A system for automatic verification of hospital medical record fields

Abstract

The application relates to the technical field of medical health data processing, and discloses a hospital medical record field automatic checking system, which comprises an original field acquisition module, a special disease knowledge dictionary retrieval module, a candidate standard item generation module, a large model field standardization module, a rule checking module and a standardization result storage module; the large model field standardization module outputs mapping data carrying uncertain reason text according to a candidate standard tree; the rule checking module dynamically calculates a field mapping variation rate according to a checking conflict number; when a set threshold is crossed, the retrieval radius of the special disease knowledge dictionary retrieval module is reversely fine-tuned to shrink the retrieval boundary; the application eliminates the backlog at the multi-source heterogeneous medical record data cleaning gateway, constructs a two-way dependent loop of dictionary preliminary screening and quality inspection correction, and improves the flow control convergence speed and the integrated automatic throughput upper limit of the health data management pipeline.

Description

Technical Field

[0001] This invention relates to a system for automatic verification of hospital medical record fields, belonging to the field of medical and health data processing technology. Background Technology

[0002] With the development of medical data sharing technology, the data integration and governance of multi-source heterogeneous clinical information systems is gradually adopting a common processing path of standardized terminology matching. Among them, using a pre-set medical knowledge dictionary in conjunction with a statistical language model to combat the problem of multi-source heterogeneity of medical record text has become the mainstream approach with industry consensus. This architecture can handle unstructured clinical records under normal working conditions and achieve standardized alignment of fields. The routine process of the standardized medical terminology matching system usually relies on a relational database to store standard dictionary codes, and uses a literal similarity algorithm to initially screen candidate terms. The subsequent processing unit uses a deep learning network to identify the semantics of the context text and predict the standard expression of the target field. In conjunction with a rigid rule verification unit, abnormal outputs that exceed the constraints of medical common sense are filtered to maintain the compliance status of data entering the database.

[0003] However, in real-world scenarios involving high-concurrency integration of multi-source medical data, interface changes and long-tail abbreviation distortions from different institutions exhibit dynamic evolution. Due to the lack of a state feedback loop between the front-end static dictionary and the back-end independent rule verification unit, they are in a fragmented state. This leads to logical conflicts between large-scale heterogeneous fields at the rule verification gateway. This situation causes massive amounts of data awaiting standardization to accumulate and stagnate in the manual review queue. The entire data processing chain stalls due to high verification costs. Not only does the front-end hardware channel of the data cleaning gateway face physical interception and throughput backlog limitations, but the control methods also have shortcomings. For example, Chinese invention patent application CN118133774A discloses an AI-based diagnostic text standardization method that uses prefix matching of a dictionary tree or, when encoding is missing, utilizes the BM25 model to retrieve diagnostic names and construct candidate diagnostic texts. The recall set is then used to predict the standard text using a short text similarity model. This method is based on a statically or offline-updated mapping knowledge base. The topological boundaries of the front-end retrieval space and the recall model are fixed and rigid. However, in the complex working conditions of high-concurrency integration of multi-source medical data and frequent interface version changes, the lack of a closed-loop control mechanism to feed back the conflict features captured by the downstream verification layer to the upstream retrieval end in real time makes it impossible to dynamically shrink the initial screening boundary when faced with large-scale deep-tail abbreviation distortion or sudden field heterogeneity. This leads to a large number of irrelevant and redundant noise points mixed into the recall set, causing the inference space of the downstream deep neural network model to diverge and produce encoding illusions. It also causes systemic logical conflicts and blockages at the rule verification gateway. The fundamental defect of the separation between the upstream retrieval and the downstream verification makes it impossible to fundamentally eliminate the stagnation of the flow in the high-concurrency environment and make it difficult to guarantee the overall throughput limit in the highly heterogeneous environment.

[0004] Therefore, the technical problem to be solved by this invention is how to use the system benchmark offset generated by the legality verification to dynamically fine-tune the topological screening boundary of the knowledge dictionary, shrink the recall radius and construct a restricted prompt word tensor to constrain the large model inference unit to complete the conditional probability optimization matching of multi-source heterogeneous medical record texts in a closed-loop self-feedback loop. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A system for automatic verification of hospital medical record fields, comprising: The raw field acquisition module is used to acquire medical record data containing fields to be standardized, field source tags, context fields, and disease-specific topics; The disease-specific knowledge dictionary retrieval module is used to retrieve candidate standard dictionary items within the retrieval space based on field source tags and disease-specific topics. When the disease-specific knowledge dictionary retrieval module receives a dictionary correction control instruction, it shrinks the retrieval radius for the disease-specific topic within the retrieval space to a constrained subset containing conflicting semantic variants based on the uncertain cause text. The candidate standard item generation module is used to construct a candidate item tree based on the candidate standard dictionary items, and assemble the candidate item tree and conversion rules into prompt word vectors; The large model field standardization module is used to input the prompt word vectors into the large language model, and output the candidate standardization results containing standard codes, standard units and text with uncertain reasons under the constraints of the candidate item tree. The rule verification module is used to calculate the field mapping variation rate within a sliding time window based on standard encoding and standard units. When the field mapping variation rate exceeds the set critical threshold, a dictionary correction control instruction is generated based on the text of uncertain cause and sent to the disease knowledge dictionary retrieval module. The standardized results storage module is used to persistently store the candidate standardized results that have passed the verification.

[0006] Preferably, the original field acquisition module includes a data anonymization unit and a tensor generation unit; the data anonymization unit is used to anonymize the patient identification identifier in the medical record data using a preset one-way hash function to extract non-private text; the tensor generation unit is used to concatenate and assemble the non-private text to generate an input tensor containing disease-specific topics, field source labels, and context fields, and to determine the input tensor as the input background for the field to be standardized.

[0007] Preferably, the candidate standard item generation module includes an attribute alignment unit and a vector assembly unit; the attribute alignment unit is used to receive candidate standard dictionary items retrieved by the disease knowledge dictionary retrieval module, and perform attribute alignment processing on each candidate standard dictionary item to generate a candidate item tree; the vector assembly unit is used to receive field definition information in medical record data and externally input pattern specifications, and splice and assemble the field definition information, pattern specifications and candidate item tree to construct prompt word vectors.

[0008] Preferably, the large-scale model field standardization module includes a large-scale language model inference unit and a result output unit; the large-scale language model inference unit is used to receive prompt word vectors, calculate word mapping conditional probabilities under the constraints of the candidate item tree contained in the prompt word vectors and conversion rules, and output a data sequence composed of standard encoding, standard units and uncertain cause text; the result output unit is used to calculate the confidence of the data sequence and assemble the data containing the data sequence and confidence into candidate standardization results.

[0009] Preferably, the rule verification module includes a numerical verification unit and an event triggering unit; the numerical verification unit is used to extract the standard code and standard unit from the candidate standardization results, and perform secondary verification of the standard code and standard unit according to the set time format, numerical range and enumeration constraints to generate verification status data; the event triggering unit is used to generate conflict item records when the numerical verification unit determines that the enumeration constraints or numerical range are not met, and determine the number of rule conflicts according to the conflict item records to update the field mapping variation rate.

[0010] Preferably, the rule verification module further includes a correction control unit; the correction control unit is used to extract the uncertain cause text stored in the conflict item record when the field mapping variation rate is detected to cross the set critical threshold, generate a dictionary correction control instruction containing the uncertain cause text, and send the dictionary correction control instruction to the disease knowledge dictionary retrieval module.

[0011] Preferably, the disease-specific knowledge dictionary retrieval module includes an initial screening retrieval unit and a logic fine-tuning unit. The initial screening retrieval unit is used to retrieve candidate standard dictionary items based on field source tags when no dictionary correction control instruction is received. The logic fine-tuning unit is used to, after receiving the dictionary correction control instruction, use the uncertain cause text in the dictionary correction control instruction as an auxiliary retrieval feature to limit the matching range of the corresponding disease-specific topic to a constraint subset within the retrieval space, so that the initial screening retrieval unit can recall conflicting semantic variants within the constraint subset.

[0012] Preferably, the standardized result storage module includes a result grading unit and a persistent storage unit; the result grading unit is used to receive the verification status data generated by the numerical verification unit, and when the verification status data indicates that the verification has passed, the corresponding data is marked as compliant; when the verification status data indicates that the verification has failed and the confidence level calculated by the result output unit is lower than the preset lower limit, the corresponding data is marked as low confidence pending review; the persistent storage unit is used to store the data marked as compliant into the medical disease database.

[0013] Preferably, the standardized result storage module further includes an audit push unit; the audit push unit is used to send data marked as low confidence pending review to an external audit terminal.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. In the hospital medical record field automatic validation system, the rule validation module captures logical conflict events between structured candidate results and the preset rule base in real time. Combined with the corresponding source system tags and disease topics, the current system mapping benchmark offset is statistically determined within the sliding time window. When the offset crosses the preset critical threshold, a dictionary correction control instruction is triggered and injected into the disease knowledge dictionary retrieval module. This allows the disease knowledge dictionary retrieval module to fine-tune the initial screening logic in the multidimensional topology retrieval space using the uncertain reasons in the conflict entries. This shrinks the recall radius of a specific disease topic into a strongly constrained subset that includes conflict semantic variants, thereby eliminating the backlog of heterogeneous clinical fields at the rule validation gateway.

[0015] 2. The original field acquisition module receives the original medical record field names, original field values, source system, and contextual text information from external heterogeneous systems. It then initiates data anonymization processing logic, using a preset one-way hash function to anonymize the patient's privacy plaintext identity identifiers. Non-privacy text representing heterogeneous clinical characteristics is extracted and assembled to generate a standardized input tensor containing the disease topic to which the field belongs, the source table label, and the context field. This provides high-purity topological constraint input for subsequent initial screening and model inference, reducing the risk of privacy leakage when processing sensitive clinical information and enhancing the security and stability of the medical heterogeneous data integration and delivery process.

[0016] 3. The candidate standard item generation module aligns the multi-dimensional attributes of the dictionary items after initial screening or correction and constructs a standardized structured candidate item tree. Combined with the large model field standardization module, the candidate item tree, field definition information, unit conversion rules, and target pattern specifications are assembled into restricted prompt word vectors, which are then input into the built-in large language model inference unit. The large language model inference unit is strongly constrained to perform conditional probability optimization only within the space provided by the candidate item tree. The output is a structured candidate result containing standard field names, standard codes, standard units, conversion values, and confidence scores, which filters out uncertain features such as abnormal output of model text and coding errors. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the module structure of the automatic verification system for medical record fields of the present invention; Figure 2 This is a verification status diagram of the automatic verification system for medical record fields of the present invention.

[0018] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A system for automatic validation of hospital medical record fields includes: The raw field acquisition module is used to acquire medical record data containing fields to be standardized, field source tags, context fields, and disease-specific topics; The disease-specific knowledge dictionary retrieval module is used to retrieve candidate standard dictionary items within the retrieval space based on field source tags and disease-specific topics. When the disease-specific knowledge dictionary retrieval module receives a dictionary correction control instruction, it shrinks the retrieval radius for the disease-specific topic within the retrieval space to a constrained subset containing conflicting semantic variants based on the uncertain cause text. The candidate standard item generation module is used to construct a candidate item tree based on the candidate standard dictionary items, and assemble the candidate item tree and conversion rules into prompt word vectors; The large model field standardization module is used to input the prompt word vectors into the large language model, and output the candidate standardization results containing standard codes, standard units and text with uncertain reasons under the constraints of the candidate item tree. The rule verification module is used to calculate the field mapping variation rate within a sliding time window based on standard encoding and standard units. When the field mapping variation rate exceeds the set critical threshold, a dictionary correction control instruction is generated based on the text of uncertain cause and sent to the disease knowledge dictionary retrieval module. The standardized results storage module is used to persistently store the candidate standardized results that have passed the verification.

[0021] Preferably, the original field acquisition module includes a data anonymization unit and a tensor generation unit; the data anonymization unit is used to anonymize the patient identification identifier in the medical record data using a preset one-way hash function to extract non-private text; the tensor generation unit is used to concatenate and assemble the non-private text to generate an input tensor containing disease-specific topics, field source labels, and context fields, and to determine the input tensor as the input background for the field to be standardized.

[0022] Preferably, the candidate standard item generation module includes an attribute alignment unit and a vector assembly unit; the attribute alignment unit is used to receive candidate standard dictionary items retrieved by the disease knowledge dictionary retrieval module, and perform attribute alignment processing on each candidate standard dictionary item to generate a candidate item tree; the vector assembly unit is used to receive field definition information in medical record data and externally input pattern specifications, and splice and assemble the field definition information, pattern specifications and candidate item tree to construct prompt word vectors.

[0023] Preferably, the large-scale model field standardization module includes a large-scale language model inference unit and a result output unit; the large-scale language model inference unit is used to receive prompt word vectors, calculate word mapping conditional probabilities under the constraints of the candidate item tree contained in the prompt word vectors and conversion rules, and output a data sequence composed of standard encoding, standard units and uncertain cause text; the result output unit is used to calculate the confidence of the data sequence and assemble the data containing the data sequence and confidence into candidate standardization results.

[0024] Preferably, the rule verification module includes a numerical verification unit and an event triggering unit; the numerical verification unit is used to extract the standard code and standard unit from the candidate standardization results, and perform secondary verification of the standard code and standard unit according to the set time format, numerical range and enumeration constraints to generate verification status data; the event triggering unit is used to generate conflict item records when the numerical verification unit determines that the enumeration constraints or numerical range are not met, and determine the number of rule conflicts according to the conflict item records to update the field mapping variation rate.

[0025] Preferably, the rule verification module further includes a correction control unit; the correction control unit is used to extract the uncertain cause text stored in the conflict item record when the field mapping variation rate is detected to cross the set critical threshold, generate a dictionary correction control instruction containing the uncertain cause text, and send the dictionary correction control instruction to the disease knowledge dictionary retrieval module.

[0026] Preferably, the disease-specific knowledge dictionary retrieval module includes an initial screening retrieval unit and a logic fine-tuning unit. The initial screening retrieval unit is used to retrieve candidate standard dictionary items based on field source tags when no dictionary correction control instruction is received. The logic fine-tuning unit is used to, after receiving the dictionary correction control instruction, use the uncertain cause text in the dictionary correction control instruction as an auxiliary retrieval feature to limit the matching range of the corresponding disease-specific topic to a constraint subset within the retrieval space, so that the initial screening retrieval unit can recall conflicting semantic variants within the constraint subset.

[0027] Preferably, the standardized result storage module includes a result grading unit and a persistent storage unit; the result grading unit is used to receive the verification status data generated by the numerical verification unit, and when the verification status data indicates that the verification has passed, the corresponding data is marked as compliant; when the verification status data indicates that the verification has failed and the confidence level calculated by the result output unit is lower than the preset lower limit, the corresponding data is marked as low confidence pending review; the persistent storage unit is used to store the data marked as compliant into the medical disease database.

[0028] Preferably, the standardized result storage module further includes an audit push unit; the audit push unit is used to send data marked as low confidence pending review to an external audit terminal.

[0029] Example 1: When the system faces high-concurrency data integration and governance from multiple heterogeneous clinical information systems, conflicts in the naming variations and dosage expressions of medical record fields between different institutions lead to passive interception and backlog of manual review queues in the static rule gateway. The original field acquisition module accesses the original medical record field names, original field values, source systems, and contextual text information containing the rheumatoid arthritis disease topic. The data desensitization unit within the original field acquisition module calls a preset one-way hash function to process the patient's privacy plaintext identity identifier, generating non-privacy text. The tensor generation unit concatenates and assembles the non-privacy text, outputting a standardized input tensor containing the disease topic, field source labels, and context fields to construct a feature storage partition. This input background belongs to the feature context data physical storage partition of the deep learning model. Its attached objectively measurable feature input parameters are a discrete word digit sequence after symbolic segmentation and a fixed-dimensional low-dimensional dense vector read from the local pre-trained embedding matrix. The specific processing dynamics are as follows. The tensor generation unit stacks and concatenates these dense vectors according to their corresponding field attribute levels and contextual physical time order, ultimately outputting a standardized input tensor with a fixed dimension of 256 as a prior constraint state for subsequent large-scale model attention mechanism operations. In the specific data processing flow, the concatenation and assembly of non-privacy text forms a standardized long string text containing multiple structured key-value pairs. The tensor generation unit calls a preset clinical medical word segmenter to symbolically segment this standardized long string text, converting it into a discrete word number sequence. Then, by reading the locally pre-trained embedding matrix, it maps each discrete word in the word number sequence into a fixed-dimensional low-dimensional dense vector. Finally, these dense vectors are stacked and concatenated according to their corresponding field attribute levels and contextual physical time order, thereby assembling and generating a standardized input tensor that can directly participate in topological retrieval and attention mechanism operations mathematically, eliminating the conversion barrier between plain text long strings and mathematical tensors.

[0030] The disease-specific knowledge dictionary retrieval module, coupled with the original field acquisition module, receives standardized input tensors within a multidimensional topological retrieval space that includes a clinical diagnosis dictionary, a laboratory dictionary, a drug dictionary, a unit conversion rule base, and an enumeration value table. The initial screening retrieval unit within the disease-specific knowledge dictionary retrieval module retrieves candidate standard dictionary items based on field source tags and disease-specific topics. The attribute alignment unit in the candidate standard item generation module receives candidate standard dictionary items and performs alignment processing on the attributes of the candidate standard dictionary items to generate a candidate entry tree. The vector assembly unit within the candidate standard item generation module concatenates and assembles the field definition information, the external input pattern specification, and the candidate entry tree to construct a prompt word vector.

[0031] The large language model inference unit within the large model field standardization module receives prompt word vectors. Under the constraints of the candidate item tree and conversion rules contained in the prompt word vectors received by the large language model inference unit, it calculates the conditional probability of word mapping and outputs a data sequence composed of standard codes, standard units, and uncertain cause text. The result output unit within the large model field standardization module calculates the confidence score of the data sequence and assembles the data containing the data sequence and confidence scores into structured candidate results. The parameters of the data sequence include those representing standard field names. , indicating standard encoding , representing standard values , representing standard units , indicating the conversion rule number , indicating confidence level With text indicating uncertain reasons .

[0032] The numerical verification unit of the rule verification module extracts structured candidate results. and Based on the set time format, numerical range, and enumeration constraints, the system generates verification status data. When the numerical verification unit determines that the structured candidate result does not meet the enumeration constraints or numerical range, the event triggering unit generates a conflict item record and captures the logical conflict event. Within the current sliding time window, the rule verification module counts the total number of field items with rule conflicts for the fields to be standardized from the same source system and the same disease topic, and calculates the proportion of the total number of field items to the total number of fields entered by the source system within the current sliding time window to determine the current system mapping benchmark offset and complete the update of the field mapping variation rate. The white-box calculation input source for this variation rate is the specific data source channel due to the failure to meet the enumeration constraints. The number of discrete conflict records generated by the event-triggered unit within the sum of numerical ranges is specifically processed by the rule verification module, which extracts data from five consecutive equal-length sampling periods within a sliding time window. Each sampling period lasts 120 seconds. The number of rule-conflicting fields captured and the total number of fields entered in each period are counted, and then multiplied by incremental time-series weighting factors of 0.10, 0.15, 0.20, 0.25, and 0.30 respectively, and then summed. The weighted total number of conflicting fields is divided by the weighted total number of entered fields. Finally, a scalar value between 0.00 and 1.00 is output as the updated field mapping mutation rate. This mutation rate is denoted both literally and logically as... ,in, The current system mapping reference offset, when detected When a preset critical threshold is crossed, the correction control unit retrieves the conflict item record stored in the record. , generate containing The dictionary correction control command is sent to the disease-specific knowledge dictionary retrieval module, enabling the logic fine-tuning unit within the disease-specific knowledge dictionary retrieval module to utilize... The matching scope is limited to a constraint subset containing conflicting semantic variants within the multidimensional topological retrieval space for specific disease topics. This causes the prompt word vectors generated by the candidate standard item generation module in the next round to change the context purity. The result grading unit in the standardized result storage module receives the verification status data. When the verification status data indicates that the verification has passed, the persistent storage unit stores the corresponding candidate standardized result in the medical disease database. When the verification status data indicates that the verification has failed and the confidence level calculated by the result output unit is consistently lower than the preset lower limit, the result grading unit marks the corresponding candidate standardized result as a low-confidence pending review state and sends the candidate standardized result in the low-confidence pending review state to the external audit terminal through the audit push unit.

[0033] Example 2: The specific implementation process of the current hospital medical record field automatic verification system in the construction of the rheumatoid arthritis-specific disease database is as follows: After the system starts, the original field acquisition module extracts medical record data from the hospital information system, including the name of the field to be standardized, the original field value, the source system label, and the specific disease topic. Taking the rheumatoid arthritis-specific disease database as an example, the original medical record data contains a field named "Rheumatoid Factor Quantitative" with an original field value of 45.2 IU / mL. The source system label is set to Laboratory Department System A, and the specific disease topic is set to "Rheumatoid Arthritis". The specific disease knowledge dictionary retrieval module matches the source system label "Laboratory Department System A" and the specific disease topic "Rheumatoid Arthritis" within the preset specific disease knowledge dictionary retrieval space. The search space stores standard medical terms through a multi-dimensional topological structure. The retrieval logic associates rheumatoid factor quantification with the standard term "rheumatoid factor measurement" and calls the corresponding standard code, unit conversion rules, and enumeration constraint information. The candidate standard item generation module receives the above retrieval results and constructs a candidate item tree. This tree structure includes candidate standard items, standard definitions, encoding mapping logic, and numerical format requirements. The candidate standard item generation module assembles the candidate item tree into prompt word vectors. These vectors explicitly limit the large language model to performing mapping inference only based on the candidate items within the candidate item tree. To achieve hard topological constraints during the decoding stage of model inference, this invention, while inputting the prompt word vectors into the large language model, utilizes the candidate item tree's inference unit... The standard encoding and candidate entry information in the selection tree are dynamically constructed in memory to form a prefix tree syntax constraint. This constraint, which uses hard constraints on lexical generation boundaries, is represented at the underlying data structure as a discrete prefix tree dictionary structure built from candidate standard encoded characters as path nodes. The specific white-box data flow execution is as follows: When the model performs the current single-step prediction, the constraint traverses the set of legal child nodes at the current level of the prefix tree and performs an explicit matrix index comparison with the global vocabulary probability distribution output by the large language model. For all vocabulary components that do not match the legal child node characters at the current level of the prefix tree, their multidimensional floating-point probability values ​​are forcibly overwritten and modified to 0. For vocabulary components that match the legal path, the system... Normalization calculations redistribute relative probability weights, thereby closing the loop and truncating the decoding divergence path of the large language model. The output is a sequence of lexical terms completely constrained by the character boundaries of the entry tree, thus achieving deterministic state transitions within the digital space. During the decoding iteration process of the model predicting each lexical term, this grammatical constraint obtains the probability distribution of all candidate lexical terms output by the model in real time, compares it with the legal branch path of the prefix tree, forces the probability of lexical terms that do not belong to the legal branch of the prefix tree to be set to zero, and recalculates the normalized probability of the remaining legal lexical terms. This forces the optimization path of the word image conditional probability of the large language model to flow only within the physical space provided by the candidate entry tree, thus avoiding erroneous encoding outputs in the decoding stage from the perspective of generation logic.

[0034] The large language model field standardization module receives prompt word vectors and outputs standardized candidate results containing standard codes, standard units, and text explaining the reasons for uncertainty. If the model's mapping confidence decreases during inference due to ambiguity in the original field values, the output includes text describing the reason for the ambiguity. The rule validation module performs rule validation on fields from the same source system and for the same disease topic within a sliding time window. The sliding time window is set to 600 seconds. Within the window, it counts the total number of field entries with rule conflicts for fields from the same source system and for the same disease topic, and calculates the ratio of the number of conflicting fields to the total number of fields entered in that source system. This ratio is defined as the current system's mapping baseline offset. The rule validation module will It is compared with a preset critical threshold, which is 0.15. When it is detected... When the value exceeds 0.15, the rule verification module triggers the dictionary correction control command and injects it back into the disease knowledge dictionary retrieval module. After receiving the control command, the disease knowledge dictionary retrieval module shrinks the recall radius of the corresponding disease topic in the multidimensional topological retrieval space into a strongly constrained subset containing conflict semantic variants, based on the uncertain cause text recorded in the conflict entry. This makes the prompt word vectors generated by the candidate standard item generation module in the next round have higher contextual purity. The standardization result storage module finally writes the verified standardization code, value, and unit into the database, and records the corresponding processing timestamp and system offset parameters for subsequent data traceability and quality control.

[0035] Example 3: This example combines Figures 1 to 2 A system description for automatic validation of hospital medical record fields, such as... Figure 1 As shown, a system for automatic field validation in hospital medical records includes: a raw field acquisition module, a disease-specific knowledge dictionary retrieval module, a candidate standard item generation module, a large-scale model field standardization module, a rule validation module, and a standardization result storage module. The raw field acquisition module connects to the disease-specific knowledge dictionary retrieval module after acquiring medical record data. The disease-specific knowledge dictionary retrieval module connects to the candidate standard item generation module after retrieving candidate standard dictionary items. The candidate standard item generation module connects to the large-scale model field standardization module after assembling prompt word vectors. The large-scale model field standardization module connects to the rule validation module after outputting candidate standardization results. The rule validation module persists the validation results and connects to the standardization result storage module. Simultaneously, the rule validation module connects back to the disease-specific knowledge dictionary retrieval module by sending dictionary correction control commands.

[0036] like Figure 2As shown, when the system does not receive a dictionary correction control instruction, it assembles a prompt word vector and then proceeds to the calculated word mapping conditional probability state. In the calculated word mapping conditional probability state, when the field mapping mutation rate exceeds a set critical threshold, the system proceeds to the constraint subset retrieval state and returns to the calculated word mapping conditional probability state by recalling conflicting semantic variants within the constraint subset. In the calculated word mapping conditional probability state, when the verification status data indicates that the verification has passed, the system proceeds to the compliance state. In the calculated word mapping conditional probability state, when the verification fails and the confidence level is lower than the preset lower limit, the system proceeds to the low-confidence pending review state.

[0037] Example 4: In the actual construction scenario of the rheumatoid arthritis disease database, to address the backlog in field verification links caused by frequent version changes of interfaces from multiple clinical information systems, the following engineering solution is adopted: In the original field acquisition module, to correct the asynchronous transmission problem between different source systems, a timestamp-based transmission delay compensation mechanism is introduced, and the data processing unit monitors the round-trip delay of network packets in real time. Based on the collected original medical record field names and their accompanying original timestamp information, the original medical record fields are timestamped by calculating the current clock offset. This ensures that the contextual text information in the standardized input tensor is absolutely synchronized with the logical timing of the medical record fields. When the preset 60.0ms engineering limit is exceeded, the system automatically switches to offline caching mode, suspends standardization processing, and waits until network jitter returns to normal, thereby shielding the standardized data stream from interference by random fluctuations in the transport layer.

[0038] In the disease-specific knowledge dictionary retrieval module, a dynamic shrinking logic for the retrieval space based on semantic distance is implemented. When the rule validation module detects the system mapping baseline offset... When the system continuously crosses the critical threshold of 0.15, the correction control unit parses the conflict term records. The logical fine-tuning unit calculates the retrieval radius correction factor based on the word vector similarity of the text. This explicitly defines the retrieval topology boundary for discrete dictionary items. This retrieval radius belongs to the physical cutoff boundary partition of the multidimensional independent semantic vector space. Its bound objective feature input parameters are a 768-dimensional dense real vector output by the local word vector model and the feature vector corresponding to the standard term. Specifically, the processing involves calculating the cosine similarity between these two vectors and subtracting this similarity from 1 to obtain a cosine distance offset value between 0.00 and 2.00. The logic fine-tuning unit multiplies this offset value by a calibration coefficient of 0.50 and adds a constant baseline term of 0.10. If the calculation result exceeds 0.40, it is forcibly truncated and output as 0.40. Finally, the closed-loop output of this scalar value serves as a retrieval radius correction factor, which is directly written into the vector retrieval engine's judgment. Specifically, the digital mapping path for converting uncertain cause text into a retrieval radius correction factor is as follows: The uncertain cause text is input into a pre-defined local semantic vectorization model to extract a low-dimensional dense vector representing the semantic features of the unstructured text; the cosine similarity between this dense vector and the feature vectors of each standard term in the disease-specific knowledge dictionary is calculated, and the target value obtained by subtracting this cosine similarity from 1 is determined as the semantic distance offset value between the uncertain cause text and the standard terms. This achieves a deterministic mapping of qualitative text features to discrete scalar values. Finally, this semantic distance offset value is used to constrain the retrieval range of the multidimensional topological space. From a data processing logic perspective, this eliminates the computational limitation that qualitative text cannot directly participate in numerical space geometric pruning. The correction factor calculation logic is set as follows: ,in, This is a retrieval radius correction factor. This is the sensitivity calibration coefficient, with a value of 0.5. for The semantic distance offset between the text and standard terms ranges from 0.0 to 0.6. The dynamic adjustment of the recall radius drives the recall scope of the corresponding disease topic in the multidimensional topology retrieval space to a constrained subset that includes conflicting semantic variants. The engineering selection basis and physical boundary description of the above specific numerical range, constant term of the calculation formula are as follows: The upper limit of communication latency is set to 60.0ms, which is based on the average network jitter benchmark measurement of the hospital LAN under peak concurrent flow conditions. If it is higher than 60.0ms, it will cause the real-time integrated cleaning gateway to experience thread suspension and retry blocking.

[0039] The sliding time window is set to 600 seconds. Shortening this window would prevent the collection of statistically significant long-term interface change samples, while extending it would cause a time lag in the response of the self-feedback flow control loop to abnormal backlogs. The critical threshold for the system mapping baseline offset is set to 0.15, representing a maximum allowed long-tail heterogeneous expression rate of 15%. Exceeding this boundary confirms a systemic version change in the upstream interface. The constant term 0.1 and the upper limit 0.4 in the retrieval radius correction factor formula ensure that the shrunken strongly constrained subset contains at least two nearest-neighbor conflicting semantic variants without excessive pruning leading to incorrect filtering of legitimate medical terms. This achieves the optimal balance between the convergence speed of the data governance pipeline flow control. This adjustment, by limiting the number of candidate entries in the prompt word vectors entering the inference unit of the large language model, physically narrows the retrieval baseline of the inference space, thereby reducing the computational path divergence when the large language model processes fuzzy fields and improving the determinism of the output standard encoding.

[0040] To eliminate the sliding time window False positives caused by occasional network transient fluctuations trigger a time-weighted smoothing operator within the rule verification module. The rule verification module then uses the formula... Calculate the smooth offset, where, This is the weighted smoothed offset. The statistical period is 3 sampling periods; The weighting coefficient for the corresponding period is [0.2, 0.3, 0.5]. The original system mapping reference offset measured in each cycle, when the event triggering unit... The dictionary correction control command is generated only when the error exceeds 0.15. This smoothing logic ensures that it can respond to long-term offsets caused by system interface changes, and at the same time has engineering fault tolerance for instantaneous communication anomalies, ensuring the operational stability of the data governance pipeline in the complex network physical environment of the hospital.

[0041] Example 5: In the multi-source medical data governance scenario of the rheumatoid arthritis specialty database, the medical record field verification system executes the following standardized operating procedures for high-concurrency heterogeneous data streams: The original field acquisition module reads the text sequence of original medical record field names and corresponding values ​​transmitted by the HIS system, removes patient identity identifiers through a preset one-way hash function, extracts non-privacy clinical attribute tensors, and in the process of generating standardized candidate results, the system introduces a multi-dimensional feature comprehensive evaluation mechanism to independently calculate the semantic matching degree, format compliance, and terminology coverage of each candidate entry tree, and generates a comprehensive confidence score based on preset weights. Specifically, the total weight normalization benchmark is set to 1.0, and the weight coefficients for semantic matching degree, format compliance degree and terminology coverage are set to 0.5, 0.3 and 0.2 respectively, to ensure that the final comprehensive result reflects the multi-dimensional comprehensive effectiveness and avoids the extreme bias of single-dimensional evaluation.

[0042] The rule validation module deploys a time-series data smoothing filter path for rule conflict events within a sliding time window. For fields to be standardized from the same source system and the same specific disease topic, it counts the number of conflicting fields in real time. Total number of fields entered The system follows the formula Calculate the smoothed system mapping reference offset, where, This is the smoothed system mapping reference offset. This represents the number of sampling periods within the sliding window, with a value of 5. The forgetting factor corresponds to the time period and takes values ​​of [0.1, 0.15, 0.2, 0.25, 0.3]. For the first Number of conflicting fields per cycle For the first The total number of fields entered in each cycle is filtered out by this smoothing logic to ensure that dictionary correction control commands are only triggered when there are systemic interface variations.

[0043] when When the error continuously exceeds the 0.15 threshold, the correction control unit reads the conflict items. Dynamic thresholds are set based on the clinical risk level within the specific disease theme. For high-risk specific diseases, such as rheumatoid arthritis, the threshold The threshold was automatically increased to 0.95 to strengthen data quality control standards; for low-risk diseases, the threshold was adjusted accordingly. Adjusted to 0.85, the resulting grading unit is based on The standardized candidate results are split, with a confidence level higher than [missing information]. Data that passes validation is directly stored in the database; fields with values ​​below this threshold are generated by the audit push unit and include... The abnormal handling work orders are pushed to the manual review terminal to ensure that abnormal items in the data integration and governance flow can be handled in a deterministic closed loop.

[0044] Example 6: In the integrated application scenario of a rheumatoid arthritis-specific database in a hospital's rheumatology and immunology department, the original medical record data exhibits frequent interface version changes, causing the standardization verification link to easily experience flow delays when handling cross-system semantic variations. This example constructs a self-feedback verification procedure based on transient conflict characteristics. This procedure is deployed between the rule verification module and the specific disease knowledge dictionary retrieval module of the hospital's automatic medical record field verification system to solve the flow control congestion problem in the standardization pipeline. In the specific deployment, the original field acquisition module accesses the original medical record field names and original field values ​​containing the rheumatoid arthritis topic and generates an input tensor containing the specific disease topic and field source labels. The specific disease knowledge dictionary retrieval module performs candidate standard item retrieval in the multidimensional topological retrieval space based on the input tensor. The candidate standard item generation module assembles the retrieved entry tree and conversion rules into a prompt word vector. The large model field standardization module outputs candidate standardization results containing standard codes, standard units, and uncertain cause text based on this vector. The rule verification module uses the built-in verification rule set to perform rigid assertions for each candidate... The standardization results are captured in real time by logic gate circuits to identify numerical range conflicts, enumeration value constraint conflicts, or time format conflicts, and the conflict events are discretized and recorded as corresponding conflict items. Data flow isolation is based on the principle of independent distribution of multiple channels. By constructing an independent addressing routing model based on data source labels, noise coupling interference between different data channels is blocked. The data splitting unit reads the field source labels contained in the medical record data and writes the conflict items into a specific memory physical buffer uniquely bound to the field source label, establishing the independent bucketing status of different data channels. Statistics and calculations within the sliding time window are started independently within their respective independent specific memory physical buffers. The number of isolated specific conflict item records in the specific memory physical buffer is divided by the total number of fields entered corresponding to the field source label to determine the bucketing image variation rate of the specific source channel. This physically eliminates the channel pollution phenomenon caused by a single high-frequency error source that causes the system to indiscriminately shrink the general retrieval space. The rule verification module maintains a sliding time window with a duration of 600 seconds. Within this window, the current system status is statistically analyzed in real time, and the current system mapping baseline offset is calculated. The calculation logic is as follows: ,in, This represents the weighting coefficient, with values ​​ranging from [0.1, 0.15, 0.2, 0.25, 0.3]. For the first The number of rule conflict fields captured within a sampling period. For the first The total number of fields entered by the source system within each sampling period.

[0045] The correction control unit sets the critical threshold. The value is 0.15, when the calculated value is detected. When the value exceeds 0.15, the error correction control unit initiates the reverse injection logic, extracts the uncertain cause text stored in the conflict item, and converts it into a dictionary error correction control instruction, which is then sent back to the disease-specific knowledge dictionary retrieval module. When processing multi-source high-concurrency data integration streams, if multiple conflict item records caused by different heterogeneous medical data sources are captured simultaneously within the current sliding time window, this invention deploys and executes macro-aggregation decision logic based on the item with the highest conflict frequency. The system continuously calculates the frequency of occurrence of each uncertain cause text in all conflict item records within the current sliding time window, automatically selects the group of uncertain cause texts with the highest frequency as global representative features, and encapsulates them into the current dictionary error correction control instruction. If two or more conflict texts have the same frequency, the event triggering unit prioritizes extracting the text with the latest timestamp after clock synchronization calibration, thereby establishing a unique causal relationship and data convergence path between the systemic macro-proportional index and the multi-source micro-text objects, ensuring the self-consistent closed loop of the flow control circuit. After receiving the instruction, the disease-specific knowledge dictionary retrieval module performs retrieval radius correction based on text features and utilizes the correction factor. Dynamic pruning is performed on the multidimensional topological retrieval space, and the calculated relationship is as follows: ,in, This is the sensitivity calibration coefficient, with a value of 0.5. The semantic distance offset between the text with uncertain causes and standard terms ranges from 0.0 to 0.6. This correction operation shrinks the recall radius of a specific disease topic into a strongly constrained subset containing conflicting semantic variants, substantially improving the contextual purity of the prompt word vectors that subsequently enter the large model field standardization module. This suppresses semantic divergence in the inference space at the physical link level. The retrieval boundary shrinkage is based on metric space pruning theory. By establishing a boundary constraint model based on the negative correlation of distance thresholds, the quantitative retrieval radius is transformed into the physical pruning boundary of a high-dimensional vector space. Upon receiving the dictionary correction control command, the logic fine-tuning unit adjusts the correction factor. As the basic subtrahend, subtract the correction factor from the number 1. The obtained target difference is used as the absolute distance threshold for the final retrieval. When retrieving term similarity in the tree-graph index database, the vector retrieval engine writes the absolute distance threshold into the distance cutoff decision register of the nearest neighbor search mechanism. This distance cutoff decision register is entirely represented as a scalar numerical conditional decision variable residing in the system's runtime memory at the underlying technical entity level. The white-box software pruning action of this boundary constraint model is entirely implemented by the conditional control flow code in memory. Specifically, when the system calls the vector retrieval tree based on the inverted index or graph topology structure to retrieve term similarity, the vector retrieval engine assigns the aforementioned absolute distance threshold to this scalar numerical conditional decision variable. As the engine traverses the nodes of the retrieval tree, the system calculates the cosine distance between the feature vector of the current candidate node and the target disease topic vector in real time, and then compares the calculated discrete scalar distance value with this variable. The absolute distance threshold is used for explicit conditional comparison. Once the current distance value is greater than the threshold, the conditional control flow immediately triggers a loop termination instruction, forcibly truncating and discarding all branch subtree indexes below the node. This achieves closed-loop pruning of the retrieval space in the hypersphere dimension, ensuring that the retrieval process completes adaptive state transitions at the pure software memory level. During the step-by-step retrieval process of the topology structure, the retrieval engine calculates the cosine distance between the current candidate node and the target disease topic vector in real time. Once the cosine distance is detected to be greater than the absolute distance threshold, the depth traversal path of the branch node is immediately forcibly truncated, limiting the recall scope to a compact hypersphere containing conflicting semantic variants. This ensures that when the span of the semantic variant indicated by the uncertain cause text increases, the actual physical space volume participating in the recall undergoes a monotonically decreasing shrinkage action, ensuring consistency between the logical calculation output and the physical shrinkage purpose.

[0046] After verification and data routing, the result grading unit marks standard data with a confidence level higher than the 0.95 risk threshold of the rheumatoid arthritis disease database as compliant and performs persistent storage. Data with a confidence level lower than this threshold is marked as low-confidence and awaits review. An exception handling work order containing text indicating an uncertain cause is then sent to the external audit terminal via the audit push unit. Through this closed-loop feedback mechanism, the system eliminates the backlog and congestion of multi-source heterogeneous medical record data at the cleaning gateway, ensuring the data governance throughput and integration stability of the medical disease database in complex network environments. Dictionary timeliness maintenance is based on the time-varying baseline decay theory. By introducing a time decay factor and a sliding time window to eliminate outdated data, the dynamic accuracy of the dictionary knowledge base is ensured, maintaining the long-term performance of the disease knowledge dictionary retrieval module. When checking the accuracy of entries under normal operating conditions, the rule verification module has a periodic trigger reconstruction unit. The periodic trigger reconstruction unit takes 168 hours of continuous operation as a single statistical cycle. At the end of each statistical cycle, it retrieves all candidate standardized results marked as compliant and their corresponding uncertain reason texts accumulated within the statistical cycle. The system uses a sliding time window to eliminate old conflict records that are more than 30 days old and sets a time decay weight factor that decreases exponentially over time for the remaining conflict semantic variants. When the weighted frequency of a specific conflict item exceeds 50 times in the current statistical cycle, the dictionary reconstruction instruction is automatically triggered, and it is permanently written into the disease knowledge dictionary retrieval space as a new static mapping entry, realizing adaptive baseline update for long-term evolution interface distortion.

[0047] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A system for automatic validation of hospital medical record fields, characterized in that, include: The raw field acquisition module is used to acquire medical record data that includes fields to be standardized, field source tags, context fields, and disease-specific topics; The disease-specific knowledge dictionary retrieval module is used to retrieve candidate standard dictionary items within the retrieval space based on field source tags and disease-specific topics. When the disease-specific knowledge dictionary retrieval module receives a dictionary correction control instruction, it shrinks the retrieval radius for the disease-specific topic within the retrieval space to a constrained subset containing conflicting semantic variants based on the uncertain cause text. The candidate standard item generation module is used to construct a candidate item tree based on the candidate standard dictionary items, and assemble the candidate item tree and conversion rules into prompt word vectors; The large model field standardization module is used to input the prompt word vectors into the large language model, and output the candidate standardization results containing standard codes, standard units and text with uncertain reasons under the constraints of the candidate item tree. The rule verification module is used to calculate the field mapping variation rate within a sliding time window based on standard encoding and standard units. When the field mapping variation rate exceeds the set critical threshold, a dictionary correction control instruction is generated based on the text of uncertain cause and sent to the disease knowledge dictionary retrieval module. The standardized results storage module is used to persistently store the candidate standardized results that have passed the verification.

2. The system for automatic verification of hospital medical record fields according to claim 1, characterized in that, The original field acquisition module includes a data anonymization unit and a tensor generation unit. The data anonymization unit is used to anonymize patient identifiers in medical record data using a preset one-way hash function to extract non-private text. The tensor generation unit is used to concatenate and assemble the non-private text to generate an input tensor containing disease-specific topics, field source labels, and context fields, and to determine the input tensor as the input background for the fields to be standardized.

3. The system for automatic verification of hospital medical record fields according to claim 1, characterized in that, The candidate standard item generation module includes an attribute alignment unit and a vector assembly unit. The attribute alignment unit is used to receive the candidate standard dictionary items retrieved by the disease knowledge dictionary retrieval module, and to perform attribute alignment processing on each candidate standard dictionary item to combine and generate a candidate item tree. The vector assembly unit receives field definition information from medical record data and pattern specifications from external input, and concatenates and assembles the field definition information, pattern specifications, and candidate item tree to construct prompt word vectors.

4. The system for automatic verification of hospital medical record fields according to claim 1, characterized in that, The large-scale language model field standardization module includes a large-scale language model inference unit and a result output unit. The large-scale language model inference unit is used to receive prompt word vectors, calculate word mapping conditional probabilities under the constraints of the candidate item tree contained in the prompt word vectors and conversion rules, and output a data sequence composed of standard encoding, standard units, and text with uncertain causes. The result output unit is used to calculate the confidence of the data sequence and assemble the data containing the data sequence and confidence into candidate standardization results.

5. The system for automatic verification of hospital medical record fields according to claim 1, characterized in that, The rule verification module includes a numerical verification unit and an event triggering unit. The numerical verification unit is used to extract the standard codes and standard units from the candidate standardization results, and performs secondary verification of the standard codes and standard units based on the set time format, numerical range and enumeration constraints to generate verification status data. The event triggering unit is used to generate conflict item records when the numerical verification unit determines that the enumeration constraints or numerical range are not met, and to determine the number of rule conflicts based on the conflict item records in order to update the field mapping variation rate.

6. The system for automatic verification of hospital medical record fields according to claim 1, characterized in that, The rule verification module also includes a correction control unit; the correction control unit is used to extract the uncertain cause text stored in the conflict item record when the field mapping variation rate is detected to cross the set critical threshold, generate a dictionary correction control instruction containing the uncertain cause text, and send the dictionary correction control instruction to the disease knowledge dictionary retrieval module.

7. The system for automatic verification of hospital medical record fields according to claim 1, characterized in that, The disease-specific knowledge dictionary retrieval module includes an initial screening retrieval unit and a logic fine-tuning unit. The initial screening retrieval unit is used to retrieve candidate standard dictionary items based on field source tags when no dictionary correction control instruction is received. The logic fine-tuning unit is used to, after receiving the dictionary correction control instruction, use the uncertain cause text in the dictionary correction control instruction as an auxiliary retrieval feature to limit the matching range of the corresponding disease-specific topic to a constraint subset within the retrieval space, so that the initial screening retrieval unit can recall conflicting semantic variants within the constraint subset.

8. The system for automatic verification of hospital medical record fields according to claim 1, characterized in that, The standardized result storage module includes a result grading unit and a persistent storage unit. The result grading unit receives the verification status data generated by the numerical verification unit. When the verification status data indicates that the verification has passed, the corresponding data is marked as compliant. When the verification status data indicates that the verification has failed and the confidence level calculated by the result output unit is lower than the preset lower limit, the corresponding data is marked as low-confidence pending review. The persistent storage unit stores the data marked as compliant into the medical disease database.

9. A system for automatic verification of hospital medical record fields according to claim 1, characterized in that, The standardized results storage module also includes an audit push unit; the audit push unit is used to send data marked as low confidence pending review to external audit terminals.

Images