Clinical information acquisition and synchronization system for digestive system department

Abstract

The invention relates to the technical field of data synchronization, in particular to a digestive system department clinical information acquisition and synchronization system, which comprises a multi-source heterogeneous acquisition module for generating a standardized diagnosis and treatment sequence, a characteristic spectrum construction module for mapping a text and an image into vector nodes and constructing a dynamic diagnosis and treatment characteristic spectrum, the state fingerprint verification module calculates a map fingerprint Hamming distance to position a difference feature node, and the incremental collaborative synchronization module constructs an incremental data packet and sends the incremental data packet to the central server. According to the method, the multi-dimensional characteristic spectrum based on the diagnosis and treatment time sequence is constructed, discrete images and texts are converted into topological association units, heterogeneous data semantic level alignment and integrity verification are achieved, version conflicts and information faults are eliminated, meanwhile, a dynamic hash fingerprint difference comparison strategy is adopted, accurate recognition is achieved, and only substantial change nodes are transmitted; and the network load is reduced, and high real-time consistency and zero-loss circulation of whole-flow information are ensured.

Description

Technical Field

[0001] This invention relates to the field of data synchronization technology, and in particular to a system for collecting and synchronizing clinical information in gastroenterology. Background Technology

[0002] The field of data synchronization technology involves the real-time consistency maintenance of data in various information management systems, as well as the aggregation and distribution mechanisms for multi-source heterogeneous data. Traditional gastroenterology clinical information collection and synchronization systems involve storing endoscopic reports, pathology results, and medical orders on a local server through manual input by doctors or one-way interface reading. These systems typically utilize scheduled scripts to scan change logs in the source database, perform batch data extraction and overwrite updates through standardized middleware, and complete data exchange between different departments based on fixed field mapping rules.

[0003] Existing data synchronization models heavily rely on timed full or incremental scanning mechanisms, which easily leads to data version conflicts when processing frequently updated endoscopic images and delayed pathology reports. Furthermore, the lack of semantic consistency verification between unstructured image data and structured text data causes information gaps and high concurrency congestion in transmission links when accessing data across departments. At the same time, the single field mapping logic cannot effectively identify the logical integrity of multi-source heterogeneous data, often resulting in the loss or misalignment of key diagnostic and treatment time-series data when facing network fluctuations. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a clinical information acquisition and synchronization system for gastroenterology.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a gastroenterology clinical information collection and synchronization system includes: A multi-source heterogeneous acquisition module is connected to the digestive endoscopy equipment and electronic medical record port to capture endoscopic image streams, diagnosis and treatment texts and form raw clinical data. The raw clinical data is cleaned and spatiotemporal label alignment is performed to generate patient identification, examination timestamps and construct standardized diagnosis and treatment sequences. The feature map construction module receives the standardized diagnosis and treatment sequence, extracts the diagnosis conclusion using semantic analysis, extracts the lesion texture features using image recognition, maps the diagnosis conclusion and the lesion texture features as multi-dimensional vector nodes, and establishes topological connections between the multi-dimensional vector nodes to construct a dynamic diagnosis and treatment feature map. The state fingerprint verification module performs hash dimensionality reduction on the dynamic diagnostic feature map to generate a local state fingerprint, calls the historical benchmark fingerprint, calculates the Hamming distance between the local state fingerprint and the historical benchmark fingerprint, and if the Hamming distance exceeds a preset threshold, it locates the difference feature nodes based on the XOR operation logic. The incremental collaborative synchronization module extracts the difference feature nodes and topological location coordinates, generates update instructions, constructs incremental synchronization data packets based on the update instructions, the difference feature nodes, and the topological location coordinates, and sends the incremental synchronization data packets to the central server.

[0006] As a further aspect of the present invention, the multi-source heterogeneous acquisition module includes: The data cleaning submodule is used to receive the raw clinical data, identify and filter out blurry frames and invalid black border areas in the endoscopic image stream, remove redundant format control characters in the diagnosis and treatment text, and generate clean clinical data. The spatiotemporal alignment submodule is used to acquire the clean clinical data, extract the image acquisition time point and the record time point of the diagnosis and treatment text in the endoscopic image stream, calculate the time deviation value between the two, and associate and bind the images and text with the time deviation value within the preset allowable range to generate spatiotemporal aligned data pairs. The sequence construction submodule is used to receive the spatiotemporally aligned data pairs, extract patient identity information to generate the patient identity identifier according to a preset data encapsulation protocol, extract the examination start time to generate the examination timestamp, and arrange the aligned image data and text data in chronological order to generate the standardized diagnosis and treatment sequence.

[0007] As a further aspect of the present invention, the feature map construction module includes: The semantic extraction submodule is used to receive the diagnostic text in the standardized diagnostic sequence, use a natural language processing model to identify the names of anatomical sites and keywords describing lesions, and output a structured diagnostic conclusion. The texture mapping submodule is used to receive image data from the standardized diagnostic sequence, perform gray-level co-occurrence matrix analysis on the region of interest, calculate energy, entropy, contrast and correlation parameters, generate the lesion texture features, and convert the diagnostic conclusion and the lesion texture features into the multidimensional vector nodes of the same dimension. The topology connection submodule is used to obtain all the multidimensional vector nodes, calculate the connection weights based on the temporal proximity and semantic relevance between nodes, establish directed edges between nodes whose relevance exceeds a preset connection threshold, and construct the dynamic diagnosis and treatment feature map.

[0008] As a further aspect of the present invention, the status fingerprint verification module includes: The fingerprint generation submodule is used to obtain the dynamic diagnostic feature map, traverse all node features and topological structures in the map, and use the locality-sensitive hashing algorithm to map the high-dimensional map features into a fixed-length binary string to generate the local state fingerprint. The distance calculation submodule is used to call the historical reference fingerprint stored in the local cache, compare the local state fingerprint with the historical reference fingerprint bit by bit, count the number of different bits to calculate the Hamming distance, and compare the Hamming distance with a preset threshold. The differential localization submodule is used to perform a bitwise XOR operation on the local state fingerprint and the historical reference fingerprint when the Hamming distance exceeds a preset threshold, identify the bit index position with a result of 1, and trace the graph node corresponding to the bit index position in reverse according to the hash mapping relationship to locate the changed differential feature node.

[0009] As a further aspect of the present invention, the incremental collaborative synchronization module includes: The node extraction submodule is used to obtain the located differential feature nodes, retrieve the current feature vector data of the nodes and their adjacency relationships in the current graph, and obtain the topological location coordinates; The instruction generation submodule is used to analyze the change type of the difference feature nodes, determine whether the change belongs to an addition, modification or deletion operation, and generate the corresponding database operation instruction as the update instruction; The package construction submodule is used to serialize and encode the update instruction, the feature vector data of the difference feature nodes, and the topological location coordinates, add check bits and version numbers, and encapsulate them into an independent incremental synchronization data packet.

[0010] As a further aspect of the present invention, the specific function of the spatiotemporal alignment submodule is as follows: Obtain the endoscopic image frame sequence and electronic medical record text paragraph from the pure clinical data, and parse the hardware clock stamp in the image frame metadata and the server record time in the text paragraph respectively; Based on the server's recorded time, linear drift correction is performed on the hardware clock stamp, and the absolute difference between the corrected image time and the text time is calculated. Image frames and text segments with an absolute difference less than a preset synchronization window threshold are selected, a unique association index key is established, and the successfully associated images and texts are combined and marked with a unified logical time tag to generate the spatiotemporal aligned data pair.

[0011] As a further aspect of the present invention, the specific functional implementation of the texture mapping submodule is as follows: The endoscopic images in the standardized diagnosis and treatment sequence are acquired, the lesion area is located and the normalized gray-level matrix is ​​calculated, the feature values ​​of the gray-level co-occurrence matrix are extracted, and the semantic encoding vector of the diagnosis conclusion is combined to construct the multi-dimensional vector node based on weighted fusion logic. The construction process of the multidimensional vector nodes must follow the following feature fusion formula: ; in, Represents the generated multidimensional vector node, The vector representation of the extracted lesion texture features. The vector norm representing the texture features of the lesion. The semantic embedding vector representing the diagnostic conclusion. Represents the preset semantic projection matrix. This represents the preset bias term. This represents the preset visual feature weighting coefficients. This represents the hyperbolic tangent activation function.

[0012] As a further aspect of the present invention, the specific function of the fingerprint generation submodule is as follows: Obtain each of the multidimensional vector nodes in the dynamic diagnostic feature map, generate a random projection vector of target length for each node, and calculate the dot product between the node vector and the random projection vector; Based on the sign of the dot product result, each node is mapped to a binary bit sequence. A weighted summation operation is performed on the binary bit sequences of all nodes to generate a comprehensive feature weight vector. Each dimension of the comprehensive feature weight vector is binarized and reduced in dimensionality. When the value is greater than zero, it is set to 1; otherwise, it is set to 0. The combined values ​​generate the local state fingerprint that can globally represent the state of the map.

[0013] As a further aspect of the present invention, the specific execution process of the differential localization submodule is as follows: Perform an XOR operation between the local state fingerprint and the historical reference fingerprint to generate a difference indicator mask string, and extract all bit index positions in the difference indicator mask string where the value is 1; Traverse the bit index positions, query the pre-built hash bucket mapping table, obtain the candidate node list that falls into the index bucket, and calculate the Euclidean distance between the current features and historical snapshot features of multiple nodes in the candidate node list. Nodes with non-zero Euclidean distances are selected and marked as target objects whose states have undergone substantial changes. The unique ID of the object is extracted as the difference feature node, and the hierarchical index of the node in the graph structure is output as the topological position coordinate.

[0014] As a further aspect of the present invention, the specific functional implementation of the package construction submodule is as follows: The update instruction, the data entity of the difference feature node, and the topological location coordinates are obtained. The data entity is compressed using a preset dictionary compression algorithm to generate a compressed data block. Construct a synchronization header, write the current system version serial number, the length information of the incremental synchronization data packet and the encryption check code into the synchronization header, and fill the compressed data block into the packet body area; Calculate the cyclic redundancy check value of the synchronization header and body regions, append the check value to the end of the packet, complete the encapsulation, and generate the incremental synchronization data packet.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by constructing a multidimensional feature map based on the diagnosis and treatment time sequence, discrete endoscopic images and pathological texts are transformed into standardized data units with topological associations. This achieves semantic-level alignment and integrity verification of heterogeneous clinical data, eliminating version conflicts and information gaps in high-concurrency scenarios. At the same time, by adopting an incremental difference comparison strategy based on dynamic hash fingerprints, the invention accurately identifies and transmits only data nodes that have undergone substantial changes, significantly reducing the network load of cross-departmental synchronization and ensuring high real-time consistency and zero-loss flow of information throughout the entire gastroenterology diagnosis and treatment process across multiple terminals. Attached Figure Description

[0016] Figure 1 This is a block diagram illustrating the principle of the gastroenterology clinical information acquisition and synchronization system of the present invention. Figure 2 This is a flowchart of the operation of the multi-source heterogeneous acquisition module of the present invention; Figure 3 This is a flowchart illustrating the construction process of the dynamic diagnostic and therapeutic feature map of this invention. Figure 4 This is a flowchart of the status fingerprint verification and difference location process of the present invention; Figure 5 This is a flowchart of the incremental collaborative synchronization module execution of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0018] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0019] Please see Figure 1 and Figure 2 This invention provides a technical solution: a gastroenterology clinical information collection and synchronization system comprising: The multi-source heterogeneous acquisition module connects to the digestive endoscopy equipment and electronic medical record port, captures endoscopic image streams and diagnostic texts and assembles them into raw clinical data, cleans the raw clinical data and performs spatiotemporal label alignment, generates patient identification, examination timestamps and constructs standardized diagnostic and treatment sequences.

[0020] The multi-source heterogeneous acquisition module includes: The data cleaning submodule is used to receive raw clinical data, identify and filter out blurry frames and invalid black border areas in the endoscopic image stream, remove redundant format control characters in the diagnosis and treatment text, and generate clean clinical data. The spatiotemporal alignment submodule is used to acquire clean clinical data, extract the image acquisition time point and the record time point of the diagnosis and treatment text in the endoscopic image stream, calculate the time deviation value between the two, and associate and bind the images and texts with time deviation values ​​within the preset allowable range to generate spatiotemporal aligned data pairs. The specific functionalities of the spatiotemporal alignment submodule are as follows: Obtain endoscopic image frame sequences and electronic medical record text segments from clean clinical data, and parse the hardware clock stamp in the image frame metadata and the server record time in the text segment respectively; Based on the server's recorded time, linear drift correction is performed on the hardware clock stamp, and the absolute difference between the corrected image time and the text time is calculated. Filter image frames and text segments whose absolute difference is less than the preset synchronization window threshold, establish a unique association index key, combine successfully associated images and texts and mark them with a unified logical time label to generate spatiotemporal aligned data pairs; The sequence construction submodule is used to receive spatiotemporally aligned data pairs, extract patient identity information to generate patient identity identifiers according to a preset data encapsulation protocol, extract examination start time to generate examination timestamps, and arrange the aligned image data and text data in chronological order to generate a standardized diagnosis and treatment sequence.

[0021] The multi-source heterogeneous acquisition module, serving as the system's sensing front-end, physically connects to mainstream endoscopy mainframes such as Olympus or Fujinon via dual-channel HD-SDI and HDMI interface cards. Simultaneously, it interfaces with the hospital's HIS (Hospital Information System) and EMR (Electronic Medical Record) server ports via HL7 / FHIR standard interfaces. The data cleaning submodule features a FPGA-based hardware-accelerated preprocessing unit at the receiving end. For endoscopy image streams, this unit receives raw YUV4:2:2 format video data at a resolution of 1920×1080 and a frame rate of 60fps. To identify blurred frames, this submodule does not rely on a black-box model but directly runs the Laplacian operator to perform second-order differential operations on the luminance component (Y channel) of each frame, calculating the grayscale variance of all pixels in the image. The system's preset fuzzy detection threshold is set to 100.0. This threshold is selected based on statistical analysis of 5000 endoscopic images under different lighting conditions. Experimental data shows that when the Laplacian variance is below 100.0, the high-frequency texture loss rate of the image exceeds 85%, making it unusable for subsequent lesion identification. For invalid black border areas, the submodule loads a preset mask matrix. This matrix is ​​set for different endoscope models. For example, for the GIF-H260 model, the mask defines a 45-pixel-wide area around the four edges of the image as an invalid area, directly setting the pixel value of this area to zero or cropping it. For diagnostic text, the submodule has a built-in regular expression engine to execute specific character filtering logic. It loads the pattern string [\x00-\x1F\x7F] to remove ASCII control characters and uses specific replacement rules to remove consecutive HTML tags (such as...). 、 ) is converted to standard newline characters or spaces to ensure the purity of text data.

[0022] The Laplacian operator mentioned above is a second-order differential operator in image processing, mainly used to detect edge and texture information in images. It quantifies the image sharpness by calculating the rate of gray-level change in the neighborhood of a pixel.

[0023] After acquiring clean data, the core task of the spatiotemporal alignment submodule is to resolve the clock asynchrony issue between heterogeneous devices. This submodule internally maintains a time correction model based on linear regression. Specifically, the submodule periodically (e.g., every 10 minutes) sends NTP time synchronization commands to the endoscope equipment, recording the correspondence between the endoscope hardware clock stamp and the server's standard time. Due to the temperature drift effect of the endoscope equipment's crystal oscillator, the submodule uses the least squares method to fit the drift curve. ,in For hardware clock, The server's base time. For drift rate, This represents the initial deviation. In actual operation, the drift rate is calculated by continuously collecting 10 sets of time synchronization data. The value is typically between 1.000002 and 1.000005. Using this fitting formula, the submodule maps the hardware timestamp of each frame of the video to a unified timeline on the server and calculates the absolute difference between it and the time of the medical record text recording. The preset synchronization window threshold is strictly set to 2000 milliseconds (2 seconds). This threshold is based on the fact that the average reaction delay of a physician in discovering a lesion and making text recordings (such as using a foot switch or voice-to-text input) is 1.5 seconds, plus a system processing jitter of 0.5 seconds. Experimental data shows that when the synchronization window is set to 2000 milliseconds, the matching accuracy between the image and the description reaches 98.2%, which is a 15% improvement in correlation accuracy compared to the traditional method of aligning only the start time.

[0024] The NTP mentioned above refers to the Network Time Protocol, which is used to synchronize clocks in computer networks through packet switching, so that the computer clocks in the network are aligned with Coordinated Universal Time (UTC).

[0025] The sequence construction submodule operates based on the DICOM-SR standard and JSON encapsulation protocol. When extracting patient identification, the submodule parses the PID segment (PatientID) in the HL7 message header and combines it with the date string of the examination day to generate a unique UUID. For the examination timestamp, the submodule locks the corrected time of the first frame of the image stream as the baseline. During the data arrangement stage, the submodule creates a doubly linked list structure, with each node containing a "data type (image / text)", a "unified logical timestamp", and a "data body pointer". The submodule traverses all successfully associated image frames and text segments, inserting them into the linked list in ascending order of timestamp. To ensure sequence compactness, the submodule uses H.265 encoding for inter-frame compression of the image data and UTF-8 encoding for the text data. Finally, the linked list is serialized into a binary stream, and a metadata header containing the patient ID, examination ID, and total duration is written to the beginning, generating a standardized diagnosis and treatment sequence file. This provides a time-series-aligned input source for subsequent feature extraction.

[0026] The aforementioned DICOM-SR refers to the structured reporting component of the medical digital imaging and communication standard, which is used to store and transmit medical diagnostic report data in a structured form to ensure data interoperability between different medical devices and systems.

[0027] Please see Figure 1 and Figure 3 The feature map construction module receives standardized diagnosis and treatment sequences, extracts diagnostic conclusions using semantic analysis, extracts lesion texture features using image recognition, maps diagnostic conclusions and lesion texture features into multi-dimensional vector nodes, and establishes topological connections between multi-dimensional vector nodes to construct a dynamic diagnosis and treatment feature map.

[0028] The feature map construction module includes: The semantic extraction submodule is used to receive the diagnostic text in the standardized diagnostic sequence, use a natural language processing model to identify the names of anatomical sites and keywords describing lesions, and output a structured diagnostic conclusion. The texture mapping submodule is used to receive image data from standardized diagnostic sequences, perform gray-level co-occurrence matrix analysis on regions of interest, calculate energy, entropy, contrast and correlation parameters, generate lesion texture features, and transform diagnostic conclusions and lesion texture features into multi-dimensional vector nodes of the same dimension. The specific functionality of the texture mapping submodule is as follows: Acquire endoscopic images from standardized diagnostic and treatment sequences, locate lesion areas and calculate normalized gray-level matrices, extract eigenvalues ​​from gray-level co-occurrence matrices, combine semantic encoding vectors of diagnostic conclusions, and construct multidimensional vector nodes based on weighted fusion logic; The construction process of multidimensional vector nodes must follow the following feature fusion formula: ; in, Represents the generated multidimensional vector nodes. The vector representation of the extracted lesion texture features. The vector norm representing the texture features of the lesion. The semantic embedding vector representing the diagnostic conclusion. Represents the preset semantic projection matrix. This represents the preset bias term. This represents the preset visual feature weighting coefficients. Represents the hyperbolic tangent activation function; The topology connection submodule is used to obtain all multi-dimensional vector nodes, calculate connection weights based on the temporal proximity and semantic relevance between nodes, establish directed edges between nodes whose relevance exceeds a preset connection threshold, and construct a dynamic diagnostic feature map.

[0029] The feature map construction module is the core component for mapping from raw data to a high-dimensional semantic space. The semantic extraction submodule internally loads a BERT-Base model fine-tuned for Chinese gastroenterology corpora. This model architecture includes a 12-layer Transformer encoder, each layer equipped with 12 self-attention heads, and a 768-dimensional hidden layer. When processing diagnostic texts, the submodule first performs word segmentation and stop word removal on the input text, with a maximum sequence length limit of 512 tokens. The model training process employs a multi-task learning strategy combining Masked Language Modeling (MLM) and Named Entity Recognition (NER) tasks, with the loss function consisting of cross-entropy loss. The optimization algorithm used is AdamW, with an initial learning rate of 2e-5 and a weight decay coefficient of 0.01. After training on 100,000 annotated endoscopic reports, the model can accurately output 768-dimensional semantic embedding vectors representing anatomical locations (such as "gastric antrum" and "duodenal bulb") and lesion descriptions (such as "congestion," "erosion," and "ulcer"), and output structured diagnostic conclusion text.

[0030] The aforementioned BERT-Base model refers to a pre-trained language model based on Transformer bidirectional encoder representation. Through unsupervised pre-training on a large-scale corpus, it can gain a deep understanding of the contextual semantic relationships of natural language.

[0031] The texture mapping submodule focuses on the quantization of image features, with its core being the construction and analysis of the gray-level co-occurrence matrix. The submodule first converts the located lesion region into an 8-bit grayscale image and then quantizes and compresses the grayscale levels from 0-255 to 0-63 (64 levels in total) to reduce computation and suppress noise. Next, the submodule... ,angle Co-occurrence matrices are calculated at four directions: 0°, 45°, 90°, and 135°. Four second-order statistics—energy, entropy, contrast, and correlation—are extracted from each matrix. The eigenvalues ​​in the four directions are averaged to form a 16-dimensional basic texture feature vector. To fuse with the semantic vector, this 16-dimensional vector is projected to 768 dimensions through a fully connected layer and then L2 normalized to obtain the final texture feature vector.

[0032] The aforementioned gray-level co-occurrence matrix refers to a method that describes texture features by statistically analyzing the spatial gray-level co-occurrence patterns of two pixels in an image. It can reflect comprehensive information about the image in terms of direction, spacing, magnitude of change, and speed.

[0033] During the construction of multi-dimensional vector nodes, the texture mapping submodule strictly adheres to the feature fusion formula. The formula is expressed as follows: ; in, Represents the generated multidimensional vector nodes. The vector representation of the extracted lesion texture features. The vector norm representing the texture features of the lesion. The semantic embedding vector representing the diagnostic conclusion. Represents the preset semantic projection matrix. This represents the preset bias term. This represents the preset visual feature weighting coefficients. This represents the hyperbolic tangent activation function.

[0034] Regarding the parameters in the formula The weight coefficient for this visual feature was set to 0.65. This value was not chosen arbitrarily, but was determined based on ROC curve analysis: in the task of distinguishing between polyps and early cancer, the contribution of visual texture features (AUC=0.88) was slightly higher than that of simple textual semantic description (AUC=0.74), so visual features were given a higher weight. It is a dimension of The semantic projection matrix, whose initial values ​​are obtained through orthogonal initialization, is updated during joint model training. Bias term. Initialize to a zero vector. (Introduced in the formula) The purpose of the activation function is to compress the numerical range of semantic features to the [-1, 1] interval, maintaining a consistent magnitude with the normalized texture features and preventing numerical divergence. An example illustrating the calculation process of this formula is provided: assuming the normalized lesion texture features... vector norm The value is 1 (because it has been normalized), and the extracted value is 1. The component in one dimension is 0.8. Semantic embedding vector. go through After projection and bias addition, the median value obtained is 0.5. Substituting this into the formula, It is approximately 0.462. At this point, the fusion feature value in this dimension is calculated as follows: The calculated result, 0.6817, represents the generated multidimensional vector node. The specific numerical values ​​in this dimension. The results show that the fused node features retain strong visual texture attributes while incorporating a moderate amount of semantic contextual information.

[0035] The topology connectivity submodule is responsible for constructing the graph structure. It traverses all generated multidimensional vector nodes, calculating the cosine similarity between nodes as a semantic relevance weight, and the reciprocal of the timestamp difference between nodes as a temporal proximity weight. The system sets a connectivity threshold of 0.75, determined through connectivity tests on known disease progression graphs. When the threshold is below 0.75, the graph contains numerous meaningless noise connections (such as connecting antral gastritis with irrelevant esophageal descriptions); when it is above 0.75, potential complication associations are lost. The submodule only establishes directed edges between node pairs with a combined weight exceeding 0.75, with the edge direction determined by temporal sequence, thereby constructing a dynamic diagnostic and treatment feature graph that characterizes the evolution of the treatment process.

[0036] Please see Figure 1 and Figure 4 The state fingerprint verification module performs hash dimensionality reduction on the dynamic diagnostic feature map to generate a local state fingerprint, calls the historical benchmark fingerprint, calculates the Hamming distance between the local state fingerprint and the historical benchmark fingerprint, and if the Hamming distance exceeds a preset threshold, it locates the difference feature nodes based on the XOR operation logic.

[0037] The status fingerprint verification module includes: The fingerprint generation submodule is used to obtain dynamic diagnostic feature maps, traverse all node features and topological structures in the map, and use the locality-sensitive hashing algorithm to map the high-dimensional map features into a fixed-length binary string to generate local state fingerprints. The specific functions of the fingerprint generation submodule are as follows: Obtain each multidimensional vector node in the dynamic diagnostic feature map, generate a random projection vector of target length for each node, and calculate the dot product between the node vector and the random projection vector. Based on the sign of the dot product result, each node is mapped to a binary bit sequence. A weighted summation operation is performed on the binary bit sequences of all nodes to generate a comprehensive feature weight vector. The numerical values ​​of each dimension in the comprehensive feature weight vector are binarized and reduced in dimensionality. When the value is greater than zero, it is set to 1, otherwise it is set to 0. The combination generates a local state fingerprint that can globally represent the state of the map. The distance calculation submodule is used to call the historical reference fingerprint stored in the local cache, compare the local state fingerprint with the historical reference fingerprint bit by bit, count the number of different bits to calculate the Hamming distance, and compare the Hamming distance with a preset threshold. The differential localization submodule is used to perform a bitwise XOR operation on the local state fingerprint and the historical baseline fingerprint when the Hamming distance exceeds a preset threshold, identify the bit index position with a result of 1, and trace the graph node corresponding to the bit index position in reverse according to the hash mapping relationship to locate the differential feature node that has changed. The specific execution process of the differential localization submodule is as follows: Perform an XOR operation between the local state fingerprint and the historical baseline fingerprint to generate a difference indicator mask string, and extract all bit index positions in the difference indicator mask string where the value is 1; Traverse the bit index positions, query the pre-built hash bucket mapping table, obtain the list of candidate nodes that fall into the index bucket, and calculate the Euclidean distance between the current features and historical snapshot features of multiple nodes in the candidate node list. Filter nodes with non-zero Euclidean distance, mark them as target objects whose state has undergone substantial changes, extract the unique ID of the object as the difference feature node, and output the node's hierarchical index in the graph structure as the topological position coordinate.

[0038] The state fingerprint verification module employs Locality Sensitive Hashing (LSH) to achieve rapid fingerprinting of high-dimensional graph features, aiming to realize state change detection with extremely low storage cost. The core mechanism of the fingerprint generation submodule is random projection. This submodule pre-defines a projection matrix containing 128 randomly generated vectors. The dimension of each vector is consistent with the dimension of the multidimensional vector node (768 dimensions), and the elements follow a Gaussian distribution with a mean of 0 and a variance of 1.

[0039] Locality-Sensitive Hash (LSH) is an algorithm technique for finding similar data in a high-dimensional space. Its core idea is to map similar data items to the same hash bucket with a high probability, thereby improving the efficiency of similarity retrieval.

[0040] In practice, the submodule obtains the vector of each node in the graph and performs a dot product operation with each of these 128 random vectors. According to the SimHash algorithm, if the dot product result is greater than 0, the corresponding bit is marked as 1; otherwise, it is marked as -1 (mapped to 0 during subsequent binarization). To integrate the topological information of the graph, the submodule does not only hash individual nodes but also employs a weighted stacking strategy: for each node in the graph, its own hash bit sequence (a sequence of 1 / -1) is multiplied by its PageRank weight, and then the weighted sequences of all nodes are summed bit by bit to generate a 128-dimensional comprehensive feature weight vector. Finally, this comprehensive vector is binarized: if the accumulated value of a certain dimension is greater than 0, the corresponding bit of the final fingerprint is 1; otherwise, it is 0. The resulting 128-bit (16-byte) binary string is the local state fingerprint.

[0041] The PageRank weight mentioned above is an algorithmic metric used to measure the importance of nodes in a graph structure. By analyzing the connection relationships between nodes, nodes with rich connections or those connected by important nodes are given higher weight values.

[0042] The distance calculation submodule is responsible for comparing the current fingerprint with the historical baseline fingerprint. The historical baseline fingerprint is stored in a high-speed Redis cache. After reading both, the submodule performs an XOR bitwise operation and counts the number of "1"s in the result, which is the Hamming distance. The system's preset Hamming distance threshold is 3. This threshold is set based on large-scale disturbance testing: experiments show that small feature fluctuations caused only by sensor thermal noise usually result in a 1-2 bit change in Hamming distance, while any substantial modification to the disease description or change in the labeling of lesions on imaging will cause at least a 4-bit change in fingerprint. Therefore, setting the threshold to 3 can effectively filter system noise and ensure the accuracy of alarms.

[0043] The Hamming distance mentioned above refers to the number of different characters at corresponding positions between two strings of equal length, and is used to measure the degree of difference between two sequences.

[0044] The difference localization submodule is triggered when the Hamming distance exceeds 3. It first extracts all bit indices whose XOR result is 1 (e.g., bits 5, 12, and 89). To trace back to the nodes causing the differences, the system pre-builds a bucketing mapping table. This table is established synchronously during fingerprint generation and records which nodes (i.e., nodes with larger absolute values ​​of the bit's projected weight) primarily influence each bit index. The submodule queries this mapping table to obtain a list of candidate nodes. Subsequently, the submodule calculates the Euclidean distance between the current feature vector of these candidate nodes and the corresponding vector in the historical snapshot. During this process, the system sets the threshold for Euclidean distance judgment to 1e-5; any node with a distance greater than this value is judged as a substantially changed node.

[0045] Table 1 shows the test data of system false alarm rate and false negative rate under different Hamming distance threshold settings.

[0046] Table 1. Hamming Distance Threshold Performance Test Table ; As shown in Table 1, when the threshold is set to 3, the system maintains an extremely low false alarm rate (0.5%) while keeping the false alarm rate below 0.8%, and the detection time is stable at around 4ms, proving the rationality of this parameter setting. The submodule finally outputs the unique ID of the determined differential feature node and its hierarchical index in the map, completing the accurate localization.

[0047] Please see Figure 1 and Figure 5 The incremental collaborative synchronization module extracts the difference feature nodes and topological location coordinates, generates update instructions, constructs incremental synchronization data packets based on update instructions, difference feature nodes, and topological location coordinates, and sends the incremental synchronization data packets to the central server.

[0048] The incremental collaborative synchronization module includes: The node extraction submodule is used to obtain the located differential feature nodes, retrieve the current feature vector data of the nodes and their adjacency relationships in the current graph, and obtain the topological location coordinates. The instruction generation submodule is used to analyze the change type of the difference feature nodes, determine whether the change belongs to the addition, modification or deletion operation, and generate the corresponding database operation instruction as the update instruction; The package construction submodule is used to serialize and encode the update command, the feature vector data of the difference feature nodes and the topological location coordinates, add check bits and version numbers, and encapsulate them into an independent incremental synchronization data package. The specific functionalities of the package building submodule are as follows: Obtain the update command, the data entity of the difference feature node and the topological location coordinates, and compress the data entity using a preset dictionary compression algorithm to generate compressed data blocks; Construct a synchronization header, write the current system version serial number, the length information of the incremental synchronization data packet, and the encryption check code into the synchronization header, and fill the compressed data block into the packet body area; Calculate the cyclic redundancy check value of the synchronization header and body regions, append the check value to the end of the packet, complete the encapsulation, and generate an incremental synchronization data packet.

[0049] The incremental collaborative synchronization module is responsible for efficiently and securely transmitting the located change data to the central server, minimizing network bandwidth consumption. The node extraction submodule directly accesses the graph object in memory based on the ID output by the difference positioning submodule. It extracts not only the node's 768-dimensional feature vector data but also its first-degree adjacency matrix, i.e., all incoming and outgoing edge information for the node. For topological location coordinates, the submodule extracts the node's level depth and its same-level sorting index; these two parameters constitute the node's spatial coordinate system.

[0050] The instruction generation submodule generates standardized SQL operation instructions or NoSQL document update instructions by comparing change types. If the difference node does not exist in the historical snapshot, it is determined as "added" and an INSERT instruction is generated; if it exists but the feature distance exceeds the limit, it is determined as "modified" and an UPDATE instruction is generated; if it exists in the past but is missing in the current map (which requires full map fingerprint verification), a DELETE instruction is generated.

[0051] The package construction submodule is responsible for data serialization and encapsulation. For optimal transmission efficiency, this submodule does not use the verbose JSON format, but instead employs Protocol Buffers for binary serialization. Before serialization, the submodule compresses the feature vector data (floating-point arrays) using the Zstandard (Zstd) algorithm. The Zstd dictionary, pre-generated based on a large amount of medical feature data, is 10MB in size and is stored on both the client and server sides.

[0052] Protocol Buffers, mentioned above, refers to a language-neutral, platform-neutral, and scalable method for serializing structured data, developed by Google. It is commonly used in communication protocols, data storage, and other fields, and has a smaller file size and faster parsing speed than XML and JSON.

[0053] The packet construction submodule strictly adheres to a custom protocol format when constructing the synchronization header: the first 4 bytes are the magic number, followed by 2 bytes for the version sequence number, and then 4 bytes for the payload length. When filling the packet body, the submodule first writes the compressed data block and then encrypts it. The encryption algorithm uses AES-256-GCM mode, and the key is dynamically derived from each hospital's root key using the HKDF algorithm to ensure the confidentiality and integrity of data transmission. Finally, the submodule calculates the CRC32 checksum of the synchronization header and the encrypted packet body. The standard CRC32 polynomial is 0x04C11DB7. Assuming the calculated CRC32 checksum is 0xA1B2C3D4, this 4-byte checksum is appended to the end of the data packet.

[0054] The aforementioned AES-256-GCM refers to an encryption algorithm that combines the Advanced Encryption Standard (AES) with the Galois / Counter Mode (GCM), capable of simultaneously providing data confidentiality, integrity, and authenticity authentication.

[0055] The following example illustrates the advantages of incremental synchronization over full synchronization. Assume the graph contains 1000 nodes, each with 4KB of data. Full synchronization would require transmitting approximately 4MB of data. If only one node experiences a feature change (the changed portion is compressed to 200 bytes), plus the header (16 bytes), instruction information (50 bytes), and CRC (4 bytes), the total size of the incremental synchronization data packet is approximately 270 bytes. The transmission ratio is then calculated. The results show that, in single-point change scenarios, the incremental synchronization data packet size generated by this scheme is only 0.06% of the total data size, greatly saving network bandwidth resources, and is particularly suitable for mobile ward round scenarios with unstable network environments. The finally generated binary data packet is pushed to a message queue, awaiting transmission to the central server.

[0056] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A clinical information acquisition and synchronization system for gastroenterology, characterized in that, The system includes: A multi-source heterogeneous acquisition module is connected to the digestive endoscopy equipment and electronic medical record port to capture endoscopic image streams, diagnosis and treatment texts and form raw clinical data. The raw clinical data is cleaned and spatiotemporal label alignment is performed to generate patient identification, examination timestamps and construct standardized diagnosis and treatment sequences. The feature map construction module receives the standardized diagnosis and treatment sequence, extracts the diagnosis conclusion using semantic analysis, extracts the lesion texture features using image recognition, maps the diagnosis conclusion and the lesion texture features as multi-dimensional vector nodes, and establishes topological connections between the multi-dimensional vector nodes to construct a dynamic diagnosis and treatment feature map. The state fingerprint verification module performs hash dimensionality reduction on the dynamic diagnostic feature map to generate a local state fingerprint, calls the historical benchmark fingerprint, calculates the Hamming distance between the local state fingerprint and the historical benchmark fingerprint, and if the Hamming distance exceeds a preset threshold, it locates the difference feature nodes based on the XOR operation logic. The incremental collaborative synchronization module extracts the difference feature nodes and topological location coordinates, generates update instructions, constructs incremental synchronization data packets based on the update instructions, the difference feature nodes, and the topological location coordinates, and sends the incremental synchronization data packets to the central server.

2. The gastroenterology clinical information acquisition and synchronization system according to claim 1, characterized in that, The multi-source heterogeneous acquisition module includes: The data cleaning submodule is used to receive the raw clinical data, identify and filter out blurry frames and invalid black border areas in the endoscopic image stream, remove redundant format control characters in the diagnosis and treatment text, and generate clean clinical data. The spatiotemporal alignment submodule is used to acquire the clean clinical data, extract the image acquisition time point and the record time point of the diagnosis and treatment text in the endoscopic image stream, calculate the time deviation value between the two, and associate and bind the images and text with the time deviation value within the preset allowable range to generate spatiotemporal aligned data pairs. The sequence construction submodule is used to receive the spatiotemporally aligned data pairs, extract patient identity information to generate the patient identity identifier according to a preset data encapsulation protocol, extract the examination start time to generate the examination timestamp, and arrange the aligned image data and text data in chronological order to generate the standardized diagnosis and treatment sequence.

3. The gastroenterology clinical information acquisition and synchronization system according to claim 1, characterized in that, The feature map construction module includes: The semantic extraction submodule is used to receive the diagnostic text in the standardized diagnostic sequence, use a natural language processing model to identify the names of anatomical sites and keywords describing lesions, and output a structured diagnostic conclusion. The texture mapping submodule is used to receive image data from the standardized diagnostic sequence, perform gray-level co-occurrence matrix analysis on the region of interest, calculate energy, entropy, contrast and correlation parameters, generate the lesion texture features, and convert the diagnostic conclusion and the lesion texture features into the multidimensional vector nodes of the same dimension. The topology connection submodule is used to obtain all the multidimensional vector nodes, calculate the connection weights based on the temporal proximity and semantic relevance between nodes, establish directed edges between nodes whose relevance exceeds a preset connection threshold, and construct the dynamic diagnosis and treatment feature map.

4. The gastroenterology clinical information acquisition and synchronization system according to claim 1, characterized in that, The status fingerprint verification module includes: The fingerprint generation submodule is used to obtain the dynamic diagnostic feature map, traverse all node features and topological structures in the map, and use the locality-sensitive hashing algorithm to map the high-dimensional map features into a fixed-length binary string to generate the local state fingerprint. The distance calculation submodule is used to call the historical reference fingerprint stored in the local cache, compare the local state fingerprint with the historical reference fingerprint bit by bit, count the number of different bits to calculate the Hamming distance, and compare the Hamming distance with a preset threshold. The differential localization submodule is used to perform a bitwise XOR operation on the local state fingerprint and the historical reference fingerprint when the Hamming distance exceeds a preset threshold, identify the bit index position with a result of 1, and trace the graph node corresponding to the bit index position in reverse according to the hash mapping relationship to locate the changed differential feature node.

5. The gastroenterology clinical information acquisition and synchronization system according to claim 1, characterized in that, The incremental collaborative synchronization module includes: The node extraction submodule is used to obtain the located differential feature nodes, retrieve the current feature vector data of the nodes and their adjacency relationships in the current graph, and obtain the topological location coordinates; The instruction generation submodule is used to analyze the change type of the difference feature nodes, determine whether the change belongs to an addition, modification or deletion operation, and generate the corresponding database operation instruction as the update instruction; The package construction submodule is used to serialize and encode the update instruction, the feature vector data of the difference feature nodes, and the topological location coordinates, add check bits and version numbers, and encapsulate them into an independent incremental synchronization data packet.

6. The gastroenterology clinical information acquisition and synchronization system according to claim 2, characterized in that, The specific functional implementation of the spatiotemporal alignment submodule is as follows: Obtain the endoscopic image frame sequence and electronic medical record text paragraph from the pure clinical data, and parse the hardware clock stamp in the image frame metadata and the server record time in the text paragraph respectively; Based on the server's recorded time, linear drift correction is performed on the hardware clock stamp, and the absolute difference between the corrected image time and the text time is calculated. Image frames and text segments with an absolute difference less than a preset synchronization window threshold are selected, a unique association index key is established, and the successfully associated images and texts are combined and marked with a unified logical time tag to generate the spatiotemporal aligned data pair.

7. The gastroenterology clinical information acquisition and synchronization system according to claim 3, characterized in that, The specific functional implementation of the texture mapping submodule is as follows: The endoscopic images in the standardized diagnosis and treatment sequence are acquired, the lesion area is located and the normalized gray-level matrix is ​​calculated, the feature values ​​of the gray-level co-occurrence matrix are extracted, and the semantic encoding vector of the diagnosis conclusion is combined to construct the multi-dimensional vector node based on weighted fusion logic. The construction process of the multidimensional vector nodes must follow the following feature fusion formula: ； in, Represents the generated multidimensional vector node, The vector representation of the extracted lesion texture features. The vector norm representing the texture features of the lesion. The semantic embedding vector representing the diagnostic conclusion. Represents the preset semantic projection matrix. This represents the preset bias term. This represents the preset visual feature weighting coefficients. This represents the hyperbolic tangent activation function.

8. The gastroenterology clinical information acquisition and synchronization system according to claim 4, characterized in that, The specific functions of the fingerprint generation submodule are as follows: Obtain each of the multidimensional vector nodes in the dynamic diagnostic feature map, generate a random projection vector of target length for each node, and calculate the dot product between the node vector and the random projection vector; Based on the sign of the dot product result, each node is mapped to a binary bit sequence. A weighted summation operation is performed on the binary bit sequences of all nodes to generate a comprehensive feature weight vector. Each dimension of the comprehensive feature weight vector is binarized and reduced in dimensionality. When the value is greater than zero, it is set to 1; otherwise, it is set to 0. The combined values ​​generate the local state fingerprint that can globally represent the state of the map.

9. The gastroenterology clinical information acquisition and synchronization system according to claim 4, characterized in that, The specific execution process of the differential localization submodule is as follows: Perform an XOR operation between the local state fingerprint and the historical reference fingerprint to generate a difference indicator mask string, and extract all bit index positions in the difference indicator mask string where the value is 1; Traverse the bit index positions, query the pre-built hash bucket mapping table, obtain the candidate node list that falls into the index bucket, and calculate the Euclidean distance between the current features and historical snapshot features of multiple nodes in the candidate node list. Nodes with non-zero Euclidean distances are selected and marked as target objects whose states have undergone substantial changes. The unique ID of the object is extracted as the difference feature node, and the hierarchical index of the node in the graph structure is output as the topological position coordinate.

10. The gastroenterology clinical information acquisition and synchronization system according to claim 5, characterized in that, The specific functional implementation of the package construction submodule is as follows: The update instruction, the data entity of the difference feature node, and the topological location coordinates are obtained. The data entity is compressed using a preset dictionary compression algorithm to generate a compressed data block. Construct a synchronization header, write the current system version serial number, the length information of the incremental synchronization data packet and the encryption check code into the synchronization header, and fill the compressed data block into the packet body area; Calculate the cyclic redundancy check value of the synchronization header and body regions, append the check value to the end of the packet, complete the encapsulation, and generate the incremental synchronization data packet.

Images