Registration number key information rapid screening and completion method

Abstract

A registration number key information rapid screening and completion method, comprising the following steps: S1, taking an identity card number of a patient as a unique index, and after patient authorization, integrating patient data from different channels under a federated learning and differential privacy fusion framework and performing privacy protection processing; S2, according to a modal feature memory library and dynamic feature engineering, outputting a completion information result and a comprehensive credibility score thereof by means of a lightweight completion model fused with a self-supervised learning architecture; S3, performing real-time diagnosis on original information of a patient registration number and the completion information result by means of a parallel three-layer rapid screening engine; S4, integrating the obtained comprehensive credibility score and a diagnosis result, generating a standardized information set and pushing same to a doctor workstation, and after doctor confirmation, synchronizing the standardized information set to an infectious disease monitoring system by means of an HL7FHIRR5 standard interface.

Description

A method for quickly screening and completing key registration information Technical Field

[0001] This invention relates to the field of medical informatics and data security and artificial intelligence, specifically to a method for rapidly screening and completing key registration information. Background Technology

[0002] Timely and accurate reporting of infectious disease outbreaks is the cornerstone of public health security. Currently, the country is vigorously promoting intelligent monitoring and early warning software for infectious diseases, placing higher demands on the quality and timeliness of reported data. However, a core bottleneck in current practice is that outpatients and emergency room patients, in order to save time, often only provide their ID number and mobile phone number when registering. This results in a large number of crucial fields required for infectious disease reporting cards, such as current address, occupation, and workplace, being missing at the source, severely delaying the reporting process.

[0003] However, existing solutions have significant drawbacks: manual data completion is inefficient and unsustainable, relying on follow-up phone calls from medical staff, which is time-consuming, labor-intensive, and has a low success rate, easily exceeding the 2-hour reporting deadline; traditional IT methods suffer from "data silos" and privacy risks, with data barriers between hospital systems and between hospitals and government / public health platforms not being effectively broken down, leading to system incompatibility and data "unrecognition," while traditional centralized data fusion methods pose serious privacy risks; existing automatic data completion technologies lack specificity and accuracy, and some smart healthcare patents focus on using deep learning and medical imaging for disease... While aiding in disease diagnosis, this approach is unrelated to addressing the lack of demographic information for patients. Furthermore, existing completion models often employ supervised learning architectures, which are weak in handling incomplete and unstructured data, lacking quantitative credibility assessment and conflict resolution logic. They also lack real-time screening and dynamic compliance verification, failing to dynamically link with the latest industry standards. They cannot complete key verifications such as address hierarchy integrity and occupational classification standardization instantly upon information entry, and no quantitative anomaly judgment criteria have been established. Therefore, there is an urgent need to invent a method that can deeply integrate multi-source data, incorporate the latest artificial intelligence and privacy computing technologies, and intelligently screen and complete key information in real time during the registration process. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to propose a method for rapid screening and completion of key registration information. This method solves problems such as the inability to simultaneously identify missing key fields for infectious disease reporting during patient registration, high privacy risks associated with multi-source data fusion, poor accuracy and scenario adaptability of key registration information completion, inability to guarantee real-time performance, and insufficient dynamics in compliance verification.

[0005] This was achieved through the following technical solutions:

[0006] A method for rapid screening and completion of key registration information includes the following steps: S1. Using the ID number provided by the patient during registration as a unique index by the registration information collection module, and with the patient's authorization, the multi-source data privacy fusion module, under a federated learning and differential privacy fusion framework, extracts and calculates features from the current patient's government data source, life data source, medical data source, and wearable device health time-series data source to obtain multiple numerical feature vectors; S2. Based on the modal feature memory and dynamic feature engineering, the multiple numerical feature vectors obtained in step S1 are processed by a lightweight completion model that integrates a self-supervised learning architecture, and the results are output for the current patient's current address, occupation, and work unit. The process involves: S3, real-time diagnosis of the patient's original registration information and the results of the missing key fields, using a parallel three-layer rapid screening engine. The three-layer rapid screening engine includes: a missing field identification layer based on bitmap and vector database indexes, a compliance intelligent verification layer based on knowledge graphs and finite state machines, and a logic anomaly insight layer based on topological potential calculation. S4, integrating the comprehensive credibility score and diagnostic results obtained from steps S2 and S3 to generate a standardized information set containing both the comprehensive credibility score and diagnostic results. This standardized information set is then pushed to the doctor's workstation. After the doctor confirms its accuracy, it is synchronized to the infectious disease monitoring system via the HL7 FHIR R5 standard interface. This invention significantly improves the efficiency and quality of infectious disease reporting and can be applied to the pre-monitoring stage of infectious disease monitoring in various medical institutions.

[0007] Preferably, in step S1, federated learning adopts a hierarchical architecture. Each data source node uses a lightweight neural network with a unified structure, and the gradient G is aggregated through a weighted averaging strategy to optimize the lightweight completion model. Here, gradient G = ∑(N_i / N_total) × G_i, where N_i is the data volume of the i-th node, N_total is the total data volume, G_i is the gradient of the i-th node, and i is the index of each node. By adopting a hierarchical architecture, federated learning enables local feature extraction and computation from multiple data sources.

[0008] Preferably, in step S1, differential privacy is used to add noise perturbation to the extracted sensitive features. It adopts a dynamic budget allocation strategy, which allocates sub-budgets in three levels—high, medium, and low—according to the feature sensitivity based on the global privacy budget ε. Through the differential privacy framework, privacy protection processing can be performed on data from different channels.

[0009] Preferably, the high-level allocation sub-budget corresponds to high-sensitivity features, the mid-level allocation sub-budget corresponds to medium-sensitivity features, and the low-level allocation sub-budget corresponds to low-sensitivity features. Simultaneously, the total cost ε_total of the extracted high-sensitivity features is calculated, and it is determined whether it exceeds a set safety threshold. The purpose of judging the total cost ε_total of high-sensitivity features is to ensure that the total cost ε_total is within the safety threshold.

[0010] Preferably, in step S1, the types of multiple numerical feature vectors include: address vector, occupation vector, time series vector, and privacy-enhancing vector. By transforming multi-source data into unified numerical feature vectors rich in semantic, spatiotemporal, and health-related information, high-quality input can be provided for model learning.

[0011] Preferably, in step S2, the method for constructing and calling the modality-specific memory is as follows: each non-standard feature combination Key and its corresponding standardized information and confidence value are stored as a key-value pair structure, and a cosine similarity association query is used. When the similarity Sim between a new input case Key and all Keys in the memory is greater than a set threshold, the most matching Value is returned; otherwise, a new key-value pair is added. By adopting a key-value pair storage structure, a domain knowledge base that can be quickly retrieved can be formed.

[0012] Preferably, in step S2, the lightweight completion model employs a lightweight neural network combining depthwise separable convolution and an attention mechanism, and uses a weighted cross-entropy loss function, the formula of which is: =-(1 / N)×∑w_i·[y_i·log(p_i)+(1-y_i)·log(1-p_i)], where, Let be the total loss value, N be the number of training samples, w_i be the weight of the i-th sample, y_i be the true label of the i-th sample, and p_i be the model prediction probability of the i-th sample. By using a lightweight model completion feature, the adaptability to incomplete data and long-tail scenarios can be significantly improved.

[0013] Preferably, in step S3, the implementation process of the logical anomaly insight layer is as follows: First, a dynamic relationship network containing five types of entity nodes—patient, address, workplace, residence, and wearable device activity points—is constructed, and edge weights w are set. Then, according to the calculation formula of the topological potential value V(x): V(x)=∑Γ(x)·w·exp(-β·d(x,y)²), the topological potential value of each entity node in the network is calculated. Simultaneously, the matching degree between the physiological time-series characteristics of the wearable device and its occupation and address is calculated. Finally, based on the completion results of low topological potential values ​​and feature mismatches, abnormal information is marked as high-risk results and prioritized for review. Here, x is any node, Γ(x) is the set of neighboring nodes of node x, d(x,y) is the semantic distance between node x and node y, and β=0.1. The logical anomaly insight layer based on topological potential value calculation can improve the accuracy of identifying abnormal information.

[0014] Preferably, in step S4, the HL7 FHIR R5 standard interface integrates a BERT large language model auxiliary verification module to achieve bidirectional conversion between structured data and natural language reports. By adopting the HL7 FHIR R5 standard interface, the final information can be synchronized to the infectious disease monitoring system with one click.

[0015] Preferably, in step S4, the standardized information set automatically triggers the blockchain audit and evidence storage module during the synchronization process, generating an immutable evidence record containing the data source, processing time, auditor ID, and audit result, and storing it encrypted using the SM4 algorithm. The blockchain audit and evidence storage module enables the traceability of sensitive personal information.

[0016] The beneficial effects of this invention compared to the prior art are:

[0017] The technical solution of this invention uses the patient's ID number as a unique index. With the patient's authorization, it can securely aggregate multi-source data under a federated learning and differential privacy fusion framework without the risk of data leakage. Based on a lightweight completion model with a fusion self-supervised learning architecture, combined with a modality-specific memory and dynamic feature engineering, it can output key missing field completion information results with comprehensive credibility scores. Simultaneously, a three-layer screening engine runs in parallel to complete real-time diagnosis, enabling proactive early warning of abnormal information. Combined with a dynamic synchronization mechanism of policies and standards, it ensures that compliance verification is consistent with the latest industry requirements, improving the in-depth defense capability of data quality. Through the HL7 FHIR R5 standard interface, the final information can be synchronized to the infectious disease monitoring system with one click, achieving seamless connection from registration, completion, review to final reporting. And through blockchain audit and evidence storage, it can achieve the traceability requirements of sensitive personal information. Attached Figure Description

[0018] Figure 1 is a flowchart of a method for quickly screening and completing key registration information;

[0019] Figure 2 is a system architecture diagram for implementing a method to quickly screen and complete key registration information.

[0020] Figure 3 is a schematic diagram of the collaboration of the privacy computing fusion engine;

[0021] Figure 4 is a flowchart of the three-layer rapid screening engine. Detailed Implementation

[0022] The technical solutions of the present invention will now be described in detail with reference to Figures 1-4 in the embodiments of the present invention.

[0023] Figure 1 shows a flowchart of a method for rapid screening and completion of key registration information. First, the patient's ID number is used as a unique index. After patient authorization, patient data from different channels is integrated and privacy-protected within a federated learning and differential privacy fusion framework. Then, based on a modal feature memory and dynamic feature engineering, a lightweight completion model with a self-supervised learning architecture outputs the completed information and its comprehensive credibility score. Next, a parallel three-layer rapid screening engine performs real-time diagnosis on the patient's original registration information and the completed information. Finally, the obtained comprehensive credibility score and diagnostic results are integrated to generate a standardized information set, which is pushed to the doctor's workstation. After the doctor confirms the accuracy, it is synchronized to the infectious disease monitoring system via the HL7 FHIR R5 standard interface. Figure 2 shows the system architecture diagram for implementing the rapid screening and completion method for key registration information. The diagram includes a registration information collection module, a multi-source data privacy fusion module, a hospital information system, a lightweight completion model, a three-layer rapid screening engine, and an HL7 FHIR system. The invention, in conjunction with Figures 1 and 2, utilizes an R5 standard interface, national infectious disease intelligent monitoring and early warning pre-processing software, a doctor's workstation, and a blockchain audit and evidence storage module. This method for rapidly screening and completing key registration information can significantly improve the efficiency and quality of infectious disease reporting.

[0024] The method specifically includes the following steps:

[0025] S1. Based on the registration information collection module, the ID number provided by the patient during registration is used as a unique index. After authorization by the patient, the multi-source data privacy fusion module extracts and calculates features from the current patient's government data source, life data source, medical data source, and wearable device health time series data source under the federated learning and differential privacy fusion framework, and obtains multiple numerical feature vectors.

[0026] The government data sources include registered addresses, previous addresses, and historical records of occupation and work unit. These are obtained through a government data sharing platform, after patient authorization, by querying the population database and social insurance database. The address information is then segmented and encoded, retaining valid data with a matching degree greater than a set threshold to obtain address vectors and occupation vectors. The residential data sources include utility bill payment addresses, travel records, and access control records. These are obtained through access to utility bill payment addresses, real-name public transportation travel records, and community smart access control records, within a framework of user authorization and data security protocols. Timestamps are extracted, and the MD5 hash algorithm is used to deduplicate the address information, retaining data with a residential activity level greater than a set threshold. Addresses are marked as high-frequency residential addresses to obtain a list of high-frequency addresses; medical data sources include historical medical addresses, occupations, and treatment information, which are obtained from hospital information systems and electronic medical record systems. Invalid records with a missing rate greater than a set threshold are removed, and the weights of the remaining data are calculated using a time decay weighting formula. The weights of recent data within a specific event period are increased to obtain a weighted historical information vector; wearable device health time-series data sources include wearable device activity trajectories and active time periods, which are obtained by accessing the time-series data of patients' wearable devices. Trajectory sampling is performed at fixed intervals, geofence matching degree is calculated, and valid trajectory data with a matching degree greater than a set threshold are filtered to obtain a time-series vector.

[0027] In this embodiment, in step S1, federated learning adopts a hierarchical architecture. Each data source node uses a lightweight neural network with a unified structure, and aggregates the gradient G through a weighted averaging strategy. The gradient G = ∑(N_i / N_total) × G_i, where N_i is the data volume of the i-th node, N_total is the total data volume, G_i is the gradient of the i-th node, and i is the index of each node. This optimizes the lightweight completion model in step S2. The initialization parameters of the local model of each data source node follow a normal distribution N(0, 0.1) and are trained based on local preprocessed data, which includes standardized... Address vectors, occupation vectors, high-frequency address lists, and time-series vectors are used. The loss function is the mean squared error (MSE) = ∑(y_true - y_pred)² / N, where y_true is the true data value and y_pred is the model's predicted value. The optimizer is Adam. After calculating the corresponding node gradient at each data source node, the local gradient is encrypted and uploaded to the federated server. The federated server uses a weighted average to aggregate the gradient G and distributes the aggregated gradient G to each data source node to update the model. The convergence condition is that the difference in the model's loss function ΔMSE after three consecutive rounds of aggregation is less than a set threshold; otherwise, the above process is repeated.

[0028] Figure 3 illustrates the collaborative process of the privacy-preserving computing fusion engine. Under the premise that each data source does not leave its domain, each edge node trains its local preprocessed data using a local model to obtain node gradients. These gradients are then encrypted and transmitted to the federated server via a differential privacy module after adding noise perturbations. The federated server aggregates the gradients of each node and transmits them to the feature fusion unit, which securely fuses the multi-source data. Here, "not leaving the domain" means that each data source is strictly restricted to flow and computation within its predefined physical or logical boundaries and cannot be transmitted or stored outside those boundaries.

[0029] In this embodiment, in step S1, differential privacy is used to add noise perturbation to the extracted sensitive features. It adopts a dynamic budget allocation strategy, which sets three levels of sub-budgets—high, medium, and low—based on the global privacy budget ε and the sensitivity of the features. The high-level sub-budget corresponds to high-sensitivity features, the medium-level sub-budget corresponds to medium-sensitivity features, and the low-level sub-budget corresponds to low-sensitivity features. At the same time, the total consumption ε_total of the extracted high-sensitivity features is calculated, and it is determined whether it exceeds the set safety threshold to ensure that the total consumption ε_total is within the safety threshold and to verify the effectiveness of privacy protection.

[0030] For numerically sensitive features, Laplace noise is added: f(x) = x + Laplace(0, λ), where x is the original feature value, λ = Δ / ε, and Δ is the sensitivity of any sensitive feature. For text-sensitive features, after being converted into a 128-dimensional vector by the Word2Vec word embedding model, Gaussian noise is added: v_new = v_old + N(0, σ²·I), where σ = Δ / ε × √2, v_new is the vector after adding noise, v_old is the original vector, and I is the identity matrix. For temporally sensitive features, noise is injected in segments according to fixed time windows, with the same noise value added in each window.

[0031] S2. Based on the modal feature memory and dynamic feature engineering, the multiple numerical feature vectors obtained in step S1 are processed by a lightweight completion model that integrates a self-supervised learning architecture, and the results of completing the key missing fields of the current patient's address, occupation and workplace, as well as their comprehensive credibility score, are output.

[0032] In this embodiment, in step S2, the method for constructing and calling the modality-specific memory bank is as follows: a key-value pair memory network is constructed for the three key fields of current address, occupation, and workplace. Each non-standard feature combination Key and its corresponding standardized information and confidence value are stored as a key-value pair structure. Cosine similarity association query is used. When the similarity Sim between a new case Key and all Keys in the memory bank is greater than a set threshold, the most matching Value is returned and the confidence Conf is updated. Otherwise, a new key-value pair is added.

[0033] In this embodiment, dynamic feature engineering is used to convert multi-source data into a unified numerical feature vector, specifically including: address vector, which parses text addresses into structured vectors of province, city, district, street, community, and house number, and adds a geolocation code centered on the address, while also incorporating features of surrounding points of interest; occupation vector, which generates industry code vectors by combining national occupational classification standards, and generates semantic vectors of occupational names using the Word2Vec word embedding model, while supplementing custom codes for emerging occupations; time-series vector, which extracts five types of time-series features from wearable devices: average daily outdoor time, trajectory volatility, distribution of active time periods, number of high-frequency stops, and overlap of commuting routes, and normalizes each type of feature to the [0,1] interval, concatenates them to obtain a 5-dimensional basic vector, and then generates a time-series feature vector through LSTM network encoding; privacy enhancement vector, which generates feature vectors after desensitizing sensitive features to ensure privacy and security.

[0034] In this embodiment, in step S2, the lightweight completion model employs a lightweight neural network combining depthwise separable convolution and an attention mechanism, and uses a weighted cross-entropy loss function, the formula of which is: =-(1 / N)×∑w_i·[y_i·log(p_i)+(1-y_i)·log(1-p_i)], where, denoted as the total loss value, N as the number of training samples, w_i as the weight of the i-th sample, y_i as the true label of the i-th sample, and p_i as the model prediction probability of the i-th sample. Simultaneously, an online learning mechanism is established, using newly added completion cases verified daily as new data to dynamically adjust the parameters of the lightweight completion model.

[0035] In this embodiment, when there are conflicts in the completion results of the same field from multiple data sources, a decision is made by weighted calculation of the overall credibility S=W×F×C; where W is the weight of any data source, and the weight is updated monthly according to the formula W_new=W×(0.9+0.1×Acc), and Acc is the accuracy of the completion result of any data source; F is the freshness factor, F=exp(-0.05×Δt), where Δt is the internal consistency score of any data source, C=0.6×C_addr+0.4×C_occ, where C_addr is the consistency score between address and other life trajectories, and C_occ is the consistency score between occupation and work unit. Finally, the result with the highest overall credibility S is selected as the final completion value.

[0036] S3. A parallel three-layer fast screening engine performs real-time diagnosis on the original information of patient registration and the results of key missing field completion information. The three-layer fast screening engine includes: a missing field identification layer based on bitmap and vector database index, a compliance intelligent verification layer based on knowledge graph and finite state machine, and a logic anomaly insight layer based on topological potential value calculation.

[0037] Figure 4 shows the workflow of the three-layer rapid screening engine. The rapid screening engine includes a missing field identification layer, a compliance intelligent verification layer, and a logical anomaly insight layer. It transmits the obtained compliance results and anomaly analysis reports to the result aggregation layer, which obtains the final screening report. This engine is used to perform second-level diagnostic verification of registration information and key missing field completion information, with a total time of less than 250ms. It can proactively discover compliant but questionable abnormal data, changing from passive reception to proactive warning. At the same time, combined with the dynamic synchronization mechanism of policies and standards, it ensures that compliance verification is consistent with the latest industry requirements, thereby improving the in-depth defense capability of data quality.

[0038] In this embodiment, in step S3, the time taken for diagnosis based on the missing field identification layer using bitmap and vector database indexes is less than 50ms. It employs a dual indexing technique of bitmap index + Milvus vector data index: First, one bit is allocated to each of the four key fields—current address, occupation, workplace, and contact number—to construct a 64-bit bit sequence: bit0 = current address, bit1 = occupation, bit2 = workplace, bit3 = contact number, and the remaining bits are reserved. Then, based on the set screening mask 0b1111, bitwise AND operations are used to analyze the patients in the registration queue. The patient data is quickly screened. If the screening result is not equal to 0b1111, it is determined that there is a missing field. The field corresponding to 0 in the result is a missing field. If the screening result is equal to 0b1111, it is determined that there is no missing field. At the same time, the patient's ID number is converted into a 64-dimensional vector and stored in the Milvus vector database. The IVF_FLAT index type is used, and nlist=1024 is used. When querying, the patient's historical records are quickly located by vector similarity matching. When the similarity is greater than a set threshold, the historical medical records are matched to obtain field completion suggestions, thereby supplementing basic information and reducing the pressure of completion.

[0039] In this embodiment, in step S3, the time taken for the compliance intelligent verification layer based on the knowledge graph and finite state machine to perform the diagnosis is less than 100ms. The knowledge graph contains three types of elements: entities, attributes, and relationships. Entities include addresses, occupations, and codes; attributes include levels, formats, and ranges; and relationships include attribution and mapping. Address entities are associated with administrative division codes, and occupation entities are associated with occupational classifications and codes. The finite state machine is used to implement the verification logic, and its state transitions are: S0 initial state → S1 format verification state → S2 code verification state → S3 association verification state → S4 compliant state or S5 abnormal state. The verification rules include using regular expressions and length checks. The mechanism validates the format of field information; it calls the coding library for mapping calculation, where the address coding mapping rate = number of successfully mapped fields / total number of fields × 100%, and requires the address coding mapping rate to be ≥ 80%, and the occupation coding mapping rate = 1, otherwise it enters an abnormal state; it calculates the reasonableness of the association between occupation and address, the association degree = co-occurrence frequency of the occupation and the region to which the address belongs in the knowledge graph / total occurrence frequency × 100%, and the association degree ≥ 30% is reasonable; at the same time, it automatically synchronizes the latest policy standards on the 1st of each month, updates the coding library and FSM rules, and verifies the accuracy through 500 historical samples after the update. If Acc ≥ 98%, it takes effect, otherwise it reverts to the previous version.

[0040] In this embodiment, in step S3, the diagnosis time of the logical anomaly insight layer based on topological potential value calculation is less than 100ms. The implementation process of the logical anomaly insight layer is as follows: First, a dynamic relationship network containing five types of entity nodes including patients, addresses, workplaces, residences, and active points of wearable devices is constructed, and edge weights w are set; then, according to the calculation formula of topological potential value V(x): V(x)=∑Γ(x)·w·exp(-β·d(x,y)²), the topological potential value of each entity node in the network is calculated. When the topological potential value V(x) is lower than the set threshold, the current node is marked as high risk; at the same time, the matching degree M_occ and M_addr of the physiological time sequence characteristics of wearable devices with occupation and address are calculated, M_occ=cosine(U,T), where U is the occupational feature vector and T is the time sequence characteristic vector. The ordered feature vector is M_addr = 0.5 × a1' + 0.3 × a2' + 0.2 × a3', where a1' is the normalized value of residential activity, a2' is the normalized value of trajectory matching, and a3' is the normalized value of payment frequency. M_occ ≥ 0.7 and M_addr ≥ 0.6 are considered reasonable. Finally, based on the completion results of low topological potential values ​​and feature mismatches, abnormal information is marked as high-risk results and prioritized for review. The risk level is quantified using the risk level formula: Risk = 0.6 × (1 - V(x) / 0.7) + 0.4 × (1 - M), where M is the average matching degree between M_occ and M_addr. Here, x is any node, Γ(x) is the set of neighboring nodes of node x, d(x,y) is the semantic distance between node x and node y, and β = 0.1.

[0041] S4. Integrate the comprehensive credibility score and diagnostic results obtained from steps S2 and S3 to generate a standardized information set containing information such as completion value, data source, comprehensive credibility score, compliance status, and abnormal risk level. Push the standardized information set to the doctor's workstation. After the doctor confirms that it is correct, synchronize it to the infectious disease monitoring system through the HL7 FHIR R5 standard interface, including the hospital information system, electronic medical record system, and national infectious disease intelligent monitoring and early warning software.

[0042] In this embodiment, in step S4, the HL7 FHIR R5 standard interface integrates the BERT large language model auxiliary verification module to achieve bidirectional conversion between structured data and natural language reports. HL7 FHIR R5 stands for Health Level Seven Fast Healthcare Interoperability Resources Release 5, which is a healthcare data exchange standard. BERT stands for Bidirectional Encoder Representations from Transformers, which is a bidirectional pre-trained language model based on the Transformers encoder architecture. During the synchronization process, the standardized information set automatically triggers the blockchain audit and evidence storage module to generate an immutable evidence record containing data source, processing time, auditor ID, and audit result, which is encrypted and stored using the SM4 algorithm, enabling traceability of sensitive personal information.

[0043] In summary, this invention uses the patient's ID number as a unique index. With the patient's authorization, it can securely aggregate multi-source data within a federated learning and differential privacy fusion framework without the risk of data leakage. Based on a lightweight completion model with a fusion self-supervised learning architecture, combined with a modality-specific memory and dynamic feature engineering, it can output key missing field completion information results with comprehensive credibility scores. Simultaneously, a three-layer screening engine runs in parallel to complete real-time diagnosis, enabling proactive early warning of abnormal information. Combined with a dynamic synchronization mechanism of policies and standards, it ensures compliance verification is consistent with the latest industry requirements, enhancing the in-depth defense capability of data quality. Through the HL7 FHIR R5 standard interface, the final information can be synchronized to the infectious disease monitoring system with one click, achieving seamless connection from registration, completion, review to final reporting. Furthermore, blockchain audit and evidence storage enables traceability requirements for sensitive personal information, demonstrating significant progress.

[0044] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for quick screening and completion of key information of a registration, characterized in that, The method includes the following steps: S1. Based on the registration information collection module, the ID number provided by the patient during registration is used as a unique index. After authorization by the patient, the multi-source data privacy fusion module extracts and calculates features from the current patient's government data source, life data source, medical data source and wearable device health time series data source under the federated learning and differential privacy fusion framework to obtain multiple numerical feature vectors. S2. Based on the modal feature memory and dynamic feature engineering, the multiple numerical feature vectors obtained in step S1 are processed by a lightweight completion model that integrates a self-supervised learning architecture, and the results of completing the key missing fields of the current patient's address, occupation and workplace, as well as their comprehensive credibility score, are output. S3. Real-time diagnosis of the patient's original registration information and key missing field completion information results through a parallel three-layer rapid screening engine; The three-layer rapid screening engine includes: a missing field identification layer based on bitmap and vector database indexes, a compliance intelligent verification layer based on knowledge graphs and finite state machines, and a logical anomaly insight layer based on topological potential value calculation. S4. Integrate the comprehensive confidence score and diagnostic results obtained from steps S2 and S3 to generate a standardized information set containing the comprehensive confidence score and diagnostic results. Push the standardized information set to the doctor's workstation. After the doctor confirms that it is correct, synchronize it to the infectious disease monitoring system through the HL7 FHIR R5 standard interface.

2. The method of claim 1, wherein, In step S1, federated learning adopts a hierarchical architecture. Each data source node uses a lightweight neural network with a unified structure, and the gradient G is aggregated through a weighted averaging strategy to optimize the lightweight completion model. Here, gradient G = ∑(N_i / N_total) × G_i, N_i is the amount of data in the i-th node, N_total is the total amount of data, G_i is the gradient of the i-th node, and i is the index of each node.

3. The method of claim 1, wherein the method further comprises: In step S1, differential privacy is used to add noise perturbation to the extracted sensitive features. It adopts a dynamic budget allocation strategy, which allocates sub-budgets in three levels (high, medium, and low) according to the feature sensitivity based on the global privacy budget ε.

4. The method of claim 3, wherein, High-level allocation sub-budgets correspond to high-sensitivity features, medium-level allocation sub-budgets correspond to medium-sensitivity features, and low-level allocation sub-budgets correspond to low-sensitivity features. At the same time, the total cost ε_total of the extracted high-sensitivity features is calculated, and it is determined whether it exceeds the set safety threshold.

5. The method of claim 1, wherein, In step S1, the types of multiple numerical feature vectors include: address vector, occupation vector, time series vector, and privacy enhancement vector.

6. The method of claim 1, wherein, In step S2, the method for constructing and calling the modality-specific memory is as follows: each non-standard feature combination Key and its corresponding standardized information and confidence value are stored as a key-value pair structure, and cosine similarity association query is used. When the similarity Sim between a new input case Key and all Keys in the memory is greater than a set threshold, the most matching Value is returned; otherwise, a new key-value pair is added.

7. The method of claim 1, wherein the method further comprises: In step S2, the lightweight completion model employs a lightweight neural network combining depthwise separable convolutions and an attention mechanism, and uses a weighted cross-entropy loss function, the formula of which is: =-(1 / N)×∑w_i·[y_i·log(p_i)+(1-y_i)·log(1-p_i)], where, denoted as the total loss value, N as the number of training samples, w_i as the weight of the i-th sample, y_i as the true label of the i-th sample, and p_i as the model prediction probability of the i-th sample.

8. The method of claim 1, wherein, In step S3, the implementation process of the logical anomaly insight layer is as follows: First, a dynamic relationship network containing five types of entity nodes, including patients, addresses, workplaces, residences, and active points of wearable devices, is constructed, and edge weights w are set; then, according to the calculation formula of topological potential V(x): V(x)=∑Γ(x)·w·exp(-β·d(x,y)²), the topological potential of each entity node in the network is calculated, and the matching degree between the physiological time sequence characteristics of wearable devices and occupation and address is calculated. Finally, based on the completion results of low topological potential and feature mismatch items, the abnormal information is marked as high-risk results and sent for review first; where x is any node, Γ(x) is the set of neighboring nodes of node x, d(x,y) is the semantic distance between node x and node y, and β=0.

1.

9. The method of claim 1, wherein, In step S4, the HL7 FHIR R5 standard interface integrates the BERT large language model auxiliary verification module to achieve bidirectional conversion between structured data and natural language reports.

10. The method of claim 1, wherein the method further comprises: In step S4, the standardized information set automatically triggers the blockchain audit and evidence storage module during the synchronization process, generating an immutable evidence storage record containing data source, processing time, auditor ID, and audit result, and storing it encrypted using the SM4 algorithm.

Images