Multi-source data fusion-based VA-ECMO data analysis method and system

Abstract

The invention relates to a VA-ECMO data analysis method and system based on multi-source data fusion. The method comprises the following steps: acquiring a first sample set based on historical patient data, and marking a sample acquisition mode; constructing a second sample set based on the first sample set; training an auxiliary classification model; acquiring patient data and inputting the patient data into the auxiliary classification model to obtain auxiliary data; and presenting the auxiliary data to the user for auxiliary decision making. According to the method, a data-driven model optimization closed loop is formed, the accuracy and reliability of auxiliary decision making are improved, and deep application of intelligent medical treatment in the VA-ECMO field is promoted.

Description

Technical Field

[0001] This invention belongs to the field of smart healthcare technology, and in particular relates to a method and system for VA-ECMO data analysis based on multi-source data fusion. Background Technology

[0002] In recent years, artificial intelligence (AI) technology has demonstrated enormous potential in areas such as medical diagnosis, treatment decision-making, and prognosis assessment. Intelligent medical systems, by integrating multi-source medical data and building predictive models, provide decision support for doctors, significantly improving the accuracy and efficiency of medical services. In the field of VA-ECMO, AI technology can handle complex physiological parameters and clinical indicators, assisting doctors in making timely and accurate auxiliary decisions. However, in vertical medical applications, especially in high-end life support technologies such as extracorporeal membrane oxygenation (VA-ECMO), insufficient sample size has become a key bottleneck restricting the effective application of AI technology. As an expensive critical care technology, VA-ECMO has relatively few cases and limited historical data accumulation, leading to serious overfitting risks and insufficient data-driven capabilities for traditional machine learning methods.

[0003] In existing technologies, data augmentation methods are typically employed to overcome the challenge of insufficient sample size. Traditional methods mainly include uniform interpolation, random oversampling, and synthetic minority class oversampling. These methods are based on the assumption of uniform data distribution, expanding the dataset by uniformly generating new samples within the feature space or balancing the number of samples in each class. However, in VA-ECMO practice, patient data distribution exhibits significant non-uniformity. Physiological parameters of patients at different disease stages show specific distribution patterns, with complex physiological correlations and pathological couplings between various indicators. For example, hemodynamic parameters are strongly correlated with oxygenation and organ function indicators, and exhibit different correlation patterns at different disease severity levels. The specific problems with existing technologies are mainly: First, traditional uniform interpolation methods ignore the inherent distribution characteristics of data, resulting in generated data that does not conform to medical principles, potentially introducing erroneous physiological patterns and reducing the reliability and practicality of the model. Second, existing methods fail to fully consider the actual bedside correlations between different data types. In VA-ECMO usage decisions, there are complex interactions between multiple key parameters such as mean arterial pressure, lactate level, and cardiac index. Simple independent interpolation disrupts these correlations, leading to a lack of physiological consistency in the generated data. Furthermore, traditional methods are inadequate in handling data unevenness. In real-world scenarios, the unevenness of distribution among different data types varies significantly, and this unevenness itself has important practical significance. Uniform interpolation methods obliterate this difference, potentially losing crucial pathological information. In addition, current technologies lack customized interpolation strategies for the specific needs of vertical medical fields. VA-ECMO usage decisions involve multi-parameter fusion analysis, requiring the maintenance of realistic plausibility and physiological consistency between data, a requirement that traditional methods cannot meet.

[0004] In summary, existing AI-assisted medical systems face severe technical limitations in small-sample vertical applications such as VA-ECMO. Developing an intelligent interpolation method that preserves the actual data distribution characteristics and parameter correlations is urgently needed and of significant value for promoting the effective application of AI in high-end medical technologies. This invention aims to solve the aforementioned technical problems, providing a more efficient data processing method for smart healthcare, forming a data-driven model optimization closed loop, improving the accuracy and reliability of assisted decision-making, and promoting the deep application of smart healthcare in the VA-ECMO field. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention proposes a VA-ECMO data analysis method and system based on multi-source data fusion, the method comprising: Step S1: Obtain the first sample set based on historical patient data and label the sample collection method; Step S2: Construct a second sample set based on the first sample set; specifically: divide the first sample set into blocks according to the collection method, and determine the distribution unevenness of each data type for each sample block in turn; expand the sample blocks based on the distribution unevenness to obtain the second sample set; the size of the second sample set is larger than that of the first sample set; unevenness reflects whether the value distribution of the data type is uniform. If some values ​​or certain ranges appear frequently in the sample block, while other values ​​rarely appear, the unevenness is high; if all values ​​are evenly distributed, the unevenness is low. The process of expanding the sample block based on distribution unevenness to obtain a second sample set involves: expanding the typical data types with high unevenness in the sample blocks of the first sample set within a local range; and generating values ​​for atypical data types in the same record based on the numerical relationship between typical and atypical data types, so that the interpolated values ​​of typical data types maintain an uneven distribution pattern, while maintaining a logical correlation between the interpolated data of atypical data types and the values ​​of typical data types. Step S3: Train the auxiliary classification model; specifically: train the auxiliary classification model using the second sample set, and perform external validation and evaluation on the auxiliary classification model; Step S4: Obtain patient data and input it into the auxiliary classification model to obtain auxiliary data; present the auxiliary data to the user to assist in decision-making.

[0006] Furthermore, the historical patient data containing multiple parameters includes one or more of the following: basic patient information, pre-heart rate and respiratory parameters, VA-ECMO related parameters, and post-use related parameters.

[0007] Furthermore, data access is achieved through the HIS interface or CDR database, automatically acquiring various types of data generated during patient hospitalization in EMR, LIS, and PACS medical information systems, and synchronizing them over time.

[0008] Furthermore, step S2 specifically includes the following steps: Step S21: Determine the sample block to be processed from the first sample set; specifically: select multiple records with the same or similar collection methods from the first sample set to form the current sample block to be processed; Step S22: Determine the distribution unevenness of typical data types in the current sample block to be processed, and sort the typical data types from largest to smallest unevenness to obtain a data type sequence; Step S23: Expand the sample block to be processed based on the non-uniformity so that the expansion of typical data types with high non-uniformity is formed in a local range; while the expansion of atypical data types with low non-uniformity and the expansion of typical data types maintain logical correlation. Step S24: Determine whether the construction deadline is met. If yes, update the first sample set as the second sample set using the expanded unprocessed data block. Otherwise, return to step S21.

[0009] Furthermore, the data collection methods include: data collection methods related to equipment and instruments, driving methods, batches of supporting consumables used, data collection methods related to medical institutions and operating procedures, data collection methods related to the experience level of the nursing team, anticoagulation monitoring methods, and data collection methods related to the time dimension.

[0010] Furthermore, the historical patient data is multi-source data.

[0011] A VA-ECMO data analysis platform based on multi-source data fusion is provided, the platform being used to implement the VA-ECMO data analysis method based on multi-source data fusion.

[0012] A server for VA-ECMO data analysis based on multi-source data fusion includes a processor coupled to a memory. The memory stores program instructions, and when the program instructions stored in the memory are executed by the processor, the VA-ECMO data analysis method based on multi-source data fusion is implemented.

[0013] A VA-ECMO data analysis system based on multi-source data fusion is provided, the system being used to implement the VA-ECMO data analysis method based on multi-source data fusion.

[0014] A computer-readable storage medium includes a program that, when run on a computer, causes the computer to perform the VA-ECMO data analysis method based on multi-source data fusion.

[0015] The beneficial effects of this invention include: (1) Based on the characteristics of limited sample size and uneven data distribution in the VA-ECMO field, a second sample set is constructed using intelligent data expansion methods of distribution unevenness calculation and cluster analysis. By identifying the natural cluster centers of each data type, it is ensured that the newly generated sample data falls within the range of real logical data clusters, avoiding the data points that may be generated by traditional mean interpolation that are out of touch with reality. Synthetic data that conforms to real logic is generated, and a data-driven model optimization closed loop is formed, which improves the accuracy and reliability of auxiliary decision-making and promotes the in-depth application of intelligent medical care in the VA-ECMO field. (2) By dividing the sample blocks and determining the interpolation basis corresponding to the collection method, a data expansion strategy adapted to reality is formed. By constructing a continuous I-dimensional space constraint mechanism, it is ensured that each newly generated sample point is located in the closed multi-dimensional space composed of actual samples, so that all interpolated data conforms to the inherent correlation law of actual data in multiple dimensions, fundamentally eliminating the problem of unreasonable numerical combination that may be caused by single-dimensional interpolation, and providing a more reliable data basis for decision support. Thus, large server resources can be used for complex calculations and model training, while terminals can be used for data verification and real-time application, improving the efficiency of medical data processing and the accuracy of decision-making, and supporting the rapid generation of personalized medical plans. Attached Figure Description

[0016] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to unduly limit the invention. In the drawings: Figure 1 A schematic diagram of the VA-ECMO data analysis method based on multi-source data fusion provided by the present invention.

[0017] Figure 2 This is a schematic diagram of the extension method when K=2 provided by the present invention.

[0018] Figure 3 This is a schematic diagram of the extension method when K=3 provided by the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0020] This invention proposes a VA-ECMO data analysis method and system based on multi-source data fusion, as shown in the appendix. Figure 1 As shown, the method includes the following steps: Step S0: Obtain multi-parameter indicators and historical patient data during VA-ECMO participation; specifically: retrieve relevant original studies on VA-ECMO, comprehensively analyze multi-dimensional data such as patients' physiological parameters, laboratory test results, and VA-ECMO parameters as historical patient data, and extract multi-parameter indicators during VA-ECMO participation; use the MIMIC-Ⅳ public database to obtain historical patient data containing multi-parameter indicators; store all acquired data uniformly on a local server or cloud server, and perform data integration, encryption, and management to ensure data security, accessibility, and integrity, providing a reliable data foundation for subsequent processing; Step S1: Obtain the first sample set based on historical patient data and label the sample collection method; Preferably, the historical patient data is multi-source data, and the multi-source data is organized according to data type to form a sample or a record; wherein: multiple data types are organized to form a sample or record; Preferably, the historical patient data containing multiple parameters includes: basic patient information, pre-cardiac and respiratory parameters, VA-ECMO related parameters, and post-use related parameters; wherein: the basic patient information includes age, gender, underlying diseases, and clinical condition; the pre-cardiac and respiratory parameters include the duration of cardiac and respiratory arrest before use, resuscitation time, and Glasgow Coma Scale scores before and after resuscitation; the VA-ECMO related parameters include: ECMO support duration, blood flow, oxygenator, ventilation volume, and anticoagulant dosage; and the post-use related parameters include complications and post-cardiac and respiratory parameters, etc. Preferred methods include: data collection methods related to equipment and instruments, differences in drive methods, batches of supporting consumables, medical institutions and operating procedures, the experience level of the nursing team, anticoagulation monitoring methods, and time-related data collection methods; further, data collection methods related to equipment and instruments, such as data from different systems like Maquet Rotaflow, Medtronic, and Cardiohelp; membrane oxygenator types, the impact of different brands and models of oxygenators on gas exchange and hemodynamics; differences in drive methods, such as the differences in data characteristics of centrifugal pumps and roller pumps in flow monitoring and hemolysis indicators; batches of supporting consumables, such as the impact of different batches of cannulas and tubing on resistance parameters and pressure monitoring; data collection methods related to medical institutions and operating procedures, such as the level and characteristics of medical institutions, the differences in ECMO management procedures between a tertiary-level hospital's cardiac center and a regional medical center; and departmental operating procedures, such as cardiac surgery... Differences between ICU and emergency ICU in sedation, analgesia, and volume management strategies; the experience level of the nursing team, such as the difference in data recording precision between ECMO-dedicated nursing teams and general ICU nursing teams; anticoagulation monitoring methods, such as data characteristics under two protocols: ACT-based monitoring and anti-Xa factor monitoring; time-related data collection methods, such as the division of treatment stages (initial stabilization period, maintenance treatment period, and preparation period); data collection time periods, such as daytime routine monitoring data, nighttime monitoring data, and shift handover data; and seasonal and time cycles, such as the impact of different seasons on patient infection indicators and thermoregulation-related parameters. The process of obtaining the first sample set based on historical patient data specifically involves: preprocessing the historical patient data; using the first part and the second part of the historical patient data as the input and output parts of the sample, respectively; wherein: the first part includes: basic patient information, pre-cardiac and respiratory parameters, and VA-ECMO related parameters; the second part includes post-use related parameters; The preprocessing specifically includes: cleaning historical patient data and processing missing values ​​using methods such as imputation and removal; handling outliers using methods such as mean correction; standardizing numerical features to make the data follow a standard normal distribution; and using the first part of the historical patient data as the input part of the sample and the second part as the output part of the sample. Step S2: Construct a second sample set based on the first sample set; specifically: divide the first sample set into blocks according to the collection method, and determine the distribution unevenness of each data type for each sample block; expand the first sample set based on the distribution unevenness to obtain the second sample set; the size of the second sample set is larger than that of the first sample set; unevenness reflects whether the value distribution of the data type is uniform. If some values ​​or ranges in the dataset appear frequently, while other values ​​rarely appear, the unevenness is high; if all values ​​are evenly distributed, the unevenness is low; obviously, both the first and second sample sets contain multiple records. The expansion of the first sample set based on distribution unevenness to obtain the second sample set involves: expanding the typical data types with high unevenness in the first sample set within a local range; and generating values ​​for atypical data types in the same record based on the numerical relationship between typical and atypical data types, so that the interpolated values ​​of typical data types maintain an uneven distribution pattern; and maintaining a logical correlation between the interpolated data of atypical data types and the values ​​of typical data types. Step S2 specifically includes the following steps: Step S21: Determine the sample block to be processed from the first sample set; specifically: select multiple records with the same or similar collection methods from the first sample set to form the current sample block to be processed; for example: put samples collected from the same organization, based on the same ECMO model, etc., into the same sample block to be processed; data obtained using the same collection method often have implicit internal logic, which may lead to different uniformity of numerical distribution of different data types between different sample blocks to be processed; thus, the expandable data positions of different sample blocks to be processed are different. By combining the non-uniformity of the data blocks to be processed and the non-uniformity of the data types, a non-uniform data expansion can be formed in the entire space of the first sample set by utilizing the characteristics of the data collection method and the non-uniformity of the distribution, and finally achieve a relatively dense expansion of the first sample set; Step S22: Determine the distribution unevenness of typical data types in the current sample block to be processed, and sort the typical data types from largest to smallest unevenness to obtain a data type sequence; Preferably, the typical data type is a preset data type selected from the data type; preferably, the typical data type comes from the first part and / or the second part; for example, the typical data type is Glasgow Coma Scale score before and after resuscitation, pre-heart rate and respiratory parameters, ECMO support duration, etc. The determination of the distribution non-uniformity of each typical data type in the current sample block to be processed is specifically as follows: Step S22A1: Cluster each typical data type i, and use the number of clusters as the number of unique values ​​of typical data type i. Specifically, clustering methods such as density-based clustering (DBSCAN), mean-shift clustering, and hierarchical clustering are used to cluster typical data types, and the number of clusters is taken as the number of unique values ​​for the typical data type i; where: ; I is the number of typical data types; Preferred approach: If the typical data type is numerical, first discretize it to convert it into a categorical distribution to ensure the consistency of the measurement; Step S22A2: Calculate the value frequency distribution for each typical data type; specifically: for each typical data type i, calculate its unique value frequency. ;in: It is the number of elements in cluster j of typical data type i, which is also the size of cluster j; Step S22A3: Calculate the Herfindahl exponent Specifically, the Herfindahl exponent is calculated based on the following formula (1). Scope Between 1 and 1, the closer the value is to 1, the more concentrated and uneven the distribution; the closer the value is to 1, the more concentrated and uneven the distribution. The more evenly distributed; (1); Step S22A4: Based on the Herfindahl exponent Calculate normalized non-uniformity Specifically, the uniformity is calculated based on the following formula (2). ;if but The distribution is completely concentrated; The numerical range is between 0 and 1: 0 represents a completely uniform distribution, and 1 represents a completely concentrated distribution; (2); Preferred approach: Select typical data types from typical data types to calculate the distribution unevenness and perform subsequent processing; that is, perform subsequent processing on non-typical data types. This avoids unnecessary computational overhead caused by an excessive number of typical data types and insufficient actual meaning contained in the typical data types themselves, which may instead lead to the generation of many bad samples. Step S23: Expand the sample block to be processed based on non-uniformity so that the expansion of typical data types with high non-uniformity is formed within a local range; while the expansion of atypical data types with low non-uniformity and the expansion of typical data types maintain logical correlation; specifically including the following steps: Step S231: Take a typical data type from the head of the typical data type sequence as the current typical data type; it can be seen that the initial value of the current typical data type is the one with the highest non-uniformity. Step S232: Cluster the values ​​of the current typical data type; select a cluster as the current cluster and determine the range of values ​​in the current cluster; the cluster selection method can be arbitrary selection or sequential selection. Step S233: Take the sample portion of the current sample block to be processed that corresponds to the numerical range of the current typical data type within the numerical span of the cluster as the updated sample block to be processed; that is, if the value of the current typical data type of a sample does not belong to the numerical span of the cluster, it will not exist in the updated sample block to be processed after the update; that is, starting from the typical data type with the highest non-uniformity, the sample block to be processed is continuously reduced to form the final interpolation basis. Step S234: Determine if there are any unprocessed typical data types. If yes, return to step S231; otherwise, proceed to the next step. Step S235: Perform interpolation on the sample block to be processed so that each typical data type in the interpolated sample is within its numerical range. In the above manner, the data blocks to be processed are reduced sequentially based on the uniformity of typical data types. In the final sample part, the value of each typical data type of each sample falls within the range of the clustering values ​​in step S232, so that the interpolation of each typical data type conforms to the logic of numerical occurrence. Step S236: Based on the numerical relationship between typical and atypical data types, supplement the values ​​of atypical data types in the interpolation sample to form a complete interpolation sample, and finally form an expanded data block to be processed. Alternatively: Step S23 specifically involves: when each sample is I data elements, that is, when it contains I data types, obtaining I+1 samples from the current sample block to be processed, and expanding based on the I+1 samples, so that the value of the i-th data element of the expanded sample is within the I-dimensional space range formed by the values ​​of the I+1 i-th data elements of the I+1 samples; in this way, the value of any data type or typical data type will not exceed the logical space range and enter an unreasonable range when interpolating, thereby causing a reverse training effect on the training of the classification model; Furthermore: when the replacement method is only for typical data types, the values ​​of atypical data types in the interpolation sample are supplemented based on the numerical relationship between typical and atypical data types, thereby forming a complete interpolation sample and finally forming an expanded data block to be processed. As attached Figure 2 As shown, when I=2, three samples are obtained for expansion. The numerical space of the expanded samples needs to be within the 2D space formed by these three samples; see attached. Figure 3 As shown, when I=3, four samples are obtained for expansion. The numerical space of the expanded samples needs to be within the 3D space formed by the three samples. In this way, when new samples are generated, they are constrained by the multi-dimensional numerical range of multiple actual samples, forming within a reasonable numerical space, rather than simply forming at the numerical boundaries. (See attached...) Figure 2 As shown, if it is only based on the expansion of a single-dimensional value, when samples S1 and S2 are obtained, the resulting value for the X-axis may lie on the straight line formed by the two. Furthermore: The above alternative interpolation method is performed for typical data types, while for atypical data types, the numerical values ​​of atypical data types in the interpolation sample are supplemented based on the numerical relationship between typical and atypical data types formed by interpolation, thereby forming a complete interpolation sample and finally forming an expanded data block to be processed. Step S24: Determine whether the construction deadline is met. If yes, update the first sample set as the second sample set using the expanded unprocessed data block; otherwise, return to step S21. The construction cutoff condition is that the number of the second sample set reaches the preset size and / or there are no sample blocks to be processed; after all the sample blocks to be processed have been processed, the data block splitting based on the acquisition method, etc., may change the organization of the samples and form a new sample block division method, thereby creating new interpolation opportunities. Preferred methods include: using SMOT or ADASYN to avoid excessive bias of the model towards a minority class; and further using LASSO to predict second samples with small contributions and potential overfitting, removing them from the second sample set. Step S3: Train the auxiliary classification model; specifically: train the auxiliary classification model using the second sample set, and perform external validation and evaluation on the auxiliary classification model; The method of using the second sample set to train the auxiliary classification model is as follows: the second sample set is randomly divided into k subsets of equal size. During training, one subset is selected as the test set, and the remaining k-1 subsets are used as the training set. This process is repeated k times, so that each data set is used as the test set once and the training set k-1 times. This method reduces the impact of sample set size on training and reduces the variation or instability of evaluation results caused by the randomness of sample set partitioning. Preferred method: The second sample set is randomly divided into 5 equal parts, and 4 parts are used as training data in turn, and the remaining part is used as validation data. This process is repeated until each part of the data has been used as validation data. The external validation involves using data not used to build the auxiliary classification model to evaluate the model's performance on new data, verifying the model's portability and generalizability; the performance evaluation uses accuracy, recall, precision, and F1 score as evaluation metrics, while also considering plotting ROC curves and calculating AUC values, and using calibration curves, Hosmer-Lemeshow goodness-of-fit tests, and other methods to evaluate the model's calibration. Step S4: Input patient data into the auxiliary classification model to obtain auxiliary data; the auxiliary data is the ECMO machine participation score or risk assessment score; standardize the output format for subsequent analysis. Preferably: the preprocessed patient data is organized into input features according to the time dimension, corresponding to the first part of the sample; the input features include static patient data and dynamic patient data; obviously, the input data format of the classification model needs to conform to the tensor format required by the classification model; a pre-trained deep learning model or ensemble machine learning model is used. Preferably, the classification model is a time series model (LSTM, GRU) or a convolutional network model; first, the completeness and rationality of the auxiliary data are checked using a knowledge base, and then the consistency is verified by expert knowledge to ensure that the auxiliary data is within the physiologically credible range; Compared with the traditional uniform interpolation method, this method improves the AUC-ROC value from 0.72 to 0.85, a relative improvement of 18.1%; accuracy from 68.3% to 82.7%, an absolute improvement of 14.4 percentage points; recall and precision achieve relative improvements of 24.1% and 17.8% respectively; and the F1 score increases from 0.682 to 0.825, an increase of 20.9%. It demonstrates stability under varying sample sizes. When the base sample size decreases from 185 to 80, the AUC-ROC value of the traditional method drops significantly to 0.63, while this method maintains an excellent performance of 0.79, reducing the performance decline by 76%, fully demonstrating its strong robustness in small sample scenarios. Regarding generalization ability, through k-fold cross-validation (k=5), the AUC fluctuation range of each fold is significantly narrowed from 0.67-0.77 in the traditional method to 0.82-0.88, reducing variance by 62%. The performance retention rate on the independent external test set (n=45) reached 94%, an improvement of 18 percentage points compared to the traditional method's 76%, demonstrating excellent generalization performance. Due to the increased sample size, the effective expansion rate reached 85%, an improvement of 124% compared to the traditional method's 38%. The quality score of the generated auxiliary data improved from 4.2 to 8.7, an increase of 107%. The parameter correlation retention rate improved from 67% to 92%, ensuring that the generated synthetic data fully conforms to clinical logic and physiological laws. Based on the same inventive concept, the present invention also provides a VA-ECMO data analysis system based on multi-source data fusion, the system being used to implement the above-mentioned VA-ECMO data analysis method based on multi-source data fusion; Based on the same inventive concept, the present invention also provides a VA-ECMO data analysis server based on multi-source data fusion, the server being used to implement the above-mentioned VA-ECMO data analysis method based on multi-source data fusion; Based on the same inventive concept, the present invention also provides a VA-ECMO data analysis device based on multi-source data fusion, the device being used to implement the above-mentioned VA-ECMO data analysis method based on multi-source data fusion; Based on the same inventive concept, the present invention also provides a VA-ECMO data analysis platform based on multi-source data fusion, the platform being used to implement the above-mentioned VA-ECMO data analysis method based on multi-source data fusion; Example For the first sample set, a sample block to be processed is started, containing 10 patient records collected from the same model of ECMO device; typical data types: Data Type A: ECMO support duration (unit: hours), Data Type B: Lactate level (unit: mmol / L); Atypical data type: Data Type C: Mean arterial pressure (unit: mmHg), the value of which will be generated based on the expansion relationship of the typical data types; by calculating the heterogeneity, the sample block is expanded to generate 5 new records, which are part of the data of the second sample set; first, the heterogeneity of the typical data types is calculated; for data type A ECMO support duration, cluster analysis is performed using DBSCAN clustering, dividing the data into 3 clusters: Cluster 1: Support duration 10-20 hours, containing 4 samples; Cluster 2: Support duration 21-30 hours, containing 5 samples; Cluster 3: Supports durations of 31-40 hours, containing 1 sample; performs frequency distribution calculations: pA1=0.4, pA2=0.5, pA3=0.1; Herfindahl exponent calculation, HA=0.42; normalized non-uniformity, uA=0.13; for data type B lactate level, uses mean-shift clustering to divide the data into 2 clusters: Cluster 1: lactate 1-2 mmol / L, containing 2 samples; Cluster 2: lactate 3-4 mmol / L, containing 8 samples; pB1=0.2, pB2=0.8; HB=0.68, uB=0.36; sorts typical data types, compares non-uniformity, data type sequence is: [B, [A], that is, first process lactate levels, then process ECMO support duration; then, expand the data based on non-uniformity; initial sample block to be processed: 10 records, containing the original values ​​of data types A and B; steps S231-S234: reduce the sample block according to the data type sequence; process data type B, select cluster 2, containing 8 records, update the sample block to be processed to these 8 records, then process data type A, in the updated 8 records, cluster data type A is also divided into 3 clusters, but the distribution changes, cluster 2 (support duration 21-30 hours) contains 4 records, further update the sample block to be processed to these 4 records; in step S235, interpolate within the reduced sample block containing 4 records, the numerical range of data type A: 21-30 hours, the numerical range of data type B: 3-4 mmol / L; use linear interpolation to generate 5 new records in two-dimensional space (A and B), namely: new record 1: A=23 hours, B=3.2 mmol / L; New record 2: A=25 hours, B=3.5 mmol / L; New record 3: A=27 hours, B=3.8 mmol / L; New record 4: A=22 hours, B=3.3 mmol / L; New record 5: A=29 hours, B=3.9 mmol / L; Finally, based on the numerical relationship between typical and atypical data types, the mean arterial pressure of atypical data type C was supplemented; due to historical data indicating that data type C has a linear relationship with A and B, C=70+0.5×A−10×BC=70+0.5×A−10×B, the C value decomposition of the new records is calculated as 49.5 mmHg, 47.5 mmHg, 45.5 mmHg, 48.0 mmHg, 45.5 mmHg; at this time, an expanded sample block is obtained, containing 5 new records, each containing the values ​​of data types A, B and C, all of which fall within a reasonable range; by calculating the non-uniformity, the lactate level distribution is identified as more concentrated, so it is processed first to ensure that the interpolation is performed within a local range. After interpolation, the typical data type values ​​of the new samples all fall within the actual data clustering range, such as lactate 3-4 mmol / L, supporting a duration of 21-30 hours, avoiding values ​​that may be deviated from clinical logic by traditional mean interpolation, such as lactate 1 The data was measured in mmol / L or supported for up to 50 hours. For atypical data types, numerical relationships were used to supplement the data, maintaining a logical connection with typical data types and improving overall data consistency. This ensured that the generated second sample set better reflected the true distribution of VA-ECMO patient data, providing a high-quality and reliable data foundation for subsequent model training. A computer program (also referred to as a program, software, software application, script, or code) can be written in any form of programming language, including assembly or interpreted languages, declarative or procedural languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, object, or other unit suitable for use in a computing environment. A computer program may, but does not necessarily, correspond to a file in a file system. A program can be stored as part of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to said program, or in multiple co-located files (e.g., a file storing one or more modules, subroutines, or code portions). A computer program can be deployed to execute on a single computer or on multiple computers located at a single site or distributed across multiple sites and interconnected by a communications network.

[0021] 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 embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0022] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0023] 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.

[0024] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0025] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A VA-ECMO data analysis method based on multi-source data fusion, characterized in that, The method includes: Step S1: Obtain the first sample set based on historical patient data and label the sample collection method; Step S2: Construct a second sample set based on the first sample set; specifically: divide the first sample set into blocks according to the acquisition method, and determine the distribution unevenness of each data type for each sample block in turn; expand the sample blocks based on the distribution unevenness to obtain the second sample set; the size of the second sample set is larger than that of the first sample set; Unevenness reflects whether the distribution of values ​​of a data type is uniform. If some values ​​or ranges appear frequently in a sample block, while other values ​​rarely appear, the unevenness is high; if all values ​​are uniformly distributed, the unevenness is low. The process of expanding the sample block based on distribution unevenness to obtain a second sample set involves: expanding the typical data types with high unevenness in the sample blocks of the first sample set within a local range; and generating values ​​for atypical data types in the same record based on the numerical relationship between typical and atypical data types, so that the interpolated values ​​of typical data types maintain an uneven distribution pattern, while maintaining a logical correlation between the interpolated data of atypical data types and the values ​​of typical data types. Step S3: Train the auxiliary classification model; specifically: train the auxiliary classification model using the second sample set, and perform external validation and evaluation on the auxiliary classification model; Step S4: Obtain patient data and input it into the auxiliary classification model to obtain auxiliary data; present the auxiliary data to the user to assist in decision-making.

2. The VA-ECMO data analysis method based on multi-source data fusion according to claim 1, characterized in that, The historical patient data containing multiple parameters includes one or more of the following: basic patient information, pre-heart rate and respiratory parameters, VA-ECMO related parameters, and post-use related parameters.

3. The VA-ECMO data analysis method based on multi-source data fusion according to claim 2, characterized in that, Data access is achieved through the HIS interface or CDR database, automatically acquiring various types of data generated during patient hospitalization in EMR, LIS, and PACS medical information systems, and synchronizing them over time.

4. The VA-ECMO data analysis method based on multi-source data fusion according to claim 3, characterized in that, Step S2 specifically includes the following steps: Step S21: Determine the sample block to be processed from the first sample set; specifically: select multiple records with the same or similar collection methods from the first sample set to form the current sample block to be processed; Step S22: Determine the distribution unevenness of typical data types in the current sample block to be processed, and sort the typical data types from largest to smallest unevenness to obtain a data type sequence; Step S23: Expand the sample block to be processed based on the non-uniformity so that the expansion of typical data types with high non-uniformity is formed in a local range; while the expansion of atypical data types with low non-uniformity and the expansion of typical data types maintain logical correlation. Step S24: Determine whether the construction deadline is met. If yes, update the first sample set as the second sample set using the expanded unprocessed data block. Otherwise, return to step S21.

5. The VA-ECMO data analysis method based on multi-source data fusion according to claim 4, characterized in that, The data collection methods include: data collection methods related to equipment and instruments, driving methods, batches of supporting consumables used, data collection methods related to medical institutions and operating procedures, data collection methods related to the experience level of the nursing team, anticoagulation monitoring methods, and data collection methods related to the time dimension.

6. The VA-ECMO data analysis method based on multi-source data fusion according to claim 5, characterized in that, The historical patient data is multi-source data.

7. A platform for VA-ECMO data analysis based on multi-source data fusion, characterized in that, The platform is used to implement the VA-ECMO data analysis method based on multi-source data fusion as described in any one of claims 1-6.

8. A server for VA-ECMO data analysis based on multi-source data fusion, characterized in that, The system includes a processor coupled to a memory, the memory storing program instructions, which, when executed by the processor, implement the VA-ECMO data analysis method based on multi-source data fusion as described in any one of claims 1-6.

9. A VA-ECMO data analysis system based on multi-source data fusion, characterized in that, The system is used to implement the VA-ECMO data analysis method based on multi-source data fusion as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, Includes a program that, when run on a computer, causes the computer to perform the VA-ECMO data analysis method based on multi-source data fusion as described in any one of claims 1-6.

Images