Intelligent Processing Method and System for Public Health Physical Examination Follow-up Home Signature Data Based on Multi-device Acquisition

Abstract

This invention provides an intelligent processing method and system for public health physical examination follow-up family signature data based on multi-device acquisition. The method includes: collecting raw data streams from multiple vendors' devices via an edge gateway; unifying fields, types, and units through protocol parsing, regular expression and syntax analysis; and promoting automatic semantic alignment of cross-device fields through a dynamic semantic benchmark template. This method employs a lightweight encoder-decoder and contrastive learning mechanism to construct a shared embedding space, achieving device-level semantic matching. It integrates contextual consistency verification and incremental knowledge distillation to support continuous model optimization with new devices and clinical feedback. This invention effectively solves the problem of inconsistent structure and semantics in multi-source medical data, improves the standardization and label consistency capabilities of health data, and promotes the automated application of intelligent risk assessment and follow-up management.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing technology for medical and health data, and in particular to a method and system for intelligent processing of public health physical examination follow-up family signature data based on multi-device collection. Background Technology

[0002] Currently, in primary healthcare information technology scenarios such as public health checkups and family doctor contract services, the business needs of health management and chronic disease follow-up require the integration and analysis of health data collected from multiple heterogeneous medical devices such as blood pressure monitors, blood glucose meters, and electrocardiographs. Existing data processing systems typically rely on static ontology libraries, hard-coded standard mapping tables, or manually maintained field mapping rules to convert raw data from multiple devices into structured health tags. These solutions can meet the tag standardization needs in environments with a small number of devices and a single vendor, and support the standardized naming and preliminary classification of some common health indicators; With the rapid increase in the types and brands of medical devices on the market, significant inconsistencies exist in the output data of devices from different manufacturers in terms of field naming, data structure, units of measurement, and numerical precision. Currently, mainstream industry practices, such as data interface solutions based on international HL7 standards and DICOM specifications, mostly focus on interoperability at the communication protocol level, while semantic layer label consistency still requires manual maintenance of terminology mapping and table structure adjustments. Other studies have attempted to use algorithms such as word vectors and clustering for semantic grouping and label normalization, partially achieving automated label alignment and generation. However, these generally rely on prior static semantic ontologies and limited preset field mapping tables, limiting their scalability to new connected devices and new health indicators. Their maintenance costs increase exponentially with the increase in devices. Especially in actual public health data scenarios, near-synonymous heterogeneous labels such as 'hypertension' and 'BP_High', and 'systolic blood pressure' and 'SystolicPressure' frequently appear, leading not only to label duplication, semantic ambiguity, and inconsistent statistical definitions, but also significantly impacting the accuracy of downstream intelligent analysis modules such as risk assessment and chronic disease management. Representative technologies in the field of intelligent processing of current medical and health data include: rule-based label normalization methods, automatic term indexing based on standard ontology dictionaries, and a small number of simple field synonym normalization techniques based on pre-trained language models. These solutions are mostly suitable for local scenarios with limited data scale, stable device access, and low demand for term expansion. However, for complex heterogeneous device environments, especially in the stage of public health and community health information technology upgrades, it is difficult to guarantee the continuous consistency and semantic scalability of label generation. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method and system for intelligent processing of public health physical examination follow-up family signature data based on multi-device collection.

[0004] The technical solution of this invention is implemented as follows: A method for intelligent processing of public health physical examination follow-up home signature data collected from multiple devices, comprising: S1: Based on the data collection needs of multi-source heterogeneous medical devices in the scenario of public health physical examination and family doctor contract service, the system acquires raw data streams from different manufacturers such as blood pressure monitors, blood glucose meters, and electrocardiographs in real time, and records the data format characteristics, field naming patterns and unit expression methods of each device to generate the raw dataset on the device side. S2: Perform structured parsing and type normalization on the original dataset on the device side, and map the numerical, textual and time series fields to a unified intermediate representation format to generate standardized data units, which serve as the basic input for subsequent semantic analysis; S3: Based on the current business task objectives, identify the set of key health indicators required for chronic disease risk assessment, select the electronic health record system from all connected devices as the initial semantic benchmark source, and generate a dynamic semantic benchmark template because it has the highest semantic coverage and clinical authority. S4: Input the standardized data units of non-benchmark devices into the lightweight encoder-decoder structure, and project their field semantics to a shared embedding space aligned with the dynamic semantic benchmark template through a contrastive learning mechanism to generate cross-device field-level semantic alignment vectors; S5: Based on the cross-device field-level semantic alignment vector, a dual-channel alignment network is constructed, wherein the first channel performs semantic similarity matching between field names and baseline labels, and the second channel combines contextual record information to perform consistency verification and generate a preliminary consistent label cluster. S6: When a new device is connected, the existing shared embedding space is used as a teacher model to guide the learning process of the new device's feature extractor. The boundary definition of the shared semantic space is updated through an incremental knowledge distillation mechanism to generate an expanded global semantic alignment model. S7: Deploy the extended global semantic alignment model on the edge inference module, perform localized label generation operation on the multi-source data flowing in in real time, and output a structured health label sequence with semantic consistency; S8: Monitor clinical staff's feedback behavior on the output label sequence, transform the correction operation into a fine-tuning signal and inject it into the incremental knowledge distillation process to continuously optimize the consistency between the dynamic semantic benchmark template and the shared semantic space.

[0005] The present invention also provides an intelligent processing system for public health physical examination follow-up family signature data based on multi-device acquisition, which uses the above-mentioned intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition to perform intelligent processing of public health physical examination follow-up family signature data.

[0006] The intelligent processing method and system for public health physical examination follow-up family signature data based on multi-device acquisition provided by this invention has the following beneficial effects: (1) This invention significantly improves the semantic consistency of multi-source health data at the field and record levels by introducing a task-oriented dynamic semantic benchmark mechanism. The data source with the strongest information carrying capacity is used as the initial semantic anchor point, and the benchmark selection is dynamically adjusted based on the real-time monitored semantic coverage and field activity to ensure that the semantic reference system always closely matches the actual data characteristics of the current business scenario. Combined with a lightweight encoder-decoder structure and contrastive learning strategy, the system can automatically project the original fields of different manufacturers' devices (such as "hypertension" and "BP_High") to a unified semantic space, effectively avoiding the label redundancy and logical contradictions caused by naming differences. Thus, without changing the existing hardware configuration, the availability and structuring level of public health physical examination and family doctor contract service related data in primary medical institutions are greatly improved. (2) This invention employs an incremental knowledge distillation mechanism and a lightweight inference architecture at the edge, achieving continuous semantic alignment capabilities with low latency and low resource consumption. Whenever a new device is connected, the system does not need to retrain the global model. Instead, it uses the existing shared semantic space as a teacher model to guide the learning process of the new device's feature extractor and simultaneously updates the definition range of the semantic boundary, greatly reducing the cost and timeliness bottleneck of system iteration. At the same time, the dual-channel alignment network design handles field-level semantic mapping and record-level contextual association respectively, taking into account both local accuracy and global consistency. Combined with the locally deployed lightweight inference module, it achieves efficient and real-time label generation, which is suitable for grassroots medical environments with weak IT operation and maintenance capabilities. In addition, by establishing a clinical feedback closed loop, the correction operations of medical staff on label results are transformed into fine-tuning signals, which in turn enhance the accuracy and robustness of the semantic alignment model, forming a self-evolving and adaptive intelligent data governance closed loop system. Attached Figure Description

[0007] Figure 1 This is a flowchart of the intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition according to the present invention. Figure 2 This is a sub-flowchart of the intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition according to the present invention; Figure 3 This is another sub-flowchart of the intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition according to the present invention. Detailed Implementation

[0008] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0009] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0010] like Figure 1 As shown, this invention provides an intelligent processing method for public health physical examination follow-up home signature data based on multi-device acquisition, specifically including: S1: Based on the data collection needs of multi-source heterogeneous medical devices in the scenario of public health physical examination and family doctor contract service, the system acquires raw data streams from different manufacturers such as blood pressure monitors, blood glucose meters, and electrocardiographs in real time, and records the data format characteristics, field naming patterns and unit expression methods of each device to generate the raw dataset on the device side. S2: Perform structured parsing and type normalization on the original dataset on the device side, and map the numerical, textual and time series fields to a unified intermediate representation format to generate standardized data units, which serve as the basic input for subsequent semantic analysis; S3: Based on the current business task objectives, identify the set of key health indicators required for chronic disease risk assessment, select the electronic health record system from all connected devices as the initial semantic benchmark source, and generate a dynamic semantic benchmark template because it has the highest semantic coverage and clinical authority. S4: Input the standardized data units of non-benchmark devices into the lightweight encoder-decoder structure, and project their field semantics to a shared embedding space aligned with the dynamic semantic benchmark template through a contrastive learning mechanism to generate cross-device field-level semantic alignment vectors; S5: Based on the cross-device field-level semantic alignment vector, a dual-channel alignment network is constructed, wherein the first channel performs semantic similarity matching between field names and baseline labels, and the second channel combines contextual record information to perform consistency verification and generate a preliminary consistent label cluster. S6: When a new device is connected, the existing shared embedding space is used as a teacher model to guide the learning process of the new device's feature extractor. The boundary definition of the shared semantic space is updated through an incremental knowledge distillation mechanism to generate an expanded global semantic alignment model. S7: Deploy the extended global semantic alignment model on the edge inference module, perform localized label generation operation on the multi-source data flowing in in real time, and output a structured health label sequence with semantic consistency; S8: Monitor clinical staff's feedback behavior on the output label sequence, transform the correction operation into a fine-tuning signal and inject it into the incremental knowledge distillation process to continuously optimize the consistency between the dynamic semantic benchmark template and the shared semantic space.

[0011] Step S1: Based on the data acquisition needs of multi-source heterogeneous medical devices in public health check-up and family doctor contract service scenarios, real-time acquisition of raw data streams from devices from different manufacturers, such as blood pressure monitors, blood glucose meters, and electrocardiographs, is performed. The data format characteristics, field naming patterns, and unit expression methods of each device are recorded to generate the device-side raw dataset. Specifically, this includes: S1.1: Based on the multi-source heterogeneous medical device access environment, the edge gateway module is used to concurrently capture the raw data streams of devices from different manufacturers such as blood pressure monitors, blood glucose meters, and electrocardiographs, in order to establish a data acquisition channel covering all device types and output the initial raw data stream set. Based on the access environment of connected blood pressure monitors, blood glucose meters, electrocardiographs and other heterogeneous medical devices, a multi-channel concurrent capture mechanism of the edge gateway module (parameters: thread pool size of 1-64, buffer queue capacity of 1-4096, I / O interrupt priority) is adopted to realize synchronous monitoring and high-speed buffering of the raw data stream of each device. Furthermore, by using the device adapter driver management component (parameters: manufacturer identification code, device model mapping table, interface type enumeration) to call the underlying I / O protocol stack for different manufacturers' communication interfaces, data packet mirroring capture of various physical links such as serial port, USB, Bluetooth or Ethernet can be achieved, while recording the capture timestamp and device identifier. Furthermore, a data stream fragmentation algorithm (parameters: fragmentation length threshold, reassembly strategy, transmission window size) is used to fragment the original data stream captured by each channel, splitting long messages into a set of structurally complete fragments, and establishing an index relationship with the channel identifier according to the capture order; Furthermore, a high-precision clock synchronization algorithm (parameters: NTP synchronization period, time drift correction coefficient) is used to uniformly adjust the time base of the data streams of each device, so as to achieve time consistency of multi-source data streams at the acquisition layer, in order to support time window analysis in the downstream semantic alignment process; Furthermore, through anomaly detection and packet loss retransmission mechanism (parameters: CRC check code length, maximum number of retransmissions, packet loss threshold), distorted segments caused by link interference or transmission errors during the acquisition process are eliminated, and targeted retransmission requests are initiated to the source device to ensure the integrity of the final output data stream set; Through the above-mentioned multi-channel concurrent capture, device driver adaptation, fragment management, time synchronization and data verification processing methods, the captured signal in the previous step is transformed into a structured set of initial raw data streams with time consistency and integrity guarantees, thereby achieving reliable data acquisition technology for all types of devices. For example, in a primary healthcare institution, the edge gateway module is configured with a thread pool of 16 and a buffer queue capacity of 1024 data packets. I / O interrupt priority is set to the highest level to ensure real-time capture even with concurrent device access. For blood pressure monitors using the USB interface protocol, blood glucose meters using the Bluetooth SPP protocol, and electrocardiographs using the Ethernet TCP / IP protocol, the corresponding adapter drivers are accurately loaded using a device model mapping table. During capture, each data packet is stamped with a millisecond-level timestamp and device ID. During data stream fragmentation, a fragment length threshold of 512 bytes and a transmission window size of 64KB are set, allowing long messages to be broken into easily stored and indexed small segments without losing field integrity. Time base correction is performed every 500 milliseconds via NTP service, with a correction coefficient of 0.0001 seconds, to eliminate time drift between different devices. Each data segment is verified using the CRC-32 algorithm, and lost packets are recovered using a maximum of three retransmissions. In this scenario, the final output initial raw data stream set contains complete, error-free, and time-standardized data fragments from three types of devices, providing a stable data foundation for subsequent protocol parsing and semantic feature extraction; S1.2: Perform protocol parsing operation on the initial raw data stream set, decompose the messages according to the communication protocol type of each device (such as HL7, DICOM, custom serial port protocol), extract the payload field, and generate raw data fragment groups divided by device granularity; For the initial raw data stream set, a protocol type identification algorithm (parameters: device message header information, predefined protocol feature library) is used to classify and process the protocol of the data transmitted by devices such as blood pressure monitors, blood glucose meters, and electrocardiographs. Furthermore, by using a layered message decomposition method (parameters: protocol parsing rule set, segment buffer size), the message header, message body, and checksum are separated, and the corresponding parser module is called to extract fields from the payload area for different protocol types. Furthermore, the extracted payload fields are validated using a field validity determination algorithm (parameters: field length range, data type constraints, and check bit results), and the validated fields are retained for subsequent semantic feature recognition. Furthermore, by using a device-level grouping processing method (parameters: device unique identifier, timestamp alignment window), the legitimate payload fields are archived according to the source device, generating a set of original data fragments subdivided down to a single device; By using protocol parsing and packet processing strategies, the results of the previous step are transformed into structured raw data fragment groups, enabling unified disassembly and device-level classification of data from devices with different communication protocols. For example, for the initial raw data stream set from three types of devices, where the blood pressure monitor uses the HL7 protocol with a total message length of 256 bytes; the blood glucose meter uses the DICOM protocol with a total message length of 512 bytes; and the electrocardiograph uses a custom serial port protocol with a total message length of 128 bytes, during protocol type identification, the protocol identifier code in the first three bytes of the message header is matched with the protocol feature library to determine that the blood pressure monitor message type ID is HL7-OBX, the blood glucose meter message type ID is DICOM-WAVEFORM, and the electrocardiograph message type ID is CUS-UART. During the message decomposition stage, for HL7 protocol devices, a layered parser based on XML Schema is called, with a CRC16 checksum; for DICOM protocol devices, a multi-segment binary slice parser is called, with a CRC32 checksum; and for custom protocol devices, a state machine-driven serial port parser is called, with an XORCheck checksum. In the field extraction stage, data type constraints were applied (numerical fields must not exceed 8 bytes in length, and text fields must be UTF-8 encoded), and erroneous fields were discarded based on the checksum result. Ultimately, 12 valid fields were retained for the blood pressure monitor, 18 for the blood glucose meter, and 45 for the electrocardiograph. In the device-level grouping stage, the unique device identifier and a unified millisecond-level timestamp were used as indexes to categorize and encapsulate the above fields into raw data fragment groups according to the device. The blood pressure monitor fragment file size was 2KB, the blood glucose meter fragment file size was 4KB, and the electrocardiograph fragment file size was 6KB. The results of this step show that the data fragments after protocol parsing and device grouping can directly enter the naming pattern recognition process in S1.3 without further message layer verification, significantly improving the processing efficiency of subsequent semantic feature extraction. S1.3: Based on the original data fragment group, the naming pattern of each device field (such as 'BP_Sys', 'systolic pressure', 'SystolicPressure') is identified by a combination of regular expression matching and syntax tree analysis, and the unit expression features are extracted by combining the unit annotation rule base (such as mmHg, kPa) to generate a device-level metadata feature matrix; S1.4: Perform structured classification processing on the device-level metadata feature matrix, taking field naming pattern, data type (numerical / textual / time-series), sampling frequency, and unit system as dimension items, and construct a device-side data format profile table with a unified description format as the basic input for device semantic feature modeling; S1.5: Based on the device-side data format profile table, the original observations of all devices within the same time window are integrated through timestamp alignment and device identification to generate a device-side original dataset containing multi-source heterogeneous attributes, which is used to support the subsequent standardized parsing and dynamic semantic benchmark construction process.

[0012] Step S2: The original dataset from the device side undergoes structured parsing and type normalization, mapping numerical, textual, and time-series fields to a unified intermediate representation format to generate standardized data units, which serve as the basic input for subsequent semantic analysis. Specifically, this includes: S2.1: Based on the raw data streams obtained from heterogeneous devices such as blood pressure monitors, blood glucose meters, and electrocardiographs, identify the data format characteristics (such as JSON, XML, CSV), field naming patterns (such as 'BP_High', 'blood glucose value'), and unit expression methods (such as mmHg, mg / dL) of each device. Use a regular expression parser and schema inference algorithm to perform syntactic-level decomposition of unstructured or semi-structured data, and generate preliminary parsing units with metadata annotations to support subsequent data type classification processing; S2.2: Perform dimension normalization and unit standardization on the numerical fields (such as systolic blood pressure and fasting blood glucose) in the preliminary analysis unit. Based on the pre-set medical measurement standard table (such as WHO unit reference standard), map different unit expressions to standard units under the unified public health indicator system, and use a linear transformation function to recalibrate the numerical values ​​to generate a standardized numerical vector with unified dimensions as an intermediate representation of numerical data. S2.3: For text-based fields (such as health advice and abnormality alerts), keyword extraction and synonym merging are performed based on a clinical terminology dictionary (such as SNOMEDCT and ICD-11). The BiLSTM-CRF model is used to identify medical entities in the context and map them to a standardized terminology coding space to generate a structured text semantic coding sequence as a unified intermediate representation of text-based data. Based on the set of text fields separated from the initial parsing unit, a keyword extraction algorithm based on a clinical terminology dictionary (parameters: dictionary range set to SNOMED CT and ICD-11 combined version, matching mode set to maximum positive matching) is used to achieve rapid location and boundary labeling of explicit medical terms in the text field content; Furthermore, through a synonym merging algorithm (parameter: the synonym mapping table is expanded from the terminology dictionary and combined with custom industry terms), a unified semantic mapping of various different expressions existing in the output text of multiple devices is achieved, and a merged term index list is obtained; Furthermore, the BiLSTM-CRF sequence labeling model (parameters: BiLSTM hidden layer dimension is 256, CRF decoding layer label set covers all standardized medical entity types) is used to realize medical entity recognition in context, and combined with part-of-speech tagging and semantic dependency parsing, medical entity sequences with location information and category labels are generated. Furthermore, through a terminology coding mapping algorithm (parameter: the mapping rule base includes the cross-library correspondence between LOINC and SNOMEDCT), the identified medical entities are established with a one-to-one coding index in the standardized terminology coding space, and a structured entity sequence with coding is generated. By using a hierarchical semantic encapsulation method, the above structured entity sequence is transformed into a text semantic encoding sequence, realizing standardized processing of text data under a unified intermediate representation format, and ensuring cross-device compatibility and label consistency in subsequent semantic analysis stages. For example, in a public health checkup scenario, the original text "The patient exhibits atrial fibrillation; a cardiac ultrasound examination is recommended" is parsed from the health advice field generated by the electrocardiogram (ECG) machine. The keyword extraction algorithm matches "atrial fibrillation" and "cardiac ultrasonography" in the SNOMEDCT dictionary. The synonym merging algorithm maps "atrial fibrillation" and "heart fibrillation" to "atrial fibrillation," and maps "cardiac ultrasound examination" and "echocardiography" to "cardiac ultrasonography." BiLSTM The CRF model identifies Atrial fibrillation as a disease entity category and Cardiac ultrasonography as an examination entity category in context analysis, and generates location indices [0,4] and [9,14]. The terminology encoding mapping algorithm maps Atrial fibrillation to SNOMEDCT code 49436004 and Cardiac ultrasonography to SNOMEDCT code 40701008, ultimately outputting a structured text semantic encoding sequence [{entity:"Atrial fibrillation", category:"disease", code:49436004},{entity:"Cardiac ultrasonography", category:"examination", code:40701008}]. After encapsulation, it forms a unified intermediate representation consistent with text fields from other devices. In cross-device semantic alignment experiments, it can significantly improve label consistency and mapping accuracy, thereby enhancing the usability of multi-source heterogeneous text data. S2.4: Perform timestamp alignment and sampling frequency normalization on time-series fields (such as ECG waveform sequences and continuous blood glucose monitoring data), use cubic spline interpolation algorithm to fill missing time points, combine sliding window segmentation strategy to extract time-series segments of equal length, and extract low-frequency trend and high-frequency detail coefficients through discrete wavelet transform (DWT) to generate a standardized time-series tensor with consistent time base and frequency domain feature representation; S2.5: The standardized numerical vectors, text semantic encoding sequences, and time series tensors generated above are uniformly encapsulated into a standardized data unit structure containing field identifiers, data type markers, semantic label candidates, and confidence scores. The structure is then compactly encoded using the Apache Avro serialization protocol to generate intermediate data objects that can be passed across modules, serving as the input basis for the subsequent lightweight encoder-decoder structure.

[0013] like Figure 2 As shown, step S3 involves: based on the current business task objectives, identifying the set of key health indicators required for chronic disease risk assessment, selecting the electronic health record system from all access devices as the initial semantic benchmark source due to its highest semantic coverage and clinical authority, and generating a dynamic semantic benchmark template. Specifically, this includes: S3.1: Based on the chronic disease management business needs in the scenario of public health physical examination and family doctor contract service, obtain the current task target definition file, which contains risk assessment standard items for key chronic diseases such as hypertension and diabetes; use natural language processing technology to perform entity recognition and relation extraction on the current task target definition file, extract the key health indicator set, and generate a standardized task semantic requirement vector. S3.2: Perform a metadata scan on the list of connected multi-source medical devices to obtain the data field list of each device and its corresponding semantic annotation information; based on the standardized task semantic requirement vector, calculate the field coverage score of each device on key health indicators and generate a device-level semantic coverage matrix. The input conditions are the standardized task semantic requirement vector generated based on step S3.1 and the list of multi-source medical devices that have been connected. Each device has completed data structure parsing and type normalization processing in step S2, and is accompanied by field identifiers, data type tags and semantic label candidate information. The device metadata scanning method (parameters: device list, metadata interface protocol) is used to read the data structure description file of each device in batches and parse the field names, field types, unit systems and additional semantic comments to form a device field list. Furthermore, through a semantic annotation parsing method (parameters: clinical terminology dictionary, named entity mapping rules), the correspondence between field names and standardized task semantic requirement vectors is matched, and the semantic matching labeling results of the fields in the set of key health indicators are obtained; Furthermore, a field coverage calculation algorithm (parameters: matching tag results, total set of key health indicators) is used to calculate the field coverage ratio of each device on the target indicator set and generate a coverage score. The coverage calculation formula is as follows:

[0014] in, The coverage score for device d. This is the flag value for the i-th health indicator matched by device d. A flag value of 1 indicates a successful match, and 0 indicates no match. The total number of indicators in the set of key health indicators; Furthermore, by using a coverage score normalization processing method (parameters: coverage score list, normalization interval [0,1]), the numerical standard of coverage scores for different devices is unified so as to form the input format for matrix operations; Furthermore, a matrix construction algorithm (parameters: device index, standardized coverage score vector) is adopted to initialize and fill the device-level semantic coverage matrix, where the matrix rows represent devices, the columns represent key health indicators, and the element values ​​are the matching tags or standardized scores of the corresponding indicators. By using coverage calculation and matrix construction processing, the metadata scanning results of the previous step are transformed into a device-level semantic coverage matrix that can be used for subsequent device benchmark source selection, thereby realizing a quantitative expression of the indicator support capabilities of each device under the semantic requirements of the task. For example, in a test scenario at a primary healthcare facility, the connected devices include a blood pressure monitor from manufacturer A, a blood glucose meter from manufacturer B, and a multi-lead electrocardiograph from manufacturer C. The key health indicator set is {systolic blood pressure, diastolic blood pressure, fasting blood glucose, glycated hemoglobin, and heart rate}. The task semantic requirement vector length is 5, and the field lists for each device are obtained in batches via the HL7 interface. The field list for manufacturer A's blood pressure monitor includes systolic blood pressure (mmHg) and diastolic blood pressure (mmHg). After semantic matching, the matching marker sequence is [1,1,0,0,0], and the coverage score is calculated as follows: =0.4; Manufacturer B's blood glucose meter field list includes fasting blood glucose (mmol / L) and glycated hemoglobin (%), with a matching marker sequence of [0,0,1,1,0], and a score of =0.4; Manufacturer C's ECG machine field list includes heart rate (bpm) and QT interval (ms), and the matching marker sequence is [0,0,0,0,1], with a score of =0.2. After normalization, the scores of each device remain relatively different. The device-level semantic coverage matrix is ​​3×5-dimensional, and the matrix elements correspond to the matching tags of each device on each indicator. It is used to select the optimal benchmark source in the subsequent step S3.3 by combining it with the external credibility score. The actual test shows that this matrix can significantly improve the accuracy of benchmark source selection. S3.3: Based on the device-level semantic coverage matrix and combined with the external credibility scoring system (including data source authority, update frequency, and clinical adoption rate), a comprehensive weighted score is given to each device; the electronic health record system is determined as the optimal semantic benchmark candidate source through a ranking mechanism, and a benchmark source selection decision record is generated; S3.4: Based on the structured data model of the electronic health record system, extract the core health indicator field names, unit systems, normal value ranges, and related term codes (such as LOINC, SNOMED CT); integrate the above elements into a hierarchical standard field directory tree to generate an initial semantic benchmark ontology framework; The input conditions are: the structured data model of the electronic health record system, including standardized field names, the unit system of intermediate numerical, text and time series data, normal value ranges and associated standard terminology codes (such as LOINC, SNOMED CT). A field parsing algorithm (parameters: field type classification table, regular expression matching pattern library) is used to extract the names of core health indicator fields from the electronic health record system. Furthermore, through a unit standardization mapping method (parameter: International System of Units / WHO unit conversion table), the identification and conversion of the field unit system are realized, and a unified set of unit representations is obtained; Furthermore, by employing a normal value range extraction method (parameters: clinical standard range library, field numerical distribution statistics), the normal value ranges for each core health indicator are determined, and a set of range boundary parameters is generated. During the range extraction process, the mean value of the numerical range is... with standard deviation It can be calculated using the following formula:

[0015] in, The mean, For the i-th sample value, The total number of samples;

[0016] in, Standard deviation is used to measure the dispersion of values ​​in a field. Furthermore, through a terminology encoding mapping algorithm (parameters: terminology dictionary index, encoding lookup table), the encoding association between fields and the standard terminology space is realized, and a semantic encoding set is generated; Furthermore, by using a hierarchical directory tree construction algorithm (parameters: set of field names, set of unit representations, set of normal value range parameters, and set of terminology codes), all extracted elements are organized into a hierarchical field directory structure, where leaf nodes represent specific measurement indicators, parent nodes represent indicator categories, and root nodes represent the semantic category of business tasks. The algorithm constructs a directory tree to transform the parsing results of the previous step into an initial semantic benchmark ontology framework, thereby realizing a structured reference system for subsequent cross-device semantic alignment. For example, in a chronic disease risk assessment task, the blood pressure field parsing results from the electronic health record system include the field names "systolic blood pressure" and "diastolic blood pressure," with the unit system being mmHg. After unit standardization mapping, a unified unit mmHg is obtained, with normal ranges of 90–139 and 60–89, respectively. The mean systolic blood pressure is calculated using the normal range extraction formula. Standard deviation Mean diastolic blood pressure Standard deviation The terminology encodings are mapped to LOINC codes “8480-6” and “8462-4”, respectively. The fields are organized into two leaf nodes under the parent node “blood pressure”, forming the blood pressure subtree structure in the initial semantic benchmark ontology. This structure significantly improves the consistency and accuracy of label generation in the subsequent semantic alignment process for non-benchmark devices. S3.5: Inject the initial semantic benchmark ontology framework into the dynamic template engine, bind the current business task cycle and user characteristic parameters, and generate a dynamic semantic benchmark template that can be adjusted as the task context evolves. This template is then output as a reference anchor point for subsequent cross-device semantic alignment to the shared embedding space construction process.

[0017] like Figure 3 As shown, step S4 involves inputting standardized data units from non-reference devices into a lightweight encoder-decoder structure, and projecting their field semantics onto a shared embedding space aligned with a dynamic semantic reference template through a contrastive learning mechanism, thereby generating a cross-device field-level semantic alignment vector. Specifically, this includes: S4.1: Based on the standardized data units and dynamic semantic benchmark templates generated in the preceding steps, a lightweight encoder-decoder structure is constructed. The encoder uses a multi-head attention mechanism to jointly embed the field names, numerical ranges and unit information of non-benchmark devices to generate an initial field semantic representation vector, which serves as the input condition for semantic alignment. Based on the standardized data units and dynamic semantic benchmark templates generated in the preceding steps, a multi-head attention mechanism encoding algorithm (parameters: number of attention heads = 8, embedding dimension = 256, dropout rate = 0.1) is used to jointly embed the field names, numerical ranges and unit information of non-benchmark devices, thereby achieving parallel capture and weight allocation of cross-modal features; Furthermore, by using the field name vectorization method (parameters: pre-trained medical term word vector model, word embedding dimension = 128), textual field names are mapped to a high-dimensional word vector space, and the position embedding module is used to preserve their order features within the standardized data units, resulting in a field name embedding matrix, which supports name association modeling in subsequent semantic alignment. Furthermore, by using the numerical range normalization embedding method (parameters: normalization interval [0,1], smoothing coefficient = 0.05), the numerical fields are converted into normalized numerical vectors according to the public health medical measurement standard table, and a linear projection layer is used to map them to the same dimensional space as the text embedding, so as to ensure the dimensional consistency of different types of features at the encoder input. Furthermore, by using a unit information encoding method (parameters: unit category mapping table, unit vector dimension = 64), the extracted unit features (such as mmHg, mg / dL) are converted into trainable unit embedding vectors, and then integrated into the query vector of field name and numerical range embedding through the key-value pair structure of the multi-head attention mechanism, thereby achieving unit-based cross-dimensional attention weighting. The formula for calculating the attention weight distribution using a multi-head attention network is as follows:

[0018] in, For attention weight output, For querying the matrix, The key matrix, For value matrices, The dimension of the key vector. This is the matrix transpose; the above formula calculates the correlation weights between field features using a scaled dot product attention mechanism. The output vectors of each attention head are connected by a multi-head concatenation mechanism and input into a linear transformation layer to generate an initial field semantic representation vector, thereby achieving deep fusion of different feature sources. By using a multi-head attention mechanism and joint embedding processing, the multi-source features of the device field in the previous step are transformed into a high-dimensional initial field semantic representation vector containing three types of information: semantics, numerical value, and unit, thereby realizing a unified semantic representation of non-baseline device fields at the encoder end. For example, in the public health examination data collection at a primary healthcare institution, non-benchmark devices included two blood pressure monitors from different manufacturers (field names "BP_Sys" and "systolic blood pressure," respectively), with values ​​ranging from [80, 180] mmHg. The field names were vectorized using a pre-trained medical terminology model, forming two 128-dimensional word embeddings; the numerical range was mapped to the [0,1] interval using a normalization function, for example, 120 mmHg corresponds to the normalized value... =0.4, then projected to a 128-dimensional numerical embedding space; unit information is encoded as a 64-dimensional vector through a unit category mapping table. The three-class embeddings are input into an 8-head attention mechanism, and the scaling dot product calculation in single-head attention is as follows: The weight matrix is ​​then obtained through Softmax, and multiplied by the V matrix to generate the attention output. The outputs of the eight attention heads are concatenated and passed through a linear transformation layer to obtain a 256-dimensional initial field semantic representation vector. This vector integrates name, value, and unit multimodal information, which can significantly improve the accuracy of semantic projection in the subsequent contrastive learning process. S4.2: Perform layer normalization and position encoding on the initial field semantic representation vector to eliminate the dimensional differences between different fields and retain the sequence order information, and output a normalized semantic embedding tensor as the processing object of the contrastive learning module; S4.3: Based on the known semantic vectors of authoritative fields in the dynamic semantic benchmark template, construct positive sample pairs and negative sample pairs, optimize the encoder-decoder parameters using the contrastive loss function, minimize the embedding distance between synonymous fields, maximize the separation between heteronymous fields, and generate a discriminative shared embedding space. Based on the known semantic vectors of authoritative fields in the dynamic semantic benchmark template, a sample pair construction method (parameters: field name embedding, numerical range embedding, unit embedding, authoritative semantic vector set) is used to generate positive sample pairs of synonymous fields and negative sample pairs of heteronymous fields, thereby realizing the construction of a supervision signal for semantic alignment. Furthermore, the semantic distance value of each pair of samples in the current shared embedding space is obtained by using the feature difference calculation method (parameter: embedding space dimension d, distance metric is cosine distance), and the expected target distance interval is set according to whether they are synonymous, which is used as the direct input for the optimization of the contrastive loss function; Furthermore, a contrastive loss function is employed, combined with a boundary constraint regularization term, to optimize the encoder-decoder parameters, thereby minimizing the embedding distance between synonymous fields and maximizing the embedding distance between heteronymous fields, forming an embedding space with enhanced semantic discriminativeness; the loss function can be expressed as:

[0019] in, The loss value. Label the sample pairs with synonyms. To embed the spatial cosine distance, For preset interval boundary parameters, The total number of sample pairs; Furthermore, by using the gradient backpropagation algorithm (parameters: learning rate η, adaptive optimizer Adam), the gradients calculated by the above loss function are used to update the encoder and decoder weights, ensuring that vectors with similar semantics converge to the same subspace in the shared embedding space, while maintaining the spatial separation of vectors with different semantics. Furthermore, batch normalization and embedding vector recalibration methods are adopted to perform scale standardization on the optimized embedding vectors, eliminating training instability caused by distribution differences between batches, and enabling the shared embedding space to have cross-device field semantic discrimination capabilities. Through the above comparative learning and space optimization processing, the normalized semantic embedding tensor of the previous step is transformed into a structurally stable and highly discriminative shared embedding space, thereby achieving accurate alignment of cross-device field semantics and enhanced conflict resolution capabilities. For example, in a chronic disease risk assessment task, the semantic vector of the "systolic blood pressure" field from an electronic health record system and the semantic vector of the "BP_Sys" field from a certain model of blood pressure monitor are selected as positive sample pairs. The cosine distance calculation result is 0.12, and the m value is set to 0.5. The semantic vector of "systolic blood pressure" and the semantic vector of the "heart rate" field from an electrocardiograph are selected as negative sample pairs, and the cosine distance calculation result is 0.68. The contrast loss function is calculated as follows: For positive sample pairs, =1, the loss sub-item is = For negative sample pairs, =0, the loss sub-item is = With the total number of batch sample pairs N set to 1024, the total loss was calculated and the model parameters were updated using the Adam optimizer with a learning rate η=0.001. After iterative training, the cosine distance between the embedding vectors of "BP_Sys" and "systolic blood pressure" in the validation set significantly decreased to 0.05, while the embedding distance between "systolic blood pressure" and "heart rate" increased to 0.82, indicating that the discriminability and alignment accuracy of the shared embedding space were significantly improved, and the label consistency generation effect was effectively guaranteed. S4.4: Map the normalized semantic embedding tensor of each field of the non-benchmark device to the shared embedding space, calculate its cosine similarity with the corresponding health index in the dynamic semantic benchmark template, and obtain the cross-device field-level semantic alignment score matrix as the quantitative basis for semantic consistency. For the normalized semantic embedding tensor after layer normalization and position encoding, the cosine similarity calculation method (parameters: reference vector set, vector set to be aligned) is used to quantify the semantic relevance between cross-device fields and corresponding health indicators in the dynamic semantic reference template. Furthermore, by performing vector normalization (parameter: L2 norm constraint), we achieve scale consistency in cosine similarity calculation for embedding vectors of different fields, and obtain aligned input data after vector norm adjustment. Furthermore, using the ratio of vector dot product to norm product, the cosine similarity between each non-reference field and the reference field is calculated according to the following formula:

[0020] in, Output the cosine similarity. The embedding vector of the field to be aligned. The embedding vector for the baseline field. For vectors L2 norm, For vectors The L2 norm; Furthermore, through a similarity matrix construction algorithm (parameters: total number of fields, total number of benchmark fields), the above similarity calculation results are mapped into a two-dimensional matrix structure, where the matrix rows represent non-benchmark device fields and the matrix columns represent benchmark health indicator fields, and a cross-device field-level semantic alignment score matrix is ​​generated. Furthermore, through matrix normalization and numerical truncation (parameters: normalization method min-max, truncation threshold setting), the effective range of similarity scores under a unified dimension is compressed, eliminating the impact of numerical scale differences on subsequent threshold selection; By using cosine similarity calculation and matrix construction, the semantic embedding tensor from the previous step is transformed into a cross-device field-level semantic alignment score matrix that can be directly used as a basis for quantifying semantic consistency, thereby achieving an accurate characterization of the degree of semantic matching between fields. For example, in the implementation of chronic disease risk assessment based on a public health data platform, the dimension of the input non-benchmark device field embedding vector is set to 128, and the dimension of the benchmark field vector is also 128. L2 norm constraints are used to normalize all vectors to unit length, and the vector dot product in the cosine similarity formula is obtained by multiplying and summing the 128 elements one by one. In one calculation, the similarity calculation result of the "systolic blood pressure" field vector in the dynamic semantic benchmark template for a certain blood pressure monitor field embedding is: The similarity with the "diastolic blood pressure" field is . After min-max normalization, all similarities are compressed to the [0,1] interval, where Mapped to normalized score In practical applications, the constructed scoring matrix is ​​50×35, with rows corresponding to 50 fields from different manufacturers and columns corresponding to 35 health indicators from the baseline template. The distribution of matrix entry values ​​is limited to the range [0.2, 1] using a numerical truncation strategy, with values ​​below the truncation threshold automatically set to 0 to facilitate subsequent high-confidence mapping and filtering. The output of this matrix is ​​used to support the subsequent dual-channel alignment network, significantly improving accuracy and stability in the field matching stage. S4.5: Based on the semantic alignment score matrix, filter high-confidence mapping relationships that are higher than a preset threshold, and generate cross-device field-level semantic alignment vectors. These vectors include the field source device identifier, target reference label, alignment confidence, and semantic offset, which are used to support the consistency verification operation of the subsequent dual-channel alignment network. Based on the cross-device field-level semantic alignment score matrix as the input data object, an adaptive threshold filtering method is adopted (parameter: alignment confidence threshold τ, the value is based on the historical device semantic matching performance distribution) to achieve the initial selection of high-confidence semantic mapping relationships; Furthermore, by using the confidence interval estimation method (parameters: normal distribution assumption, significance level α), the statistical verification of the alignment confidence of the mapping relationship is achieved, and the filtered high-confidence field-label mapping set is obtained; Furthermore, by using a semantic offset calculation method (parameter: shared embedding space vector coordinate system), the offset distance of each mapping relationship in the embedding space is quantified, and a semantic offset matrix is ​​generated, as shown in the following formula:

[0021] in, Embed a vector for the source device field. Embed vectors for target baseline labels. This is the semantic offset; Furthermore, by encapsulating the mapping relationship (parameters: field source device identifier, target baseline label, alignment confidence, semantic offset), a structured cross-device field-level semantic alignment vector is constructed, and a complete vector set for subsequent consistency verification modules is generated. By using high-confidence mapping filtering and semantic offset encapsulation processing, the results of the previous step are transformed into cross-device field-level semantic alignment vectors that can be directly called by the dual-channel alignment network, thereby significantly improving the semantic consistency and mapping reliability of the tag generation process. For example, in a public health physical examination data processing, the input semantic alignment score matrix comes from the embedding similarity calculation results of a blood pressure monitor (field 'BP_Sys'), a blood glucose meter (field 'Glucose_Level'), and an electrocardiograph (field 'ECG_Rhythm'). An alignment confidence threshold τ=0.85 is set, and adaptive threshold filtering is performed on the matrix, retaining only field-label pairs with scores above 0.85, such as 'BP_Sys'-'systolic blood pressure' and 'Glucose_Level'-'fasting blood glucose'. Confidence interval validation is performed on the filtering results, with a significance level α=0.05. The calculated alignment confidence CI interval for 'BP_Sys'-'systolic blood pressure' is [0.88, 0.93], meeting the high confidence standard. Using semantic offsets, if the 'BP_Sys' vector v=(0.12,0.35,0.78) and the 'systolic pressure' vector u=(0.10,0.33,0.80), then:

[0022] The calculated result Δ≈0.0346 corresponds to a low offset, indicating high semantic alignment accuracy. The source device identifier, target baseline label, alignment confidence (0.90), and semantic offset (0.0346) are encapsulated into a structured vector and output to the consistency verification module. In subsequent dual-channel network inference, this field mapping relationship showed no conflict or ambiguity, verifying the applicability and efficiency of this method in generating consistent labels for multi-source heterogeneous devices.

[0023] Step S5: Based on the cross-device field-level semantic alignment vector, a dual-channel alignment network is constructed. The first channel performs semantic similarity matching between field names and baseline labels, while the second channel combines contextual record information for consistency verification, generating a preliminary consistent label cluster. Specifically, this includes: S5.1: Based on the cross-device field-level semantic alignment vector, construct a field-level semantic similarity calculation model, use the cosine similarity algorithm to perform pairwise matching between the embedding vector of the non-benchmark device field and the standard label vector in the dynamic semantic benchmark template, calculate their semantic relevance score, and generate a field-level semantic similarity matrix as the input basis for the first channel. Based on cross-device field-level semantic alignment vectors as input conditions, a vector similarity calculation method (parameters: dynamic semantic benchmark template, non-benchmark device field embedding vector) is used to achieve pairwise matching between fields and standard labels. Furthermore, the similarity between the field embedding vector and the standard label embedding vector is measured by the cosine similarity algorithm (parameters: embedding vector dimension d=256, vector normalization method L2), and the semantic relevance score matrix is ​​obtained. Furthermore, a normalization method (parameter: maximum value normalization) is used to unify the dimensions of the similarity score matrix, achieving comparability under different vector scales, and generating a normalized semantic similarity matrix; Furthermore, by using a matrix batch calculation method (parameters: GPU floating-point parallel calculation, batch size = 512), parallel similarity calculations between all non-benchmark device fields and benchmark labels are achieved, and a field-level semantic similarity matrix is ​​generated as output. By using vector similarity calculation methods, the cross-device field-level semantic alignment vectors from the previous step are transformed into a data matrix with quantifiable semantic matching capabilities, thereby enabling the first channel's input to build the expected technical effect. For example, in a primary healthcare physical examination scenario, the input condition is a set of cross-device field-level semantic alignment vectors. This set contains 200 non-benchmark device field embedding vectors (dimension 256) and 50 dynamic semantic benchmark template standard label embedding vectors. The cosine similarity algorithm is used, and its calculation formula is as follows:

[0024] in, This is the embedding vector for the non-reference device field. The embedding vector for the dynamic semantic baseline label. Representing vectors L2 norm, Representing vectors The L2 norm is used. In practice, a 200×50 similarity matrix is ​​calculated in batches using GPU parallel computing, resulting in a 200-row, 50-column field-level semantic similarity matrix. This matrix is ​​then normalized by the maximum value of each column, ensuring all scores fall within the [0,1] range, effectively eliminating the scaling effect before field embedding normalization. During the validation phase, this matrix is ​​input into the first-channel clustering and segmentation module for subsequent candidate mapping relationship filtering. The number of output field-label matching scores is significantly increased under high semantic relevance, thus improving the accuracy of label consistency generation. S5.2: Perform thresholding clustering on the field-level semantic similarity matrix, use an adaptive threshold segmentation algorithm to identify strongly associated field-label pairs that are higher than the preset semantic similarity threshold, filter out the candidate mapping relationship set, mark the unmatched fields as items to be resolved, generate preliminary field mapping results, and form the output results of the first channel; For the generated field-level semantic similarity matrix, an adaptive threshold segmentation algorithm (parameters: mean of similarity distribution μ, standard deviation σ, dynamic adjustment factor λ) is used to realize the flexible boundary determination function of semantic matching degree; Furthermore, a probability density curve for the similarity score is fitted using a Gaussian kernel density estimation method (parameter: bandwidth h depends on the variance of field similarity samples), and an adaptive threshold is calculated based on the inflection point of the curve. This yields the boundary value used to distinguish between strong and weak association pairs; Furthermore, by using a density peak-based clustering algorithm (parameters: cosine distance is used as the distance metric, and the clustering stopping condition is that the mean similarity within the class is greater than T), field-label embedding vectors with values ​​above the threshold are quickly aggregated, and a set of candidate mapping relationships is generated. Furthermore, by using the association pair filtering method (parameter: only retain the one with the highest similarity score among duplicate class labels in the candidate relationship set), the set is simplified and a preliminary mapping list without duplicate labels is generated; The set complementation method is used (parameters: the set of unmatched fields F, and the status marker code to be resolved). Fields that do not reach the threshold during the initial mapping process are marked as items to be resolved, forming the output of the first channel, which serves as the input for subsequent context consistency verification. For example, in a cluster of devices including blood pressure monitors, blood glucose meters, and electrocardiographs, the field-level semantic similarity matrix has a dimension of 20×15, a mean μ of 0.78, a standard deviation σ of 0.05, and an adaptive adjustment factor λ of 0.4. The calculated threshold... The similarity score was 0.8. The stability of the threshold was verified at the inflection point of the similarity distribution curve using Gaussian kernel density estimation (bandwidth h=0.02). Density peak clustering was employed to aggregate field-label pairs with similarity greater than 0.8 into 7 candidate relation classes, each containing 2-4 field mappings. After filtering for duplicate class labels, 3 low-confidence duplicate mappings were removed, resulting in a preliminary mapping list containing 7 unique mappings. The remaining 5 fields were then labeled with [missing information - likely a label or identifier]. The state is further resolved in the second channel. Execution results show that in this embodiment, the alignment rate of high-confidence fields is significantly improved, and the label redundancy of the mapping set is significantly reduced, providing an optimized basic input for subsequent context consistency correction; S5.3: Obtain the contextual information stream associated with the current data record, including other physiological indicator values, historical diagnostic labels and time series trends of the same patient in this physical examination session. Extract the logical dependency relationship between multidimensional features based on the context-aware graph neural network model, and generate a context-related embedding vector as the consistency verification input of the second channel. For standardized data units from devices such as blood pressure monitors, blood glucose meters, and electrocardiographs in the current physical examination session, a multi-source data fusion method (parameters: field identifier, numerical range, sampling timestamp) is used to initially construct the context information flow and associate multiple physiological indicator values ​​for the same patient. Furthermore, based on the patient health record management system call interface (parameters: patient unique ID, historical medical record index), the historical diagnostic tags are obtained in batches and aligned with the physiological indicators of the current session in chronological order to obtain a joint time series. Furthermore, a time series feature extraction algorithm (parameters: sliding window width, window step size) is applied to the joint time series to extract trend features and fluctuation features. The differences between different indicator units are eliminated through standardization to form a multidimensional time series feature tensor. Furthermore, a context-aware graph neural network model (parameters: node type = physiological index / diagnostic label / time node, edge type = time dependency / logical dependency, number of layers = 3) is used to construct a feature graph structure, and physiological index nodes and diagnostic label nodes are associated through edge weight functions to generate a high-dimensional graph embedding vector that reflects the logical dependency relationship between features. Furthermore, through a message passing mechanism (iteration count = 2), multiple rounds of node feature updates are implemented in the graph structure to obtain a context-related embedding vector containing the trend relationship of physiological indicators and the dependency pattern of diagnostic labels. Through the above context-aware graph neural network model, the time series features and historical diagnostic label dependency information of the previous step are transformed into context-related embedding vectors that can be called by the second channel consistency verification module, thereby realizing the label consistency reasoning capability across fields and across time. For example, in a patient's physical examination session, the blood pressure monitor samples at a frequency of 1Hz, the mean systolic blood pressure in the current window is 135 mmHg, the blood glucose meter samples every 5 minutes, the mean fasting blood glucose is 6.8 mmol / L, the historical diagnostic labels include "hypertension" and "diabetes", and there are physical examination records for three consecutive years. When constructing the joint time series data, the sliding window width is set to 10 minutes, the step size is 5 minutes, and the mean, standard deviation, and linear trend slope are extracted for each window to form a time-series feature vector of length N. When constructing the graph structure, the physiological indicator nodes (systolic blood pressure, fasting blood glucose) and diagnostic label nodes (hypertension, diabetes) are connected through logical dependency edges, with edge weights... The calculation formula is as follows:

[0025] in, The cosine similarity between the trend sequence of indicator values ​​and the historical occurrence pattern of the label is used. The time distance is a standardized value (in years). Edge weights are used as input to a graph neural network, and after two layers of message passing to update node embeddings, the resulting context-related embedding vector has a length of 128 dimensions and possesses highly cohesive logical dependency information. After this embedding vector is input into the consistency verification module, it can automatically adjust the label priority when a high correlation is detected between the current blood pressure index and the historical "hypertension" label, significantly improving the semantic consistency and rationality of label generation. S5.4: Jointly reason with the candidate mapping relationship set and the context-related embedding vector, perform consistency constraint judgment by constructing a differentiable logic rule layer, detect potential label conflicts or redundant combinations (such as simultaneously labeling 'hypertension' and 'hypotension'), and perform priority arbitration on contradictory labels to generate an optimized label set corrected by context verification; S5.5: Integrate the preliminary field mapping results output from the first channel with the optimized label set generated by the second channel, perform label cluster merging and deduplication, and use a label integration algorithm based on set coverage to generate a preliminary consistent label cluster with complete structure and consistent semantics, and use it as the input basis for the next stage of global semantic alignment model update.

[0026] Step S6: When a new device is added, the existing shared embedding space is used as a teacher model to guide the learning process of the new device's feature extractor. The boundary definition of the shared semantic space is updated through an incremental knowledge distillation mechanism to generate an expanded global semantic alignment model. Specifically, this includes: S6.1: Based on the shared embedding space constructed in the previous steps, obtain its high-dimensional semantic parameter distribution as a teacher model. The shared embedding space is trained by contrastive learning using a lightweight encoder-decoder structure and has the ability to represent fields from multiple sources such as blood pressure monitors and blood glucose meters in a unified vector format. The output embedding trajectory of the teacher model is used as a soft label target to provide transfer learning supervision signals for the feature extractor of the new device, so as to preserve the existing semantic alignment relationship. S6.2: Field-level sampling is performed on the standardized data units of newly accessed devices to generate the initial input batch, and forward propagation is performed through the student feature extractor to be trained to obtain the corresponding student embedding vector; the KL divergence loss between the student embedding vector and the soft label generated by the teacher model is calculated as the core optimization objective of knowledge distillation to measure the consistency deviation between the new device representation and the existing semantic space. S6.3: Based on the KL divergence loss, the adaptive gradient descent algorithm is used to fine-tune the parameters of the student feature extractor of the new device and optimize its encoding weights to reduce the distribution difference with the teacher model in the shared embedding space; in this process, the teacher model parameters are frozen and only the student network is updated, thereby achieving low-overhead incremental learning and outputting a preliminarily aligned new device embedding mapping function. Based on the KL divergence loss value calculated in the previous steps, an adaptive gradient descent algorithm is used (parameter: initial learning rate). 0, attenuation factor γ, momentum μ), to achieve iterative update of parameters of the student feature extractor for newly added equipment to optimize encoding weights; Furthermore, the gradient calculation module performs analytical backpropagation on the KL divergence loss function to obtain the gradient tensor of the weight matrix of each encoding layer. Combined with the adaptive learning rate update strategy, the effective step size of each layer is dynamically adjusted according to the gradient magnitude to achieve fast convergence in the high gradient direction and fine approximation in the low gradient direction. Furthermore, the gradient vector generated in the previous step is weighted and combined with the momentum term using the weight update module to suppress oscillations and accelerate convergence. During this process, the teacher model parameter matrix is ​​frozen, and only the student network parameters are kept updatable, ensuring that the existing shared semantic space structure remains stable during the iteration process. Furthermore, a normalization regularization operator is used to apply norm constraints to the weight matrix after each update to prevent parameter explosion and keep the distribution of the embedding space vector within a reasonable range. The balance between the model's generalization ability and fitting accuracy is controlled by adjusting the regularization coefficient. Through the optimized gradient descent iterative process, the soft label supervision signal of the previous step is transformed into the embedding mapping function parameter set of the student feature extractor, which significantly reduces the difference between the semantic vector of the new device field and the embedded space distribution shared by the teacher model, and outputs the newly aligned new device embedding mapping function.

[0027] in, For KL divergence loss; The number of samples; This represents the probability distribution of the teacher model output; This represents the probability distribution of the student model output; The label is the label corresponding to the i-th sample; For example, in a scenario where new heart rate monitoring devices are added to primary healthcare institutions, the initial learning rate η0 of the student feature extractor is set to... The attenuation factor γ is set to The momentum μ is set to Batch size is Performing KL divergence loss gradient backpropagation, the gradient norms of encoding layers one through three are obtained as follows: , , After adaptive adjustment, the update step size was increased to [number missing]. , , After 15 iterations, the cosine similarity between the student model's embedding vector for the heart rate field on the new device and the corresponding field vector of the teacher model increased from the initial... Increase to The differences in the embedded space distribution were significantly reduced, and consistent alignment between the heart rate monitor and existing devices in the shared semantic space was achieved, laying a stable foundation for the subsequent expansion of the global semantic alignment model. S6.4: Perform joint clustering analysis on the output of multiple new device embedding mapping functions and the original shared embedding space to identify potential semantic boundary shift regions; calculate the semantic space expansion coefficient based on the cluster center drift, and use this coefficient to dynamically adjust the boundary definition threshold of the shared semantic space to generate an expanded global semantic alignment model that covers a wider range of device types; S6.5: Perform a consistency verification test on the extended global semantic alignment model. Input cross-sample data streams from new and old devices and detect the changes in F1-score in field-level semantic matching and record-level context verification. If the performance degradation exceeds the preset tolerance, trigger the feedback adjustment mechanism to inject the error residuals back into the attention layer of the student network to further optimize the feature selection sensitivity and ensure that the model maintains high-precision label generation capability after expansion.

[0028] Step S7: Deploy the extended global semantic alignment model to the edge inference module, perform localized label generation operations on the real-time inflow multi-source data, and output a structured health label sequence with semantic consistency. Specifically, this includes: S7.1: Based on the extended global semantic alignment model generated in the previous steps, extract its shared embedding space parameters, lightweight encoder-decoder weights and dual-channel alignment network configuration information, and use model compression technology to perform pruning and quantization processing to generate a compact inference model suitable for edge computing environments, so as to reduce the resource consumption at the edge and improve response efficiency. S7.2: Deploy the compact inference model at the edge, build a localized label generation engine, receive standardized data units from devices such as blood pressure monitors, blood glucose meters, and electrocardiographs as input, perform field-level semantic projection based on dynamic semantic benchmark templates, and generate an initial semantic embedding vector as the basic representation for subsequent label mapping; S7.3: Input the initial semantic embedding vector into the first channel of the edge-end dual-channel alignment network, execute the field name and benchmark label matching algorithm based on cosine similarity, calculate the semantic similarity score between each field and the standard terms in the dynamic semantic benchmark template, filter the candidate label set that is higher than the preset threshold, and generate the preliminary field-level mapping result; The initial semantic embedding vector is input into the first channel of the constructed edge-end dual-channel alignment network, and a field name and benchmark label matching algorithm based on cosine similarity (parameter: the embedding vector set of each standard term in the dynamic semantic benchmark template) is used to realize the semantic distance measurement between non-benchmark device fields and benchmark labels. Furthermore, the initial semantic embedding vector and the benchmark template vector are normalized by vector normalization method (parameter: L2 norm normalization factor) to achieve the unification of semantic vector scale and provide a numerical basis for cosine similarity calculation. Furthermore, the semantic similarity score is calculated using the cosine similarity formula:

[0029] in Embed a vector for the field to be matched. Embed a vector into the baseline template field. Describing the L2 norm, Let be the cosine similarity between the two. Furthermore, by setting a semantic similarity threshold (the parameter is automatically adjusted based on the task cycle of the dynamic template and the statistical distribution of the validation set), a set of candidate tags with a higher threshold is filtered out, thereby achieving automatic identification of field-tag pairs with good semantic matching. Furthermore, a preliminary field-level mapping result is generated using the association mapping table between the field source device identifier and the candidate label set, which includes the field name, the corresponding baseline label and its similarity score; Through the above matching algorithm and filtering process, the initial semantic embedding vector of the previous step is transformed into structured field-level preliminary mapping data, realizing high-precision semantic tag preliminary matching of multi-source heterogeneous devices in the edge environment; For example, in a deployment scenario at the edge of a primary healthcare institution, the initial semantic embedding vector dimension is 128, and the standard label vector dimension in the dynamic semantic baseline template is the same. The L2 norm normalization factor is set to... The cosine similarity threshold is set to In the actual calculation process, the embedding vectors of the blood pressure monitor field "BP_Sys" and the baseline template field "systolic blood pressure" are normalized, and the similarity is calculated according to the cosine similarity formula: , The similarity was obtained as If the value is higher than the threshold, "BP_Sys" is mapped to "systolic blood pressure" and a similarity score of 0.92 is recorded. After batch calculation, the blood glucose meter field "Glucose_Level" is matched as "fasting blood glucose" with a score of 0.88, and the electrocardiogram field "HR" is matched as "heart rate" with a score of 0.90. Finally, a preliminary field-level mapping result containing the device source identifier, standard label name, and matching score is formed for subsequent context consistency verification. S7.4: Input the initial semantic embedding vector into the second channel of the dual-channel alignment network simultaneously, combine it with the current patient context record (such as historical diagnosis, contracted service status, follow-up period), execute the context-aware consistency verification logic, perform conflict detection and ambiguity resolution on the candidate label set output by the first channel, and generate a localized consistent label cluster with context correction. S7.5: Perform structured encapsulation processing on the locally consistent tag cluster, generate a structured health tag sequence containing timestamp, device source, semantic baseline version number and confidence score according to the output format specifications required by the public health business system, and push it to the local application interface or central platform through a secure transmission protocol for downstream intelligent analysis modules such as follow-up management and risk warning to call.

[0030] Step S8: Monitor the feedback behavior of clinical personnel on the output label sequence, convert the correction operation into a fine-tuning signal and inject it into the incremental knowledge distillation process to continuously optimize the consistency maintenance capability between the dynamic semantic benchmark template and the shared semantic space. Specifically, this includes: S8.1: Obtain the correction operation log of the structured health label sequence output by the clinical staff on the edge end. The correction operation includes label replacement, label deletion or label addition. Capture the operation record with timestamp based on the human-computer interaction audit interface and generate a feedback event stream with labeled intent as the original input for subsequent fine-tuning signal extraction. S8.2: Perform semantic parsing on the labeled feedback event stream, use the rule engine to identify the mapping relationship between the original label and the target label corresponding to the correction behavior, combine the contextual patient health record data to calculate the semantic offset, and generate a set of label correction vectors with clinical significance as the basic data for model bias analysis. S8.3: Based on the label correction vector set, calculate its distribution density and clustering characteristics in the current shared semantic space, use an incremental clustering algorithm to identify high-frequency correction pattern clusters, determine whether there are any new semantic conflict types that have not been covered, and if so, trigger the boundary expansion mechanism of the dynamic semantic benchmark template to generate a set of semantic anchor candidates to be updated. S8.4: The label correction vector set is used as a supervision signal input to the lightweight encoder-decoder structure. The existing shared embedding space is used as the teacher model. The backpropagation calculation of the knowledge distillation loss function is performed to obtain the gradient update direction of the feature extractor parameters and generate fine-tuned weight increments for the new semantic alignment relationship, which are used to locally optimize the field projection capability of the student model. S8.5: Integrate the semantic anchor candidate set with the fine-tuned weight increment, perform version iteration operation on the dynamic semantic benchmark template and the global semantic alignment model, update the boundary definition of the shared semantic space through a weighted fusion strategy, and push the new version model to the edge inference module to complete hot replacement, so as to realize the continuous evolution of the label consistency generation capability.

[0031] The present invention also provides an intelligent processing system for public health physical examination follow-up family signature data based on multi-device acquisition, which uses the above-mentioned intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition to perform intelligent processing of public health physical examination follow-up family signature data.

[0032] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent processing of public health physical examination follow-up home signature data based on multi-device acquisition, characterized in that, Includes the following steps: S1: Based on the data collection needs of multi-source heterogeneous medical devices in public health physical examination and family doctor contract service scenarios, the system acquires raw data streams from different manufacturers' devices in real time, records the data format characteristics, field naming patterns, and unit expression methods of each device, and generates raw datasets on the device side. S2: Perform structured parsing and type normalization on the original dataset on the device side to generate standardized data units; S3: Based on the current business task objectives, identify the set of key health indicators required for chronic disease risk assessment, select the electronic health record system from all access devices as the initial semantic benchmark source, and generate a dynamic semantic benchmark template based on the initial semantic benchmark source; S4: Input the standardized data units of non-reference devices into the encoder-decoder structure, project their field semantics to a shared embedding space aligned with the dynamic semantic reference template, and generate a cross-device field-level semantic alignment vector; S5: Based on the cross-device field-level semantic alignment vector, a dual-channel alignment network is constructed, wherein the first channel performs semantic similarity matching between field names and baseline labels, and the second channel combines contextual record information to perform consistency verification and generate a preliminary consistent label cluster. S6: When a new device is added, the existing shared embedding space is used as a teacher model to guide the learning process of the new device's feature extractor. The boundary definition of the shared semantic space is updated through a knowledge distillation mechanism to generate an expanded global semantic alignment model.

2. The intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition as described in claim 1, characterized in that, Following step S6, the following is also included: S7: Deploy the global semantic alignment model on the edge inference module to perform localized label generation operation on the multi-source data flowing in in real time and output a structured health label sequence; S8: Monitor the feedback behavior of clinical staff on the structured health tag sequence, convert the correction operation into fine-tuning signal and inject it into the knowledge distillation process, and continuously optimize the consistency between the dynamic semantic benchmark template and the shared semantic space.

3. The intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition as described in claim 1, characterized in that, In step S1, the raw data streams from different manufacturers' devices are acquired in real time through the edge gateway module. The edge gateway module uses a multi-channel concurrency mechanism for concurrent data capture and is configured with a thread pool size of 1-64 and a buffer queue capacity of 1-4096.

4. The intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition as described in claim 1, characterized in that, Step S3 specifically includes: Based on the chronic disease management business needs in the context of public health physical examination and family doctor contract service, the current task target definition file is obtained, entity recognition and relation extraction are performed on the current task target definition file, key health indicator set is extracted, and standardized task semantic requirement vector is generated. Perform metadata scanning on the list of connected multi-source medical devices to obtain the data field list of each device and its corresponding semantic annotation information; based on the standardized task semantic requirement vector, calculate the field coverage score of each device on key health indicators and generate a device-level semantic coverage matrix. Based on the device-level semantic coverage matrix and combined with the external credibility scoring system, each device is comprehensively weighted and scored to determine the electronic health record system as the optimal semantic benchmark candidate source, and a benchmark source selection decision record is generated. Based on the structured data model of the electronic health record system, the core health indicator field names, unit systems, normal value ranges and related term codes are extracted and integrated into a standard field directory tree to generate an initial semantic benchmark ontology framework. The initial semantic benchmark ontology framework is injected into the dynamic template engine, bound to the current business task cycle and the population characteristic parameters, to generate a dynamic semantic benchmark template.

5. The intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition as described in claim 4, characterized in that, The external credibility scoring system includes the authority of the data source, the frequency of updates, and the clinical adoption rate.

6. The intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition as described in claim 1, characterized in that, Step S4 specifically includes: Based on the standardized data unit generated in step S2 and the dynamic semantic benchmark template generated in step S3, an encoder-decoder structure is constructed. The encoder uses a multi-head attention mechanism to jointly embed the field name, numerical range and unit information of the non-benchmark device to generate an initial field semantic representation vector. Perform layer normalization and positional encoding on the initial field semantic representation vector to output a normalized semantic embedding tensor; Based on the known semantic vectors of the authoritative fields in the dynamic semantic benchmark template, positive sample pairs and negative sample pairs are constructed. The encoder-decoder parameters are optimized using the contrastive loss function to minimize the embedding distance between synonymous fields and maximize the separation between heteronymous fields, thereby generating a shared embedding space. The normalized semantic embedding tensor of each field of the non-benchmark device is mapped to the shared embedding space, and its cosine similarity with the corresponding health index in the dynamic semantic benchmark template is calculated to obtain the semantic alignment score matrix. Based on the semantic alignment score matrix, high-confidence mapping relationships with scores above a preset threshold are selected to generate cross-device field-level semantic alignment vectors.

7. The intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition as described in claim 6, characterized in that, The cross-device field-level semantic alignment vector includes the field source device identifier, target baseline label, alignment confidence, and semantic offset.

8. The intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition as described in claim 1, characterized in that, Step S5 specifically includes: Based on the cross-device field-level semantic alignment vector generated in step S4, a field-level semantic similarity calculation model is constructed. The embedding vectors of non-benchmark device fields are matched one by one with the standard label vectors in the dynamic semantic benchmark template, and their semantic relevance scores are calculated to generate a field-level semantic similarity matrix. A thresholding clustering operation is performed on the field-level semantic similarity matrix to identify strongly associated field-label pairs that are higher than a preset semantic similarity threshold, a candidate mapping relationship set is selected, and unmatched fields are marked as items to be resolved, generating preliminary field mapping results; Obtain the contextual information stream associated with the current data record, extract the logical dependencies between multidimensional features based on the context-aware graph neural network model, and generate context-related embedding vectors; The candidate mapping relationship set and the context-related embedding vector are jointly reasoned. By constructing a differentiable logic rule layer to perform consistency constraint judgment, potential label conflicts or redundant combinations are detected, and priority arbitration is performed on contradictory labels to generate an optimized label set that has been corrected by context verification. The preliminary field mapping results are merged with the optimized tag set, and tag cluster merging and deduplication are performed to generate a preliminary consistent tag cluster.

9. The intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition as described in claim 8, characterized in that, The contextual information stream includes other physiological indicator values, historical diagnostic labels, and time series trends for the same patient during this physical examination session.

10. A smart data processing system for public health physical examination follow-up home signature data based on multi-device acquisition, characterized in that: The intelligent processing method for public health physical examination follow-up family signature data based on multi-device acquisition, as described in any one of claims 1-9, is used to perform intelligent processing of public health physical examination follow-up family signature data.

Images