Medical big data management system and method based on cloud edge collaboration

Abstract

The invention discloses a medical big data management system and method based on cloud edge collaboration, and relates to the technical field of medical data management. Efficient data management is achieved through cooperation of multiple modules, and the heterogeneous data fusion module collects multi-source data in a unified mode and constructs a knowledge graph; the privacy protection calculation module guarantees data security by using homomorphic encryption and differential privacy; the edge intelligent analysis module deploys a lightweight model to realize real-time analysis; the cloud edge collaborative decision-making module dynamically allocates tasks according to the data and the equipment state; in addition, data quality evaluation, knowledge graph updating and other mechanisms are included, and the data reliability and timeliness are improved. Efficient and safe management of medical data is realized, the data processing efficiency is greatly improved, and the cloud cost and energy consumption are reduced; the data privacy is guaranteed, and the leakage risk is almost zero; the model diagnosis accuracy is improved, and clinical decision making is assisted; the diagnosis time is shortened, the error rate is reduced, and the medical service quality is remarkably optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data management, and in particular to a medical big data management system and method based on cloud-edge collaboration. BACKGROUND

[0002] In the current high-speed development of medical informatization, medical data is growing explosively, covering multi-modal data such as electronic medical records, medical images, and genetic testing. However, the traditional medical data management mode faces many bottlenecks. On the one hand, data is scattered and stored in different medical institutions and department systems. Due to the use of different data standards and formats by each system, such as the use of HL7 protocol to transmit medical record data by some hospitals and the use of DICOM standard by some image systems, data is difficult to interconnect and share, forming a large number of "data islands", which seriously hinders the integration of medical resources and collaborative diagnosis and treatment.

[0003] On the other hand, the efficiency and security requirements of medical data processing are in sharp conflict. Medical data is highly sensitive and contains patient privacy information. In cross-institutional collaborative analysis, the traditional mode is difficult to guarantee data privacy and security. At the same time, with the application of AI technology in the medical field, higher requirements are put forward for real-time analysis and processing of medical data, but existing technologies mostly rely on centralized processing in the cloud, which faces problems such as high network delay and high transmission cost. For example, in the emergency scenario, if medical image analysis relies entirely on the cloud, it may delay the diagnosis opportunity due to network congestion, affecting patient treatment.

[0004] In addition, the quality of medical data is uneven and the knowledge mining capability is limited. In the data collection process, missing values, outliers and other problems are prone to occur, affecting the accuracy of analysis; traditional data mining techniques are difficult to deeply mine potential laws from massive medical data, and cannot provide strong support for clinical decision-making. Moreover, the existing medical data management system lacks intelligent management of the whole life cycle of data, and is inefficient in data storage, updating, and tracing, which is difficult to meet the needs of precision medicine and smart medicine development. SUMMARY

[0005] The present application proposes a medical big data management system and method based on cloud-edge collaboration to solve the problems mentioned in the existing technology.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme: a medical big data management system based on cloud-edge collaboration, comprising: A heterogeneous data fusion module: deploy edge data collection terminals, use the Transformer architecture for feature extraction, build a medical knowledge graph to realize data semantic association and conflict resolution; introduce a medical terminology standardization engine integrating UMLS and SNOMEDCT terminology library to support cross-language data mapping; Privacy protection computing module: design a federated learning framework based on homomorphic encryption, adopt BGV homomorphic encryption algorithm, support linear operation to realize cross-institution model training; introduce differential privacy mechanism ε=0.5, δ=10 -6 , add Laplace noise when querying data to protect patient information; deploy zero-knowledge proof system zk-SNARKs to verify the correctness of the calculation result; Edge analysis module: deploy CNN model on the edge, realize automatic detection of lung nodules for CT images; use RNN-LSTM hybrid network to analyze electrocardiogram data and real-time warning of arrhythmia; integrate explainable AI components to generate heat maps through Grad-CAM to explain model decisions; Cloud-edge collaborative decision-making module: establish a task offloading model: C=α・C e +β・C emo +γ・C m , α=0.4, β=0.3, γ=0.3, where C e is the edge computing cost, C emo is the edge-to-cloud transmission cost, and C m is the cloud computing cost; design an adaptive scheduling algorithm to dynamically allocate computing tasks according to network conditions. Introduce edge caching mechanism LRU strategy to cache hot data and reduce cloud access pressure.

[0007] Medical data quality evaluation mechanism: construct a data quality evaluation index system: Q=w1・C+w2・A+w3・I+w4・C o , w1=0.3, w2=0.25, w3=0.25, w4=0.2, where C is integrity, A is accuracy, I is consistency, and C o is compliance; use Bayesian network for quality evaluation, train prior probability distribution through historical data to realize real-time monitoring of data quality; deploy anomaly detection engine, use isolation forest algorithm to identify data anomalies and automatically trigger manual review process.

[0008] Further, it also includes: Medical knowledge graph dynamic updating mechanism: design incremental learning algorithm, trigger update when new data confidence > 0.8; establish knowledge conflict detection rule base, use D-S evidence theory to fuse knowledge, and introduce knowledge distillation technology to compress medical knowledge base to model.

[0009] Further, in the heterogeneous data fusion module, use Transformer architecture, introduce medical field pre-training model BioBERT, adapt to medical text understanding task through fine-tuning, and achieve F1 value of 0.93 in medical named entity recognition task; design cross-modal alignment mechanism, realize semantic alignment of image and text features through contrastive learning.

[0010] Further, in the privacy protection computing module, a federated learning framework based on a block chain is designed, a PBFT consensus algorithm is adopted to realize model parameter trust sharing, a contract is used to automatically execute a data use protocol, a violation penalty coefficient λ=10 is introduced, a secure computing MPC protocol is introduced to support cross-institutional data joint analysis while ensuring data privacy.

[0011] Further, in the edge analysis module, for resource-limited edge devices, a model compression strategy is designed: a pruning technique is used to reduce redundant connections, a quantization technique INT8 is used to reduce storage requirements, and distillation learning is used to transfer teacher model knowledge to student models; an adaptive inference engine is developed to dynamically adjust model complexity according to device resources.

[0012] Further, in the cloud-edge collaborative decision module, a task scheduling strategy based on reinforcement learning is established, the state space includes device load, network delay, and data priority dimensions, and the action space includes edge processing, cloud processing, and partial offloading; a reward function is designed: R=r1・(1-D / D max )+r2・(1-E / E max )-r3・C(r1=5, r2=3, r3=1), where D is the delay, D max is the maximum acceptable delay, E is the energy consumption, E max is the maximum energy consumption, and C is the cost; a predictive scheduling algorithm is introduced to predict network congestion through an ARIMA model for time series analysis and adjust the task allocation strategy in advance.

[0013] A method of applying the cloud-edge collaborative medical big data management system, comprising: Data acquisition and preprocessing: deploying data acquisition terminals at the edge of medical institutions to support CT, MRI, ECG, electronic medical records, etc.; using ETL tools for data cleaning, multiple imputation method for missing value processing, IsolationForest algorithm for anomaly detection, Z-score normalization for standardization, and converting to a unified format FHIR standard; deploying edge computing gateways to realize local data caching and preprocessing acceleration; Edge analysis: deploying computing models at the edge, using a U-Net++ network with attention gate mechanism for lesion segmentation for image data; using TemporalConvolutionalNetwork for disease prediction for time series data, with AUC of 0.89 in heart failure prediction tasks; developing an edge inference engine to realize real-time inference and support online model updates; Cloud edge collaborative processing: Establish a task offloading decision model to dynamically determine the processing location according to the data type, device state and network condition; when K>0.7, C<0.5 and B>100Mbps, the task is offloaded to the cloud; otherwise, it is processed at the edge; deploy an edge-cloud collaborative scheduling system to support task priority scheduling and resource elastic allocation; Cloud deep analysis: Deploy a computing cluster in the cloud, use graph neural network GNN to analyze the medical knowledge graph, and find disease association rules; combine genome, proteome and metabolome to build a disease molecular mechanism model to support medical decision-making; develop a parallel computing framework to support node knowledge graph analysis; Decision support: Generate personalized medical recommendations based on analysis results; use natural language generation technology to convert analysis results into doctor-readable reports; visualize the disease development process through virtual reality technology to assist doctors in developing treatment plans; deploy a clinical decision support system CDSS that integrates diagnosis and treatment path recommendations and drug interaction warning functions.

[0014] Further, in the data collection and preprocessing step, a blockchain node is deployed to record the data flow process, and RSA-2048 asymmetric encryption is used to ensure data integrity; a data traceability algorithm is designed to realize data lifecycle traceability through a hash chain; a data quality traceability system is developed to record the quality indicators of each link in data collection, transmission and processing, and to support quality problem root cause analysis.

[0015] Further, in the decision support step, a clinical decision support system CDSS is constructed to integrate guideline knowledge, Cochrane systematic review of evidence-based medicine and world data. A decision analysis MCDM method is used to set the doctor's preference weight , , i=1, i is the medical criterion, patient willingness weight , , j=1, j is the individual demand, and a scheme utility formula is constructed: Generate personalized schemes, fi(x) is the treatment benefit corresponding to criterion i, and gj(x) is the value corresponding to demand j; deploy a SHAP explanation module to quantify the contribution of features Using SHAP values , visualize risk-reward to assist doctors in decision-making.

[0016] Compared with the existing technology, the beneficial effects of the present application are: In data processing, through heterogeneous data fusion and edge intelligent analysis, the data preprocessing and image analysis time is greatly shortened, the data processing efficiency is improved by more than 66%, and the clinical real-time diagnosis demand is met. In terms of privacy protection, homomorphic encryption, differential privacy and other technologies are used to reduce the data leakage risk to a very low level, while ensuring the safe and efficient cooperation of cross-institutional data, and the convergence time of federated learning is shortened by 24%.

[0017] The system resource utilization efficiency is significantly improved, the intelligent unloading of tasks is realized through cloud-edge collaborative decision-making, the cloud computing cost is reduced by 35%, and the energy consumption of edge devices is reduced by 25%. The model performance and interpretability are enhanced, the disease detection accuracy, model explanation accuracy and other indicators are significantly improved, and more reliable auxiliary diagnosis results and clear decision-making basis are provided for doctors. The clinical application effect is outstanding, the diagnosis and decision-making time is shortened by 83.3%, the treatment scheme quality is improved, the medical error rate is reduced by 58.3%, and the patient satisfaction is improved by 11%, which effectively promotes the upgrading of medical service quality and helps the development of intelligent medical treatment. BRIEF DESCRIPTION OF DRAWINGS

[0018] Fig. 1 A schematic block diagram of a medical big data management system based on cloud-edge collaboration is provided for the present application; Fig. 2 A schematic block diagram of a medical big data management method based on cloud-edge collaboration is provided for the present application; Fig. 3 A data processing efficiency comparison diagram of a medical big data management method based on cloud-edge collaboration is provided for the present application; Fig. 4 A system resource utilization rate and cost comparison diagram of a medical big data management method based on cloud-edge collaboration is provided for the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0020] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0021] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited. In addition, the terms "mounting", "connecting", "connecting" should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the communication between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances, and the present application will be further described in detail with reference to the accompanying drawings.

[0022] Reference Figs. 1 to 4 A cloud edge collaboration based medical big data management method and system The present application relates to a cloud edge collaboration based medical big data management method and system, which realizes efficient collection, intelligent analysis and precise application of medical data through multi-module collaborative work. The present application will be described in detail below in combination with the drawings and specific implementation cases.

[0023] System architecture and hardware deployment The present system adopts a three-level hierarchical architecture, including an edge layer, a network layer and a cloud layer.

[0024] The edge layer is deployed inside the medical institution and includes a multi-modal data acquisition terminal, an edge computing gateway, and an edge inference engine. The data acquisition terminal supports CT (such as Siemens SOMATOM Force), MRI (such as GE SIGNA Pioneer 3.0T), electrocardiogram monitor (such as Philips IntelliVue MX800), and other medical equipment data acquisition, and realizes standardized data transmission through DICOM, HL7, and other protocols. The edge computing gateway uses an industrial-grade server (such as Huawei TaiShan 2280) with an 8-core CPU, 16GB RAM, and 1TB SSD, supporting local data caching and preprocessing. The edge inference engine integrates an NVIDIA Jetson AGXXavier GPU, providing 21TOPS computing power, and supports real-time inference of lightweight AI models.

[0025] The network layer uses a 5G + private line hybrid networking, and a 5G base station (such as Huawei 5GCPE Pro2) is deployed inside the medical institution to provide a peak rate of 2.5Gbps wireless access. Key data is directly connected to the cloud through a private line network (such as China Unicom MPLS-VPN), ensuring a transmission delay of ≤10ms. An SD-WAN controller (such as DeepinSoft SD-WAN) is deployed to realize intelligent scheduling of network traffic and prioritize key medical data transmission.

[0026] The cloud layer is deployed in a medical cloud data center and is configured with a high-performance computing cluster (such as Sunway TC6600) containing 100 computing nodes (each node is configured with 2x Intel Xeon Gold 6348R CPU, 512GB RAM, and 8x NVIDIA A100 GPU). The storage system uses a distributed file system (such as Ceph) with a total capacity of ≥10PB, supporting 100,000+ IOPS per second. A container orchestration platform (such as Kubernetes) is deployed to realize resource elastic scaling and support smooth expansion of service load from 100 to 10,000 users.

[0027] Detailed implementation of each step 1. Heterogeneous data fusion module This module realizes the collection, cleaning, and fusion of multi-source heterogeneous medical data. In the data collection stage, a medical Internet of Things gateway (such as Advantech UNO-2483G) is deployed to support 20+ communication protocols (such as Modbus, OPCUA, DICOM, HL7, etc.), realizing unified access of data from different medical devices. The collection frequency is dynamically adjusted according to the data type, such as real-time collection of electrocardiogram data (sampling rate 250Hz) and on-demand collection of image data (automatic transmission after CT scanning is completed).

[0028] Data cleaning was performed using ETL tools (such as TalendDataIntegration), with over 100 cleaning rules configured. For missing values, the MICE algorithm was used to generate 5-10 imputed datasets based on the correlation between variables. For outliers, the IsolationForest algorithm (100 isolation trees, subsample size 256) was used to identify and label data points that deviate from the normal distribution (outlier score threshold 0.6).

[0029] Data standardization adopts the FHIR (Fast Healthcare Interoperability Resources) standard, and a medical terminology mapping engine has been developed (integrating terminology databases such as UMLS, SNOMEDCT, and ICD-10) to support cross-language data conversion. For example, the Chinese diagnostic term "myocardial infarction" is mapped to the SNOMEDCT code "410429000" with a mapping accuracy of 98.5%.

[0030] Heterogeneous data fusion employs an improved Transformer architecture, incorporating prior knowledge from the medical domain into the standard multi-head attention mechanism. The model's input layer contains 2048 neurons, corresponding to standardized medical features; the hidden layer comprises a 6-layer Transformer encoder, each with 12 attention heads; and the output layer generates a 768-dimensional feature vector for subsequent analysis. Model training utilizes BioBERT pre-trained weights for initialization and fine-tuning on a medical text dataset (learning rate 2e-5, batch size=16, training epochs 10), achieving an F1 score of 0.932 in the medical named entity recognition task.

[0031] When constructing the medical knowledge graph, the Neo4j graph database is used for storage. Node types include diseases, symptoms, examinations, treatments, and drugs, while relationship types include "cause," "diagnosis," and "treatment." Knowledge is extracted from unstructured text through entity recognition (e.g., the BERT-BiLSTM-CRF model, with an accuracy of 95.3%) and relation extraction (e.g., the RE-BERT model, with an accuracy of 90.1%). Currently, the graph contains 1.2 million nodes and 3.5 million relationships.

[0032] 2. Privacy-preserving computing module The module ensures the privacy and security of medical data and supports cross-institutional data collaboration. In homomorphic encryption federated learning, the BGV homomorphic encryption algorithm is used with a key length of 2048 bits and supports up to 10 multiplication depths. During model training, each medical institution encrypts the data and trains the model locally, and only uploads the encrypted gradient parameters to the federated server. The federated server aggregates the gradients (using the FedAvg algorithm with a learning rate of 0.01) and distributes the updated encrypted model to each institution. After testing, in joint training involving 10 hospitals, the model convergence time increased by only 25% compared to traditional centralized training, while the accuracy remained above 97%.

[0033] The differential privacy mechanism is implemented in the data query phase, adding Laplace noise to the query results. The sensitivity parameter Δf is dynamically set according to the query type, such as 1 for count queries and the maximum data value for sum queries. The privacy budget ε = 0.5 and δ = 10 -6 While ensuring data privacy, the query result error is controlled within an acceptable range (e.g., the relative error of count queries is ≤5%).

[0034] The blockchain federated learning framework uses the Hyperledger Fabric platform, deploying 21 consensus nodes (2 per institution and 1 for the third-party audit institution). The smart contract is written in Solidity, enabling automatic execution of data usage agreements. For example, when an institution violates data usage rules, the system automatically freezes its access rights and deducts the deposit (10% of the initial guarantee). After testing, the smart contract execution delay is ≤500ms, supporting 100+ transactions per second.

[0035] The zero-knowledge proof system uses the zk-SNARKs algorithm, allowing the verifier to confirm the correctness of the calculation result without accessing the original data. In the genetic data analysis scenario, the zero-knowledge proof verifies that the genetic test results meet the characteristics of a specific disease, with a verification time of only 87ms, while traditional methods require access to the original data and take ≥10s.

[0036] 3. Edge intelligent analysis module This module implements real-time analysis of medical data on the edge, reducing the burden on the cloud. For CT image analysis, a lightweight U-Net++ model is deployed with a parameter size of 3.8M, a reduction of 85% compared to the original model. The model is trained using knowledge distillation technology, with the teacher model being a full-size U-Net++ (parameter size 25.6M) and the student model improving performance by learning the teacher model's soft labels (temperature T = 3) and feature maps (using attention transfer loss). In the lung nodule detection task, the student model achieved an accuracy of 95.3%, close to the teacher model's 96.2%, while the inference speed improved by 4 times (single CT scan processing time reduced from 2.8s to 0.7s).

[0037] The ECG data analysis uses a TCN (Temporal Convolutional Network) model, the input layer receives 12-lead ECG signals (sampling rate 250 Hz, window length 10 s), the hidden layer contains 5 convolutional blocks (convolution kernel size 7, expansion factor 1, 2, 4, 8, 16), and the output layer predicts 8 types of arrhythmia. The model is trained on the MIT-BIH arrhythmia database, with an F1 value of 0.925, an improvement of 3.2 percentage points over the traditional LSTM model, and a reduction of 60% in parameter quantity.

[0038] To improve the inference efficiency of edge devices, an adaptive inference engine is developed to dynamically adjust the model complexity according to device resources. When CPU usage is > 80%, the calculation precision of some computationally intensive layers (such as convolutional layers) is automatically reduced from FP32 to INT8, while maintaining an accuracy loss of < 1%. In the case of extreme resource constraints (such as memory usage > 90%), model pruning technology (threshold τ = 0.02) is used to dynamically remove unimportant connections, temporarily reducing model parameter quantity by 30% and reducing inference delay by 40%.

[0039] Integrating an interpretable AI component, Grad-CAM algorithm is used to generate heat maps to explain the basis for model decision-making. In the diagnosis of pneumonia CT, the heat map can accurately mark the infected area of the lungs, with an explanation accuracy of 87.2%. Doctors can quickly verify the model's judgment through the heat map, improving the confidence of diagnosis.

[0040] 4. Cloud-edge collaborative decision-making module This module realizes intelligent allocation of computing tasks and optimizes the overall performance of the system. The task offloading decision model considers the data criticality, device status and network condition. The data criticality K is determined by expert scoring and historical importance statistics, for example, emergency imaging data K = 0.9, routine physical examination data K = 0.3. The device status C is obtained through real-time monitoring, including CPU usage, memory usage, GPU utilization, etc., and is normalized (0-1 range). Network condition B is calculated by 30 consecutive ping test results, including average delay, packet loss rate, etc.

[0041] The adaptive scheduling algorithm is trained using reinforcement learning (PPO algorithm), with a state space containing 10 dimensions (device load, network delay, data priority, task type, time slice, etc.) and an action space containing 3 choices (edge processing, cloud processing, partial offloading). The reward function is designed as: R = r1・(1-D / D max )+r2・(1-E / E max )-r3・C, where r1 = 5, r2 = 3, r3 = 1, D is the task completion delay, D maxFor the maximum acceptable latency (e.g., <5 minutes for emergency image analysis), E is the energy consumption, E max For the maximum energy consumption, C is the cost (e.g., cloud computing fees). After 10 6 steps of training, the algorithm converges to the optimal strategy, with an average reward improvement of 42% in the test environment.

[0042] An edge caching mechanism is introduced, using the LRU (Least Recently Used) strategy to manage cached data. The cache capacity is dynamically allocated based on the storage size of the edge device (default allocation of 30% storage space), and the current average hit rate reaches 72.5%. For example, frequently accessed common medical image templates and diagnostic report templates are cached in priority, and the next time they are accessed, they are read directly from the local device, reducing the response time from an average of 2.3s to 0.15s.

[0043] The predictive scheduling algorithm uses an ARIMA(2,1,2) model to predict network congestion in the next 30 minutes, and adjusts the task allocation strategy in advance. The model is trained based on the network traffic data of the past 7 days, and the prediction accuracy reaches 86.7%. In the network congestion warning scenario, the system can transfer non-urgent tasks to edge processing 15 minutes in advance to avoid service interruption caused by cloud overload.

[0044] 5. Medical data quality evaluation mechanism This mechanism ensures the reliability and availability of medical data. The data quality evaluation index system includes 4 first-level indicators and 12 second-level indicators. Completeness C evaluates the missing situation of data fields, including record completeness (e.g., the completeness rate of patient basic information), field completeness (e.g., the completeness rate of diagnosis field), etc.; Accuracy A evaluates the degree of conformity between data and the true situation, including numerical accuracy (e.g., the error rate of test results), logical accuracy (e.g., the consistency of age and birth date), etc.; Consistency I evaluates the consistency of data from different sources and at different times, including format consistency (e.g., the uniformity rate of date format), term consistency (e.g., the standardization rate of diagnosis terms), etc.; Compliance C o evaluates whether the data complies with regulations and internal policies, including privacy compliance (e.g., the desensitization rate of sensitive information), security compliance (e.g., the data access audit coverage rate), etc.

[0045] The improved Bayesian network contains 50 nodes and 120 edges, with nodes representing data quality indicators and edges representing the dependency relationship between indicators. The network parameters are trained through historical data to construct a conditional probability table. For example, when the "record completeness" indicator is observed to be declining, the network can predict the probability of "diagnosis accuracy" being affected as 0.65. In real-time monitoring, the system collects data quality indicators every hour, and calculates the overall quality score Q through Bayesian inference. When Q < 0.8, the warning mechanism is triggered, and a quality analysis report is automatically generated and pushed to the data administrator.

[0046] The anomaly detection engine uses the Isolation Forest algorithm to monitor data in real-time. The algorithm builds 100 isolation trees, each randomly selecting features and split points. Anomaly data usually requires fewer splits to be isolated. In an application in a top-tier hospital, the engine detects about 2000 abnormal data records per month, with a recall rate of 92.3%, of which about 15% are true medical errors (such as abnormal high values of test results), and 85% are data entry errors (such as negative age input).

[0047] 6. Dynamic updating mechanism of medical knowledge graph This mechanism ensures the timeliness and accuracy of the knowledge graph. The incremental learning algorithm uses a small batch training strategy, and triggers an update when the cumulative number of new data reaches 1000 and the confidence is greater than 0.8. The confidence calculation considers data source reliability, data consistency, and model prediction probability. For example, data from authoritative medical guidelines has a data source reliability weight of 0.9, while data from social media has a weight of 0.2.

[0048] The knowledge conflict detection rule base contains 1000+ rules, divided into semantic conflict, logical conflict, and fact conflict. For example, the rule "Disease A and Disease B are mutually exclusive, if a patient is diagnosed with both A and B, there is a conflict" belongs to logical conflict detection. When a conflict is detected, the D-S evidence theory is used to fuse multi-source knowledge and calculate the trust degree of each proposition. For example, when two diagnosis results conflict, the system will consider evidence such as test results and doctor experience to calculate the trust degree of each diagnosis and select the result with higher trust degree.

[0049] Knowledge distillation technology migrates knowledge from large medical knowledge base (such as UMLS, parameter quantity 12B) to lightweight model (parameter quantity 85M). A two-stage training strategy is adopted, the first stage trains the student model to imitate the output distribution of the teacher model (soft label), and the second stage fine-tunes on real medical data. In the medical terminology standardization task, the F1 value of the lightweight model reaches 0.915, close to the teacher model's 0.932, and the reasoning delay is reduced from an average of 3.2s to 0.4s.

[0050] 7. Data collection and preprocessing This step realizes the standardized collection and preprocessing of medical data. The multi-modal data collection terminal supports the access of 20+ medical devices, and realizes unified data collection through standardized interfaces (such as DICOM, HL7). The collection frequency is dynamically adjusted according to the device type and data importance, for example, real-time collection of electrocardiogram monitoring data (sampling rate 250Hz), and body temperature data collected every 5 minutes.

[0051] ETL tool configures 100+ cleaning rules, with specific cleaning strategies for different types of data. For example, image data cleaning includes denoising (using non-local mean filter, h parameter = 10), enhancement (using CLAHE algorithm, clipLimit = 2.0, tileGridSize = (8, 8)), and normalization (adjusting to 0-255 grayscale range); text data cleaning includes deduplication (using SimHash algorithm, threshold 3), segmentation (using jieba segmentation, custom medical dictionary containing 100,000+ terms), and normalization (e.g., "heart attack" unified as "myocardial infarction").

[0052] Edge computing gateway is deployed in each department of the hospital, with 8-core CPU, 16GB RAM, and 1TB SSD, supporting local data caching and preprocessing acceleration. The gateway uses containerized deployment, with each data processing task running in an independent container, ensuring stability through resource isolation. Data preprocessing uses parallel computing frameworks (such as Dask), reducing the time for processing 1000 patient records from 15 minutes to 3 minutes on a 10-core CPU.

[0053] 8. Edge intelligent analysis This step implements real-time analysis of medical data on the edge. Lightweight AI models use a three-stage model compression strategy: first, network architecture optimization (e.g., replacing ResNet50 with MobileNetV3), reducing parameters by 75%; second, pruning (threshold τ = 0.01), removing 30% redundant connections; and finally, quantization (INT8), reducing storage requirements and computational complexity. In CT image analysis, the compressed model parameter size is reduced from 25.6M to 3.8M, the inference speed is increased from 0.35 per second to 1.5 per second, and the accuracy is only reduced by 0.9 percentage points.

[0054] The edge inference engine supports GPU acceleration, using TensorRT to optimize model inference. On an NVIDIA Jetson AGXXavier, the processing time for a single CT scan (512x512x300) is reduced from 2.8s for the original TensorFlow to 0.7s, meeting the clinical real-time diagnosis requirements. The engine also supports online model updates, using differential update technology (only transmitting the changed part of the model parameters), reducing the update time from 5 minutes for the complete model to 30 seconds, reducing the impact on clinical work.

[0055] 9. Cloud-edge collaborative processing This step realizes the intelligent allocation and collaborative processing of computing tasks. The task offloading decision model considers the data criticality, device status, and network conditions. For example, in an emergency scenario, the CT image data of a patient with chest pain has a criticality K = 0.9. If the edge device CPU usage C = 0.8 and the network bandwidth B = 150 Mbps, the system automatically offloads the analysis task to the cloud. In a regular physical examination scenario, the chest X-ray data K = 0.3, C = 0.3, and B = 50 Mbps, then the edge processing is performed.

[0056] The edge-cloud collaborative scheduling system supports task priority scheduling, defining 9 priorities (1-9 levels, 9 being the highest). For example, the examination data of an emergency patient is by default at level 9 and needs to be analyzed within 5 minutes; the regular examination of an inpatient is at level 5 and needs to be completed within 2 hours; the physical examination data of an outpatient is at level 3 and needs to be completed within 24 hours. The system dynamically adjusts resource allocation according to priority to ensure that high-priority tasks are processed first.

[0057] The resource elasticity allocation mechanism automatically adjusts resources according to load. During regular working hours (8:00-17:00), the system automatically increases the cloud resource allocation ratio from 50% to 80%; during the night (0:00-6:00), the resource allocation ratio is reduced to 30%. During public health emergencies, the system can quickly increase the resource scaling ratio from the default 1:5 to 1:10 to meet the surge in analysis demand.

[0058] 10. Cloud deep analysis This step realizes the deep mining and analysis of large-scale medical data in the cloud. The high-performance computing cluster uses a distributed computing framework (such as Spark) to support billion-node knowledge graph analysis. In a certain provincial medical data joint analysis project, the system processes a dataset containing 50 million patient records and 1 billion medical events, and completes full-data analysis (including disease prevalence trends and treatment effect evaluation) in 24 hours, down from 1 month using traditional methods.

[0059] The graph neural network uses a GCN (Graph Convolutional Network) architecture for deep analysis of medical knowledge graphs. The model input layer receives node features (such as disease symptom features and drug molecular features), the hidden layer contains 3 layers of graph convolution (convolution kernel size 16), and the output layer predicts the potential relationship between nodes. In the drug repositioning task, the model predicts the potential efficacy of 5 marketed drugs against the novel coronavirus, of which 2 drugs are verified effective in subsequent clinical trials.

[0060] Multi-omics data analysis platform integrates genomic, proteomic, and metabolomic data to construct disease molecular mechanism models. In cancer research, the platform analyzed multi-omics data from 1000 patients, discovered 12 new driver gene mutations, and 5 key signaling pathways, providing a theoretical basis for developing personalized treatment plans.

[0061] 11. Intelligent decision support This step provides intelligent support for clinical decision-making based on analysis results. The clinical decision support system (CDSS) integrates 500+ clinical guidelines, Cochrane systematic reviews, and real-world data, uses multi-criteria decision analysis (MCDM) methods, combines physician preferences (weight ω) and patient wishes (weight φ), and generates personalized treatment plans. In the application of a certain tertiary hospital, CDSS assisted physicians in developing treatment plans for about 5000 cases per month, of which about 70% of the plans were consistent with expert consensus, and the decision-making time was shortened from an average of 30 minutes to 5 minutes.

[0062] Natural language generation technology uses a T5 model fine-tuning to convert analysis results into readable reports for doctors. The model is trained on a medical text dataset (learning rate 1e-4, batch size 8, training rounds 20), and the generated reports are fluent, logical, and consistent with medical professional expression standards. In the CT image diagnosis report generation task, the similarity between the model-generated report and the report written by a radiologist reached 85%, and the generation time was only 2 seconds.

[0063] Virtual reality (VR) technology is used for disease development process visualization to assist physicians in developing treatment plans. The system reconstructs three-dimensional models based on patients' image data (such as CT, MRI), combines disease progression models (such as tumor growth models), and simulates the development trend of diseases under different treatment plans. In neurosurgery planning, doctors can immerse themselves in observing the relationship between tumors and surrounding neural tissues through VR devices, develop the optimal surgical path, and the success rate of surgery is increased from 85% to 92% compared to traditional methods.

[0064] Experimental data and effect analysis To verify the effectiveness of the invention, a series of comparative experiments were conducted. The experiments were conducted on a provincial medical cloud platform, covering 20 tertiary hospitals and 50 primary medical institutions, with a test period of 6 months.

[0065] 1. Data processing efficiency comparison

[0066] From the above data, it can be seen that the present application has a significant improvement in data processing efficiency. The data preprocessing time is reduced from 15.2 minutes of the traditional method to 3.0 minutes, mainly due to edge computing and parallel processing technology. The image analysis delay is reduced from 2.8 seconds to 0.7 seconds, which is the combined effect of model compression and GPU acceleration. The efficiency of knowledge graph updating and the time of multi-omics data analysis are greatly shortened, which reflects the advantages of distributed computing and incremental learning. These improvements enable medical data to be converted into clinical value more quickly, improving the efficiency of medical services.

[0067] 2. Privacy protection and security comparison

[0068] In terms of privacy protection and security, the present application has obvious advantages. The risk of data leakage is reduced from "high" of the traditional method to "very low", which benefits from the comprehensive application of homomorphic encryption, differential privacy and blockchain technology. The convergence time of federated learning is shortened by 24.0%, indicating that the aggregation algorithm of the present application is more efficient. The zero-knowledge proof verification time is reduced from 10.2 seconds to 0.087 seconds, with an increase of nearly 100 times, making real-time verification possible. The reduction of smart contract execution delay improves the response speed and availability of the system. These improvements provide reliable protection for the safe sharing and cooperation of medical data.

[0069] 3. Model performance and interpretability comparison

[0070] The present application has improved both model performance and interpretability. The accuracy of lung nodule detection is increased from 92.5% to 95.3%, and the F1 value of arrhythmia prediction is increased from 0.893 to 0.925, which benefits from model architecture optimization and knowledge distillation technology. The accuracy of drug repositioning is increased by 7.1%, showing the advantage of graph neural network in medical knowledge graph analysis. The model explanation accuracy is increased from 75.3% to 87.2%, enabling doctors to better understand the basis of AI decision-making and enhancing trust. These improvements make the application of AI models in the medical field more reliable and practical.

[0071] 4. System resource utilization and cost comparison

[0072] In terms of resource utilization and cost, the invention achieves significant reduction. Cloud computing cost is reduced by 35%, which is the result of task intelligent offloading and resource elastic allocation. Edge device energy consumption is reduced by 25%, thanks to model compression and adaptive inference engine. Data storage cost is reduced by 40%, mainly because of data preprocessing and cache mechanism optimization. System maintenance workload is reduced by 55%, which is the embodiment of automated management and remote update function. These improvements make the medical big data management system more economical and efficient, reducing the burden of medical institutions.

[0073] 5. Clinical application effect comparison

[0074] In terms of clinical application effect, the invention brings significant improvement. Diagnosis decision time is shortened from 30 minutes to 5 minutes, improving medical service efficiency, especially for emergency patients. Treatment plan quality score is improved from 75 to 85, indicating that the intelligent decision support system can provide better treatment recommendations. Patient satisfaction is improved from 82% to 91%, reflecting the improvement of overall medical service experience. Medical error rate is reduced from 1.2% to 0.5%, which is the common result of data quality improvement and AI assisted diagnosis. These improvements directly improve the quality of medical services and improve patient prognosis.

[0075] In summary, the invention achieves significant improvement in data processing efficiency, privacy protection, model performance, resource utilization and clinical application through multi-technology fusion and innovation strategy, and can provide a full range of intelligent solutions for medical big data management.

[0076] The above is only the preferred embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can make equivalent replacement or change within the technical scope disclosed by the present application according to the technical scheme and inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. A cloud-edge collaborative medical big data management system, characterized in that, include: Heterogeneous data fusion module: Deploys edge data acquisition terminals, uses Transformer architecture for feature extraction, constructs a medical knowledge graph to achieve semantic association and conflict resolution of data; introduces a medical terminology standardization engine to integrate UMLS and SNOMEDCT terminology databases, and supports cross-language data mapping; Privacy-preserving computation module: A federated learning framework based on homomorphic encryption is designed, employing the BGV homomorphic encryption algorithm to support linear operations for cross-institutional model training; a differential privacy mechanism ε=0.5, δ=10 is introduced. -6 Add Laplace noise during data querying to protect patient information; Deploy zero-knowledge proof systems (zk-SNARKs) to verify the correctness of computation results; Edge analysis module: Deploys CNN models at the edge to automatically detect lung nodules in CT images; uses RNN-LSTM hybrid network to analyze ECG data and provide real-time warnings of arrhythmias; integrates interpretive AI components to generate heatmaps through Grad-CAM to explain model decisions; Cloud-edge collaborative decision-making module: Establishing a task offloading model: C=α・C e +β・C emo +γ・C m , α=0.4, β=0.3, γ=0.3, where C e For edge computing costs, C emo For edge-to-cloud transmission costs, C m To reduce cloud computing costs, an adaptive scheduling algorithm is designed to dynamically allocate computing tasks based on network conditions. An edge caching mechanism and an LRU strategy are introduced to cache hot data and reduce cloud access pressure.

2. The cloud-edge collaborative medical big data management system according to claim 1, characterized in that, Also includes: Medical data quality assessment mechanism: Constructing a data quality evaluation index system: Q=w1・C+w2・A+w3・I+w4・C o w1=0.3, w2=0.25, w3=0.25, w4=0.2, where C represents completeness, A represents accuracy, I represents consistency, and C o For compliance purposes, a Bayesian network is used for quality assessment, and a prior probability distribution is trained using historical data to achieve real-time monitoring of data quality. An anomaly detection engine is deployed, and the isolated forest algorithm is used to identify data anomalies and automatically trigger a manual review process.

3. The cloud-edge collaborative medical big data management system according to claim 1, characterized in that, Also includes: Dynamic update mechanism for medical knowledge graph: Design an incremental learning algorithm to trigger an update when the confidence of new data is >0.8; A knowledge conflict detection rule base is established, knowledge is integrated using DS evidence theory, and knowledge distillation technology is introduced to compress the medical knowledge base into the model.

4. The cloud-edge collaborative medical big data management system according to claim 1, characterized in that, The heterogeneous data fusion module adopts the Transformer architecture and introduces the pre-trained medical model BioBERT. By fine-tuning, it adapts to the medical text understanding task and achieves an F1 score of 0.93 in the medical named entity recognition task. A cross-modal alignment mechanism is designed to achieve semantic alignment of image and text features through contrastive learning.

5. The cloud-edge collaborative medical big data management system according to claim 1, characterized in that, The privacy-preserving computing module is designed with a blockchain-based federated learning framework and uses the PBFT consensus algorithm to achieve trusted sharing of model parameters; it automatically executes data usage agreements through contracts, with a violation penalty coefficient λ=10; and it introduces a secure computing MPC protocol to support cross-institutional data joint analysis while ensuring data privacy.

6. The cloud-edge collaborative medical big data management system according to claim 1, characterized in that, In the edge analysis module, a model compression strategy is designed for resource-constrained edge devices: pruning technology is used to reduce redundant connections, INT8 quantization technology is used to reduce storage requirements, and distillation learning is used to transfer knowledge from the teacher model to the student model. Develop an adaptive inference engine to dynamically adjust model complexity based on device resources.

7. The cloud-edge collaborative medical big data management system according to claim 1, characterized in that, In the cloud-edge collaborative decision-making module, a task scheduling strategy based on reinforcement learning is established. The state space includes device load, network latency, and data priority dimensions, and the action space includes three options: edge processing, cloud processing, and partial offloading. The reward function is designed as: R = r1・(1-D / D) max )+r2・(1-E / E max )-r3・C (r1=5, r2=3, r3=1), where D is the delay, D max The maximum acceptable delay is E, where E is the energy consumption. max C represents the maximum energy consumption, and C represents the cost. A predictive scheduling algorithm is introduced, which uses time series analysis and the ARIMA model to predict network congestion and adjust the task allocation strategy in advance.

8. A method for using a cloud-edge collaborative medical big data management system as described in any one of claims 1-7, characterized in that, include: Data Acquisition and Preprocessing: Deploy data acquisition terminals at the edge of medical institutions to support the acquisition of data such as CT, MRI, ECG, and electronic medical records; ETL tools were used for data cleaning, multiple imputation was used for missing value handling, the IsolationForest algorithm was used for outlier detection, and Z-score normalization was used for standardization to convert the data into the unified FHIR standard format. Deploy edge computing gateways to accelerate local data caching and preprocessing; Edge analysis: The computational model is deployed at the edge. For image data, an attention gating mechanism is added to the U-Net++ network for lesion segmentation. For time series data, TemporalConvolutionalNetwork is used for disease prediction. The AUC is 0.89 in the heart failure prediction task. Develop an edge inference engine to enable real-time inference and support online model updates; Cloud-edge collaborative processing: Establish a task offloading decision model to dynamically determine the processing location based on data type, device status, and network conditions; when K > 0.7, C < 0.5, and B > 100Mbps, the task is offloaded to the cloud; otherwise, it is processed at the edge. Deploy an edge-cloud collaborative orchestration system to support task priority scheduling and elastic resource allocation; Deep cloud analytics: Deploy computing clusters in the cloud and use graph neural networks (GNNs) to analyze medical knowledge graphs and discover disease association rules; combine genomics, proteomics, and metabolomics to build molecular mechanism models of diseases to support medical decision-making; Develop a parallel computing framework to support node knowledge graph analysis; Decision support: Generate personalized medical recommendations based on analysis results; Natural language generation technology is used to transform analysis results into reports for doctors; virtual reality technology is used to visualize the disease development process to assist doctors in formulating treatment plans; and a clinical decision support system (CDSS) is deployed, integrating treatment pathway recommendations and drug interaction warning functions.

9. The medical big data management method based on cloud-edge collaboration according to claim 8, characterized in that, In the data acquisition and preprocessing steps, blockchain nodes are deployed to record the data flow process, and asymmetric encryption RSA-2048 is used to ensure data integrity; a data traceability algorithm is designed to achieve full lifecycle traceability of data through a hash chain; and a data quality traceability system is developed to record the quality indicators of each stage of data acquisition, transmission, and processing, and to support root cause analysis of quality problems.

10. The medical big data management method based on cloud-edge collaboration according to claim 8, characterized in that, In the decision support step, a Clinical Decision Support System (CDSS) is constructed, integrating guideline knowledge, evidence-based medicine (Cochrane systematic reviews), and world data; the Decision Analysis Method (MCDM) is employed, with physician preference weights assigned. , i=1, where i represents the weight of medical criteria and patient wishes. , Let j=1, where j represents individual needs. The utility formula for constructing the solution is: Generate a personalized plan, where fi(x) is the treatment benefit corresponding to criterion i, and gj(x) is the value corresponding to demand j; Deploy the SHAP interpretation module and use SHAP values. Quantitative feature contribution Visualize the risks and benefits to assist doctors in decision-making.

Images