segment. Through the preset segmentation mapping mechanism of the three-level model chain, the first sub-hash value is associated with the on-chain node storage area of ​​the first subsystem, the second sub-hash value is associated with the on-chain node storage area of ​​the second subsystem, and the third sub-hash value is associated with the on-chain node storage area of ​​the third subsystem, thus achieving a unique binding between the sub-hash value and the corresponding subsystem. The hash values ​​are converted into hexadecimal strings and written sequentially to the on-chain verification fields of the three subsystems using the segmented write function of the smart contract, forming a subsystem-level hash verification identifier. When the first user data packet is uploaded for verification, the smart contract calls the on-chain verification field of the first subsystem to extract the first sub-hash value, the on-chain verification field of the second subsystem to extract the second sub-hash value, and the on-chain verification field of the third subsystem to extract the third sub-hash value. The three sub-hash values ​​are then concatenated in their original order using a bit-concatenation algorithm to form a complete 256-bit hash value. This hash value is then compared with the hash label string in the verification field of the data packet header. If they match, the verification is successful, and the data packet upload permission is confirmed. This implements a cross-subsystem joint verification mechanism based on segmented hashing, ensuring the consistency verification of data transmission and storage in the three-level model chain.

[0098] It should be further explained that the specific implementation process of uploading the labeled first user data packet to the distributed database corresponding to the second subsystem for classification and storage through the smart contract in the three-level model chain in this embodiment includes:

[0099] A smart contract is constructed based on a pre-defined on-chain access control policy and data encryption protocol. The smart contract includes a data packet verification module, a permission management module, an on-chain storage module, and a distributed synchronization module. The data packet verification module pre-stores the hash label verification rules of the first user data packet, and the permission management module records the public key lists of nodes in the first subsystem and the second subsystem. The user basic information in the first user data packet is encrypted using an asymmetric encryption algorithm to generate an encrypted data segment. The user target representation information keyword set and the plaintext index segment of the initial evaluation score are retained to obtain the encrypted first user data packet.

[0100] When the first subsystem calls the smart contract, it digitally signs the encrypted first user data packet using its own node private key, and sends the signature information, node public key, and encrypted first user data packet to the consensus node of the three-level mode chain. After receiving the data, the smart contract's data packet verification module verifies the legality of the digital signature using the node public key, extracts the hash label in the encrypted first user data packet and compares it with the preset verification rules. If the comparison is successful, the permission management module is triggered to check whether the sending node is in the authorized list. If the verification is successful, the process of uploading to the chain for storage begins.

[0101] The on-chain storage module is based on the chain structure of the blockchain. It writes the encrypted first user data packet and the verification result as a transaction record into a new block. The block consensus is completed through the practical Byzantine fault-tolerant consensus algorithm, realizing on-chain data storage.

[0102] The distributed synchronization module reads the encrypted first user data packet stored on the chain, decrypts the user's basic information using the private key of the second subsystem node, and, based on the disease type and VAS initial assessment score in the plaintext index segment, calls the classification index mechanism of the distributed database to establish a storage structure with the disease type as the first-level directory and the VAS initial assessment score as the second-level subdirectory. The decrypted complete first user data packet is written to the corresponding directory, and a distributed storage address hash is generated and fed back to the smart contract to complete the storage confirmation.

[0103] During transmission, an encrypted transmission channel is established through the TLS protocol to perform real-time integrity verification of the transmitted data, preventing data tampering and leakage, and ultimately achieving secure on-chain storage and distributed classification of the first user data packet.

[0104] The information push module is used to push first strategy information based on the user target representation information keyword set, the user status information after evaluation and treatment corresponding to the second and third subsystems, the final round evaluation score and the evaluation score change trend, and in combination with a preset push strategy library, through a matching algorithm. It should be further noted that the push strategy library in this embodiment is constructed by those skilled in the art based on historical push data, corresponding type of symptom pain representation, and VAS evaluation score combined with a knowledge graph. It should also be noted that the final round evaluation score in this embodiment is the evaluation score corresponding to the user when the corresponding evaluation score is less than 4 points for the first time after multiple rounds of treatment. When the corresponding user's VAS evaluation score is less than 4 points, the corresponding user information is removed from the pain monitoring list. It should be noted that the user status information here represents the corresponding user's pain status information.

[0105] It should be further explained that the process of personalized education content push and push strategy optimization performed by the information push module in this embodiment includes:

[0106] By connecting to the data interfaces of the cyclic acquisition and adjustment module, the second subsystem, and the third subsystem, the system obtains pain fluctuation prediction results, including pain fluctuation range, pain peak time, and probability distribution of pain fluctuation causes; a set of user target representation information keywords, including keywords describing symptoms and pain, keywords for treatment response, and keywords for pain attack time periods; and assessment and treatment data, including user pain status information after treatment, final assessment score, and assessment score change trend. Using a data fusion algorithm with the user's unique identifier as the core association key, the system matches the probability distribution of pain fluctuation causes with user target representation information keywords (e.g., matching treatment response keywords with corresponding treatment-related fluctuation causes) and matches the assessment score change trend with pain fluctuation ranges (e.g., matching a slow downward trend with a higher pain level range), generating a push decision dataset. This push decision dataset provides real-time user data support for the subsequent push strategy library construction and serves as the core data input for personalized education content matching, realizing the transformation of data from multi-source collection to a standardized and usable dataset.

[0107] Based on a medical pain management knowledge graph, a multi-dimensional association system is constructed, linking disease types, causes of pain fluctuations, educational content, and delivery channels. Disease types include postoperative pain and tumor-related pain; causes of pain fluctuations include insufficient drug onset time and activity-induced pain; educational content includes corresponding intervention guidance in text or video format; and the default delivery channel is the user's mobile device. Combining historical delivery data, including past educational content, user click-through rates, and satisfaction feedback, a gradient boosting tree algorithm is used to train a content matching model. This model takes user target representation keywords, causes of pain fluctuations, and assessment score trends as input features and optimal educational content as the output label. The content matching rules generated by the model are embedded to construct a pre-set delivery strategy library. This pre-set delivery strategy library serves as the basis for personalized educational content matching rules and content invocation, realizing the transformation from knowledge accumulation and historical experience to standardized strategies, providing a basis for subsequent content matching.

[0108] Using a push decision dataset as input and a pre-defined push strategy library as rules and content sources, the algorithm first calculates the similarity between keywords representing user targets and tags of educational content, filtering out a preliminary set of highly relevant educational content. Then, it extracts high-probability causes of pain fluctuation from the probability distribution of pain fluctuation causes, retrieves corresponding educational content from the pre-defined push strategy library, forming a candidate set of educational content. The intersection of the preliminary set and the candidate set is taken, selecting highly similar educational content. Content optimization algorithms are then used (e.g., if the same cause of pain occurs multiple times, frequency-related prompts are added; if it occurs for the first time, basic cognitive explanations are added) to adjust content details, generating personalized educational content. This personalized educational content serves as the core object for subsequent push execution, realizing the transformation from standardized strategies to user-specific content.

[0109] Using personalized educational content as input, a priority ranking algorithm is employed to set educational content associated with peak pain times as the highest priority and the rest as ordinary priority. Through a push interface connected to users' mobile devices, the highest priority content is pushed first, and ordinary priority content is pushed later. Then, push effect data, including clicks, reading completion rate, and satisfaction rating, is obtained through a user feedback interface. The effect data is synchronized to a preset push strategy library, and a reinforcement learning algorithm is used to update the content matching model parameters, such as increasing the matching weight of high satisfaction content. At the same time, the effect data is fed back to the cyclical collection and adjustment module as a supplementary basis for judging the probability distribution of pain fluctuation causes, forming a closed loop of data collection, content matching, push, feedback, and optimization.

[0110] The cyclic collection and adjustment module is used to adjust the time interval and frequency of cyclic collection of target users in real time according to the round of evaluation and treatment of the corresponding user after evaluation and treatment, the evaluation score of each target characterization, the trend of evaluation score change and the screening keywords. It also obtains the cyclic screening evaluation result based on the cyclic collection information. If the cyclic screening evaluation result meets the data screening rules again, the corresponding user information is uploaded and saved again. If it does not meet the rules, the corresponding user information is deleted from the distributed database.

[0111] It should be further explained that the implementation process of real-time adjustment of the time interval and frequency of cyclical collection of target user data in this embodiment includes:

[0112] Based on a pre-set basic acquisition parameter library, which is constructed by those skilled in the art according to the pain characteristics and monitoring needs of different disease types, the initial acquisition interval is set to a longer duration for postoperative pain patients and a relatively shorter duration for tumor pain patients. The initial acquisition frequency for both types of patients and other disease types is set to one acquisition per acquisition interval. Through the data acquisition interface built into the cyclic acquisition adjustment module, the system receives target user cyclic acquisition data uploaded in real time from front-line acquisition devices (such as nurses' mobile acquisition apps). The cyclic acquisition data includes the target user's target representation assessment score (i.e., VAS score) for each assessment, the assessment timestamp corresponding to each assessment, the type of disease diagnosed by the user, such as postoperative incision pain, tumor-related pain, chronic strain pain, etc.; and the parameters of the current treatment plan being implemented for the user, including the specific type of analgesic drug, the dosage of the drug each time, and the prescribed drug administration time, etc.

[0113] The collected cyclic data undergoes preprocessing: the VAS scores are converted into values ​​conforming to a standard normal distribution using the Z-Score normalization algorithm to eliminate the impact of differences in VAS scoring scales among different users; an independent binary classification feature vector is generated for each disease type using the One-Hot encoding algorithm, allowing non-numerical disease types to be used as input features for the model; and the VAS scores are arranged sequentially according to the evaluation timestamps using a time series alignment algorithm to ensure that each time point corresponds to a unique VAS score, avoiding time series disorder, and finally forming a well-organized VAS time series dataset, providing a standardized data foundation for subsequent model training and prediction.

[0114] A pain fluctuation prediction model is constructed based on a Long Short-Term Memory (LSTM) network. The model's input features include a preprocessed regularized VAS time-series dataset, a disease type encoding vector generated through One-Hot encoding, and a treatment parameter vector reflecting the treatment plan. In the model structure design, the input layer dimension is determined based on the total number of input feature categories and the feature dimension to ensure complete reception of all input information. The hidden layers are multi-layered, with the number of neurons in each layer rationally configured according to feature complexity and model prediction accuracy requirements to fully exploit the nonlinear correlations between features. The output layer outputs the pain fluctuation prediction results for the next 24 hours through a linear activation function. These prediction results specifically include three core pieces of information:

[0115] Pain fluctuation range, which clearly reflects the possible range of changes in the user's pain level in the next 24 hours, must cover the complete range from the lowest possible pain level to the highest possible pain level;

[0116] The peak pain time needs to be accurately pinpointed to the time period within the next 24 hours when the user's pain level is expected to reach its highest value. For example, for postoperative pain patients, it can be predicted to be a specific time period at night, and for cancer patients after chemotherapy, it can be predicted to be a specific time period on a specific number of days after the end of chemotherapy.

[0117] The probability distribution of causes of pain fluctuations lists the main causes that may lead to future pain fluctuations, such as insufficient onset time of medication, pain induced by daily activities, changes in physiological rhythms, and reactions during the adaptation period of treatment plan adjustments, and marks the probability value of each cause.

[0118] During model training, historically accumulated VAS time-series data is used as training labels. The Adam optimizer is used to continuously adjust the model parameters, with the training objective being to minimize the mean squared error between the model's predicted values ​​and the actual historical label values. The model's hyperparameters (such as the number of hidden layers, the number of neurons, and the learning rate) are adjusted and optimized using a 5-fold cross-validation method to ensure that the model maintains a high level of accuracy in predicting pain fluctuations for users with different disease types and at different treatment stages, thus guaranteeing the reliability of the prediction results.

[0119] Based on the pain fluctuation prediction results output by the above pain fluctuation prediction model, targeted collection and adjustment rules are generated through the rule engine built into the cyclic collection and adjustment module. The specific rule logic is as follows:

[0120] When the prediction results show a clear peak pain time, a specific time period before and after the peak pain time is set as a high-attention acquisition window. The duration of the high-attention acquisition window needs to cover the preparation and observation phase before the peak pain occurs and the continuous monitoring phase after the peak pain occurs, to ensure that the user's pain changes during the peak period can be fully captured. Within the high-attention acquisition window, the original initial acquisition interval is shortened to half of the initial acquisition interval, while the acquisition frequency is increased to twice the initial frequency, so as to encrypt the acquisition frequency and accurately obtain pain data during the peak period.

[0121] When the prediction results show that the pain fluctuation range is in the high-mid range of VAS, that is, it meets the pain level criteria that need to be monitored in the data screening rules, and no clear pain peak appears, the collection interval is shortened to 80% of the initial collection interval, and the collection frequency is maintained to perform one collection at each collection interval. By appropriately increasing the collection frequency, the pain fluctuation trend can be grasped in a timely manner.

[0122] When the prediction results show that the pain fluctuation range is consistently in the low VAS range, i.e., it does not meet the pain level criteria that need to be monitored in the data screening rules, the collection interval will be extended to 1.5 times the initial collection interval, while the collection frequency will remain at once per collection interval. If the VAS scores collected consecutively are consistently in the low VAS range, and it is confirmed that the user's pain status has been continuously relieved, a user information deletion command will be triggered to remove the monitoring information of the corresponding user from the distributed database and stop the cyclic collection operation for that user.

[0123] Through the parameter distribution interface built into the cyclic acquisition and adjustment module, the generated acquisition adjustment rules, including the adjusted acquisition interval, acquisition frequency, and specific time periods of high-attention acquisition windows, are synchronized to the front-line acquisition equipment in real time. The front-line acquisition equipment needs to display the adjusted acquisition time nodes and frequency reminders on the operation interface to ensure that acquisition personnel can accurately obtain the adjustment information and perform acquisition operations according to the new rules.

[0124] Meanwhile, through the module's built-in real-time feedback interface, it continuously receives adjusted VAS data uploaded by front-line acquisition devices. It performs the same preprocessing operations as before on these newly acquired data, and inputs the preprocessed new data as new training samples into the pain fluctuation prediction model. Through incremental training algorithms, it merges the new samples with historical training samples to dynamically update the model parameters and avoid prediction deviations caused by untimely data updates.

[0125] The model's prediction accuracy is evaluated regularly. If the evaluation results show that the accuracy drops below the preset standard, the model is retrained using all historical samples and new samples to ensure that the model always has stable prediction performance. This forms a closed-loop mechanism of data collection and adjustment, data feedback, model optimization, and readjustment, ensuring that the cyclical data collection strategy is always accurately matched with the user's real-time pain status.

[0126] The first analysis module is used to record in real time the data collection, filtering, strategy push, and cyclic collection process information corresponding to the first subsystem, and to perform horizontal anomaly analysis and localization on the preset first subsystem chain in conjunction with a preset node anomaly analysis model to obtain first anomaly feature information. This first anomaly feature information is then fed back to the preset third subsystem chain in the third subsystem through the three-level pattern chain. It should be further noted that the third subsystem chain in this embodiment is constructed by combining the corresponding single-function submodules within the third subsystem with the interaction information between different submodules and a graph neural network.

[0127] It should be further explained that the construction and training process of the node anomaly analysis model in this embodiment includes: a full log dataset based on the historical data collection, filtering, strategy push, and cyclic collection process of the first subsystem. This dataset contains normal operation samples and abnormal samples. Abnormal samples are labeled with anomaly type tags such as data collection timeout, filtering rule failure, push strategy mismatch, and abnormal cyclic collection parameters; time series algorithms are used to extract time-series features and correlation features from the log data. Time-series features cover the execution time, frequency fluctuation, and result feedback delay of each process, while correlation features cover the data interaction success rate between submodules, parameter transmission consistency, and rule call matching degree, forming a multi-dimensional feature matrix; a model infrastructure is constructed based on a graph neural network, with each functional node in the first subsystem chain as a graph node, and the data flow relationship between nodes is... Using graph convolutional layers as edges, node association features are aggregated, and temporal features are modeled sequentially using a long short-term memory network to construct an anomaly detection network that integrates node interaction and temporal evolution. Labeled samples are used to train the model, with anomaly type identification accuracy and anomaly localization precision as optimization objectives. Backpropagation algorithm is used to adjust network weights, and the feature weight allocation is dynamically corrected using a validation set, giving the model higher attention to high-frequency anomaly features. A sliding window mechanism is introduced during training to adapt features to real-time log fragments, and confusion matrix analysis is used to optimize the anomaly type classification boundary. Finally, a node anomaly analysis model is obtained that can accurately identify anomaly patterns such as data acquisition delays, filtering logic conflicts, push strategy failures, and cyclic acquisition parameter offsets, achieving horizontal correlation analysis and accurate localization of node anomalies on the first subsystem chain.

[0128] The first subsystem employs spectral subtraction, Mel-frequency cepstral coefficients, fundamental frequency extraction, and speech rate feature analysis to perform multi-dimensional, fine-grained feature deconstruction on the original speech signal, obtaining a comprehensive multi-dimensional dialect feature set that integrates spectrum, pitch, and rhythm information. This provides the model with a richer information base than that of a single feature. Based on this, a deep local spectral pattern is extracted using a ResNet-based convolutional neural network, and then combined with a bidirectional LSTM network to capture its temporal dependencies. This gives the model a powerful representational ability to depict complex dialect pronunciation phenomena. Furthermore, an attention mechanism is introduced to weighted convergence of key features, enabling the model to focus on highly discriminative speech segments. This significantly improves the speech recognition accuracy of user complaints, especially pain descriptions, in complex dialect environments, providing the system with high-quality and stable raw data input. Secondly, the efficient data structure of a dual-array Trie tree enables rapid mapping and replacement of dialect vocabulary with standard medical terminology, solving the terminology standardization problem. Subsequently, a finite-state transcription machine based on a pronunciation variation rule base was used to phonologically correct systematic errors in initials and finals, ensuring the accuracy of the conversion from the perspective of pronunciation mechanisms. Finally, through the collaborative work of a BiLSTM-CRF model and a maximum entropy model, medical entity extraction and semantic disambiguation were performed on the corrected text, ensuring that the final set of user target representation information keywords accurately reflects the patient's true condition, completely eliminating semantic bias caused by dialect ambiguity, and laying a solid data foundation for subsequent evaluation and decision-making. Third, a distributed joint verification mechanism was constructed by segmenting the hash value of the complete data packet and binding it to a three-level subsystem. No single subsystem can independently complete data integrity verification; consistency among the three parties' data is required for verification to pass, greatly increasing the difficulty of data tampering and ensuring the credibility of cross-level data transmission from a mechanism perspective. Combined with the automated verification and execution capabilities of blockchain smart contracts, and the application of asymmetric encryption and digital signature technologies, the entire process of data collection, uploading, notarization, and storage was made secure, controllable, and transparently traceable, establishing a foundation of trust and collaboration across medical institutions.

[0129] It should be further explained that the second subsystem in this embodiment includes a first processing module, a first evaluation module, an upgrade judgment module, and a second analysis module;

[0130] The first processing module is used to match the corresponding type processing strategy based on the user type and user target representation information keywords in the labeled first user data packet, combined with a preset classification processing node network and a preset fast index, and to process the corresponding user using the matching processing strategy. It also collects, records and preprocesses the user target representation information keywords within a preset time length after the current user is processed, and records the average processing rate of the current processing strategy within the preset time length. Each node in the classification processing node network corresponds to a processing strategy.

[0131] It should be further explained that the treatment strategy in this embodiment is constructed from the information of the department and the doctors and nurses under each disease type; matching the treatment strategy for the corresponding type involves matching the corresponding department or the doctor under the department to treat the pain of the corresponding disease, and then using the matched nurse to collect real-time monitoring data and record VAS evaluation scores; it should be further explained that the average treatment rate in this embodiment is the number of users with the corresponding type of disease treated by the corresponding doctor within a preset length of time, which is used for resource allocation adjustment of the third subsystem and performance evaluation of the corresponding doctor;

[0132] It should be further explained that the construction process of the classification and disposal node network in this embodiment includes:

[0133] S21. Based on the historical case library, extract all independent handling strategies and assign a unique node identifier to each handling strategy to obtain an initial node set. Each node stores the complete content, applicable conditions and execution parameters of the corresponding handling strategy.

[0134] S22. Based on the logical correlation between strategies and the frequency of historical joint use, a directed graph model is used to construct the connection edges between nodes. Each edge is assigned an initial weight according to the historical treatment evaluation score, treatment success rate, conversion time length and the number of rounds of reprocessing after treatment, so as to obtain the treatment strategy network topology with weighted relationship.

[0135] S23. Based on a multi-dimensional feature index system, a retrieval entry point is constructed for each node through a hybrid index structure. The index key includes user type, target representation information keyword weight, handling urgency level, resource demand type, historical handling success rate and average handling rate within a preset time length, to obtain a strategy index that supports multi-condition combined queries.

[0136] It should be further explained that the process of constructing the urgency level in this embodiment includes:

[0137] A multidimensional assessment index library is constructed based on user target representation information keyword set, symptom duration parameter, historical treatment response time data and complication risk coefficient;

[0138] A keyword weight mapping algorithm converts core symptom keywords in user target representation information into severity scores, with the score range corresponding to a gradient quantification of symptoms from mild to critical. A time decay function processes the duration parameter of symptoms, assigning a higher time coefficient to longer durations to reflect the urgency of symptom prolongation. Based on the correlation data between response delays and adverse consequences for similar symptoms in historical treatment cases, a response timeliness weight is constructed, with higher weight values ​​for more severe consequences caused by response delays. A complication risk prediction model analyzes user basic information and symptom characteristics, outputting the probability of complication occurrence as a risk coefficient. The severity score, time coefficient, response timeliness weight, and risk coefficient are weighted and summed according to preset weight proportions to obtain a comprehensive urgency score.

[0139] The urgency level is determined by matching the comprehensive urgency score with a preset threshold range. At the same time, a dynamic adjustment mechanism is established to trigger a recalculation of the comprehensive urgency score based on real-time symptom changes (such as increased pain or the appearance of new symptoms), thereby dynamically updating the urgency level and ensuring an accurate match between the urgency level and the user's real-time status.

[0140] The multi-dimensional feature index system in this embodiment specifically includes user type, executor skill level, historical index frequency, the decrease in evaluation score after policy handling, the number and probability of re-evaluation after handling, and the satisfaction rate of post-handling evaluation.

[0141] S24. Based on the real-time collected medical and nursing resource status data of the township / community subsystem, the resource availability coefficient of each node is calculated through the resource load perception layer; the resource status data includes the current number of patients seen by doctors, the saturation of nurses' monitoring tasks, and the remaining amount of medicine inventory; the resource availability coefficient is obtained by converting the amount of each resource status data into a value between 0 and 1 through a normalization algorithm and by weighted fusion.

[0142] S25. Introduce the resource availability coefficient into the node jump weight calculation. Multiply the resource availability coefficient with the original edge weight through the weight adjustment algorithm to obtain the resource awareness adjustment weight. When the resource availability coefficient is lower than the preset threshold, automatically reduce the jump weight of the corresponding node and trigger the resource warning mechanism.

[0143] S26. The retrieval priority of the fast index is dynamically adjusted based on the resource availability coefficient. The resource consumption index is associated with the resource availability coefficient through the priority recalculation algorithm, and the nodes of the disposal strategy with low resource consumption are retrieved first.

[0144] S27. Based on real-time feedback data after real-time processing, the node index weight and edge relationship weight are adjusted through simulation algorithm, and the failure mode is located by association rule mining to obtain the classified processing node network.

[0145] S28. Based on the blockchain's evidence storage requirements, the structural hash and decision log of the classification and disposal node network are synchronized to the chain via smart contracts to obtain disposal strategy records. These records are then uploaded to the abnormal decision-making node for real-time abnormal simulation decision adjustment until the corresponding abnormal adjustment is completed. The resource load perception layer monitors the medical resource status of each township / community subsystem in real time. When resource shortages are detected, a resource warning signal containing the resource type, degree of scarcity, and expected duration is automatically sent to the third subsystem. The retrieval priority adjustment of the fast index adopts a dynamic weight mechanism, updating the index sorting hourly based on the real-time resource availability coefficient.

[0146] The preset fast index is constructed by combining the user type, skill level, index frequency, evaluation score decrease of real-time user target representation information after the corresponding handling strategy, the number of rounds and probability of re-evaluation after the current handling strategy, and the evaluation satisfaction rate after handling with a hash algorithm.

[0147] It should be further explained that the construction process of the fast index in this embodiment includes:

[0148] Based on the handling strategies corresponding to each node in the classification and handling node network, the following are extracted: user type, executor skill level, historical index frequency, the decrease in evaluation score of real-time user target representation information after strategy handling, the number of rounds and probability of re-evaluation after the current strategy handling, and the post-handling evaluation satisfaction rate. A convolutional neural network is used to convert user type into classification codes, skill level into numerical levels, index frequency into original count values, score decrease into percentage values, rounds into average values, probability into percentage values, and satisfaction rate into percentage values, obtaining a standardized feature set. Each feature in the standardized feature set is then encoded according to user type, skill level, index frequency, score decrease, and re-evaluation. The sequence of round number, re-evaluation probability, and satisfaction rate is concatenated into a feature string. The SHA-1 hash algorithm is used to operate on this feature string to generate a fixed-length hash index key. A mapping mechanism is used to associate the hash index key with the node identifier of the corresponding disposal strategy in the classification and disposal node network, establishing a direct pointing relationship from the index key to the node. Based on the dynamic changes of each feature in the feature string, a hash index key recalculation mechanism is triggered periodically to update the association relationship between the index key and the node identifier, ensuring real-time synchronization between the fast index and the classification and disposal node network. This fast index serves as the retrieval entry point for the classification and disposal node network, enabling rapid location of the target strategy node in the classification and disposal node network through the fast matching capability of the hash index key.

[0149] The first evaluation module is used to perform real-time evaluation based on the keywords of the user target representation information after treatment, combined with preset evaluation indicators and evaluation algorithms, to obtain the first evaluation score and the trend of the first evaluation score change of the real-time user target representation information, and at the same time, based on the trend of the first evaluation score change, to obtain the relief rate of the target user within a preset time period.

[0150] The escalation judgment module is used to perform escalation judgment based on the target user's remission rate within a preset time period and preset escalation judgment conditions. Specifically:

[0151] When the mitigation rate of the target user within a preset time period meets the preset escalation judgment condition, the first user data packet corresponding to the target user and the second user data packet constructed by combining the treatment strategy information of each round with the hash algorithm are fed back to the preset third subsystem chain in the third subsystem through the three-level mode chain. It should be further noted that the escalation judgment condition in this embodiment is: when the change trend of the first evaluation score of the corresponding user within the preset time period is on the upward trend or the first evaluation score is still greater than 4, the corresponding user is upgraded for consultation, treatment and evaluation.

[0152] If the relief rate of the target user within a preset time period does not meet the preset escalation criteria, the first satisfaction level of the target user with the current treatment strategy information is assessed. This assessment is then fed back to the third subsystem and the first subsystem through the three-level model chain. The anomaly analysis and decision-making module of the third subsystem generates a follow-up decision plan for the corresponding user, and the corresponding follow-up decision plan is fed back to the corresponding follow-up personnel through the first subsystem. This allows for real-time adjustments to the follow-up frequency, interval, and pain management strategy for the current user. It should be further noted that the first satisfaction level in this embodiment is used to measure the performance of the corresponding doctor or nurse.

[0153] The second analysis module is used to perform anomaly analysis and location on the corresponding user who meets the upgrade judgment condition within the corresponding preset time length in the classification and disposal node network, obtain the second anomaly feature information, and feed the second anomaly feature information back to the third subsystem through the three-level pattern chain.

[0154] It should be further explained that the process of obtaining the second abnormal feature information in this embodiment is as follows:

[0155] Based on the full log of the treatment strategy execution for users who meet the escalation judgment conditions within a preset time period in the classification and treatment node network, the full log includes the node identifier associated with the strategy, execution parameters, target representation evaluation score before treatment, target representation evaluation score after treatment, nurse monitoring data, specific reasons for escalation trigger, doctor strategy parameter setting record, nurse monitoring data entry record, and timestamp of medical staff operation.

[0156] The TF-IDF algorithm was used to extract features from textual information in the full log, converting textual content such as the specific reasons for escalation triggers and medical staff operation notes into feature vectors. A sliding window algorithm was used to process the strategy execution time series data, dividing the data into windows at preset time intervals and extracting the execution start time, parameter adjustment nodes, and effect feedback times within each window as strategy execution time series features. A statistical feature extraction algorithm was used to calculate the numerical data in the logs, obtaining effect features and node correlation features. These effect features included the mean of the evaluation score change trajectory, the variance of the evaluation score change trajectory, and the duration of failure to achieve the expected effect. The length and node association characteristics include the frequency of occurrence of jump paths between strategy nodes and the Euclidean distance between the jump paths between strategy nodes and the preset paths in the classification and treatment node network; the medical and nursing operation data are counted and the proportion is calculated by statistical feature extraction algorithm to obtain medical and nursing operation characteristics, which include the total number of doctor strategy parameter mismatches, the proportion of doctor strategy parameter mismatches to the total number of operations, the total frequency of missing nurse monitoring data, the proportion of missing nurse monitoring data to the total number of monitoring, and the time difference between the time of medical and nursing operation and the time specified in the standard procedure; the above feature vectors and various features are integrated to form a multi-dimensional feature set to be analyzed;

[0157] A causal association model is constructed based on the Bayesian network algorithm. The upgrade result is used as the target variable of the model, and each feature in the multidimensional feature set to be analyzed is used as the influencing variable of the model. The conditional probability relationship between each influencing variable and the target variable is determined through model training. The conditional probability data output by the causal association model is processed by the C4.5 decision tree algorithm. Nodes are split according to the information gain ratio to generate decision paths. The decision path analysis is used to locate feature combinations that have a strong causal relationship with the upgrade result. The feature combinations with a strong causal relationship include: the execution parameter and user type matching deviation and the conditional probability corresponding to the deviation is greater than a preset probability threshold; the correlation coefficient between node jump delay and effect decay is greater than a preset correlation coefficient threshold; the support of the association rule between the number of times the doctor's strategy parameter mismatch is greater than a preset support threshold; and the linear correlation coefficient between the duration of missing nurse monitoring data and the effect evaluation deviation is greater than a preset correlation threshold.

[0158] The aforementioned strong causal relationship feature combinations are compared with the preset normal execution features and preset medical and nursing operation standards of the corresponding strategy nodes in the classification and treatment node network. The preset normal execution features include the standard parameter range of the strategy node, the expected effect threshold of the strategy node, and the association weight range of the strategy node. The preset medical and nursing operation standards include the error threshold of the doctor's strategy parameter setting, the completeness threshold of the nurse's monitoring data, and the time difference threshold of the medical staff's operation. The difference between each feature value in the strong causal relationship feature combination and the corresponding preset standard value is calculated by the threshold judgment algorithm to extract the deviation features. The deviation features include the specific magnitude of the execution parameters exceeding the standard parameter range, the number of times the effect index is lower than the expected effect threshold, the specific degree of the node association weight deviating from the association weight range, the specific value of the number of times the doctor's strategy parameter mismatch exceeds the error threshold, the specific proportion of the nurse's monitoring data missing time exceeding the completeness threshold, and the specific duration of the medical staff's operation time difference exceeding the time difference threshold.

[0159] A data storage mechanism is constructed based on Ethereum smart contract algorithms. Extracted medical and nursing operation deviation features are bound to the unique identifiers of corresponding medical and nursing personnel. These unique identifiers include doctor's employee ID and nurse's employee ID. The bound data is written to the blockchain for storage through the immutability of smart contracts, forming a basic database of medical and nursing operations. The Apriori association rule algorithm is used to perform frequent itemset mining on the data in the basic database of medical and nursing operations. Minimum support and minimum confidence are set to filter out high-frequency medical and nursing operation deviation itemsets associated with specific disease types, thus identifying the core competency gaps of medical and nursing personnel. These core competency gaps include frequent mismatches in strategy parameter settings for specific disease types by specific doctors and frequent missing monitoring data for specific disease types by specific nurses.

[0160] A training course recommendation model is constructed based on a collaborative filtering recommendation algorithm. The core competency gaps of medical staff are used as input labels for the model. A pre-set medical training resource library is invoked, which contains training course resources corresponding to various competency gaps. By calculating the similarity between the competency gap label and the training course resource label, personalized training courses are generated for the corresponding medical staff. The personalized training courses include a course on matching standard parameters for disease treatment to address parameter setting gaps and a course on standard pain monitoring data collection to address monitoring data gaps.

[0161] The data on the handling operations of medical staff before and after training are processed by the difference percentage calculation algorithm to obtain the strategy parameter mismatch rate and monitoring data missing rate of medical staff before training and the strategy parameter mismatch rate and monitoring data missing rate of medical staff after training. The difference between the training and the training is calculated, and then the difference is divided by the value before training to obtain the capability improvement coefficient. The capability improvement coefficient includes the percentage decrease in parameter mismatch rate and the percentage decrease in monitoring data missing rate.

[0162] A capability improvement coefficient assessment model is constructed using a logistic regression algorithm. The capability improvement coefficient is used as the model input, and a preset capability improvement threshold is used as the model output assessment standard. If the capability improvement coefficient is greater than or equal to the preset capability improvement threshold, a KPI weight adjustment algorithm is triggered. This algorithm calls the medical staff performance evaluation system to increase the weight ratio of the two performance indicators: treatment accuracy and monitoring data completeness. If the capability improvement coefficient is less than the preset capability improvement threshold, a cyclic triggering algorithm is triggered. This algorithm restarts personalized training targeting the medical staff's core capability shortcomings according to the preset number of training cycles or training effect achievement conditions.

[0163] Principal component analysis (PCA) is used to reduce the dimensionality of deviation features, core competency gaps of medical staff, and competency improvement coefficients, retaining principal component features with variance contribution greater than a preset contribution threshold. A feature concatenation algorithm is then used to bind the dimensionality-reduced principal component features with corresponding strategy node identifiers and the influence range of associated nodes. The influence range of associated nodes includes upstream and downstream associated strategy nodes of the abnormal strategy node and the resource allocation node corresponding to the abnormal strategy node. The binding results are integrated to obtain a second set of abnormal feature information, including the location of the abnormal strategy node, the anomaly type, the influence of associated nodes, the core competency gaps of medical staff, the competency improvement coefficient, and training needs. The anomaly type includes parameter mismatch, execution delay, unsatisfactory results, and non-standard medical operations. This process uses existing algorithms to accurately locate anomalies in escalation-related treatment strategies and provides data support for the third subsystem's anomaly analysis and decision-making module to generate training strategies, ensuring that the training strategies can effectively address the execution problems of treatment strategies caused by abnormal medical operations.

[0164] It should be further explained that the third subsystem in this embodiment includes an escalation handling matching module, an escalation evaluation module, an anomaly analysis and decision-making module, an anomaly adjustment module, and an assessment and training module;

[0165] The upgraded treatment matching module is used to obtain the corresponding upgraded treatment strategy and matching accuracy based on the second user data packet and the upgraded treatment strategy library configured in the third subsystem, and to collect the status information of the second user after the upgraded treatment strategy has been applied to the corresponding user in real time. It should be further noted that the upgraded treatment strategy library in this embodiment was constructed by those skilled in the art based on the pain management strategies of tertiary hospitals for the corresponding types of diseases, combined with a tree database and historical management strategy index information.

[0166] The upgrade evaluation module is used to obtain a second evaluation score after upgrade processing based on the second user status information and the second evaluation algorithm. When the second evaluation score is less than the preset processing completion evaluation threshold, the processing of the corresponding user ends. At the same time, based on the evaluation information of the corresponding user on the matched upgrade processing strategy during the upgrade processing, the second satisfaction rate of the current user is obtained, and the second satisfaction rate of the corresponding upgrade processing strategy is fed back to the upgrade processing strategy library to adjust the index information of the current upgrade processing strategy in real time.

[0167] When the second evaluation score is greater than or equal to the preset disposal completion evaluation threshold, the third subsystem presets the abnormal decision node in the first subsystem chain to perform cyclic matching evaluation of the escalation disposal strategy until the second evaluation score is less than the preset disposal completion evaluation threshold, and records the corresponding cycle number and the corresponding escalation disposal strategy information that does not meet the condition.

[0168] It should be further explained that the exemplary implementation process of the hierarchical classification evaluation criteria in this embodiment includes:

[0169] The chief complaint recognition module identifies the corrected user voice text information, which includes pain representation information. For users with pain representation information, a VAS forced assessment is performed through the initial screening module to obtain the initial assessment score of the user's target representation. At the same time, the corrected user voice text information is matched with the screening keywords to determine whether the corresponding user's pain representation information contains the screening keywords. If it exists or the corresponding user's initial assessment score is greater than 4, the first user data package of the corresponding user is added to the pain monitoring list according to the corresponding disease type.

[0170] The first treatment module matches the corresponding type of treatment strategy with the first user data packet of the user in the pain monitoring list, combined with the classification treatment node network and the preset fast index, and uses the matching treatment strategy to treat the corresponding user. It also collects, records and preprocesses the target representation information keywords within a preset time length after the current user is treated. Based on the collected and preprocessed target representation information keywords within the preset time length, it performs real-time evaluation with preset evaluation indicators and evaluation algorithms to obtain the first evaluation score and the trend of the first evaluation score of the real-time user target representation information. At the same time, based on the trend of the first evaluation score, it obtains the relief rate of the target user within the preset time length.

[0171] The escalation decision is made based on the target user's mitigation rate within a preset time period and the escalation criteria. If the target user's mitigation rate within the preset time period meets the preset escalation criteria, the first user data packet corresponding to the target user and the second user data packet constructed by combining the information of each round of handling strategy with a hash algorithm are fed back to the preset third subsystem chain in the third subsystem through the three-level mode chain. If the target user's mitigation rate within the preset time period does not meet the preset escalation criteria, the first satisfaction level of the target user with the current handling strategy information is evaluated and fed back to the third subsystem and the first subsystem through the three-level mode chain.

[0172] The system is based on the second user data packet and the escalation handling matching module. The escalation handling process is evaluated in real time. When the VAS evaluation score after handling is less than 4, the user satisfaction rate of the corresponding handling process is evaluated. At the same time, the frequency and interval of follow-up tasks are adjusted according to the satisfaction rate and the fluctuation range and direction of the VAS evaluation score during the handling process.

[0173] The anomaly analysis and decision-making module is used to obtain anomaly feature association information and difference information by combining the first anomaly feature information and the second anomaly feature information with an association algorithm. Based on the anomaly feature association information and difference information, the average handling rate of the current handling strategy within a preset time length, the first satisfaction rate, and the relief rate of the target user within a preset time length, and combined with the anomaly decision-making nodes in the preset third subsystem chain of the third subsystem, an associated anomaly adjustment strategy is obtained.

[0174] The detailed implementation process of the anomaly analysis and decision-making module is based on the first anomaly feature information (including data acquisition timeout, filtering rule failure, push strategy mismatch, and abnormal cyclic acquisition parameters in the first subsystem) and the second anomaly feature information (including parameter mismatch, execution delay, substandard effect, and node association deviation in the classification and handling node network of the second subsystem). An improved Apriori algorithm is used to divide the two types of anomaly features into itemsets. The occurrence time, associated user identifier, symptom type, and impact range of each anomaly are used as itemset attributes. The support of any two itemsets co-occurring (the ratio of the frequency of itemsets occurring simultaneously to the total sample size) and the confidence (the conditional probability of a subsequent itemset appearing when the preceding itemset appears) are calculated. Itemet combinations with both support and confidence higher than a preset threshold are selected to obtain strong association rule-based anomaly feature association information, such as data acquisition timeout - strategy parameter mismatch - filtering rule failure - node association deviation. A sequence difference algorithm based on dynamic time warping is used to align the two types of anomaly features according to the time series and calculate... The Euclidean distance of features in the dimensions of occurrence time, duration, and impact intensity is used to find the optimal matching path through dynamic programming, quantify the degree of offset in spatiotemporal distribution, and obtain information on the differences in abnormal features. For example, the anomaly of the first subsystem is more discrete in the time dimension, while the anomaly of the second subsystem has a larger impact range in the spatial dimension. Based on the rule strength of the association information of abnormal features and the offset quantification value of the difference information, combined with the average treatment rate of the current treatment strategy (the number of similar diseases treated by corresponding doctors), the first satisfaction (user evaluation of the doctor and nurse's treatment), and the target user relief rate (the decrease in pain assessment score), a weighted fuzzy logic fusion algorithm is used to assign weights to each indicator. Among them, the rule strength of the association information has the highest weight, followed by the relief rate. The average treatment rate and the first satisfaction are weighted at 1:1. The indicator values ​​are mapped to fuzzy sets (high, medium, low), and the membership degree is calculated by the membership function. After fuzzy inference and defuzzification processing, a multi-dimensional comprehensive anomaly assessment index is generated, which includes the severity of the anomaly, the tightness of the association, and the breadth of the impact.

[0175] The comprehensive anomaly assessment indicators are input into the preset anomaly decision node in the third subsystem chain. The decision tree classifier of this node uses anomaly severity, correlation tightness, and impact breadth as input features. The leaf nodes correspond to predefined anomaly adjustment strategies: when a delay in treatment caused by a lack of medical staff skills is detected, and the strategy push in the first anomaly feature is mismatched, such as inaccurate push of training resources, a tiered training strategy is triggered. By connecting to the medical education platform, the skill requirement library of the corresponding strategy node (such as the anesthesiology-postoperative analgesia node) in the classification and treatment node network is called, and training courses containing analgesia protocol operation specifications and VAS assessment standards are automatically generated. Learning tasks are assigned according to the doctor's professional title level, and the learning progress is synchronized to the performance record area of ​​the three-level model chain; when anomalies caused by deviations in workflow execution are identified... Normally, if a nurse fails to record VAS scores according to the frequency set by the cyclic collection and adjustment module, and the average treatment rate and first satisfaction rate decrease, the dynamic performance adjustment strategy is triggered. This involves modifying the KPI weights of monitoring data integrity and user satisfaction in the performance evaluation model, generating a performance adjustment list in conjunction with the salary calculation system, and synchronizing it to the nurse management module of the second subsystem. When an anomaly caused by resource allocation imbalance is detected, such as excessive workload of orthopedic surgeons leading to delayed postoperative rehabilitation guidance, and associated node jump delay characteristics, the intelligent scheduling strategy is triggered. This involves recalculating the optimal manpower allocation plan using a genetic algorithm with the resource requirements of each strategy node in the classification treatment node network as constraints, pushing it to the scheduling system, and updating the allocation records of the resource pool in the third subsystem. The resource requirements include the number of doctors and working hours.

[0176] For equipment resource anomalies, such as data loss due to pain assessment terminal malfunction, the system associates this with the first anomaly feature of data collection timeout and triggers a preventative maintenance strategy. It trains a prediction model using equipment operation data, including fault frequency and runtime, generates a maintenance schedule, automatically pushes work orders containing terminal model and maintenance items to the equipment management system, and writes maintenance records into the three-level pattern chain. For drug supply anomalies, such as interrupted analgesia strategy execution due to insufficient non-steroidal anti-inflammatory drug (NSAID) inventory, the system associates this with the second anomaly feature of unsatisfactory efficacy and triggers an inventory optimization strategy. It obtains drug demand by inputting historical medication usage and the current number of pain patients into the demand prediction model, adjusts the safety stock level, initiates an application to the procurement system containing drug specifications and quantities, and synchronizes the resource constraints of the corresponding strategy node in the classification and treatment node network. For insufficient patient compliance, such as slow VAS score decline due to failure to take analgesics as instructed, and associated with low relief rate, the system triggers a multimodal education strategy. It generates personalized solutions including text guidance, video demonstrations, and Q&A interactions using content recommendation algorithms combined with user target representation information keyword sets, coordinates nurse resources for regular follow-ups, and feeds compliance data back to the information push module of the first subsystem.

[0177] When an anomaly is detected due to a conflict in the logic of strategy nodes in the classification and treatment node network, such as a contradiction in the strategy parameters between the anesthesiology and orthopedics departments, a node logic reconstruction strategy is triggered. This involves retraining the weights of the connecting edges between nodes using a graph neural network, correcting the applicable conditions of the strategy, and synchronizing it to the fast index. When a deviation in the extraction of target representation information is found due to errors in the dialect recognition model of the first subsystem, such as misjudgment of pain level, a model iteration strategy is triggered. This involves adding the error samples to the training set, retraining the MFCC feature extraction layer and LSTM network of the dialect recognition model, and updating it to the chief complaint recognition module. When a data transmission delay across subsystems is detected, such as excessively long hash verification time in the three-level model chain, an on-chain optimization strategy is triggered. This involves adjusting the segmented verification logic of the smart contract to shorten the hash value splicing and comparison time and improve data synchronization efficiency.

[0178] Furthermore, the anomaly analysis and decision-making module monitors in real time the anomaly repair rate (e.g., the reduction rate of handling delays after training), resource consumption (e.g., changes in human resource costs after scheduling adjustments), and user feedback (e.g., the degree of satisfaction recovery) after the execution of each strategy. Based on reinforcement learning algorithms, it uses these indicators as reward signals to dynamically optimize the classification threshold and strategy weights of the decision tree in the anomaly decision-making node. At the same time, it feeds back the optimized rule parameters to the rule base of the first subsystem (e.g., updating the cyclic collection parameters), the classification and handling node network of the second subsystem (e.g., adjusting the strategy node weights), and the resource allocation module of the third subsystem, forming a closed-loop control. Finally, it outputs a set of associated anomaly adjustment strategies that include specific execution objects (e.g., doctors, nurses, and equipment in specific departments), operation procedures (e.g., training course arrangements, scheduling adjustment steps), and effect evaluation standards (e.g., the anomaly repair rate must reach the threshold, and the satisfaction level must recover to the threshold).

[0179] The anomaly adjustment module is used to adjust the anomaly node information located in the first subsystem and the second subsystem in real time according to the associated anomaly adjustment strategy and the anomaly decision node, reinforcement adjustment node and historical anomaly adjustment index channel information in the first subsystem chain, with the goal of minimizing delay and maximizing anomaly adjustment rate and adjustment speed, until the preset evaluation index threshold of the corresponding anomaly node is met.

[0180] The assessment and training module is used to train personnel based on the anomalies located by the first subsystem and the technical anomalies of the personnel located by the second subsystem. It then uses the trained personnel in each subsystem to perform simulations using a three-level model chain and simulation algorithms to obtain the anomaly elimination rate, personnel completion rate, user satisfaction with the personnel, and performance score for each management level. Based on these metrics, and combined with preset threshold values, the module performs cyclical training simulations until the personnel completion rate, user satisfaction with the personnel, and performance score for each management level meet the preset threshold values.

[0181] It should be further explained that the process corresponding to the three-level management mode in this embodiment includes: the first-level management mode generates a first user data packet with hash label through the initial screening module and the main complaint identification module, and automatically uploads it to the distributed database managed by the second-level management mode in combination with the blockchain smart contract to complete data initialization and trusted evidence storage; the second-level management mode matches the first user data packet with a handling strategy based on the classification handling node network and fast index through the first handling module, and after the first evaluation module and the escalation judgment module judge, the cases to be escalated are constructed into second user data packets, and fed back to the third-level management mode through the same blockchain on-chain channel; the third-level management mode matches and handles the second user data packet through the escalation handling matching module, and uses the anomaly analysis decision module to integrate the real-time logs and anomaly feature information of the first subsystem and the second subsystem to generate associated anomaly adjustment strategies. Finally, the anomaly adjustment module performs real-time parameterized adjustment of resources and strategies for the anomaly nodes in the first subsystem or the second subsystem based on the historical index channel and the enhanced adjustment node, until they meet the preset evaluation index threshold.

[0182] It should be further explained that the specific workflow of the three-level pattern chain in this embodiment includes:

[0183] S1. Based on the multimodal input interface of the first subsystem, the user's original complaint data is obtained through the embedded voice acquisition device and the mobile terminal form. The pre-trained dialect recognition model is used to perform spectrum analysis and speech-to-text processing on the voice signal. Then, standardized medical term keywords are extracted from the text through the entity extraction algorithm to obtain a set of structured user target representation information keywords.

[0184] S2. Based on the preset screening rules and keyword weight table, the initial evaluation score is calculated by the weighted scoring algorithm. When the score exceeds the preset threshold or hits the preset high-risk keyword, the data packet assembly process is triggered. The user identifier, keyword set and evaluation score are serialized according to the preset format. The data packet digital fingerprint is generated by the hash algorithm. Then, the data packet is encrypted by the asymmetric encryption algorithm to obtain the encrypted first user data packet.

[0185] S3. Based on the predefined data upload rules in the blockchain smart contract, the first subsystem digitally signs the encrypted first user data packet using the node's private key, combines the signature data, node certificate, and encrypted first user data packet into a transaction request, and broadcasts it to the blockchain network.

[0186] S4. Based on the practical Byzantine fault-tolerant consensus mechanism, the consensus node verifies the legality of the digital signature and the node's authority, calculates the hash value of the data packet and compares it with the digital fingerprint. After the verification is successful, the transaction record is written into a new block and appended to the chain after verification by a majority of nodes, thus completing the trusted data storage.

[0187] S5. The on-chain data monitoring module based on the second subsystem subsystem obtains encrypted data packets in the new block in real time by subscribing to data arrival events in the smart contract. After decrypting the data packets with the node's private key, it uses a multi-dimensional feature index matching and processing strategy for classifying and processing node networks.

[0188] S6. The execution process of the disposal strategy is recorded, including the execution time, operators, disposal parameters and effect evaluation indicators, forming a disposal log. This log is written to the blockchain through a smart contract after being hashed.

[0189] S7. If the effect evaluation does not reach the preset threshold, the smart contract will automatically trigger the upgrade rule, assemble the case identifier, initial data packet hash, treatment log and evaluation results into a second user data packet, and send it to the third subsystem of the tertiary hospital after re-encryption.

[0190] S8. The escalation and matching module based on the third subsystem of a tertiary hospital parses the second user data packet and calls the expert matching algorithm to generate a treatment plan. The treatment results and process data are synchronized to the blockchain.

[0191] S9. Based on the full on-chain log data, identify abnormal patterns in lower-level subsystems through time-series analysis and association rule mining algorithms, including resource allocation deviations, execution delays, and skill deficiencies.

[0192] S10. Based on the identification results, generate resource adjustment instructions, training plans or performance adjustment schemes through the dynamic rule engine, and issue the instructions to the corresponding subsystems through smart contracts;

[0193] S11. The execution effect of the instruction is quantitatively evaluated and then fed back to the blockchain to form a closed-loop control.

[0194] All data transmission between subsystems uses encrypted channels, and keys are dynamically distributed through the blockchain key management module; all data operations and status changes at all levels are recorded on the blockchain to ensure that operations are tamper-proof and fully traceable.

[0195] It should be further explained that the consensus node verification process in this embodiment includes: first, verifying the legality of the digital signature and node certificate; then, calculating the hash value of the received encrypted data packet; comparing the hash value with the digital fingerprint carried in the data packet to verify the data integrity; after all verifications are passed, the consensus node packages the transaction records into a new block, completes the consensus process based on the practical Byzantine fault-tolerant consensus algorithm, and appends the new block to the blockchain.

[0196] It should be further explained that the process of the smart contract triggering the upgrade rule in this embodiment includes: the smart contract continuously monitors the effect evaluation indicators written in the disposal log on the blockchain. When it finds that the effect evaluation indicators of a specific case have not reached the preset threshold within a preset time window, it automatically extracts the initial data packet hash, all relevant disposal log hashes and evaluation results of the case from the blockchain storage, assembles them into a second user data packet, encrypts it using the public key of the third subsystem, and finally sends an upgrade request event to the third subsystem.

[0197] It should be further explained that the abnormal pattern recognition process in this embodiment includes: obtaining the full operation log from the blockchain, extracting feature data including resource allocation time, execution timestamp, and skill certification records, using time series analysis algorithms to detect resource allocation deviations and execution delays, and using association rule mining algorithms to analyze the correlation between skill defects and handling effects to identify abnormal patterns in lower-level subsystems.

[0198] It should be further explained that the process of the dynamic rule engine generating the adjustment scheme in this embodiment includes: according to the identified abnormal pattern type, the dynamic rule engine calls the corresponding resource allocation rules, training management rules or performance evaluation rules to generate an adjustment scheme that includes specific parameter adjustments, training course allocation or performance coefficient modification, and sends the scheme to the corresponding subsystem for execution through the smart contract call interface.

[0199] It should be further explained that the closed-loop control process in this embodiment includes: after receiving the adjustment instruction, the lower-level subsystem executes the corresponding operation and converts the effect data after execution into a transaction form and writes it into the blockchain; the smart contract monitors these effect data again and compares them with the expected target. If the expected effect is still not achieved, a new round of analysis and adjustment process is triggered until the abnormal mode is eliminated.

[0200] In this embodiment, the first treatment module of the second subsystem relies on a classified treatment node network and a fast index. It extracts independent treatment strategies from a historical case database to construct a node set, establishes a weighted edge network topology using a directed graph model, and builds a multi-condition retrieval index based on a multi-dimensional feature index system. Combined with a fast index based on a hash algorithm, it achieves rapid location of strategy nodes. This allows for accurate matching of disease types with corresponding departments and medical personnel, such as matching postoperative incision pain with anesthesiology analgesia and orthopedic rehabilitation guidance strategies. It also records the average treatment rate of doctors in real time, providing data support for resource allocation and performance evaluation in the third subsystem, effectively avoiding mismatched treatment strategies and improving pain management efficiency. The urgency level of treatment is assessed through a multi-dimensional evaluation index database (keywords, symptoms, etc.). The system employs a weighted calculation and dynamic adjustment based on factors such as duration of symptoms, response time, and risk of complications. This allows for precise prioritization of treatment based on the patient's real-time pain status, preventing delays in treatment for high-urgency patients and excessive resource consumption for low-urgency patients, thus ensuring a rational allocation of treatment resources. The first assessment module calculates post-treatment evaluation scores and relief rates in real time. The escalation judgment module triggers an escalation mechanism based on the relief rate, promptly identifying ineffective treatment cases and securely feeding patient data back to the third subsystem. This prevents treatment delays due to insufficient primary care capacity. Simultaneously, the first satisfaction feedback can be directly linked to medical staff performance evaluations, improving the service quality of medical personnel. The second analysis module extracts features from the escalation case treatment logs and locates strong correlation anomalies through causal association analysis algorithms. The system combines various approaches, including a classification and treatment node network with pre-set normal feature comparison deviations, to accurately identify anomalies such as strategy parameter mismatch and execution delays, along with associated nodes. This provides precise node-level data for anomaly adjustments in the third subsystem, reducing the recurrence of similar anomalies. The third subsystem's escalation treatment matching module, based on a tertiary hospital pain management strategy library, can match professional solutions (such as multidisciplinary consultation strategies) to complex pain cases where primary care treatment is ineffective, significantly improving the success rate of treating difficult pain cases. The escalation assessment module uses a cyclical matching strategy until the assessment score reaches the target, ensuring patients ultimately receive effective treatment and preventing abandonment of intervention due to a single inappropriate strategy, thus guaranteeing patient treatment outcomes. The anomaly analysis and decision-making module employs an improved API. The Riori algorithm mines anomalous feature association rules, the dynamic time warping algorithm quantifies the spatiotemporal differences of anomalies, and the weighted fuzzy logic fuses multi-dimensional indicators to generate comprehensive evaluation results. Combined with a decision tree classifier to match predefined adjustment strategies, it can comprehensively identify anomalies across subsystems, such as lack of medical staff skills, imbalanced resource allocation, equipment failure, drug shortage, insufficient patient compliance, node logic conflicts, model recognition errors, and on-chain transmission delays. It can also generate targeted solutions, such as triggering tiered training for lack of skills, intelligent scheduling for imbalanced resource allocation, preventive maintenance for equipment failure, inventory optimization for drug shortage, and multimodal education for insufficient compliance. This effectively solves operational obstacles in each link and improves the overall stability of the system.The anomaly adjustment module, relying on enhanced adjustment nodes and historical index channels, optimizes anomalies in lower-level subsystems with the goal of minimizing latency and maximizing adjustment rate. It can quickly correct issues such as parameter deviations in the first subsystem and strategy node anomalies in the second subsystem, reducing the impact of anomalies on patient management procedures. The assessment and training module combines a three-level model chain with simulation algorithms to conduct cyclical training. Training effectiveness is verified through indicators such as anomaly elimination rate, personnel completion rate, and user satisfaction. This not only specifically improves the anomaly resolution capabilities of medical staff at different levels (e.g., screening skills for village-level medical staff and treatment skills for township-level medical staff), but also reduces the risk of operational errors through simulation. Combined with a performance incentive mechanism, this creates a virtuous cycle, continuously improving the professional level of the medical team. The three-level model chain utilizes technologies such as the practical Byzantine fault-tolerant consensus algorithm, asymmetric encryption, and TLS encrypted transmission. This system enables trusted collaboration across the entire process, from encrypted on-chain data collection in the village-level subsystem, log storage in the township-level subsystem, to escalation and adjustment instructions from the tertiary hospital subsystem. Data transmission is encrypted throughout and operations are traceable, preventing data tampering and leakage, protecting patient privacy, and ensuring smooth data synchronization across subsystems. For example, the first user data package from the village level is securely synchronized to the township subsystem, and the second user data package from the township level is accurately fed back to the tertiary hospital subsystem. Simultaneously, the smart contract triggers escalation rules, the dynamic rule engine generates adjustment plans, and the closed-loop control continuously optimizes the process design for anomalies. This ensures that anomalies are detected promptly, resolved accurately, and continuously improved. For instance, escalation rules automatically identify and handle ineffective cases, the effects of adjustment plans are fed back to the blockchain, and if expectations are not met, a new round of optimization is triggered, forming a closed-loop management system for the entire process. In summary, the modules in this embodiment work synergistically through technical means to achieve comprehensive management of pain patients, from precise enrollment and screening, efficient primary care, timely escalation of intervention, precise adjustment of abnormalities, professionalization of medical and nursing capabilities, and reliable data flow. This effectively improves the quality of pain management services, reduces resource waste, protects patients' treatment rights, and provides stable, safe, and efficient technical support for cross-level medical collaboration, promoting the standardized development of pain management systems.

[0201] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

[0202] If the technical solution disclosed herein involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of explicit consent. For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. An information monitoring and circulation platform supporting a three-tier management model, characterized in that, include: The three-level model chain is constructed by combining the first subsystem, the second subsystem, and the third subsystem with blockchain algorithms; The first subsystem is used to collect different types of user target representation information, and to filter and classify the user target representation information according to preset data filtering rules; At the same time, based on the filtered user target representation information and preset key filtering fields, the preset push strategy library is called to perform information push and fixed-point cyclic collection operations. The second subsystem performs hierarchical classification cyclical evaluation and disposal based on the user target representation information stored in the classification system, combined with the preset hierarchical classification evaluation criteria and disposal strategy library; when the disposal result within the preset time does not reach the preset evaluation threshold, the corresponding user information and disposal strategy information are recorded. The third subsystem executes a preset escalation assessment process based on the recorded user information and handling strategy information combined with the deep Q network; at the same time, it performs anomaly analysis based on the escalation assessment results and the log records collected in real time by the first and second subsystems; and adjusts the resources of the first and second subsystems according to the anomaly analysis results until the preset assessment indicator thresholds of each subsystem are met.

2. The information monitoring and transfer platform supporting a three-tier management model as described in claim 1, characterized in that, The first subsystem includes a chief complaint identification module and an initial screening module; The chief complaint recognition module is used to extract and correct user speech text information according to a preset dialect recognition model, and to identify and extract user target representation information based on the corrected user speech text information combined with an entity extraction algorithm to obtain a set of user target representation information keywords; the data filtering rules are constructed by combining an initial filtering score threshold with a filtering keyword library and a matching algorithm. The initial screening module is used to evaluate user target representation information and screen keywords based on the acquired user target representation information keyword set and a comprehensive evaluation algorithm. When the initial evaluation score of the corresponding user target representation is greater than the preset initial screening score threshold or a screening keyword is matched, a first user data package is constructed based on the user basic information, user target representation information keyword set and initial evaluation score that meet the conditions. The first user data package is then hashed using a hash algorithm and uploaded to the distributed database corresponding to the second subsystem through the smart contract in the three-level model chain for classified storage.

3. The information monitoring and transfer platform supporting a three-tier management model as described in claim 2, characterized in that, The first subsystem also includes an information push module, a cyclic collection and adjustment module, and a first analysis module; The information push module is used to push first strategy information based on the user target representation information keyword set, the user status information after evaluation and processing corresponding to the second subsystem and the third subsystem, the final evaluation score and the evaluation score change trend, and in combination with the preset push strategy library, through a matching algorithm. The cyclic collection and adjustment module is used to adjust the time interval and frequency of cyclic collection of target users in real time according to the round of evaluation and treatment of the corresponding user after evaluation and treatment, the evaluation score of each target characterization, the trend of evaluation score change and the screening keywords. It also obtains the cyclic screening evaluation result based on the cyclic collection information. If the cyclic screening evaluation result meets the data screening rules again, the corresponding user information is uploaded and saved again. If it does not meet the rules, the corresponding user information is deleted from the distributed database. The first analysis module is used to record in real time the data collection, filtering, strategy push and cyclic collection process information corresponding to the first subsystem, and to perform horizontal anomaly analysis and location on the preset first subsystem chain in combination with the preset node anomaly analysis model to obtain the first anomaly feature information, and to feed back the first anomaly feature information to the preset third subsystem chain in the third subsystem through the three-level mode chain.

4. The information monitoring and transfer platform supporting a three-tier management model as described in claim 3, characterized in that, The second subsystem includes a first processing module; The first processing module is used to match the corresponding type processing strategy based on the user type and user target representation information keywords in the labeled first user data packet, combined with a preset classification processing node network and a preset fast index, and to process the corresponding user using the matching processing strategy. It also collects, records and preprocesses the user target representation information keywords within a preset time length after the current user is processed, and records the average processing rate of the current processing strategy within the preset time length. Each node in the classification processing node network corresponds to a processing strategy. The preset fast index is constructed by combining the user type, skill level, index frequency, evaluation score decrease of real-time user target representation information after the corresponding handling strategy, the number of rounds and probability of re-evaluation after the current handling strategy, and the evaluation satisfaction rate after handling with a hash algorithm.

5. An information monitoring and transfer platform supporting a three-tier management model as described in claim 4, characterized in that, The second subsystem also includes a first evaluation module and an upgrade discrimination module; The first evaluation module is used to perform real-time evaluation based on the keywords of the user target representation information after treatment, combined with preset evaluation indicators and evaluation algorithms, to obtain the first evaluation score and the trend of the first evaluation score change of the real-time user target representation information, and at the same time, based on the trend of the first evaluation score change, to obtain the relief rate of the target user within a preset time period. The escalation judgment module is used to perform escalation judgment based on the target user's remission rate within a preset time period and preset escalation judgment conditions. Specifically: When the mitigation rate of the target user within a preset time period meets the preset escalation judgment condition, the first user data packet corresponding to the target user and the second user data packet constructed by combining the information of each round of handling strategy with the hash algorithm are fed back to the preset third subsystem chain in the third subsystem through the three-level mode chain. If the relief rate of the target user within the preset time period does not meet the preset escalation criteria, the first satisfaction level of the corresponding target user with the current treatment strategy information is evaluated and fed back to the third subsystem and the first subsystem through the three-level mode chain.

6. The information monitoring and transfer platform supporting a three-tier management model as described in claim 5, characterized in that, The second subsystem further includes a second analysis module; the second analysis module is used to perform anomaly analysis and localization on the handling strategies of corresponding users in the classification and handling node network that meet the upgrading judgment conditions within the corresponding preset time length, through causal association analysis algorithm, to obtain second anomaly feature information, and to feed back the second anomaly feature information to the third subsystem through the three-level pattern chain.

7. The information monitoring and transfer platform supporting a three-tier management model as described in claim 6, characterized in that, The third subsystem includes an upgraded handling matching module and an upgraded evaluation module; The escalation handling matching module is used to obtain the corresponding escalation handling strategy and matching accuracy by combining the second user data packet with the escalation handling strategy library configured by the third subsystem through a matching algorithm, and to collect the second user status information after the corresponding user is handled according to the escalation handling strategy in real time. The upgrade evaluation module is used to obtain a second evaluation score after upgrade processing based on the second user status information and the second evaluation algorithm. When the second evaluation score is less than the preset processing completion evaluation threshold, the processing of the corresponding user ends. At the same time, based on the evaluation information of the corresponding user on the matched upgrade processing strategy during the upgrade processing, the second satisfaction rate of the current user is obtained, and the second satisfaction rate of the corresponding upgrade processing strategy is fed back to the upgrade processing strategy library to adjust the index information of the current upgrade processing strategy in real time. When the second evaluation score is greater than or equal to the preset disposal completion evaluation threshold, the third subsystem presets the abnormal decision node in the first subsystem chain to perform cyclic matching evaluation of the escalation disposal strategy until the second evaluation score is less than the preset disposal completion evaluation threshold, and records the corresponding cycle number and the corresponding escalation disposal strategy information that does not meet the condition.

8. The information monitoring and transfer platform supporting a three-tier management model as described in claim 7, characterized in that, The third subsystem also includes an anomaly analysis and decision-making module and an anomaly adjustment module; The anomaly analysis and decision-making module is used to obtain anomaly feature association information and difference information by combining the first anomaly feature information and the second anomaly feature information with an association algorithm. Based on the anomaly feature association information and difference information, the average handling rate of the current handling strategy within a preset time length, the first satisfaction rate, and the relief rate of the target user within a preset time length, and combined with the anomaly decision-making nodes in the preset third subsystem chain of the third subsystem, an associated anomaly adjustment strategy is obtained. The anomaly adjustment module is used to adjust the anomaly node information located in the first subsystem and the second subsystem in real time according to the associated anomaly adjustment strategy and the anomaly decision node, reinforcement adjustment node and historical anomaly adjustment index channel information in the first subsystem chain, with the goal of minimizing delay and maximizing anomaly adjustment rate and adjustment speed, until the preset evaluation index threshold of the corresponding anomaly node is met.

9. An information monitoring and transfer platform supporting a three-tier management model as described in claim 8, characterized in that, The first subsystem executes the first-level management mode, the second subsystem executes the second-level management mode, and the third subsystem executes the third-level management mode. The first-level management mode generates a first user data packet with hash label through the initial screening module and the chief complaint identification module, and automatically uploads it to the distributed database managed by the second-level management mode in conjunction with the blockchain smart contract; The second-level management mode uses the first disposal module to match the disposal strategy for the first user data packet based on the classification disposal node network and fast index. After the first evaluation module and the escalation judgment module make judgments, the cases to be escalated are constructed into second user data packets and fed back to the third-level management mode through the same blockchain on-chain channel. The third-level management mode uses the upgraded handling matching module to match and handle the second user data packets, and uses the anomaly analysis and decision module to integrate the real-time logs and anomaly feature information of the first and second subsystems to generate associated anomaly adjustment strategies. The anomaly adjustment module then uses historical index channels and enhanced adjustment nodes to perform real-time parameterized adjustments to the anomaly nodes in the first or second subsystems based on historical index channels and enhanced adjustment nodes, until they meet the preset evaluation index thresholds.

10. An information monitoring and transfer platform supporting a three-tier management model as described in claim 9, characterized in that, The construction process of the classification and disposal node network includes: Based on the historical case database, an initial set of nodes is obtained by extracting all independent handling strategies and assigning a unique node identifier to each strategy. Each node stores the complete content, applicable conditions and execution parameters of the corresponding handling strategy. Based on the logical correlation between disposal strategies and the frequency of historical joint use, a directed graph model is used to construct the connection edges between nodes. Each edge is assigned an initial weight according to the historical disposal evaluation score, disposal success rate, conversion time length and the number of rounds of reprocessing after disposal, so as to obtain a disposal strategy network topology with weighted relationships. Based on a multi-dimensional feature index system, a retrieval entry point is constructed for each node through a hybrid index structure. The index key includes user type, target representation information keyword weight, urgency level of handling, resource demand type, historical handling success rate and average handling rate within a preset time period, to obtain a strategy index that supports multi-condition combined queries. Based on real-time feedback data after real-time processing, the node index weight and edge relationship weight are adjusted through simulation algorithm, and the failure mode is located by mining association rules to obtain the classified processing node network. Based on the blockchain's need for evidence storage, the structure hash and decision log of the classification and disposal node network are synchronized to the chain through smart contracts to obtain disposal strategy records. These records are then uploaded to the abnormal decision nodes for real-time abnormal simulation decision adjustments until the corresponding abnormal adjustment is completed.

Images