Hospital data integration and intelligent management system based on multi-source data

Abstract

The invention discloses a hospital data integration and intelligent management system based on multi-source data, and particularly relates to the field of medical data management. Comprising a multi-source heterogeneous data acquisition layer, a self-adaptive data fusion engine, a privacy enhanced distributed storage layer, a multi-modal intelligent analysis center, a clinical decision enhancement module, a self-adaptive resource scheduling module and a full-process auditing and tracing module. Through cooperation of multi-modal model fusion, the clinical decision support accuracy is improved, the self-adaptive learning capability is achieved, the risk of data decryption leakage is avoided, 100% coverage of audit tracing and full-process operation of block chain evidence storage recording can be completed, the abnormal access detection rate is greatly improved, resource scheduling optimization is carried out, the equipment vacancy rate is reduced, and the method is suitable for large-scale popularization and application. A particle swarm optimization algorithm is combined with a digital twinborn technology, resource allocation is dynamically adjusted, scheduling is optimized through a genetic algorithm, invalid overtime is reduced, and the satisfaction degree of medical staff is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical data management, and more particularly, to a hospital data integration and intelligent management system based on multi-source data. BACKGROUND

[0002] With the deepening of medical informatization construction, hospitals have accumulated massive multi-source heterogeneous data, covering electronic medical records (EMR), medical images (such as CT / MRI / ultrasound), inspection results, wearable device monitoring data, medical equipment operation logs, and external data such as medical insurance and public health. These data are scattered and stored in different manufacturers' heterogeneous systems (such as HIS, PACS, LIS, EMR, etc.), and there are problems such as non-uniform protocol standards, data format conflicts, weak privacy protection, and single analysis means, which seriously restrict the deep utilization and value mining of medical data.

[0003] Medical devices follow SNMP, DICOM, HL7 FHIR, etc. protocols, and data collection needs to adapt to more than 200 protocol standards. Traditional static mapping rule libraries are difficult to cover new devices (such as intelligent monitors, AI-assisted diagnosis devices), and different systems have inconsistent coding methods for the same data item (such as patient ID, diagnosis code) (such as the term difference between ICD-10 and SNOMED CT), resulting in semantic ambiguity during data fusion. Traditional ETL tools (such as Informatica, Talend) rely on batch processing mode, and the data integration delay is high, which cannot meet the real-time decision-making needs of emergency, intensive care, etc. At the same time, medical data contains sensitive patient information (such as genetic data, mental illness history), and traditional encryption schemes (such as AES symmetric encryption) need to be decrypted when sharing data, which has the risk of secondary leakage. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides a hospital data integration and intelligent management system based on multi-source data.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: a hospital data integration and intelligent management system based on multi-source data, comprising a multi-source heterogeneous data acquisition layer, an adaptive data fusion engine, a privacy-enhanced distributed storage layer, a multimodal intelligent analysis center, a clinical decision enhancement module, an adaptive resource scheduling module, and a full-process audit and tracing module, wherein the input and output ends of the adaptive data fusion engine are respectively connected to the output end of the multi-source heterogeneous data acquisition layer and the input end of the privacy-enhanced distributed storage layer, the output end of the privacy-enhanced distributed storage layer is connected to the input end of the multimodal intelligent analysis center, the multimodal intelligent analysis center is bidirectionally connected to the clinical decision enhancement module and the adaptive resource scheduling module, and the output end of the clinical decision enhancement module is connected to the input end of the full-process audit and tracing module, wherein:

[0006] The multi-source heterogeneous data acquisition layer collects data from medical devices, electronic medical records, and external databases through standardized protocol interfaces. It includes a medical device data acquisition submodule, an electronic medical record acquisition submodule, and an external database acquisition submodule. The medical device data acquisition submodule supports SNMP v3 and DICOM protocol parsing and image metadata extraction; the electronic medical record acquisition submodule integrates the HL7 FHIR interface and the BioBERT entity recognition model; and the external database acquisition submodule is compatible with JDBC / ODBC and BSON parsers.

[0007] The adaptive data fusion engine includes a dynamic protocol parser, a conflict resolution rule base, and an incremental learning mechanism to achieve dynamic parsing of medical data protocols and format conflict resolution;

[0008] The privacy-enhanced distributed storage layer uses a hybrid encryption architecture, a dynamic desensitization engine, and blockchain evidence storage technology to build a secure and reliable medical data storage environment.

[0009] The multimodal intelligent analysis hub integrates a heterogeneous model orchestrator, a medical knowledge graph, and an adaptive feature selection module to support joint analysis of time series, image, and text data.

[0010] The clinical decision enhancement module includes a bimodal interpretation engine and an adversarial verification mechanism to provide explainable AI-assisted diagnostic recommendations;

[0011] The adaptive resource scheduling module combines a multi-objective optimization engine with digital twin technology to achieve dynamic configuration of medical resources;

[0012] The full-process audit and traceability module is used to ensure that the entire operation is traceable through a behavioral analysis engine and blockchain evidence records.

[0013] As a further improvement to the technical solution of the present invention, the multi-source heterogeneous data acquisition layer implements protocol adaptation and data preprocessing through the following technologies:

[0014] (1) Medical equipment data acquisition submodule:

[0015] a. Use SNMP v3 protocol to parse time-series data from monitors and other devices, and match OID identifiers (e.g., 1.3.6.1.4.1.21367.2010.1.2.100 corresponds to patient temperature) using a regular expression library.

[0016] b. Parse CT / MRI image data using the DICOM protocol, extract pixel data (0x7FE0,0010 tag) and metadata (patient ID, device serial number), and perform DICOM Tag Snooping to verify data integrity.

[0017] (2) Electronic medical record collection submodule:

[0018] a. Obtain structured medical records through the HL7 FHIR interface, parse the Resource type (such as Observation, MedicationRequest), and verify the required data elements;

[0019] b. Use the BioBERT-base model to perform entity recognition on unstructured text and extract key information from diagnosis and surgical records, with an F1-score ≥ 0.92;

[0020] (3) External database acquisition submodule:

[0021] a. Connect to relational databases such as Oracle and MySQL through JDBC / ODBC drivers and execute parameterized SQL queries to obtain test result data;

[0022] b. Use the BSON parser to extract gene sequencing data from NoSQL databases such as MongoDB and convert it to the standard FASTQ format.

[0023] As a further improvement to the technical solution of the present invention, the adaptive data fusion engine implements dynamic protocol parsing and conflict resolution through the following technologies:

[0024] (1) Dynamic protocol parser:

[0025] a. Loading a protocol signature library based on the Drools rule engine, which contains protocol fingerprints for 128 types of medical devices and systems (e.g., Modbus TCP function code 0x03 corresponds to read holding registers);

[0026] b. Two-way LSTM network (hidden layer 256 units, sequence length 512) is used for classification of unknown protocols, and the training data set contains 2150 protocol PCAP samples;

[0027] (2) Conflict resolution rule base:

[0028] a. Construct an OWL format medical ontology to define the hierarchical mapping relationship of ICD-10, ATC, LOINC terms;

[0029] b. Realize term standardization through SWRL rules (such as mapping SNOMED CT code 73211009 to ICD-10 E11.9);

[0030] Incremental learning mechanism:

[0031] a. Online learning strategy is adopted, and model fine-tuning is triggered every 1000 new data, and the learning rate is exponentially decayed (initial value 0.01, decay rate 0.7);

[0032] b. The resolution effect is evaluated by double indicators of F1-score (≥0.85) and protocol coverage (≥95%);

[0033] As a further improvement of the technical scheme of the application, the privacy-enhanced distributed storage layer realizes secure storage through the following technologies:

[0034] (1) Hybrid encryption architecture:

[0035] a. Structured data encryption: CKKS homomorphic encryption scheme (ring dimension 16384, modulus chain L=5) is adopted to support SQL aggregation query (SELECT AVG(lab_result) WHERE...) in encrypted state;

[0036] b. Image data encryption: DICOM images are divided into 512x512 pixel blocks, each block is encrypted by AES-256-GCM, and IV is generated by patient ID (first 16 bytes of SHA-256) and timestamp;

[0037] (2) Dynamic de-sensitization engine:

[0038] a. Define 5-level de-sensitization strategy (L0-L4), and L4 level adopts k-anonymity (k≥5) combined with differential privacy (ε=0.5);

[0039] b. According to the user role (doctor / research / operation and maintenance), the de-sensitization level is dynamically selected, and the role permission is verified by Hyperledger Fabric smart contract;

[0040] (3) Block chain storage sub-module:

[0041] a. Using the Hyperledger Fabric 2.4 framework, a 2-organization 3-node architecture is configured, and the endorsement policy requires signatures from at least 2 nodes;

[0042] b. Data operation logs are stored in the form of key-value pairs (Key = data hash, Value = operation metadata), and the consensus algorithm is improved PBFT (view switching timeout 5s, batch size 1000).

[0043] As a further improvement to the technical solution of the present invention, the multimodal intelligent analysis hub implements heterogeneous data analysis through the following technologies:

[0044] (1) Heterogeneous model orchestrator:

[0045] a. Time Series Analysis: A bidirectional LSTM network (128 hidden units) was deployed to predict the 72-hour ICU discharge probability. Input features included vital signs (heart rate, blood pressure) and laboratory indicators (lactate, creatinine).

[0046] b. Image Analysis: 3D U-Net++ network (residual connections 8) was used to segment lung CT images, with a Dice coefficient ≥ 0.92 and an inference speed ≤ 0.5 s / case.

[0047] c. Text analysis: Extracting electronic medical record risk factors using the SciBERT-v1.1 model, with a ROC-AUC of ≥ 0.88;

[0048] (2) Medical Knowledge Graph:

[0049] a. Extract entity relationships from the UMLS 2023AA version and construct a medical graph containing 120,000 entities and 800,000 relationships;

[0050] b. Implementing reasoning based on a graph attention network (GAT, with 8 attention heads) to predict complication risks (e.g., the probability of diabetic retinopathy).

[0051] (3) Adaptive feature selection:

[0052] a. Use SHAP value to calculate feature importance and filter out features with importance less than 0.05;

[0053] b. Use UMAP to reduce the dimensionality of high-dimensional data (such as gene expression profiles) to 50 dimensions, retaining 95% variance.

[0054] As a further improvement to the technical solution of the present invention, the clinical decision enhancement module achieves the explainability of AI recommendations through the following technologies:

[0055] (1) Dual-modal interpretation engine:

[0056] a. Local interpretation: LIME algorithm generates feature importance heatmap, sampling 10,000 times, kernel width σ = 0.05;

[0057] b. Global interpretation: Decision tree visualization showing rule chains (e.g., "If age > 65 and D-dimer > 500, then VTE risk is high"), with a maximum depth of 8 layers;

[0058] (2) Adversarial Verification Mechanism:

[0059] a. The GAN generator uses the StyleGAN2-ada architecture to generate realistic images of rare cases (such as pediatric congenital heart disease CTA);

[0060] b. The discriminator is trained using the WGAN-GP loss function with a FID score ≤ 20;

[0061] (3) Human-computer collaborative interface:

[0062] a. Doctors can use Label Studio to annotate model error cases, and select the most informative samples through uncertainty sampling of the labeled data;

[0063] b. The model iteration cycle is ≤ 7 days, and knowledge distillation technology (teacher model accuracy ≥ 98%) is used to accelerate training.

[0064] As a further improvement to the technical solution of the present invention, the adaptive resource scheduling module implements dynamic optimization through the following technologies:

[0065] (1) Multi-objective optimization engine:

[0066] a. Particle swarm optimization: inertia weight ω = 0.729, learning factor c1 = c2 = 1.494, fitness function is weighted sum (0.4 × equipment utilization + 0.3 × medical staff satisfaction + 0.3 × cost);

[0067] b. Genetic algorithm: Chromosomes are encoded as an operating room-time period matrix, crossover probability is 0.8, mutation probability is 0.1, and selection pressure is maintained by a tournament method (size 5);

[0068] (2) Digital twin technology:

[0069] a. Build a hospital operations simulation model based on AnyLogic 8.8, including dynamic interactions between the emergency department, operating room, and wards;

[0070] b. Resource allocation plan for a simulated surge in emergency department volume (3 times the baseline), with a response time of ≤ 2 minutes;

[0071] (3) Real-time fault tolerance mechanism:

[0072] a. When a device failure is detected (e.g., ventilator offline), automatically trigger the backup resource pool (redundancy ≥ 20%);

[0073] b. Update the scheduling policy through Q-learning, with a reward function of -1 x failure duration.

[0074] As a further improvement of the technical scheme of the present application, the full-process audit and traceability module is realized through the following technical operations:

[0075] (1) Behavior analysis engine:

[0076] a. Feature extraction includes access period (entropy value > 4.0 marked as abnormal), data volume (more than mean ± 2σ), operation sequence (LSTM anomaly detection, loss function is cross-entropy);

[0077] b. The classifier uses LightGBM (tree depth 10, learning rate 0.05), and the F1-score on the UNSW-NB15 dataset is 0.94;

[0078] (2) Blockchain storage record:

[0079] a. Transaction data includes operation type (read / write / delete), data hash, timestamp, and user certificate (X.509 format);

[0080] b. The smart contract implements access control policies (e.g., "attending physicians can only query data of patients in the group within 3 years"), which is written in Solidity 0.8.20;

[0081] (3) Abnormal response mechanism:

[0082] a. When a malicious operation is detected (e.g., 5 consecutive login failures), automatically freeze the account and trigger an alarm (SMS + email, response time < 30 seconds);

[0083] b. Audit logs support time travel queries (trace operation history through blockchain height, query depth ≥ 10000 blocks).

[0084] The beneficial effects of the present application are:

[0085] The present invention significantly shortens the data collection and fusion time through dynamic protocol analysis + incremental learning mechanism, and adopts a hybrid encryption architecture combined with block compression technology to reduce storage space occupancy while ensuring security. The collaboration of multimodal model fusion improves the accuracy of clinical decision support and also has adaptive learning capabilities. The model explains the importance of features through SHAP values ​​and makes real-time adjustments based on doctor feedback, shortening the iteration cycle and reducing privacy computing performance loss. CKKS homomorphic encryption supports data analysis in an encrypted state, avoiding the risk of data decryption leakage, and can complete 100% audit traceability coverage. The blockchain records the entire process operation, greatly improving the abnormal access detection rate, optimizing resource scheduling, and reducing equipment idle rate. The particle swarm optimization algorithm is combined with digital twin technology to dynamically adjust resource allocation, and the genetic algorithm optimizes scheduling to reduce ineffective overtime and improve the satisfaction of medical staff. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is a system framework diagram of the present invention. DETAILED DESCRIPTION

[0087] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0088] As attached Figure 1 A hospital data integration and intelligent management system based on multi-source data is shown, comprising a multi-source heterogeneous data acquisition layer, an adaptive data fusion engine, a privacy-enhanced distributed storage layer, a multimodal intelligent analysis hub, a clinical decision enhancement module, an adaptive resource scheduling module, and a full-process audit and tracing module. The input and output of the adaptive data fusion engine are respectively connected to the output of the multi-source heterogeneous data acquisition layer and the input of the privacy-enhanced distributed storage layer. The output of the privacy-enhanced distributed storage layer is connected to the input of the multimodal intelligent analysis hub. The multimodal intelligent analysis hub is bidirectionally connected to the clinical decision enhancement module and the adaptive resource scheduling module. The output of the clinical decision enhancement module is connected to the input of the full-process audit and tracing module.

[0089] The multi-source heterogeneous data collection layer collects data from medical devices, electronic medical records, and external databases through standardized protocol interfaces. It includes a medical device data collection submodule, an electronic medical record collection submodule, and an external database collection submodule. The medical device data collection submodule supports SNMP v3 and DICOM protocol parsing and image metadata extraction; the electronic medical record collection submodule integrates the HL7 FHIR interface and the BioBERT entity recognition model; and the external database collection submodule is compatible with JDBC / ODBC and BSON parsers.

[0090] Adaptive data fusion engine, including a dynamic protocol parser, conflict resolution rule base, and incremental learning mechanism, for dynamic parsing of medical data protocols and resolution of format conflicts;

[0091] A privacy-enhanced distributed storage layer, using a hybrid encryption architecture, a dynamic desensitization engine, and blockchain evidence storage technology, is used to build a secure and reliable medical data storage environment;

[0092] A multimodal intelligent analysis hub that integrates a heterogeneous model orchestrator, a medical knowledge graph, and an adaptive feature selection module to support joint analysis of time series, imaging, and text data.

[0093] A clinical decision enhancement module, including a bimodal explanation engine and adversarial verification mechanism, is used to provide explainable AI-assisted diagnostic recommendations;

[0094] Adaptive resource scheduling module, combining a multi-objective optimization engine with digital twin technology to achieve dynamic allocation of medical resources;

[0095] The full-process audit and traceability module uses a behavioral analysis engine and blockchain evidence records to ensure that the entire operation is traceable.

[0096] Preferably, the multi-source heterogeneous data acquisition layer implements protocol adaptation and data preprocessing through the following technologies:

[0097] (1) Medical equipment data acquisition submodule:

[0098] a. Use SNMP v3 protocol to parse time-series data from monitors and other devices, and match OID identifiers (e.g., 1.3.6.1.4.1.21367.2010.1.2.100 corresponds to patient temperature) using a regular expression library.

[0099] b. Parse CT / MRI image data using the DICOM protocol, extract pixel data (0x7FE0,0010 tag) and metadata (patient ID, device serial number), and perform DICOM Tag Snooping to verify data integrity.

[0100] (2) Electronic medical record collection submodule:

[0101] a. Obtain structured medical records through the HL7 FHIR interface, parse the Resource type (such as Observation, MedicationRequest), and verify the required data elements;

[0102] b. Use the BioBERT-base model to perform entity recognition on unstructured text and extract key information from diagnosis and surgical records, with an F1-score ≥ 0.92;

[0103] (3) External database acquisition submodule:

[0104] a. Connect to relational databases such as Oracle and MySQL through JDBC / ODBC drivers and execute parameterized SQL queries to obtain test result data;

[0105] b. Use the BSON parser to extract gene sequencing data from NoSQL databases such as MongoDB and convert it to the standard FASTQ format.

[0106] Preferably, the adaptive data fusion engine implements dynamic protocol parsing and conflict resolution through the following technologies:

[0107] (1) Dynamic protocol parser:

[0108] a. Loading a protocol signature library based on the Drools rule engine, which contains protocol fingerprints for 128 types of medical devices and systems (e.g., Modbus TCP function code 0x03 corresponds to read holding registers);

[0109] b. A bidirectional LSTM network (256 hidden units, 512 sequence length) is used to classify unknown protocols. The training dataset contains PCAP samples of 2150 protocols.

[0110] (2) Conflict resolution rule base:

[0111] a. Construct a medical ontology in OWL format and define the hierarchical mapping relationship between ICD-10, ATC, and LOINC terms;

[0112] b. Standardize terminology using SWRL rules (e.g., mapping SNOMED CT code 73211009 to ICD-10 E11.9);

[0113] Incremental learning mechanism:

[0114] a. Adopt an online learning strategy, triggering model fine-tuning every time 1000 new data are received, and the learning rate decays exponentially (initial value 0.01, decay rate 0.7);

[0115] b. The parsing effect is evaluated by the dual indicators of F1-score (≥0.85) and protocol coverage (≥95%).

[0116] Preferably, the privacy-enhanced distributed storage layer achieves secure storage through the following technologies:

[0117] (1) Hybrid encryption architecture:

[0118] a. Structured data encryption: Utilizes the CKKS homomorphic encryption scheme (ring dimension 16384, modulus chain L = 5), supporting SQL aggregation queries (SELECT AVG(lab_result) WHERE...) in an encrypted state.

[0119] b. Image data encryption: DICOM images are segmented into 512×512 pixel blocks, each of which is encrypted using AES-256-GCM. The IV is generated by combining the patient ID (the first 16 bytes of SHA-256) and the timestamp.

[0120] (2) Dynamic desensitization engine:

[0121] a. Define a five-level desensitization strategy (L0-L4). L4 uses a combination of k-anonymization (k ≥ 5) and differential privacy (ε = 0.5).

[0122] b. Dynamically select the desensitization level based on the user role (doctor / scientific researcher / operation and maintenance), and role permissions are verified through Hyperledger Fabric smart contracts;

[0123] (3) Blockchain evidence submodule:

[0124] a. Using the Hyperledger Fabric 2.4 framework, a 2-organization 3-node architecture is configured, and the endorsement policy requires signatures from at least 2 nodes;

[0125] b. Data operation logs are stored in the form of key-value pairs (Key = data hash, Value = operation metadata), and the consensus algorithm is improved PBFT (view switching timeout 5s, batch size 1000).

[0126] Preferably, the multimodal intelligent analysis center implements heterogeneous data analysis through the following technologies:

[0127] (1) Heterogeneous model orchestrator:

[0128] a. Time Series Analysis: A bidirectional LSTM network (128 hidden units) was deployed to predict the 72-hour ICU discharge probability. Input features included vital signs (heart rate, blood pressure) and laboratory indicators (lactate, creatinine).

[0129] b. Image analysis: 3D U-Net++ network (residual connection number 8) is used to segment lung CT images, Dice coefficient ≥ 0.92, and inference speed ≤ 0.5 s / case;

[0130] c. Text analysis: SciBERT-v1.1 model is used to extract electronic medical record risk factors, ROC-AUC ≥ 0.88;

[0131] (2) Medical knowledge graph:

[0132] a. Extract entity relationships from UMLS2023AA version to build a medical graph containing 120,000 entities and 800,000 relationships;

[0133] b. Realize inference based on graph attention network (GAT, attention head number 8) to predict complication risk (such as the probability of diabetic retinopathy);

[0134] (3) Adaptive feature selection:

[0135] a. SHAP value is used to calculate feature importance, and features with importance <0.05 are filtered out;

[0136] b. UMAP is used to reduce high-dimensional data (such as gene expression profiles) to 50 dimensions, retaining 95% variance.

[0137] Preferably, the clinical decision enhancement module realizes the explainability of AI suggestions through the following technologies:

[0138] (1) Dual-modal explanation engine:

[0139] a. Local explanation: LIME algorithm generates feature importance heat map, sampling number 10000 times, kernel width σ=0.05;

[0140] b. Global explanation: Decision tree visualizes rule chain (such as "if age>65 and D-dimer>500, then VTE risk is high"), maximum depth 8 layers;

[0141] (2) Adversarial verification mechanism:

[0142] a. GAN generator adopts StyleGAN2-ada architecture to generate realistic rare case images (such as pediatric congenital heart disease CTA);

[0143] b. Discriminator is trained by WGAN-GP loss function, FID score ≤ 20;

[0144] (3) Human-computer cooperation interface:

[0145] a. Doctors can use Label Studio to annotate model error cases, and select the most informative samples through uncertainty sampling of the labeled data;

[0146] b. The model iteration cycle is ≤ 7 days, and knowledge distillation technology (teacher model accuracy ≥ 98%) is used to accelerate training.

[0147] Preferably, the adaptive resource scheduling module implements dynamic optimization through the following technologies:

[0148] (1) Multi-objective optimization engine:

[0149] a. Particle swarm optimization: inertia weight ω = 0.729, learning factor c1 = c2 = 1.494, fitness function is weighted sum (0.4 × equipment utilization + 0.3 × medical staff satisfaction + 0.3 × cost);

[0150] b. Genetic algorithm: Chromosomes are encoded as an operating room-time period matrix, crossover probability is 0.8, mutation probability is 0.1, and selection pressure is maintained by a tournament method (size 5);

[0151] (2) Digital twin technology:

[0152] a. Build a hospital operations simulation model based on AnyLogic 8.8, including dynamic interactions between the emergency department, operating room, and wards;

[0153] b. Resource allocation plan for a simulated surge in emergency department volume (3 times the baseline), with a response time of ≤ 2 minutes;

[0154] (3) Real-time fault tolerance mechanism:

[0155] a. Automatically trigger the backup resource pool (redundancy ≥ 20%) when a device failure is detected (e.g., a ventilator goes offline);

[0156] b. Update the scheduling policy through Q-learning, with the reward function being -1×fault duration.

[0157] Preferably, the full-process audit and traceability module implements operation monitoring through the following technologies:

[0158] (1) Behavior Analysis Engine:

[0159] a. Feature extraction includes access time (anomaly is marked if entropy value > 4.0), data volume (exceeds mean ± 2σ), and operation sequence (LSTM anomaly detection, with cross entropy as the loss function);

[0160] b. The classifier uses LightGBM (tree depth 10, learning rate 0.05), with F1-score = 0.94 on the UNSW-NB15 dataset;

[0161] (2) Blockchain evidence records:

[0162] a. Transaction data includes the operation type (read / write / delete), data hash, timestamp, and user certificate (X.509 format);

[0163] b. Smart contracts implement access control policies (e.g., "attending physicians can only query data from patients in this group within the past three years"), written in Solidity 0.8.20;

[0164] (3) Abnormal response mechanism:

[0165] a. When malicious operations are detected (e.g., five consecutive login failures), the account will be automatically frozen and an alert will be triggered (SMS + email, response time < 30 seconds);

[0166] b. Audit logs support time travel queries (retrace operation history through the blockchain, query depth ≥ 10,000 blocks).

[0167] The differences and advantages of this invention over the prior art are as follows:

[0168]

[0169] In summary, the data collection and fusion time is greatly shortened through dynamic protocol parsing + incremental learning mechanism, and the hybrid encryption architecture combined with block compression technology is used to reduce storage space usage while ensuring security. The synergy of multimodal model fusion improves the accuracy of clinical decision support and also has adaptive learning capabilities. The model explains the importance of features through SHAP values ​​and fine-tunes in real time based on doctor feedback, shortening the iteration cycle and reducing the performance loss of privacy computing. CKKS homomorphic encryption supports data analysis in an encrypted state, avoiding the risk of data decryption leakage, and can complete 100% audit traceability. The blockchain records the entire process of operation, greatly improving the abnormal access detection rate, optimizing resource scheduling, and reducing the equipment idle rate. The particle swarm optimization algorithm is combined with digital twin technology to dynamically adjust resource allocation, and the genetic algorithm optimizes scheduling to reduce ineffective overtime and improve the satisfaction of medical staff.

[0170] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A hospital data integration and intelligent management system based on multi-source data, characterized by: The system includes a multi-source heterogeneous data acquisition layer, an adaptive data fusion engine, a privacy-enhanced distributed storage layer, a multimodal intelligent analysis center, a clinical decision enhancement module, an adaptive resource scheduling module, and a full-process audit and tracing module. The input and output ends of the adaptive data fusion engine are respectively connected to the output end of the multi-source heterogeneous data acquisition layer and the input end of the privacy-enhanced distributed storage layer. The output end of the privacy-enhanced distributed storage layer is connected to the input end of the multimodal intelligent analysis center. The multimodal intelligent analysis center is bidirectionally connected to the clinical decision enhancement module and the adaptive resource scheduling module. The output end of the clinical decision enhancement module is connected to the input end of the full-process audit and tracing module. The multi-source heterogeneous data acquisition layer collects data from medical devices, electronic medical records, and external databases through standardized protocol interfaces. It includes a medical device data acquisition submodule, an electronic medical record acquisition submodule, and an external database acquisition submodule. The medical device data acquisition submodule supports SNMP v3 and DICOM protocol parsing and image metadata extraction; the electronic medical record acquisition submodule integrates the HL7 FHIR interface and the BioBERT entity recognition model; and the external database acquisition submodule is compatible with JDBC / ODBC and BSON parsers. The adaptive data fusion engine includes a dynamic protocol parser, a conflict resolution rule base, and an incremental learning mechanism to achieve dynamic parsing of medical data protocols and format conflict resolution; The privacy-enhanced distributed storage layer uses a hybrid encryption architecture, a dynamic desensitization engine, and blockchain evidence storage technology to build a secure and reliable medical data storage environment. The multimodal intelligent analysis hub integrates a heterogeneous model orchestrator, a medical knowledge graph, and an adaptive feature selection module to support joint analysis of time series, image, and text data. The clinical decision enhancement module includes a bimodal interpretation engine and an adversarial verification mechanism to provide explainable AI-assisted diagnostic recommendations; The adaptive resource scheduling module combines a multi-objective optimization engine with digital twin technology to achieve dynamic configuration of medical resources; The full-process audit and traceability module is used to ensure that the entire operation is traceable through a behavioral analysis engine and blockchain evidence records.

2. The hospital data integration and intelligent management system based on multi-source data according to claim 1, characterized in that: The multi-source heterogeneous data acquisition layer implements protocol adaptation and data preprocessing through the following technologies: (1) Medical equipment data acquisition submodule: a. Use SNMP v3 protocol to parse the monitor's time series data and match OID identifiers using a regular expression library; b. Parse CT / MRI image data using the DICOM protocol, extract pixel data (0x7FE0,0010 tag) and metadata (patient ID, device serial number), and perform DICOM Tag Snooping to verify data integrity. (2) Electronic medical record collection submodule: a. Obtain structured medical records through the HL7 FHIR interface, parse the Resource type, and verify the required data elements; b. Use the BioBERT-base model to perform entity recognition on unstructured text and extract key information from diagnosis and surgical records, with an F1-score ≥ 0.92; (3) External database acquisition submodule: a. Connect to Oracle or MySQL relational databases through JDBC / ODBC drivers and execute parameterized SQL queries to obtain test result data; b. Use the BSON parser to extract gene sequencing data from the MongoDB NoSQL database and convert it to the standard FASTQ format.

3. The hospital data integration and intelligent management system based on multi-source data according to claim 1, characterized in that: The adaptive data fusion engine implements dynamic protocol parsing and conflict resolution through the following technologies: (1) Dynamic protocol parser: a. Loading the protocol signature library based on the Drools rule engine, including protocol fingerprints for 128 types of medical devices and systems; b. A bidirectional LSTM network (256 hidden units, 512 sequence length) is used to classify unknown protocols. The training dataset contains PCAP samples of 2150 protocols. (2) Conflict resolution rule base: a. Construct a medical ontology in OWL format and define the hierarchical mapping relationship between ICD-10, ATC, and LOINC terms; b. Standardize terminology through SWRL rules; Incremental learning mechanism: a. Adopt an online learning strategy, triggering model fine-tuning every time 1000 new data are received, and the learning rate decays exponentially (initial value 0.01, decay rate 0.7); b. The parsing effect is evaluated by the dual indicators of F1-score (≥0.85) and protocol coverage (≥95%).

4. The hospital data integration and intelligent management system based on multi-source data according to claim 1, characterized in that: The privacy-enhanced distributed storage layer achieves secure storage through the following technologies: (1) Hybrid encryption architecture: a. Structured data encryption: Utilizes the CKKS homomorphic encryption scheme (ring dimension 16384, modulus chain L = 5), supporting SQL aggregation queries (SELECT AVG(lab_result) WHERE...) in an encrypted state. b. Image data encryption: DICOM images are segmented into 512×512 pixel blocks, each of which is encrypted using AES-256-GCM. The IV is generated by combining the patient ID (the first 16 bytes of SHA-256) and the timestamp. (2) Dynamic desensitization engine: a. Define a five-level desensitization strategy (L0-L4). L4 uses a combination of k-anonymization (k ≥ 5) and differential privacy (ε = 0.5). b. Dynamically select the desensitization level based on the user role (doctor / scientific researcher / operation and maintenance), and the role permissions are verified through the Hyperledger Fabric smart contract; (3) Blockchain evidence submodule: a. Using the Hyperledger Fabric 2.4 framework, a 2-organization 3-node architecture is configured, and the endorsement policy requires signatures from at least 2 nodes; b. Data operation logs are stored in the form of key-value pairs (Key = data hash, Value = operation metadata), and the consensus algorithm is improved PBFT (view switching timeout 5s, batch size 1000).

5. The hospital data integration and intelligent management system based on multi-source data according to claim 1, characterized in that: The multimodal intelligent analysis center implements heterogeneous data analysis through the following technologies: (1) Heterogeneous model orchestrator: a. Time Series Analysis: A bidirectional LSTM network (128 hidden units) was deployed to predict the 72-hour ICU discharge probability. Input features included vital signs (heart rate, blood pressure) and laboratory indicators (lactate, creatinine). b. Image Analysis: 3D U-Net++ network (residual connections 8) was used to segment lung CT images, with a Dice coefficient ≥ 0.92 and an inference speed ≤ 0.5 s / case. c. Text analysis: Extracting electronic medical record risk factors using the SciBERT-v1.1 model, with a ROC-AUC of ≥ 0.88; (2) Medical Knowledge Graph: a. Extract entity relationships from the UMLS 2023AA version and construct a medical graph containing 120,000 entities and 800,000 relationships; b. Implementing reasoning based on the Graph Attention Network (GAT, with 8 attention heads) to predict complication risk; (3) Adaptive feature selection: a. Use SHAP value to calculate feature importance and filter out features with importance less than 0.05; b. Use UMAP to reduce the dimensionality of high-dimensional data to 50 dimensions, retaining 95% variance.

6. The hospital data integration and intelligent management system based on multi-source data according to claim 1, characterized in that: The clinical decision enhancement module achieves explainability of AI recommendations through the following technologies: (1) Bimodal interpretation engine: a. Local interpretation: LIME algorithm generates feature importance heatmap, sampling 10,000 times, kernel width σ = 0.05; b. Global interpretation: Decision tree visualization shows rule chains with a maximum depth of 8 layers; (2) Adversarial Verification Mechanism: a. The GAN generator uses the StyleGAN2-ada architecture to generate realistic rare case images; b. The discriminator is trained using the WGAN-GP loss function with a FID score ≤ 20; (3) Human-computer collaborative interface: a. Doctors can use Label Studio to annotate model error cases, and select the most informative samples through uncertainty sampling of the labeled data; b. The model iteration cycle is ≤ 7 days, and knowledge distillation technology (teacher model accuracy ≥ 98%) is used to accelerate training.

7. The hospital data integration and intelligent management system based on multi-source data according to claim 1, characterized in that: The adaptive resource scheduling module achieves dynamic optimization through the following technologies: (1) Multi-objective optimization engine: a. Particle swarm optimization: inertia weight ω = 0.729, learning factor c1 = c2 = 1.494, fitness function is weighted sum (0.4 × equipment utilization + 0.3 × medical staff satisfaction + 0.3 × cost); b. Genetic algorithm: Chromosomes are encoded as an operating room-time period matrix, crossover probability is 0.8, mutation probability is 0.1, and selection pressure is maintained by a tournament method (size 5); (2) Digital twin technology: a. Build a hospital operations simulation model based on AnyLogic 8.8, including dynamic interactions between the emergency department, operating room, and wards; b. Resource allocation plan for a simulated surge in emergency department volume (3 times the baseline), with a response time of ≤ 2 minutes; (3) Real-time fault tolerance mechanism: a. Automatically trigger the backup resource pool (redundancy ≥ 20%) when a device failure is detected; b. Update the scheduling policy through Q-learning, with the reward function being -1×fault duration.

8. The hospital data integration and intelligent management system based on multi-source data according to claim 1, characterized in that: The full-process audit and traceability module implements operation monitoring through the following technologies: (1) Behavior Analysis Engine: a. Feature extraction includes access time (anomaly is marked if entropy value > 4.0), data volume (exceeds mean ± 2σ), and operation sequence (LSTM anomaly detection, with cross entropy as the loss function); b. The classifier uses LightGBM (tree depth 10, learning rate 0.05), with F1-score = 0.94 on the UNSW-NB15 dataset; (2) Blockchain evidence records: a. Transaction data includes the operation type (read / write / delete), data hash, timestamp, and user certificate (X.509 format); b. Smart contracts implement access control policies, written in Solidity 0.8.20; (3) Abnormal response mechanism: a. Automatically freeze the account and trigger an alarm when malicious operations are detected; b. Audit logs support time travel queries.

Images