A Distributed Feature Mining Clinical Data Acquisition and Analysis System Based on Federated Learning

Abstract

This invention discloses a distributed feature mining clinical data acquisition and analysis system based on federated learning, relating to the fields of medical big data processing and artificial intelligence. The system includes: a clinical data acquisition module that collects data from various medical nodes according to standards and transmits it to a quality preprocessing module; this module evaluates and processes the data before transmitting it to a federated temporal feature mining module; this module generates a temporal causal feature set and transmits it to a federated learning model training module; the training module generates a global model and transmits it to an application and update module, which deploys and updates the model. This invention achieves efficient integration of data from multiple medical nodes by constructing a distributed system based on federated learning. It improves data reliability and feature accuracy through quality preprocessing and causal decoupling algorithms, protecting data privacy. Simultaneously, it employs a dual-dimensional driving algorithm to dynamically adjust data weights, combined with an adaptive update mechanism, ensuring the model is fair, effective, and continuously adapts to clinical changes, providing stable and reliable predictions.

Description

Technical Field

[0001] This invention relates to the fields of medical big data processing and artificial intelligence technology, specifically a distributed feature mining clinical data acquisition and analysis system based on federated learning. Background Technology

[0002] With the rapid development of medical informatization and artificial intelligence technologies, the collection, integration, and analysis of clinical data have become key links in improving the quality of medical services. Medical institutions accumulate massive amounts of multi-source heterogeneous data during the diagnosis and treatment process, including information on vital signs, disease diagnoses, and treatment records. However, the traditional data silo model makes cross-institutional data sharing difficult, and the data sample size of a single node is limited, making it difficult to support the accurate analysis of complex diseases. At the same time, clinical time-series data has the characteristics of high dimensionality, dynamics, and nonlinearity, and traditional feature extraction methods are difficult to effectively capture the causal relationships and temporal dependencies of disease development. In addition, data privacy protection and compliance requirements are becoming increasingly stringent. How to achieve multi-center collaborative modeling without leaking the original data has become an important challenge for clinical research.

[0003] Traditional clinical data analysis techniques have several shortcomings: First, the problem of data silos is prominent, with each medical institution storing data independently and lacking a secure sharing mechanism, resulting in insufficient data coverage and sample bias. Second, feature engineering relies on human experience, making it difficult to automatically mine strongly correlated features from high-dimensional time-series data, especially neglecting the impact of causal relationships on disease prognosis. Third, centralized modeling requires the aggregation of raw data, posing a risk of privacy leaks, and model updates lag behind changes in clinical practice, leading to a decline in predictive performance. For example, traditional methods typically only use basic statistical features, ignoring the dynamic trends of vital signs and the interaction between treatment procedures, thus limiting the model's ability to explain complex diseases.

[0004] Therefore, we developed a distributed feature mining clinical data acquisition and analysis system based on federated learning. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies by providing a distributed feature mining clinical data acquisition and analysis system based on federated learning. This system achieves secure and efficient data processing through multi-module collaboration. A clinical data quality preprocessing module ensures data reliability, while a federated time-series feature mining module, combined with a causal decoupling algorithm, automatically extracts strongly correlated features, addressing the shortcomings of traditional methods in capturing causal relationships. Simultaneously, through a dual-dimensional driven federated weight algorithm and an encrypted model aggregation mechanism, a dynamically updated global prediction model is generated while protecting data privacy. This technology effectively integrates cross-institutional clinical data resources, providing scientific and accurate decision support for scenarios such as chronic disease management and critical illness prediction.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a distributed feature mining clinical data acquisition and analysis system based on federated learning, comprising:

[0007] Clinical data acquisition module: Connects to the information systems of various medical nodes, collects patients' vital signs, disease diagnosis and treatment records according to preset standards, records data acquisition timestamps synchronously, and transmits the collected data to the clinical data quality preprocessing module;

[0008] Clinical data quality preprocessing module: Receives data transmitted from clinical data acquisition module, uses clinical data quality comprehensive assessment algorithm to assess the quality of data at each medical node, cleans, completes or corrects data that does not meet the standards, and transmits data that meets the standards to federated time series feature mining module;

[0009] The federated time-series feature mining module receives qualified data, uses a time-series feature dynamic mining algorithm to calculate basic features, marks local known causal relationships for each medical node and builds a library, generates causal feature hashes and exchanges them, identifies common, differentiated causal relationships and confounding factors, and uses a causal decoupling algorithm to decouple causal relationships and filter strongly correlated features to form a time-series causal feature set, which is then pushed to the federated learning model training module.

[0010] The federated learning model training module receives a time-series causal feature set, each medical node conducts local model training, dynamically adjusts the data contribution weights of each node through a dual-dimensional driven federated weight algorithm, exchanges model parameters through an encryption mechanism, securely aggregates and generates a global model containing structure and encryption parameters on the central server, and transmits it to the model application and update module.

[0011] Model Application and Update Module: Receives the transmitted global model, deploys the model in clinical scenarios, records model application feedback information, monitors changes in the distribution of newly collected clinical data, and adjusts the global model parameters through a model adaptive update algorithm.

[0012] Furthermore, in the preset clinical data collection standards of the clinical data collection module, the vital signs data collection indicators include heart rate, body temperature, blood oxygen saturation, blood pressure, and respiratory rate; in the vital signs data collection frequency, the collection interval for heart rate and blood oxygen saturation is 5 minutes, and the collection interval for body temperature, blood pressure, and respiratory rate is 1 hour; the disease diagnosis information recording content includes disease name, diagnosis time, diagnosing doctor, laboratory test result number corresponding to the diagnosis basis, and imaging report number.

[0013] Furthermore, in the clinical data quality preprocessing module, the calculation formula for the comprehensive clinical data quality assessment algorithm is as follows: ,in, This is a comprehensive data quality score, with a value range of [0, 1]. This is the theoretical total number of time-series data points that should be collected. It represents the number of missing data points. It represents the number of erroneous data points. It is the total number of data points. This refers to the number of valid data pairs that are logically matched. It is the total data pair quantity. It is the standard deviation of the data collection time interval. It is the average time interval. It is a weighting coefficient, and Adjustments will be made dynamically based on the medical scenario.

[0014] Furthermore, in the federated time-series feature mining module, the calculation formula for the dynamic time-series feature mining algorithm is as follows: ,in yes The temporal characteristic value at time point, This is the current time node for calculating the time series features. yes Real-time clinical data, It is a time index for time-series data, used to iterate through the data at various time points within the sliding window. It's the size of the sliding window. It is the trend slope. It is the window start time, used to define the time range for feature calculation.

[0015] Furthermore, in the federated time-series feature mining module, the specific steps for marking local known causal relationships and constructing a library for each medical node are as follows: From the qualified data received by each medical node, a subset of clinical time-series data related to the target disease is selected. This subset covers vital sign time-series features, treatment operation records, and disease prognosis results, forming a local causal mining dataset. For the local causal mining dataset, association rule mining parameters are configured, including: setting a minimum support threshold of at least 10% for the frequency of causal associations in the dataset and a minimum confidence threshold of at least 80% for the probability of the causal feature leading to the effect event. Based on the configured parameters, association rule mining is performed on the local causal mining dataset, traversing the combinations of time-series features and disease results, and identifying association relationships that satisfy the support and confidence thresholds. The mined causal relationships are organized into structured entries, each containing: a causal relationship description, a support value, a confidence value, the calculation conditions of the time-series features involved in the association, and the time range corresponding to the disease result. A local causal relationship library is constructed based on all entries.

[0016] Furthermore, in the federated time-series feature mining module, the identification of common, differentiated causal relationships and confounding factors is as follows: the federated coordinator counts the causal feature hash values ​​uploaded by each node, defines the causal relationship corresponding to the hash value appearing on three or more nodes as a common causal relationship, and defines the causal relationship appearing on only one or two nodes as a differentiated causal relationship; for differentiated causal relationships, by comparing with the medical common sense database, and combining the data sample size of its source node and the special characteristics of the collection scenario, those that conflict with medical common sense or have logical contradictions are identified as associations affected by confounding factors.

[0017] Furthermore, in the federated time-series feature mining module, the calculation formula for the causal decoupling algorithm is as follows: ,in, It is a feature With results The strength of the true causal relationship These are characteristics found in clinical data. These are results from clinical data. It is a feature With results mutual information, It is a potential source of confusion. It is a set of potential confounding factors, containing all variables that interfere with the association between features and outcomes. It is a feature Potential Confounding Factors Mutual information.

[0018] Furthermore, in the federated learning model training module, the calculation formula for the two-dimensional driven federated weight algorithm is as follows: ,in, These are the weights of the data contributions of medical nodes during the training of the federated model. It is a node data quality score. It is the contribution of node causal features. It is the sum of the causal feature contributions of all medical nodes participating in federated learning.

[0019] Furthermore, in the model application and update module, the global model is a clinical prediction model generated by the central coordination server aggregating encrypted parameters from various medical nodes. It includes two parts: model structure information and model parameters. The model structure information includes the input layer feature dimension, the number and hierarchical structure of hidden layer neurons, and the output layer prediction target type. The model parameters include the connection weight matrix between the input and hidden layers, the bias terms of hidden layer neurons, and the activation function parameters of the output layer. All parameters are dynamically encrypted and decrypted for prediction calculation only during clinical deployment. The clinical prediction model supports receiving preprocessed patient temporal causal feature data and outputting corresponding clinical prediction results. The clinical prediction model comprises an input layer, a feature processing layer, a prediction layer, and an output layer. The input layer receives preprocessed patient temporal causal feature data, including dynamic trends in vital signs, treatment operation correlation features, and disease development causal correlation features. The feature processing layer performs dimensional transformation and importance weighting on the input features, with weighting coefficients determined based on the causal correlation strength of each feature. The prediction layer employs a multi-layer nonlinear transformation structure, mapping the processed features to a probability distribution of clinical outcomes through hierarchical calculations. The output layer outputs specific prediction results, including the prognostic status of the target disease, risk values ​​at key time points, and the core temporal causal feature identifiers used for prediction.

[0020] Furthermore, in the model application and update module, the calculation formula for the model adaptive update algorithm is as follows: ,in, It is the adjustment amount of the global model parameters, used to update the model parameters to adapt to new data. This is the learning rate, which ranges from (0, 1) and is used to control the step size of model updates. This is the start time of the new data. This is the current time point for clinical data. It is a time variable used for iterating over the time interval. arrive At various points in time within the time period, Is the model in The prediction results for the time period, yes Actual clinical outcomes at any given time.

[0021] Compared with existing technologies, this distributed feature mining clinical data acquisition and analysis system based on federated learning has the following advantages:

[0022] I. This invention achieves seamless connection and efficient integration of clinical data among multiple medical nodes by constructing a distributed feature mining system based on federated learning. The system employs a clinical data quality preprocessing module, which uses a comprehensive evaluation algorithm to automatically filter and clean low-quality data, ensuring the accuracy and reliability of the data input to the model. At the same time, the federated temporal feature mining module can dynamically mine temporal features and use a causal decoupling algorithm to identify strongly correlated features, effectively improving the accuracy and depth of feature extraction. This distributed processing method not only accelerates data processing speed but also protects data privacy through the federated learning mechanism, providing more scientific and accurate data support for clinical decision-making.

[0023] Second, this invention uses a dual-dimensional driven federated weight algorithm to dynamically adjust the data contribution weights of each medical node, ensuring the fairness and effectiveness of the global model during training. The model application and update module monitors the distribution changes of newly collected clinical data and uses an adaptive update algorithm to adjust the model parameters in a timely manner, enabling the model to continuously adapt to changes in the clinical environment. This flexible model update mechanism not only improves the model's prediction accuracy but also extends the model's effective lifespan, providing more stable and reliable prediction services for clinical practice and helping to improve the overall quality of medical services.

[0024] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0026] Figure 1 This is a schematic diagram of the workflow of a distributed feature mining clinical data acquisition and analysis system based on federated learning.

[0027] Figure 2 A diagram illustrating the data and model interaction framework of a distributed feature mining clinical data acquisition and analysis system based on federated learning.

[0028] Figure 3 This is a schematic diagram of the entire operation nodes of a distributed feature mining clinical data acquisition and analysis system based on federated learning. Detailed Implementation

[0029] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0030] Example 1:

[0031] A scenario for analyzing clinical data of cardiovascular disease patients in a top-tier hospital.

[0032] The Department of Cardiology at this top-tier hospital has launched a clinical data acquisition module for a distributed feature mining clinical data acquisition and analysis system based on federated learning for hospitalized patients with cardiovascular diseases such as coronary heart disease and heart failure. This module first achieves deep integration with the hospital's electronic medical record system, bedside vital sign monitoring equipment, laboratory information system, and image archiving and communication system, breaking down data barriers between different information systems and realizing efficient integration of multi-source clinical data.

[0033] During data collection, the process strictly adhered to pre-defined clinical data collection standards. For patient vital signs, heart rate, body temperature, blood oxygen saturation, blood pressure, and respiratory rate were precisely collected. Different indicators were continuously recorded at their corresponding frequencies, capturing real-time fluctuations in the patient's physiological state. This provided continuous and dynamic data support for subsequent analysis of changes in the patient's condition. Simultaneously, detailed disease diagnosis information was recorded, including the specific name of the cardiovascular disease, the exact time of diagnosis, the name of the responsible physician, and the corresponding laboratory test result numbers (e.g., myocardial enzyme spectrum, coagulation function test numbers) and imaging report numbers (e.g., coronary angiography, echocardiography report numbers). This ensured the completeness and traceability of diagnostic information. Each piece of collected data was simultaneously timestamped, clearly marking the time node of data generation and ensuring the accuracy of the data's temporal dimension. Subsequently, all collected complete data were transmitted to the clinical data quality preprocessing module, laying a comprehensive and reliable data foundation for subsequent data processing stages. Figure 1 As shown.

[0034] After receiving cardiovascular disease patient data transmitted from the clinical data acquisition module, the clinical data quality preprocessing module immediately starts the comprehensive clinical data quality assessment algorithm. The calculation formula for the comprehensive clinical data quality assessment algorithm is as follows: ,in, This is a comprehensive data quality score, with a value range of [0, 1]. This is the theoretical total number of time-series data points that should be collected. It represents the number of missing data points. It represents the number of erroneous data points. It is the total number of data points. This refers to the number of valid data pairs that are logically matched. It is the total data pair quantity. It is the standard deviation of the data collection time interval. It is the average time interval. It is a weighting coefficient, and The algorithm dynamically adjusts according to medical scenarios, comprehensively verifying data from multiple dimensions such as data integrity, accuracy, logic, and time consistency. It accurately identifies potential issues such as missing data, errors, logical contradictions, and abnormal time records, ensuring that only data that meets the standards can proceed to subsequent stages and avoiding interference from low-quality data with subsequent analysis results.

[0035] After evaluation, the system performs targeted processing on substandard data. For missing data, it supplements the data by combining patients' historical medical data and clinical treatment guidelines for cardiovascular diseases. For erroneous data, it corrects the data deviation based on laboratory test standard ranges, imaging diagnostic standards, and doctors' clinical experience. For non-standard or redundant data, it performs standardized cleaning to unify the data format and remove useless information, improving data standardization. After completing all processing operations, the data is verified again by the comprehensive clinical data quality assessment algorithm to confirm that all data meets the qualified standards. The qualified data is then transmitted to the federated time-series feature mining module to provide high-quality data input for subsequent feature mining work.

[0036] After receiving the target data, the federated time-series feature mining module first applies a dynamic time-series feature mining algorithm. The calculation formula for the dynamic time-series feature mining algorithm is as follows: ,in yes The temporal characteristic value at time point, This is the current time node for calculating the time series features. yes Real-time clinical data, It is a time index for time-series data, used to iterate through the data at various time points within the sliding window. It's the size of the sliding window. It is the trend slope. The window start time is used to define the time range for feature calculation. This algorithm extracts basic temporal features that reflect the changes of data over time by analyzing the patient's vital signs and treatment response data over time, such as the trend of heart rate fluctuations and the pattern of blood pressure changes. It clearly presents the development trend of the patient's condition in the time dimension, providing basic feature support for subsequent causal relationship analysis.

[0037] Next, the hospital, acting as an independent medical node, identifies locally known causal relationships related to cardiovascular diseases and constructs a local causal relationship database. Specifically, it filters a subset of clinical time-series data directly related to cardiovascular diseases such as myocardial infarction and arrhythmia from the qualified data. This subset covers the time-series characteristics of patients' vital signs, treatment operation records (such as the use of anticoagulants and records of cardiac interventional treatment), and disease prognostic results (such as postoperative recovery and the occurrence of complications), forming a highly focused local causal mining dataset. Then, it configures association rule mining parameters and performs association rule mining on the local causal mining dataset based on these parameters. It traverses various combinations of time-series features and disease outcomes, identifies relationships with significant correlations, and organizes them into structured entries containing causal relationship descriptions and calculation conditions of the time-series features involved in the correlation. This constructs a local causal relationship database, providing local data support for subsequent cross-node causal feature interactions.

[0038] Subsequently, causal feature hashes are generated and exchanged with other medical nodes participating in federated learning, enabling the sharing and interaction of causal features among nodes. The federated coordinator statistically analyzes the causal feature hash values ​​uploaded by each node, accurately distinguishing common causal relationships that are prevalent across multiple nodes and differentiated causal relationships that appear only in a few nodes. Simultaneously, by comparing with a medical common sense database and considering the data sample size and specific collection scenarios of the nodes from which differentiated causal relationships originate, associations that conflict with or contradict medical common sense are identified as associations influenced by confounding factors, providing clear targets for subsequent causal decoupling. Finally, a causal decoupling algorithm is used to decouple the aforementioned causal relationships, eliminating confounding factors and selecting features that are truly strongly associated with cardiovascular diseases, forming a highly accurate and targeted time-series causal feature set. The calculation formula for the causal decoupling algorithm is as follows: ,in, It is a feature With results The strength of the true causal relationship These are characteristics found in clinical data. These are results from clinical data. It is a feature With results mutual information, It is a potential source of confusion. It is a set of potential confounding factors, containing all variables that interfere with the association between features and outcomes. It is a feature Potential Confounding Factors The mutual information is pushed to the federated learning model training module to provide high-quality feature input for model training.

[0039] After receiving the temporal causal feature set, the tertiary hospital, as a key node participating in the federated learning model training module, uses the local cardiovascular disease patient temporal causal feature set as training data to conduct local model training. During the training process, a two-dimensional driven federated weight algorithm is adopted. The calculation formula of the two-dimensional driven federated weight algorithm is as follows: ,in, These are the weights of the data contributions of medical nodes during the training of the federated model. It is a node data quality score. It is the contribution of node causal features. It is the sum of the causal feature contributions of all medical nodes participating in federated learning. The algorithm comprehensively considers the local data quality score of the hospital and the contribution of causal features, and dynamically adjusts the data contribution weight of the hospital in the federated model training. This allows nodes with high data quality and high causal feature value to play a more important role in the model training, ensuring that the model training can fully integrate the high-quality data resources of each node and improve the overall performance of the model.

[0040] To strictly protect patient privacy and data security, the hospital encrypts model parameters when exchanging them with other medical nodes. This prevents parameters from being leaked or tampered with during transmission, ensuring data privacy and transmission security. Each medical node transmits the encrypted model parameters to the central server, which securely aggregates these encrypted parameters, effectively integrating the training results from each node to generate a global model containing model structure information and encrypted parameters. This global model incorporates high-quality data features from multiple nodes, possessing stronger generalization ability and prediction accuracy. It is then passed to the model application and update module, preparing it for clinical application.

[0041] After receiving the global model, the model application and update module deploys it to the clinical setting of the cardiology department of the tertiary hospital for predicting the condition of hospitalized cardiovascular patients, such as... Figure 2 As shown, the model can receive preprocessed patient temporal causal feature data and output the patient's prognostic status (such as recovery level, whether there is a risk of disease deterioration), risk values ​​of key time nodes (such as the risk value of complications within 72 hours after surgery), and the core temporal causal feature identifiers of the prediction basis. This provides doctors with accurate and reliable reference information to understand the development trend of the patient's condition and formulate personalized treatment plans, assisting doctors in optimizing treatment decisions and improving the efficiency and quality of diagnosis and treatment.

[0042] During model application, the system records real-time feedback information, including the consistency between predicted and actual clinical outcomes, and physician evaluations of the predicted results. This provides a basis for model performance assessment. Simultaneously, it continuously monitors changes in the distribution of newly collected clinical data from cardiovascular disease patients. When a significant change in data distribution is detected, leading to a decrease in model prediction accuracy, the adaptive update algorithm is immediately activated. The calculation formula for the adaptive update algorithm is as follows: ,in, It is the adjustment amount of the global model parameters, used to update the model parameters to adapt to new data. This is the learning rate, which ranges from (0, 1) and is used to control the step size of model updates. This is the start time of the new data. This is the current time point for clinical data. It is a time variable used for iterating over the time interval. arrive At various points in time within the time period, Is the model in The prediction results for the time period, yes Based on the actual clinical results at any given time, the algorithm makes reasonable adjustments to the global model parameters according to the difference between the newly collected data and the model's prediction results, so that the model can adapt to the dynamic changes in clinical data in a timely manner, maintain good predictive performance, and continuously provide stable and reliable support for the clinical diagnosis and treatment of cardiovascular diseases.

[0043] In summary, in the clinical data analysis scenario of cardiovascular disease patients in a tertiary hospital, the system sequentially operates five modules: the clinical data acquisition module connects to multiple systems and integrates multi-source data, laying the foundation for subsequent analysis; the clinical data quality preprocessing module uses a comprehensive clinical data quality assessment algorithm to screen and optimize data, ensuring high-quality input data; the federated time-series feature mining module uses various algorithms to mine features, construct a causal database, and screen for strongly correlated features; the federated learning model training module uses a two-dimensional driven federated weight algorithm to adjust weights and safely aggregate to generate a global model; and the model application and update module deploys the model and maintains model performance through a model adaptive update algorithm, providing data and decision support for the diagnosis and treatment of cardiovascular diseases throughout the entire process.

[0044] Example 2:

[0045] A scenario for analyzing clinical data of pneumonia patients in the respiratory department of a top-tier hospital.

[0046] A tertiary hospital's respiratory medicine department launched a distributed feature mining clinical data acquisition and analysis system based on federated learning for hospitalized patients with different types of pneumonia, such as bacterial and viral pneumonia. The first step was to run the clinical data acquisition module. This module first completed deep integration with the hospital's electronic medical record system, bedside multi-parameter vital signs monitor, laboratory information management system, and medical imaging system (such as chest CT and chest X-ray storage systems), opening up data flow channels between the various systems and realizing the centralized aggregation of multi-dimensional clinical data of pneumonia patients, avoiding the problem of incomplete analysis caused by data dispersion.

[0047] During data collection, strictly adhered to pre-defined clinical data collection standards. For patient vital signs, heart rate, body temperature, blood oxygen saturation, blood pressure, and respiratory rate were precisely collected, with different indicators continuously recorded at their corresponding frequencies. For example, real-time tracking of patient temperature trends allowed for timely detection of fever fluctuations, providing dynamic data for assessing the effectiveness of inflammation control. Simultaneously, detailed disease diagnosis information was recorded, including the specific type of pneumonia, the exact diagnosis time, the name of the responsible physician, and the corresponding laboratory test result numbers (e.g., sputum culture, complete blood count, C-reactive protein test numbers) and imaging report numbers (e.g., chest CT, chest X-ray report numbers). This ensured a complete and traceable diagnostic chain, facilitating subsequent review of diagnostic logic. Each collected data point was simultaneously timestamped, clearly marking the time of data generation and ensuring data accuracy over time. Subsequently, all collected complete clinical data for pneumonia patients was transmitted to the clinical data quality preprocessing module, providing a comprehensive and coherent data foundation for subsequent stages, such as... Figure 3 As shown.

[0048] After receiving pneumonia patient data transmitted from the clinical data acquisition module, the clinical data quality preprocessing module immediately starts the comprehensive clinical data quality assessment algorithm. The calculation formula for the comprehensive clinical data quality assessment algorithm is as follows: The algorithm comprehensively examines data from the dimensions of data integrity, accuracy, logic, and time consistency. For example, it checks whether there are missing blood oxygen saturation data, whether the body temperature value is within a reasonable physiological range, whether there are logical contradictions in the test results of the same patient at the same time point, and whether the data collection interval conforms to the preset rules. It accurately identifies problems in the data and avoids the reliability of subsequent analysis conclusions due to data defects.

[0049] After evaluation, the system performs targeted processing on substandard data. For missing data, it supplements the data by combining the patient's concurrent medical records, typical data characteristics of patients with similar types of pneumonia, and clinical treatment guidelines in respiratory medicine. For erroneous data, it corrects the data based on laboratory test reference ranges, imaging diagnostic standards, and doctors' clinical experience. For data with non-standard formats or redundancy, it performs standardized cleaning to unify the data format and remove useless information, improving data standardization and readability. After completing all processing operations, the data is verified again by the comprehensive clinical data quality assessment algorithm to confirm that all data meets the qualification standards. The qualified data is then transmitted to the federated time-series feature mining module to ensure that the data entering the next stage is of high quality and provides reliable input for feature mining.

[0050] After receiving the target data, the federated time-series feature mining module first applies a dynamic time-series feature mining algorithm. The calculation formula for the dynamic time-series feature mining algorithm is as follows: This algorithm performs time-series analysis on vital sign data (such as changes in blood oxygen saturation over time and body temperature fluctuation curves) and treatment operation records (such as changes in patient symptoms after antibiotic use) of pneumonia patients, extracting basic time-series features that reflect changes in data over time. For example, it can clearly show the downward trend of body temperature and the recovery pattern of blood oxygen saturation after the use of antibiotics, intuitively demonstrating the manifestation of treatment effects in the time dimension, and laying the foundation for subsequent causal relationship analysis.

[0051] Next, the respiratory medicine department, acting as an independent medical node, identifies locally known causal relationships related to pneumonia and constructs a local causal relationship database. Specifically, it filters a subset of clinical time-series data related to pneumonia onset, treatment effects, and prognosis from the qualified data. This subset covers the time-series characteristics of patients' vital signs, treatment operation records (such as different types of antibiotic usage regimens and oxygen therapy parameter adjustment records), and disease prognosis results, forming a local causal mining dataset focusing on the core of pneumonia diagnosis and treatment. Then, it configures association rule mining parameters and performs association rule mining on the local causal mining dataset based on these parameters, traversing various combinations of time-series features and disease outcomes. For example, it analyzes the association between specific antibiotic use and the time for body temperature to return to normal, and the association between changes in blood oxygen saturation and the occurrence of complications, identifying associations that meet the conditions. These associations are then organized into structured entries containing causal relationship descriptions, support correlation descriptions, confidence correlation descriptions, calculation conditions of the time-series features involved in the association, and the time range corresponding to the disease outcome, constructing a local causal relationship database to provide local data support for cross-node causal feature interaction.

[0052] Next, causal feature hashes are generated and exchanged with other medical nodes participating in federated learning to achieve the sharing and complementarity of pneumonia-related causal features across nodes. The federated coordinator statistically analyzes the causal feature hash values ​​uploaded by each node, defining causal relationships corresponding to hash values ​​appearing on three or more nodes as common causal relationships, such as the widespread association of "body temperature decreases after antibiotic use"; defining causal relationships appearing on only one or two nodes as differentiated causal relationships, such as the association of "pneumonia treatment response in patients in a specific region" unique to a particular node; for differentiated causal relationships, by comparing them with the medical common sense database, and considering the sample size of the data from the source nodes and the specificity of the collection scenario, associations that conflict with medical common sense or are logically contradictory are identified as associations affected by confounding factors, such as the association of "a certain drug and its efficacy" due to insufficient sample size. Finally, a causal decoupling algorithm is used to decouple these causal relationships, eliminating confounding factors and selecting features that are truly strongly associated with the onset, treatment, and prognosis of pneumonia. The calculation formula for the causal decoupling algorithm is as follows: This generates an accurate and clinically valuable set of temporal causal features, which is then pushed to the federated learning model training module to provide high-quality feature input for model training.

[0053] After receiving the temporal causal feature set, the respiratory medicine department of the tertiary hospital used the local pneumonia patient temporal causal feature set as training data to conduct local model training. During the training process, a two-dimensional driven federated weight algorithm was adopted. This algorithm combines the hospital's local data quality score (i.e., the score obtained by the comprehensive clinical data quality assessment algorithm) and causal feature contribution (i.e., the quantity and quality of effective associations in the local causal relationship database) to dynamically adjust the hospital's data contribution weight in the federated model training. The calculation formula of the two-dimensional driven federated weight algorithm is as follows: This allows nodes with high-quality data and significant causal value to occupy more reasonable weights during model training, ensuring that high-quality data resources from each node can be fully integrated into the model training process and improving the model's ability to capture patterns related to pneumonia diagnosis and treatment.

[0054] To strictly protect the privacy and data security of pneumonia patients, the hospital encrypts model parameters when exchanging them with other medical nodes. This prevents the parameters from being leaked or tampered with during transmission, ensuring data privacy and transmission security. Each medical node transmits the encrypted model parameters to the central server, which securely aggregates these encrypted parameters, effectively integrating the training results of each node to generate a global model containing model structure information and encrypted parameters. This global model incorporates high-quality features from multi-node pneumonia diagnosis and treatment data. Compared to models trained on a single node, it can more comprehensively cover different pneumonia diagnosis and treatment scenarios, possessing stronger generalization ability and prediction accuracy. It is then passed to the model application and update module to prepare for clinical application.

[0055] After receiving the global model, the model application and update module deploys it into the clinical setting of the respiratory department of the tertiary hospital for the analysis and prediction of pneumonia patients' conditions. The model can receive preprocessed temporal causal feature data of patients and output specific prediction results, including predictions of the patient's treatment effect (such as the expected efficacy of using a certain type of antibiotic), the prognostic status of whether complications such as respiratory failure will occur, the risk values ​​of key time nodes (such as 72 hours of treatment, before discharge), and the core temporal causal feature identifiers on which the prediction is based (such as the prediction conclusion based on the trend of body temperature change and the recovery of blood oxygen saturation). This provides doctors with accurate references for developing personalized treatment plans, such as helping doctors determine whether it is necessary to adjust the type of antibiotics or whether it is necessary to strengthen oxygen therapy support, assisting in optimizing treatment decisions and improving treatment efficiency and patient treatment outcomes.

[0056] During model application, the system continuously records model application feedback information, including the degree of agreement between predicted results and actual clinical outcomes, and patient responses after doctors adjust treatment plans based on predicted results, providing a basis for evaluating model performance. Simultaneously, it closely monitors changes in the distribution of newly collected clinical data from pneumonia patients. For example, when a novel pneumonia pathogen causes changes in patient symptom characteristics and treatment responses, resulting in significant differences in data distribution, the model's adaptive update algorithm is immediately activated. The calculation formula for the model's adaptive update algorithm is as follows: The algorithm adjusts the global model parameters reasonably based on the difference between newly collected data and model prediction results, enabling the model to adapt to the dynamic changes in clinical data in a timely manner, maintain good predictive performance, and continuously provide stable and reliable support for the diagnosis and treatment of pneumonia patients in the respiratory medicine department.

[0057] In summary, in the clinical data analysis of pneumonia patients in the respiratory department of a tertiary hospital, the various modules of the system worked collaboratively. The clinical data acquisition module broke down system barriers and aggregated multi-dimensional data; the clinical data quality preprocessing module verified and processed the data using a comprehensive clinical data quality assessment algorithm to ensure data reliability; the federated time-series feature mining module extracted features, identified causal relationships, and screened key features through a series of algorithms; the federated learning model training module optimized training using a two-dimensional driven federated weight algorithm to generate a global model with strong generalization ability; and the model application and update module deployed the model and adapted it to data changes using a model adaptive update algorithm, providing accurate and continuous support for pneumonia diagnosis and treatment.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A distributed feature mining clinical data acquisition and analysis system based on federated learning, characterized in that, The system includes: Clinical data acquisition module: Connects to the information systems of various medical nodes, collects patients' vital signs, disease diagnosis and treatment records according to preset standards, records data acquisition timestamps synchronously, and transmits the collected data to the clinical data quality preprocessing module; Clinical data quality preprocessing module: Receives data transmitted from clinical data acquisition module, uses clinical data quality comprehensive assessment algorithm to assess the quality of data at each medical node, cleans, completes or corrects data that does not meet the standards, and transmits data that meets the standards to federated time series feature mining module; Federated Time-Series Feature Mining Module: Receives qualified data and uses a dynamic time-series feature mining algorithm to calculate basic features. The calculation formula for the dynamic time-series feature mining algorithm is as follows: ,in yes The temporal characteristic value at time point, This is the current time node for calculating the time series features. yes Real-time clinical data, It is a time index for time-series data, used to iterate through the data at various time points within the sliding window. It's the size of the sliding window. It is the trend slope. The window start time defines the time range for feature calculation. Simultaneously, local known causal relationships are marked for each medical node and a library is built. Causal feature hashes are generated and exchanged to identify common, differentiated causal relationships and confounding factors. A causal decoupling algorithm is then used to decouple causal relationships and filter strongly correlated features, forming a time-series causal feature set. The calculation formula for the causal decoupling algorithm is as follows: ,in, It is a feature With results The strength of the true causal relationship These are characteristics found in clinical data. These are results from clinical data. It is a feature With results mutual information, It is a potential source of confusion. It is a set of potential confounding factors, containing all variables that interfere with the association between features and outcomes. It is a feature Potential Confounding Factors The mutual information is pushed to the federated learning model training module; The federated learning model training module receives a time-series causal feature set, each medical node conducts local model training, dynamically adjusts the data contribution weights of each node through a dual-dimensional driven federated weight algorithm, exchanges model parameters through an encryption mechanism, securely aggregates and generates a global model containing structure and encryption parameters on the central server, and transmits it to the model application and update module. Model Application and Update Module: Receives the transmitted global model, deploys the model in clinical scenarios, records model application feedback information, monitors changes in the distribution of newly collected clinical data, and adjusts the global model parameters through a model adaptive update algorithm.

2. The distributed feature mining clinical data acquisition and analysis system based on federated learning according to claim 1, characterized in that, The preset clinical data collection standards of the clinical data collection module include vital sign data collection indicators such as heart rate, body temperature, blood oxygen saturation, blood pressure, and respiratory rate; the collection frequency of vital sign data is 5 minutes for heart rate and blood oxygen saturation, and 1 hour for body temperature, blood pressure, and respiratory rate; the disease diagnosis information recording content includes disease name, diagnosis time, diagnosing doctor, laboratory test result number corresponding to the diagnosis basis, and imaging report number.

3. The distributed feature mining clinical data acquisition and analysis system based on federated learning according to claim 1, characterized in that, In the clinical data quality preprocessing module, the calculation formula for the comprehensive clinical data quality assessment algorithm is as follows: ,in, This is a comprehensive data quality score, with a value range of [0, 1]. This is the theoretical total number of time-series data points that should be collected. It represents the number of missing data points. It represents the number of erroneous data points. It is the total number of data points. This refers to the number of valid data pairs that are logically matched. It is the total data pair quantity. It is the standard deviation of the data collection time interval. It is the average time interval. It is a weighting coefficient, and Adjustments will be made dynamically based on the medical scenario.

4. The distributed feature mining clinical data acquisition and analysis system based on federated learning according to claim 1, characterized in that, In the federated time-series feature mining module, the specific steps for marking local known causal relationships and building a library for each medical node are as follows: from the qualified data received by each medical node, a subset of clinical time-series data related to the target disease is selected. The subset covers vital sign time-series features, treatment operation records, and disease prognosis results to form a local causal mining dataset. For the local causal mining dataset, association rule mining parameters are configured, including: setting a minimum support threshold of at least 10% for the frequency of causal associations in the dataset, and a minimum confidence threshold of at least 80% for the probability that the causal feature leads to the occurrence of the effect event; based on the configured parameters, association rule mining is performed on the local causal mining dataset, traversing the combinations of time-series features and disease outcomes, and identifying associations that satisfy the support and confidence thresholds; the mined causal associations are organized into structured entries, each entry containing: a description of the causal relationship, a support value, a confidence value, the calculation conditions of the time-series features involved in the association, and the time range corresponding to the disease outcome, and a local causal relationship library is constructed based on all entries.

5. The distributed feature mining clinical data acquisition and analysis system based on federated learning according to claim 1, characterized in that, In the federated time-series feature mining module, the identification of common, differentiated causal relationships and confounding factors is as follows: The federated coordinator counts the causal feature hash values ​​uploaded by each node, and defines the causal relationship corresponding to the hash value appearing on three or more nodes as a common causal relationship, and the causal relationship appearing on only one or two nodes as a differentiated causal relationship; for differentiated causal relationships, by comparing with the medical common sense database, and combining the data sample size of its source node and the special characteristics of the collection scenario, those that conflict with medical common sense or have logical contradictions are identified as associations affected by confounding factors.

6. The distributed feature mining clinical data acquisition and analysis system based on federated learning according to claim 1, characterized in that, In the federated learning model training module, the calculation formula for the two-dimensional driven federated weight algorithm is as follows: ,in, These are the weights of the data contributions of medical nodes during the training of the federated model. It is a node data quality score. It is the contribution of node causal features. It is the sum of the causal feature contributions of all medical nodes participating in federated learning.

7. The distributed feature mining clinical data acquisition and analysis system based on federated learning according to claim 1, characterized in that, In the model application and update module, the global model is a clinical prediction model generated by the central coordination server after aggregating encrypted parameters from various medical nodes. It includes two parts: model structure information and model parameters. The model structure information includes the input layer feature dimensions, the number and hierarchical structure of hidden layer neurons, and the output layer prediction target type. The model parameters include the connection weight matrix between the input and hidden layers, the bias terms of hidden layer neurons, and the activation function parameters of the output layer. All parameters are dynamically encrypted and decrypted only during clinical deployment for prediction calculations. The clinical prediction model supports receiving preprocessed patient temporal causal feature data and outputting corresponding clinical prediction results. The clinical prediction model comprises an input layer, a feature processing layer, a prediction layer, and an output layer. The input layer receives preprocessed patient temporal causal feature data, including dynamic trends in vital signs, treatment-related features, and causal correlation features in disease development. The feature processing layer performs dimensional transformation and importance weighting on the input features, with weighting coefficients determined based on the causal correlation strength of each feature. The prediction layer employs a multi-layer nonlinear transformation structure, mapping the processed features to a probability distribution of clinical outcomes through hierarchical calculations. The output layer outputs specific prediction results, including the prognostic status of the target disease, risk values ​​at key time points, and the core temporal causal feature identifiers used for prediction.

8. The distributed feature mining clinical data acquisition and analysis system based on federated learning according to claim 1, characterized in that, The calculation formula for the model adaptive update algorithm in the model application and update module is as follows: ,in, It is the adjustment amount of the global model parameters, used to update the model parameters to adapt to new data. This is the learning rate, which ranges from (0, 1) and is used to control the step size of model updates. This is the start time of the new data. This is the current time point for clinical data. It is a time variable used for iterating over the time interval. arrive At various points in time within the time period, Is the model in The prediction results at that time yes Actual clinical outcomes at any given time.

Images