A multi-source data-oriented medical hidden danger analysis system

Abstract

This invention discloses a medical hazard analysis system for multi-source data, belonging to the field of medical data analysis technology. It includes a storage module, an extraction module, and an analysis module. The system segments and stores multi-source data, and based on a Master-Slaves cluster architecture, it calls the segmented multi-source data for feature extraction to obtain feature data. Based on a preset medical safety rule base, it matches the feature data with the judgment conditions in the rule base, performing medical hazard feature matching and anomaly judgment, and finally outputting the hazard analysis results. By designing differentiated data types and collection frequencies, it optimizes storage and efficiency. Relying on the Master-Slaves cluster and Spark, it extracts features, mines the value of multi-source data, and shortens the processing time to adapt to medical needs. Multi-dimensional verification avoids false positives and false negatives, accurately outputting anomalies, reducing clinical false alarms, and minimizing redundant checks by medical staff, providing reliable support for medical safety identification and treatment decisions.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and in particular to a medical risk analysis system for multi-source data. Background Technology

[0002] In recent years, with the deepening of medical informatization construction, hospital full-service systems and medical hardware equipment have continuously generated massive amounts of multi-source data, including unstructured text, semi-structured images and structured numerical data. These data contain key clues about potential medical safety hazards. Mining data value through technical means to achieve early identification of potential hazards and reduce the incidence of adverse medical events has become a core requirement for ensuring patient safety and improving the level of medical quality management, and it is also an important research direction in the current medical and health field.

[0003] However, existing medical data analysis technologies have significant shortcomings: First, data storage lacks differentiated design and is not classified and processed according to data type and collection frequency, resulting in high latency in high-frequency real-time data access, making it difficult to meet real-time analysis needs, and low efficiency in retrieving low-frequency data. Second, feature extraction capabilities are insufficient, failing to efficiently convert unstructured medical records into structured labels, and the quantitative feature mining of lesions in medical images is inadequate. Furthermore, there is a lack of systematic analysis schemes for the dynamic trends of high-frequency data. Third, the rules for hazard judgment are simplistic, relying heavily on single threshold judgments without combining multi-dimensional features such as parameter fluctuation amplitude and duration, which easily leads to misjudgments or omissions. Summary of the Invention

[0004] The technical problem solved by this invention is that existing medical data analysis technologies have obvious shortcomings: First, data storage lacks differentiated design and is not classified and processed according to data type and acquisition frequency, resulting in high latency in high-frequency real-time data access, making it difficult to meet real-time analysis needs, and low efficiency in retrieving low-frequency data. Second, feature extraction capabilities are insufficient, making it impossible to efficiently convert unstructured medical records into structured labels, and the quantitative feature mining of lesions in medical images is insufficient. Furthermore, there is a lack of systematic analysis schemes for the dynamic trends of high-frequency data. Third, the rules for judging potential risks are simplistic, relying heavily on single threshold judgments without combining multi-dimensional features such as parameter fluctuation amplitude and duration, which easily leads to misjudgments or omissions.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a medical risk analysis system for multi-source data, comprising a storage module, an extraction module, and an analysis module;

[0006] The storage module stores multi-source data in segments;

[0007] The extraction module is based on a Master-Slaves cluster architecture and uses multi-source data stored in the storage module in shards to extract features and obtain feature data.

[0008] The analysis module, based on a preset medical safety rule base, matches the feature data with the judgment conditions in the rule base, performs feature matching and anomaly judgment of medical risks, and finally outputs the risk analysis results.

[0009] As a preferred embodiment of the medical risk analysis system for multi-source data described in this invention, the multi-source data is collected through the hospital's full-service system, and the multi-source data includes unstructured data, structured data, and semi-structured data.

[0010] The multi-source data is first divided according to type to obtain image data corresponding to the semi-structured data, text data corresponding to the unstructured data, and numerical data corresponding to the structured data;

[0011] The multi-source data is further divided according to the acquisition frequency to obtain high-frequency data and low-frequency data.

[0012] The first partition and the second partition are processed in parallel.

[0013] As a preferred embodiment of the medical risk analysis system for multi-source data described in this invention, the image data and text data are both low-frequency data, and the text data includes electronic medical record text.

[0014] The numerical data is divided into high-frequency data and low-frequency data;

[0015] The numerical data in the high-frequency data includes heart rate data collected by medical real-time monitoring equipment and real-time operating parameters of the ventilator;

[0016] The real-time operating parameters include airway pressure data and tidal volume data, and the numerical data in the low-frequency data are the test indicators in the test report.

[0017] The high-frequency data is dynamic data with a sampling frequency of greater than or equal to 1 time / second, and the dynamic data includes the real-time operating parameters of the ventilator;

[0018] The low-frequency data is static data or batch data with a sampling frequency of less than or equal to 1 time / hour. The static data includes electronic medical record text, and the batch data includes test reports.

[0019] As a preferred embodiment of the medical risk analysis system for multi-source data described in this invention, the high-frequency data is preferentially transmitted to a Redis temporary cache cluster for storage, and a cache validity period is set, and the high-frequency data access latency is less than or equal to 1 second.

[0020] The low-frequency data is transmitted to the HDFS distributed file system for fragmented storage.

[0021] Among them, the text data in the low-frequency data is segmented using the patient's unique identifier and the text data generation timestamp as the segment identifier;

[0022] The numerical data in the low-frequency data is segmented according to the inspection item code and the collection date, and the image data in the low-frequency data is segmented according to the image equipment number and the inspection order number.

[0023] As a preferred embodiment of the medical risk analysis system for multi-source data described in this invention, the storage module supports the HL7 FHIR protocol, DICOM 3.0 protocol and MQTT protocol, respectively connecting to the hospital business system and medical hardware equipment;

[0024] The hospital's business system includes an electronic medical record system, a laboratory information system, and a medical imaging system; the medical hardware includes imaging equipment and real-time medical monitoring equipment.

[0025] The HL7 FHIR protocol is used to interface with the electronic medical record system and the laboratory information system to jointly parse and obtain unstructured and structured data;

[0026] The DICOM 3.0 protocol is used to interface with imaging equipment and medical imaging systems to receive and store semi-structured data;

[0027] The MQTT protocol is used to connect to medical real-time monitoring equipment and collect high-frequency real-time data.

[0028] As a preferred embodiment of the medical risk analysis system for multi-source data described in this invention, the extraction module is based on a Master-Slaves cluster architecture and uses the multi-source data stored in the storage module in segments to perform feature extraction to obtain feature data, specifically including:

[0029] The extraction module retrieves the text data of low-frequency data stored in the HDFS distributed file system by patient unique identifier and generation timestamp, and performs keyword extraction and template matching in parallel based on Spark RDD to generate structured labels of keys and values ​​as the first feature data.

[0030] The extraction module retrieves the test indicators of the low-frequency data stored in the HDFS distributed file system according to the test item code and the fragment identifier of the collection date, and processes them in parallel to form a standardized test dataset, which serves as the second feature data.

[0031] The extraction module retrieves the complete image sequence of a single inspection from the low-frequency data stored in the HDFS distributed file system in segments, according to the segment identifier of the image device number and the inspection order number.

[0032] The complete imaging sequence includes CT tomographic images and MRI sequence images;

[0033] The complete image sequence is allocated as the smallest segmentation unit to the Slaves nodes of the Master-Slaves cluster. Each Slaves node performs lesion region segmentation in parallel based on preset medical image segmentation rules to obtain the segmented lesion region.

[0034] The medical image segmentation rules are set based on the grayscale threshold and edge gradient features of lesions in the clinical image diagnostic standards, and quantitative features with clinical risk analysis significance are extracted from the segmented lesion areas.

[0035] The quantitative features include the maximum diameter of the lesion, the volume of the lesion, the average density of the lesion, the irregularity of the lesion boundary, the shortest distance between the lesion and adjacent blood vessels, and the shortest distance between the lesion and adjacent organs. These features generate an examination report number and a key-value pair of quantitative features of the lesion, which serve as the third feature data.

[0036] The extraction module extracts the numerical data of high-frequency data temporarily cached in the Redis temporary cache cluster. Based on Spark Streaming, it retrieves the real-time running parameters of the numerical data in real time and performs dynamic trend analysis in parallel to obtain dynamic trend features, which are used as the fourth feature data.

[0037] The feature data includes first feature data, second feature data, third feature data, and fourth feature data.

[0038] As a preferred embodiment of the medical hazard analysis system for multi-source data described in this invention, the dynamic trend analysis is a trend anomaly determination bound to a medical safety rule base, specifically including:

[0039] A medical safety rule base is constructed, which includes the determination conditions for heart rate trends and the determination conditions for real-time operating parameter trends of ventilators;

[0040] The criteria for determining the heart rate trend are as follows: for adults in a non-sedated state, if the increase in heart rate data at the first time threshold is greater than or equal to the first count threshold and the absolute value is greater than the second count threshold, or the decrease is greater than or equal to the third count threshold and the absolute value is less than the fourth count threshold, then it is determined to be an abnormal trend.

[0041] The criteria for determining the real-time operating parameter trend of the ventilator are as follows: if the airway pressure data of the ventilator is continuously greater than the second time threshold and the tidal volume data decreases simultaneously by a factor greater than or equal to the first proportional threshold, then it is determined to be an abnormal trend of ventilation resistance.

[0042] As a preferred embodiment of the medical risk analysis system for multi-source data described in this invention, the trend characteristics of the real-time operating parameters are calculated in real time, and the trend characteristics include the rate of change, duration, and fluctuation amplitude.

[0043] The rate of change is calculated based on the parameter type of the high-frequency data, which includes heart rate data, airway pressure data, and tidal volume data.

[0044] The heart rate data is periodized with a third time threshold, and the airway pressure data and tidal volume data are periodized with a fourth time threshold. The difference between the final value and the initial value of the corresponding parameter in the high-frequency data in each period is calculated, and then the difference is divided by the initial value to obtain the parameter change rate of the corresponding period.

[0045] The duration is determined by counting the number of consecutive samples that meet the corresponding rule threshold conditions for the real-time operating parameters, based on a sampling frequency of 1 time / second for the real-time operating parameters, and multiplying the number of samples by a 1-second sampling interval.

[0046] As a preferred embodiment of the medical risk analysis system for multi-source data described in this invention, the rule threshold conditions include threshold conditions for heart rate data and threshold conditions for real-time operating parameters of the ventilator.

[0047] The threshold condition for the heart rate data is that the increase in amplitude within the first time threshold is greater than or equal to the first count threshold or the decrease is greater than or equal to the third count threshold, and the absolute value exceeds the first count interval threshold.

[0048] The threshold conditions for the real-time operating parameters of the ventilator are that the airway pressure data is greater than a first numerical threshold and the tidal volume decrease is greater than or equal to a first proportional threshold.

[0049] The fluctuation amplitude is calculated by taking n consecutive 1-second sampling periods as a statistical window and calculating the difference between the maximum and minimum values ​​of the corresponding parameters in the high-frequency data within the window.

[0050] As a preferred embodiment of the medical hazard analysis system for multi-source data described in this invention, the system involves: matching the fourth feature data in the feature data with the judgment conditions in the medical safety rule base, performing feature matching and anomaly judgment for medical hazards, and finally outputting the hazard analysis results, specifically including:

[0051] Based on real-time operating parameters, the feature verification conditions of the corresponding target judgment conditions are retrieved from the medical safety rule base;

[0052] The target determination criteria include the determination criteria for heart rate trends and the determination criteria for real-time operating parameter trends of the ventilator.

[0053] The feature verification conditions include:

[0054] For heart rate data, the rate of change of the heart rate data meets the criteria for determining the heart rate trend, namely, the amplitude change within the first time threshold, the absolute value exceeding the first number interval threshold, and the duration greater than or equal to the fifth time threshold.

[0055] Regarding the real-time operating parameters of the ventilator, the rate of change of the airway pressure data of the real-time operating parameters of the ventilator meets the judgment conditions of the real-time operating parameter trend being greater than the first numerical threshold, the rate of change of the tidal volume data meets the judgment conditions of the decrease being greater than or equal to the first proportional threshold, the duration being greater than or equal to the sixth time threshold, and the fluctuation amplitude being less than or equal to the second proportional threshold.

[0056] If all the trend features satisfy the feature verification conditions of the corresponding target determination conditions, then the anomaly determination result is output.

[0057] If any trend feature does not meet the feature verification condition of the corresponding target determination condition, it is determined to be a normal trend;

[0058] The abnormality determination results, as part of the hidden danger analysis, include abnormal trends in heart rate and abnormal trends in ventilation resistance.

[0059] The beneficial effects of this invention are as follows: By differentiating data types and collection frequencies, real-time performance and retrieval efficiency are significantly improved, and storage resource allocation is optimized. Redis is used for lightweight storage of high-frequency data, while HDFS is used for large-capacity storage of low-frequency data, avoiding resource waste. In feature extraction, relying on Master-Slaves clusters and Spark technology, the value of multi-source data is fully explored, and parallel processing shortens the feature extraction time, adapting to the efficiency requirements of medical scenarios. In terms of hazard identification, multi-dimensional verification is combined to avoid false negatives and omissions due to a single threshold, which not only accurately outputs abnormal results but also reduces the clinical false alarm rate and reduces redundant verification work for medical staff. Ultimately, it provides reliable technical support for the identification of medical safety hazards and subsequent diagnosis and treatment decisions. Attached Figure Description

[0060] Figure 1 This is a basic flowchart of a medical risk analysis system for multi-source data, provided as an embodiment of the present invention. Detailed Implementation

[0061] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0062] Example, refer to Figure 1As an embodiment of the present invention, a medical risk analysis system for multi-source data is provided, including a storage module, an extraction module and an analysis module;

[0063] The storage module performs fragmented storage of multi-source data;

[0064] The extraction module is based on a Master-Slaves cluster architecture and uses multi-source data stored in the storage module in shards to extract features and obtain feature data.

[0065] The analysis module, based on a pre-set medical safety rule base, matches feature data with the judgment conditions in the rule base, performs feature matching and anomaly judgment of medical hazards, and finally outputs the hazard analysis results.

[0066] In one embodiment, the storage module uses a sharded storage method to classify and manage multi-source data, including hospital electronic medical records, real-time medical monitoring data, and medical images, ensuring that different types of data can be accurately located and retrieved. The extraction module, based on a Master-Slaves cluster architecture, relies on the collaborative work of master and slave nodes within the cluster to call the sharded multi-source data in the storage module to perform feature extraction, forming standardized feature data. This cluster architecture can significantly improve the parallel processing efficiency of multi-source data and shorten the feature extraction time. The analysis module pre-builds a rule base including medical safety judgment logic (such as the judgment condition of continuous excessive airway pressure and synchronous decrease in tidal volume of ventilator), matches the extracted feature data with the judgment conditions in the rule base, completes the feature matching and anomaly judgment of medical risks, and outputs the risk analysis results. Through the collaborative operation of the three modules, the system achieves efficient storage of multi-source data, rapid feature extraction, and accurate risk judgment, providing reliable technical support for the timely identification and early warning of medical safety risks.

[0067] Multi-source data is collected through the hospital's full-service system and includes unstructured data, structured data, and semi-structured data.

[0068] The multi-source data is first divided according to its type, resulting in image data corresponding to semi-structured data, text data corresponding to unstructured data, and numerical data corresponding to structured data.

[0069] The multi-source data is further divided according to the acquisition frequency to obtain high-frequency data and low-frequency data.

[0070] The first and second partitions are processed in parallel.

[0071] Both image data and text data are low-frequency data, with text data including electronic medical record text;

[0072] Numerical data is divided into high-frequency data and low-frequency data;

[0073] Numerical data in high-frequency data includes heart rate data collected by medical real-time monitoring equipment and real-time operating parameters of ventilators;

[0074] Real-time operating parameters include airway pressure data and tidal volume data. The numerical data in the low-frequency data are the test indicators in the test report.

[0075] High-frequency data refers to dynamic data with a sampling frequency of 1 time / second or higher. Dynamic data includes the real-time operating parameters of the ventilator.

[0076] Low-frequency data refers to static or batch data with a sampling frequency of less than or equal to 1 time per hour. Static data includes electronic medical record texts, while batch data includes test reports.

[0077] In one embodiment, a dual parallel partitioning mechanism is implemented for multi-source data. The first partitioning is based on data type, dividing the data into semi-structured image data, unstructured text data (such as electronic medical record text), and structured numerical data. The second partitioning is based on acquisition frequency, using the sampling frequency as the criterion to divide the data into high-frequency data (dynamic data with a sampling frequency of ≥1 time / second, such as heart rate data collected by medical real-time monitoring equipment, and real-time operating parameters of airway pressure and tidal volume of ventilators) and low-frequency data (static data or batch data with a sampling frequency of ≤1 time / hour, such as electronic medical record text and batch data such as test indicators in test reports). The two partitioning processes are processed synchronously and in parallel. Image data and text data are both classified as low-frequency data, while numerical data is classified as high-frequency or low-frequency based on the acquisition frequency. Through this dual parallel partitioning, the attribute characteristics of different data can be quickly identified, providing an accurate basis for subsequent targeted storage and processing, effectively improving data classification efficiency, and ensuring that the data processing link is compatible with the data characteristics.

[0078] High-frequency data is preferentially transmitted to the Redis temporary cache cluster for storage, and the cache validity period is set, with the high-frequency data access latency being less than or equal to 1 second;

[0079] Low-frequency data is transferred to the HDFS distributed file system for fragmented storage.

[0080] Among them, the text data in the low-frequency data is segmented using the patient's unique identifier and the timestamp of the text data generation;

[0081] Numerical data in low-frequency data are segmented by inspection item code and collection date, while image data in low-frequency data are segmented by image equipment number and inspection order number.

[0082] In one embodiment, differentiated storage is implemented for the divided high-frequency and low-frequency data. High-frequency data (such as heart rate data collected by real-time medical monitoring equipment, airway pressure data and tidal volume data from ventilators) is preferentially transmitted to a Redis temporary cache cluster for storage. To avoid data redundancy and resource consumption, a cache validity period is set, and the high-frequency data access latency is strictly controlled to be less than or equal to 1 second to meet the need for rapid data retrieval in real-time monitoring scenarios. Low-frequency data (such as electronic medical record text, test indicators in laboratory reports, and CT / MRI image data) is transmitted to the HDFS distributed file system. The system implements sharded storage, with different types of low-frequency data using dedicated shard identifiers. Low-frequency text data is identified by the patient's unique identifier plus the text data generation timestamp; low-frequency numerical data is identified by the test item code plus the collection date; and low-frequency image data is identified by the imaging device number plus the examination order number. This storage solution adapts Redis to the low-latency characteristics of high-frequency data and HDFS to the large-capacity characteristics of low-frequency data. Combined with dedicated shard identifiers, it ensures both the real-time access efficiency of high-frequency data and the orderly management and accurate positioning of low-frequency data, laying an efficient foundation for data retrieval in the subsequent feature extraction stage.

[0083] The storage module supports HL7 FHIR protocol, DICOM 3.0 protocol and MQTT protocol, respectively connecting to hospital business systems and medical hardware devices;

[0084] The hospital's operational systems include electronic medical record systems, laboratory information systems, and medical imaging systems; its medical hardware includes imaging equipment and real-time medical monitoring equipment.

[0085] The HL7 FHIR protocol is used to interface with the electronic medical record system and the laboratory information system to jointly parse and obtain unstructured and structured data;

[0086] It uses the DICOM 3.0 protocol to interface with imaging equipment and medical imaging systems, receiving and storing semi-structured data;

[0087] The system uses the MQTT protocol to connect to medical real-time monitoring equipment and collects high-frequency real-time data.

[0088] In one embodiment, the storage module supports the HL7 FHIR protocol, DICOM 3.0 protocol, and MQTT protocol to achieve comprehensive integration with hospital business systems and medical hardware devices. The hospital business systems include electronic medical record systems, laboratory information systems, and medical imaging systems. The medical hardware devices include imaging equipment (such as CT scanners and MRI machines) and real-time medical monitoring equipment (such as heart rate monitors and ventilators). Specifically, the integration logic involves using the HL7 FHIR protocol to interface with the electronic medical record system and the laboratory information system, obtaining unstructured data (such as electronic medical record text) and structured data (such as test indicators in test reports) through protocol parsing, and using DICOM... The 3.0 protocol interfaces with imaging equipment and medical imaging systems, receiving and storing semi-structured image data (such as CT tomographic images and MRI sequence images). It uses the MQTT protocol to interface with real-time medical monitoring equipment, collecting high-frequency real-time data (such as heart rate data, ventilator airway pressure data, and tidal volume data) with a sampling frequency of greater than or equal to 1 time per second. This multi-protocol interface solution can achieve coverage of data sources in the hospital's full range of scenarios, ensuring compatibility and accuracy of different types of data during transmission and parsing, while meeting the high-efficiency acquisition requirements of high-frequency real-time data, providing complete and reliable raw data support for subsequent data segmentation and storage.

[0089] The extraction module, based on a Master-Slaves cluster architecture, uses multi-source data stored in shards within the storage module to extract features, resulting in feature data, specifically including:

[0090] The extraction module retrieves the text data of low-frequency data stored in the HDFS distributed file system by patient unique identifier and generation timestamp, and performs keyword extraction and template matching in parallel based on Spark RDD to generate structured labels of keys and values ​​as the first feature data.

[0091] The extraction module extracts numerical data of low-frequency data stored in fragmented HDFS distributed file system, according to the fragment identifier of the test item code and the collection date, retrieves the test indicators of the numerical data and processes them in parallel to form a standardized test dataset, which serves as the second feature data.

[0092] The extraction module retrieves the complete image sequence of a single inspection from the low-frequency data stored in the HDFS distributed file system by the image device number and inspection order number.

[0093] The complete imaging sequence includes CT tomographic images and MRI sequence images;

[0094] The complete image sequence is allocated as the smallest segmentation unit to the Slaves nodes of the Master-Slaves cluster. Each Slaves node performs lesion region segmentation in parallel based on preset medical image segmentation rules to obtain the segmented lesion region.

[0095] The medical image segmentation rules are set based on the grayscale threshold and edge gradient features of lesions in the clinical imaging diagnostic standards. At the same time, quantitative features with clinical risk analysis significance are extracted from the segmented lesion areas.

[0096] The quantitative features include the maximum diameter of the lesion, the volume of the lesion, the mean density of the lesion, the irregularity of the lesion boundary, the shortest distance between the lesion and adjacent blood vessels, and the shortest distance between the lesion and adjacent organs. These features generate examination report number and lesion quantitative feature key-value pairs as the third feature data.

[0097] The extraction module extracts numerical data of high-frequency data from the Redis temporary cache cluster's temporary cache. Based on Spark Streaming, it retrieves real-time running parameters of the numerical data in real time, performs dynamic trend analysis in parallel, and obtains dynamic trend features, which serve as the fourth feature data.

[0098] The feature data includes first feature data, second feature data, third feature data, and fourth feature data.

[0099] In one embodiment, the extraction module, based on a Master-Slaves cluster architecture, calls multi-source data stored in shards in the storage module to perform feature extraction and generate feature data. The specific implementation is as follows:

[0100] For low-frequency text data (such as electronic medical record text) stored in the HDFS distributed file system, data is retrieved by using the patient's unique identifier plus the generation timestamp of the fragment identifier, relying on Spark. RDD performs keyword extraction and template matching in parallel to generate key-value structured labels, which serve as the first feature data. For low-frequency numerical data (such as test report indicators) stored in fragmented HDFS, data is retrieved by fragment identifier (test item code plus collection date), and processed in parallel to form a standardized test dataset, serving as the second feature data. For low-frequency image data (such as CT scan images and MRI sequences) stored in fragmented HDFS, the complete image sequence of a single examination is retrieved by fragment identifier (image equipment number plus examination order number), and this is used as the smallest fragment unit and allocated to the Slaves nodes of the cluster. Each node performs lesion region segmentation in parallel based on medical image segmentation rules set by grayscale thresholds and edge gradient features of lesions in clinical imaging diagnostic standards. From the segmented regions, quantitative features such as the maximum diameter, volume, mean density, boundary irregularity, and shortest distance to adjacent blood vessels / organs of the lesion are extracted, generating key-value pairs between the examination order number and the lesion's quantitative features, serving as the third feature data. For high-frequency numerical data (such as heart rate data and real-time ventilator operating parameters) in the Redis temporary cache cluster, Spark is used to perform segmentation. Streaming retrieves data in real time and performs dynamic trend analysis in parallel to obtain dynamic trend features, which serve as the fourth feature data. The first, second, third, and fourth feature data together constitute the feature data output by the extraction module. Through the parallel processing capabilities of the Master-Slaves cluster and the targeted extraction logic, the efficiency of feature extraction from multi-source data is ensured, and the features are accurately adapted to the needs of medical risk analysis.

[0101] Dynamic trend analysis is the determination of trend anomalies bound to the medical safety rule base, specifically including:

[0102] Construct a medical safety rule base, which includes the criteria for determining heart rate trends and the criteria for determining the trends of real-time operating parameters of ventilators;

[0103] The criteria for determining heart rate trends are for adults in a non-sedated state. If the increase in heart rate data at the first time threshold is greater than or equal to the first count threshold and the absolute value is greater than the second count threshold, or the decrease is greater than or equal to the third count threshold and the absolute value is less than the fourth count threshold, then it is determined to be an abnormal trend.

[0104] The criteria for determining the real-time operating parameter trend of the ventilator are as follows: if the airway pressure data of the ventilator is continuously greater than the second time threshold and the tidal volume data decreases simultaneously by a factor greater than or equal to the first proportional threshold, then it is determined to be an abnormal trend of ventilation resistance.

[0105] In one embodiment, the threshold for determining heart rate trends is set as follows: the first time threshold is set to 10 minutes, based on the physiological regulation of heart rate in adults under non-sedation. Simultaneously, the first threshold for the number of heart rate increases is set to 30 beats / min, and the third threshold for the number of heart rate decreases is set to 20 beats / min. These thresholds are based on the normal adult heart rate range (clinical consensus 60-100 beats / min). This range exceeds the natural fluctuation range of heart rate under physiological conditions. Such a change suggests a potential increase in circulatory system load (e.g., dehydration, infection) or functional inhibition (e.g., drug effects, cardiac conduction abnormalities). Furthermore, the second threshold for the absolute heart rate is set to 120 beats / min, and the fourth threshold is set to 50 beats / min. These thresholds correspond to the clinical diagnostic thresholds for adult tachycardia and bradycardia in internal medicine.

[0106] Regarding the threshold values ​​for judging real-time operating parameters of the ventilator, the first numerical threshold for airway pressure data is set at 30 cmH2O. This threshold is set according to the clinical safety guidelines for ventilator use and is a key safety upper limit to avoid barotrauma (such as overinflation of lung tissue and alveolar rupture) during invasive ventilation in adults. Exceeding this value will significantly increase the risk of pulmonary complications. At the same time, the first percentage threshold for the tidal volume decrease is set at 20%. This percentage is set based on the ventilation efficiency assessment standards in mechanical ventilation. If the tidal volume decreases to this percentage, it means that the gas exchange efficiency is significantly reduced, which may indicate airway obstruction and decreased lung compliance. It is a core indicator for clinically judging abnormal ventilation function. In addition, the duration of airway pressure exceeding the threshold (the second time threshold) is set at 30 seconds. This duration is set in conjunction with the fluctuation characteristics of ventilator parameters and can effectively filter out instantaneous parameter fluctuations caused by patient turning over and temporary compression of the tubing, avoiding misjudgment. Clinical monitoring data has verified that it can control the false alarm rate within a clinically acceptable range while ensuring the sensitivity of abnormality identification.

[0107] Calculate the trend characteristics of real-time operating parameters, including the rate of change, duration, and fluctuation amplitude.

[0108] The rate of change is calculated based on the parameter type of the high-frequency data, which includes heart rate data, airway pressure data, and tidal volume data.

[0109] Heart rate data is periodized with the third time threshold, while airway pressure and tidal volume data are periodized with the fourth time threshold. The difference between the final value and the initial value of the corresponding parameter in the high-frequency data within each period is calculated, and then the difference is divided by the initial value to obtain the parameter change rate of the corresponding period.

[0110] Duration is determined by counting the number of consecutive samples that meet the corresponding rule threshold conditions for the real-time operating parameters, based on a sampling frequency of 1 time / second for the real-time operating parameters. The number of samples is then multiplied by the 1-second sampling interval to obtain the duration.

[0111] The rule threshold conditions include threshold conditions for heart rate data and threshold conditions for real-time operating parameters of the ventilator;

[0112] The threshold condition for heart rate data is that the increase in amplitude within the first time threshold is greater than or equal to the first count threshold or the decrease is greater than or equal to the third count threshold, and the absolute value exceeds the first count interval threshold.

[0113] The threshold conditions for the real-time operating parameters of the ventilator are that the airway pressure data is greater than the first numerical threshold and the tidal volume decrease is greater than or equal to the first proportional threshold.

[0114] The fluctuation amplitude is calculated by taking n consecutive 1-second sampling periods as a statistical window and calculating the difference between the maximum and minimum values ​​of the corresponding parameters in the high-frequency data within the window.

[0115] In one embodiment, the trend characteristics of real-time operating parameters are calculated in real time. These trend characteristics include rate of change, duration, and fluctuation amplitude. The rate of change calculation is based on the parameter types of high-frequency data (heart rate data, airway pressure data, and tidal volume data), with corresponding periods set according to the data type. Heart rate data is calculated in 10-minute periods (the third time threshold is set to 10 minutes, based on the physiological regulation of heart rate in adults under non-sedation conditions, to cover the physiological fluctuation window and eliminate transient interference). Airway pressure and tidal volume data are calculated in 30-second periods (the fourth time threshold is set to 30 seconds, based on the requirement for rapid response to ventilation abnormalities in ventilator parameters, matching the immediacy requirements of safety monitoring). The rate of change for each period is obtained by calculating the difference between the final value and the initial value of the corresponding parameter within each period, and then dividing the difference by the initial value. The duration is calculated based on a sampling frequency of 1 time / second for the real-time operating parameters. The number of consecutive samples that satisfy the corresponding rule threshold conditions is counted, and the duration is obtained by multiplying the number of samples by the 1-second sampling interval. The criteria include threshold conditions for heart rate data (a change in amplitude within 10 minutes of the first time threshold of 30 bpm or greater than or equal to the first count threshold of 30 bpm or a decrease in amplitude within 20 bpm or greater than or equal to the third count threshold of 20 bpm, with the absolute value exceeding the first count interval threshold of 50-120 bpm) and threshold conditions for real-time ventilator operating parameters (airway pressure data greater than the first numerical threshold of 30 cmH2O and tidal volume decrease greater than or equal to the first proportional threshold of 20%). The fluctuation amplitude is calculated using 5 consecutive 1-second sampling cycles (n is set to 5 based on a 1-second sampling frequency, balancing noise filtering and real fluctuation capture through 5 consecutive cycles to avoid calculation deviations caused by instantaneous interference in a single cycle, all in line with parameter characteristics and clinical monitoring needs) as a statistical window. The difference between the maximum and minimum values ​​of the corresponding parameters within the calculation window is used to determine the amplitude. This implementation method sets calculation rules for each parameter and combines the threshold conditions verified in the previous clinical trial to ensure that the trend feature calculation is compatible with the parameter characteristics and to provide accurate quantitative dynamic indicators for subsequent abnormality judgment, thereby improving the scientificity and reliability of real-time monitoring.

[0116] The fourth feature data in the feature data is matched with the judgment conditions in the medical safety rule base to perform feature matching and anomaly judgment of medical hazards, and finally output the hazard analysis results, which include:

[0117] Based on real-time operating parameters, retrieve the feature verification conditions of the corresponding target judgment conditions from the medical safety rule base;

[0118] The criteria for determining the target include the criteria for determining the heart rate trend and the criteria for determining the trend of the ventilator's real-time operating parameters.

[0119] Feature verification conditions include:

[0120] For heart rate data, the rate of change of heart rate data meets the criteria for determining heart rate trend, including amplitude change within the first time threshold, absolute value exceeding the first time interval threshold, and duration greater than or equal to the fifth time threshold.

[0121] For the real-time operating parameters of the ventilator, the rate of change of the airway pressure data of the real-time operating parameters of the ventilator meets the judgment conditions of being greater than the first numerical threshold, the rate of change of the tidal volume data meets the judgment conditions of decreasing greater than or equal to the first proportional threshold, and the duration is greater than or equal to the sixth time threshold, and the fluctuation amplitude is less than or equal to the second proportional threshold.

[0122] If all trend features meet the feature verification conditions of the corresponding target determination conditions, then the anomaly determination result is output.

[0123] If any trend feature does not meet the feature verification condition of the corresponding target determination condition, it is determined to be a normal trend;

[0124] The abnormality assessment results, as part of the hazard analysis, include abnormal trends in heart rate and ventilation resistance.

[0125] In one embodiment, when performing medical hazard assessment, the fourth feature data (high-frequency data dynamic trend feature) in the feature data is first matched with the assessment conditions in the medical safety rule base. Based on the real-time operating parameters (heart rate data, ventilator real-time operating parameters), the corresponding target assessment conditions (heart rate trend assessment conditions, ventilator real-time operating parameter trend assessment conditions) and feature verification conditions are retrieved from the medical safety rule base. Among them, the feature verification conditions for heart rate data are that the rate of change of heart rate data must meet the amplitude change within 10 minutes of the first time threshold in the heart rate trend assessment conditions, the absolute value must exceed the first number interval threshold of 50-120 beats / min, and the duration must be greater than or equal to 10 minutes (the fifth time threshold is set to 10, which is set according to the physiological regulation law of heart rate in adults in non-sedated state, consistent with the heart rate change rate calculation cycle, to ensure the capture of stable abnormal trends). The feature verification conditions for ventilator real-time operating parameters are that the rate of change of airway pressure data must meet the ventilator parameter trend assessment conditions of 30 cmH2O greater than the first value threshold, and the rate of change of tidal volume data must meet the first value threshold of 30 cmH2O. The rate of change must meet the following criteria: a decrease greater than or equal to the first proportional threshold of 20%, a duration greater than or equal to 30 seconds (the sixth time threshold is set at 30 seconds, based on the characteristic of ventilator abnormalities requiring rapid response and adapted to the ventilator parameter calculation cycle), and a fluctuation amplitude less than or equal to 5% (the second proportional threshold is set at 5%, based on the threshold being used to filter instantaneous parameter fluctuations to ensure that the abnormality is stable). The judgment logic is as follows: if all trend features meet the feature verification conditions of the corresponding target judgment conditions, an abnormal judgment result is output; if any trend feature does not meet the conditions, it is judged as a normal trend. The final output of the hidden danger analysis results includes abnormal heart rate trends (such as a heart rate increase of 30 beats / min and exceeding 120 beats / min for 10 minutes and lasting for 10 minutes) and abnormal ventilation resistance trends (such as airway pressure exceeding 30 cmH2O, tidal volume decrease of 20% for 30 seconds and fluctuation less than or equal to 5%). This implementation method, through multi-dimensional feature verification and clinically adapted threshold settings, avoids false alarms from single-condition judgments and ensures the accuracy of hidden danger identification, providing reliable results support for medical safety early warning.

[0126] This invention significantly improves real-time performance and retrieval efficiency through differentiated design of data types and collection frequencies, optimizes storage resource allocation, uses lightweight Redis for high-frequency data and large-capacity HDFS for low-frequency data to avoid resource waste, and fully explores the value of multi-source data by relying on Master-Slaves clusters and Spark technology for feature extraction, while shortening feature extraction time through parallel processing to meet the efficiency requirements of medical scenarios. In terms of risk assessment, it combines multi-dimensional verification to avoid false negatives and omissions due to single thresholds, accurately outputting abnormal results, reducing clinical false alarm rates, and reducing redundant verification work for medical staff, ultimately providing reliable technical support for the identification of medical safety risks and subsequent diagnosis and treatment decisions.

[0127] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A medical risk analysis system for multi-source data, characterized in that, It includes a storage module, an extraction module, and an analysis module; The storage module stores multi-source data in segments; The extraction module is based on a Master-Slaves cluster architecture and uses multi-source data stored in the storage module in shards to extract features and obtain feature data. The analysis module, based on a preset medical safety rule base, matches the feature data with the judgment conditions in the rule base, performs feature matching and anomaly judgment of medical risks, and finally outputs the risk analysis results. The multi-source data is collected through the hospital's full-service system, and includes unstructured data, structured data, and semi-structured data. The multi-source data is first divided according to type to obtain image data corresponding to the semi-structured data, text data corresponding to the unstructured data, and numerical data corresponding to the structured data; The multi-source data is further divided according to the acquisition frequency to obtain high-frequency data and low-frequency data; The first partition and the second partition are processed in parallel; The high-frequency data is preferentially transmitted to a Redis temporary cache cluster for storage, and the cache validity period is set, and the high-frequency data access latency is less than or equal to 1 second; The low-frequency data is transmitted to the HDFS distributed file system for sharded storage. Among them, the text data in the low-frequency data is segmented using the patient's unique identifier and the text data generation timestamp as the segment identifier; The numerical data in the low-frequency data is segmented according to the inspection item code and the collection date, and the image data in the low-frequency data is segmented according to the image equipment number and the inspection order number. The extraction module, based on a Master-Slaves cluster architecture, calls multi-source data stored in shards within the storage module to extract features, obtaining feature data, specifically including: The extraction module retrieves the text data of low-frequency data stored in the HDFS distributed file system by patient unique identifier and generation timestamp, and performs keyword extraction and template matching in parallel based on Spark RDD to generate structured labels of keys and values ​​as the first feature data. The extraction module retrieves the test indicators of the low-frequency data stored in the HDFS distributed file system according to the test item code and the fragment identifier of the collection date, and processes them in parallel to form a standardized test dataset, which serves as the second feature data. The extraction module retrieves the complete image sequence of a single inspection from the low-frequency data stored in the HDFS distributed file system in segments, according to the segment identifier of the image device number and the inspection order number. The complete imaging sequence includes CT tomographic images and MRI sequence images; The complete image sequence is allocated as the smallest segmentation unit to the Slaves nodes of the Master-Slaves cluster. Each Slaves node performs lesion region segmentation in parallel based on preset medical image segmentation rules to obtain the segmented lesion region. The medical image segmentation rules are set based on the grayscale threshold and edge gradient features of lesions in the clinical image diagnostic standards, and quantitative features with clinical risk analysis significance are extracted from the segmented lesion areas. The quantitative features include the maximum diameter of the lesion, the volume of the lesion, the average density of the lesion, the irregularity of the lesion boundary, the shortest distance between the lesion and adjacent blood vessels, and the shortest distance between the lesion and adjacent organs. These features generate an examination report number and a key-value pair of quantitative features of the lesion, which serve as the third feature data. The extraction module extracts the numerical data of the high-frequency data temporarily cached in the Redis temporary cache cluster. Based on Spark Streaming, it retrieves the real-time running parameters of the numerical data in real time and performs dynamic trend analysis in parallel to obtain dynamic trend features, which are used as the fourth feature data. The feature data includes first feature data, second feature data, third feature data, and fourth feature data.

2. The medical risk analysis system for multi-source data as described in claim 1, characterized in that: Both the image data and the text data are low-frequency data, and the text data includes electronic medical record text. The numerical data is divided into high-frequency data and low-frequency data; The numerical data in the high-frequency data includes heart rate data collected by medical real-time monitoring equipment and real-time operating parameters of the ventilator; The real-time operating parameters include airway pressure data and tidal volume data, and the numerical data in the low-frequency data are the test indicators in the test report. The high-frequency data is dynamic data with a sampling frequency of greater than or equal to 1 time / second, and the dynamic data includes the real-time operating parameters of the ventilator; The low-frequency data is static data or batch data with a sampling frequency of less than or equal to 1 time / hour. The static data includes electronic medical record text, and the batch data includes test reports.

3. The medical risk analysis system for multi-source data as described in claim 2, characterized in that: The storage module supports the HL7 FHIR protocol, DICOM 3.0 protocol and MQTT protocol, respectively connecting to the hospital's business system and medical hardware equipment; The hospital's business system includes an electronic medical record system, a laboratory information system, and a medical imaging system; the medical hardware includes imaging equipment and real-time medical monitoring equipment. The HL7 FHIR protocol is used to interface with the electronic medical record system and the laboratory information system to jointly parse and obtain unstructured and structured data; The DICOM 3.0 protocol is used to interface with imaging equipment and medical imaging systems to receive and store semi-structured data; The MQTT protocol is used to connect to medical real-time monitoring equipment and collect high-frequency real-time data.

4. The medical risk analysis system for multi-source data as described in claim 3, characterized in that: The dynamic trend analysis refers to the determination of trend anomalies bound to the medical safety rule base, specifically including: A medical safety rule base is constructed, which includes the determination conditions for heart rate trends and the determination conditions for real-time operating parameter trends of ventilators; The criteria for determining the heart rate trend are as follows: for adults in a non-sedated state, if the increase in heart rate data at the first time threshold is greater than or equal to the first count threshold and the absolute value is greater than the second count threshold, or the decrease is greater than or equal to the third count threshold and the absolute value is less than the fourth count threshold, then it is determined to be an abnormal trend. The criteria for determining the real-time operating parameter trend of the ventilator are as follows: if the airway pressure data of the ventilator is continuously greater than the second time threshold and the tidal volume data decreases simultaneously by a factor greater than or equal to the first proportional threshold, then it is determined to be an abnormal trend of ventilation resistance.

5. The medical risk analysis system for multi-source data as described in claim 4, characterized in that: The trend characteristics of the real-time operating parameters are calculated in real time, and the trend characteristics include the rate of change, duration, and fluctuation amplitude. The rate of change is calculated based on the parameter type of the high-frequency data, which includes heart rate data, airway pressure data, and tidal volume data. The heart rate data is periodized with a third time threshold, and the airway pressure data and tidal volume data are periodized with a fourth time threshold. The difference between the final value and the initial value of the corresponding parameter in the high-frequency data in each period is calculated, and then the difference is divided by the initial value to obtain the parameter change rate of the corresponding period. The duration is determined by counting the number of consecutive samples that meet the corresponding rule threshold conditions for the real-time operating parameters, based on a sampling frequency of 1 time / second for the real-time operating parameters, and multiplying the number of samples by a 1-second sampling interval.

6. The medical risk analysis system for multi-source data as described in claim 5, characterized in that: The rule threshold conditions include threshold conditions for heart rate data and threshold conditions for real-time operating parameters of the ventilator. The threshold condition for the heart rate data is that the increase in amplitude within the first time threshold is greater than or equal to the first count threshold or the decrease is greater than or equal to the third count threshold, and the absolute value exceeds the first count interval threshold. The threshold conditions for the real-time operating parameters of the ventilator are that the airway pressure data is greater than a first numerical threshold and the tidal volume decrease is greater than or equal to a first proportional threshold. The fluctuation amplitude is calculated by taking n consecutive 1-second sampling periods as a statistical window and calculating the difference between the maximum and minimum values ​​of the corresponding parameters in the high-frequency data within the window.

7. The medical risk analysis system for multi-source data as described in claim 6, characterized in that: The fourth feature data in the feature data is matched with the judgment conditions in the medical safety rule base to perform feature matching and anomaly judgment of medical hazards, and finally output the hazard analysis results, which specifically include: Based on real-time operating parameters, the feature verification conditions of the corresponding target judgment conditions are retrieved from the medical safety rule base; The target determination criteria include the determination criteria for heart rate trends and the determination criteria for real-time operating parameter trends of the ventilator. The feature verification conditions include: For heart rate data, the rate of change of the heart rate data meets the criteria for determining the heart rate trend, namely, the amplitude change within the first time threshold, the absolute value exceeding the first number interval threshold, and the duration greater than or equal to the fifth time threshold. Regarding the real-time operating parameters of the ventilator, the rate of change of the airway pressure data of the real-time operating parameters of the ventilator meets the judgment conditions of the real-time operating parameter trend being greater than the first numerical threshold, the rate of change of the tidal volume data meets the judgment conditions of the decrease being greater than or equal to the first proportional threshold, the duration being greater than or equal to the sixth time threshold, and the fluctuation amplitude being less than or equal to the second proportional threshold. If all the trend features satisfy the feature verification conditions of the corresponding target determination conditions, then the anomaly determination result is output. If any trend feature does not meet the feature verification condition of the corresponding target determination condition, it is determined to be a normal trend; The abnormality determination results, as part of the hidden danger analysis, include abnormal trends in heart rate and abnormal trends in ventilation resistance.

Images