A multi-dimensional medical data analysis and decision-making support system and method

Abstract

The present invention provides a multidimensional medical data analysis and decision support system and method. The system includes modules for data acquisition, preprocessing, feature extraction, model training, multidimensional analysis, decision generation, and feedback optimization. The data acquisition module acquires data from multiple sources, the preprocessing module performs data cleaning and standardization, the feature extraction module extracts multidimensional features, the model training module trains an adaptive decision model, the multidimensional analysis module performs data analysis, the decision generation module generates and verifies decisions, and the feedback optimization module adjusts the model based on feedback. This invention can improve the scientificity, accuracy, and reliability of medical decision-making, providing strong support for clinical decision-making.

Description

Technical Field

[0001] The present invention relates to the field of medical data analysis technology, and more specifically, to a multi-dimensional medical data analysis and decision-making support system and method. Background Art

[0002] In the field of medical data analysis and decision support, with the continuous development of medical technology, the scale and complexity of medical data are increasing. How to effectively extract valuable information from massive data and provide a scientific basis for clinical decision-making has become a hot topic in current research. Existing medical data analysis methods usually rely on single-dimensional data, such as electronic medical records or medical images, and lack the ability to comprehensively analyze multi-source heterogeneous data. In addition, traditional data analysis methods have shortcomings in handling noisy data, missing values, and data standardization, resulting in limited accuracy and reliability of analysis results. In terms of model training and decision generation, existing technologies often use static weight allocation and fixed parameter optimization strategies, which are difficult to adapt to complex and changing clinical scenarios. In addition, there is a lack of effective decision verification mechanisms, making it difficult to ensure the reliability and practicality of decisions.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the existing technology: the existing medical data analysis system cannot effectively integrate multi-source heterogeneous data, lacks dynamic weight allocation and optimization mechanisms, and the decision-making process lacks multi-stage logical judgment and expert rule verification, making it difficult to meet the high reliability and practicality requirements of clinical decision-making. Summary of the Invention

[0004] The present invention provides a multi-dimensional medical data analysis and decision support system, comprising:

[0005] Data acquisition module, used to obtain structured and unstructured data from multiple source medical devices and databases;

[0006] A data preprocessing module is used to perform noise filtering, missing value filling and standardization on the data output by the data acquisition module;

[0007] Feature extraction module, used to extract multidimensional features from preprocessed data through dynamic weight allocation algorithm;

[0008] Model training module, used to train the adaptive decision model based on the extracted multi-dimensional features and update the model parameters using a dynamic weight optimization algorithm;

[0009] Multidimensional analysis module, used to perform spatial distribution analysis, temporal evolution analysis, and correlation analysis on feature data;

[0010] The decision generation module is used to generate clinical decision recommendations based on multidimensional analysis results and verify the reliability of the decision through multi-stage logical judgment;

[0011] Feedback optimization module, used to adjust model parameters and weight distribution strategies based on actual application feedback;

[0012] The feedback optimization module includes:

[0013] The parameter correction unit reversely adjusts the penalty term weight of the model loss function according to the decision execution results;

[0014] The weight decay unit exponentially decays the feature weights that have not been updated for a long time according to the decay rate and the time interval of non-update;

[0015] The version iteration unit generates a new model version when the cumulative error exceeds a preset threshold.

[0016] As a further improvement of the present application, the data acquisition module includes:

[0017] Multi-source interface units for connecting to electronic medical record systems, medical imaging equipment, and wearable sensors;

[0018] Dynamic cleaning unit, used to filter invalid data in real time based on preset data quality thresholds;

[0019] Real-time storage unit, used to establish a distributed index storage structure based on data type and timestamp.

[0020] As a further improvement of the present application, the data preprocessing module includes:

[0021] A timing alignment unit is used to interpolate and align multi-source asynchronous data according to a unified time base;

[0022] Anomaly correction unit, used to dynamically correct data that exceeds the range of the current data window mean plus or minus three times the standard deviation. The corrected data is generated based on the relationship between the dynamic adjustment coefficient and the mean and standard deviation;

[0023] Normalization unit, used to map data to the range of zero to one.

[0024] As a further improvement of the present application, the specific steps of the dynamic weight allocation algorithm include:

[0025] The sum of the absolute values ​​of the correlation coefficients between each feature dimension and the clinical indicator was calculated and divided by the absolute value of the maximum correlation coefficient to obtain the feature importance score;

[0026] According to the feature importance score, dynamic weights are assigned through exponential function normalization;

[0027] The weighted features are input into the convolutional attention network for fusion.

[0028] As a further improvement of the present application, the parameter updating process of the dynamic weight optimization algorithm includes:

[0029] Based on the adaptive learning rate, momentum factor and regularization coefficient, the model parameters are dynamically adjusted in combination with historical parameter changes and loss function gradient.

[0030] As a further improvement of the present application, the multi-stage logical judgment includes:

[0031] In the initial judgment stage, matching the preset clinical pathway rules;

[0032] In the confidence evaluation phase, the decision confidence is calculated by the average difference between the model prediction probability and the historical verification label;

[0033] In the expert rule verification stage, when the confidence level is lower than the preset threshold, the expert knowledge base is called for secondary verification.

[0034] As a further improvement of the present application, the spatial distribution analysis includes:

[0035] Identify data abnormal clustering areas through density clustering algorithm that dynamically adjusts neighborhood radius;

[0036] Generate a 3D visualization map of feature distribution density;

[0037] Establish a pathological feature transfer matrix between different spatial regions.

[0038] As a further improvement of the present application, the density clustering algorithm includes:

[0039] Dynamically adjust the neighborhood radius based on the average distance, distance standard deviation, and number of data points;

[0040] A bidirectional linkage strategy was used to merge overlapping clusters;

[0041] Make probabilistic attribution judgment on boundary points.

[0042] A multi-dimensional medical data analysis and decision-making support method includes the following steps:

[0043] Acquire structured and unstructured data from multiple source medical devices and databases;

[0044] Perform noise filtering, missing value filling and standardization on the acquired structured and unstructured data;

[0045] Extract multidimensional features from preprocessed data through dynamic weight assignment algorithm;

[0046] The adaptive decision model is trained based on the extracted multi-dimensional features, and the model parameters are updated using a dynamic weight optimization algorithm;

[0047] Conduct spatial distribution analysis, temporal evolution analysis, and correlation analysis on feature data;

[0048] Generate clinical decision recommendations based on multidimensional analysis results, and verify the reliability of the decisions through multi-stage logical judgment;

[0049] Adjust model parameters and weight distribution strategies based on actual application feedback.

[0050] The above-described embodiments of the present invention have at least the following beneficial effects: the system can effectively integrate multi-source heterogeneous medical data, accurately extract multidimensional features through dynamic weight allocation and optimization algorithms, and train adaptive decision models, thereby improving the accuracy and reliability of medical data analysis. The system can also analyze spatial distribution, temporal evolution, and correlation, providing a comprehensive and scientific basis for clinical decision-making and effectively assisting medical staff in developing more precise treatment plans.

[0051] The system also verifies the reliability of its decisions through multi-stage logical judgment based on actual application feedback, dynamically adjusts model parameters and weight distribution strategies, and continuously optimizes system performance. Furthermore, the system boasts efficient data acquisition, preprocessing, and storage capabilities, enabling real-time filtering of invalid data, filling in missing values, performing standardization, and establishing a distributed index storage structure to ensure data integrity and availability, providing strong support for medical data analysis and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0053] Figure 1 A schematic diagram of the structure of a multi-dimensional medical data analysis and decision support system provided by one embodiment of the present invention;

[0054] Figure 2 A flowchart of a multi-dimensional medical data analysis and decision support method provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0055] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0056] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0057] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0058] Example 1: Please refer to Figure 1 , Figure 1 This is a schematic diagram of the structure of a multi-dimensional medical data analysis and decision support system provided by an embodiment of the present invention. Figure 1 As shown, a multi-dimensional medical data analysis and decision support system 100 includes:

[0059] The data acquisition module 101 is used to obtain structured and unstructured data from multiple medical devices and databases;

[0060] The data preprocessing module 102 is used to perform noise filtering, missing value filling and standardization on the data output by the data acquisition module;

[0061] A feature extraction module 103 is used to extract multidimensional features from the preprocessed data using a dynamic weight allocation algorithm;

[0062] A model training module 104 is used to train an adaptive decision model based on the extracted multi-dimensional features and update model parameters using a dynamic weight optimization algorithm;

[0063] Multidimensional analysis module 105, used to perform spatial distribution analysis, temporal evolution analysis and correlation analysis on feature data;

[0064] Decision generation module 106, used to generate clinical decision recommendations based on multi-dimensional analysis results and verify the reliability of the decision through multi-stage logical judgment;

[0065] The feedback optimization module 107 is used to adjust the model parameters and weight distribution strategy according to actual application feedback.

[0066] It should be noted that embodiments of the present invention provide a multidimensional medical data analysis and decision support system, which primarily includes a data acquisition module, a data preprocessing module, a feature extraction module, a model training module, a multidimensional analysis module, a decision generation module, and a feedback optimization module. The data acquisition module is used to acquire structured and unstructured data from multiple medical devices and databases, including electronic medical record systems, medical imaging devices, and wearable sensors. The databases encompass various medical-related databases. The data preprocessing module performs noise filtering, missing value filling, and standardization on the collected data to ensure data quality and usability. The feature extraction module uses a dynamic weight allocation algorithm to extract multidimensional features from the preprocessed data, providing a foundation for subsequent model training. The model training module trains an adaptive decision model based on the extracted multidimensional features and uses a dynamic weight optimization algorithm to update model parameters to improve model accuracy and adaptability. The multidimensional analysis module performs spatial distribution analysis, temporal evolution analysis, and correlation analysis on the feature data, providing comprehensive data support for decision generation. The decision generation module generates clinical decision recommendations based on multidimensional analysis results and verifies the reliability of the decisions through multi-stage logical judgment to ensure the scientific and practical nature of the decisions. The feedback optimization module adjusts model parameters and weight allocation strategies based on actual application feedback to continuously optimize system performance.

[0067] Specifically, the multi-source interface unit in the data acquisition module is responsible for connecting various medical devices and databases to ensure stable data transmission. The dynamic cleaning unit filters invalid data in real time based on preset data quality thresholds to ensure data accuracy and integrity. The real-time storage unit establishes a distributed index storage structure based on data type and timestamp to facilitate rapid data retrieval and management. The timing alignment unit in the data preprocessing module interpolates and aligns multi-source asynchronous data to a unified time base to ensure temporal consistency. The anomaly correction unit uses a specific formula to correct data that deviates from the normal range. The mean and standard deviation of the current data window are used to determine whether the data is abnormal, and the dynamic adjustment coefficient is adjusted based on the actual situation. The normalization unit maps the data to the [0, 1] interval to eliminate dimensional differences between different features. The dynamic weight assignment algorithm in the feature extraction module calculates the importance score of each feature dimension, dynamically assigns weights based on the score, and inputs the weighted features into the convolutional attention network for fusion to extract more representative features. The dynamic weight optimization algorithm in the model training module uses a specific parameter update formula combined with adaptive learning rate, momentum factor, and regularization coefficient to optimize model parameters and improve model training performance. The spatial distribution analysis in the multidimensional analysis module uses an improved DBSCAN algorithm to identify areas of abnormal data clustering and generate a three-dimensional visualization map. It also establishes a pathological feature transfer matrix between different spatial regions, providing intuitive data support for clinical decision-making. The multi-stage logical judgment in the decision-making module includes an initial judgment stage, a confidence assessment stage, and an expert rule verification stage, ensuring the reliability of decisions through step-by-step verification. The parameter correction unit in the feedback optimization module reversely adjusts the penalty term weight of the model loss function based on the decision execution results. The weight decay unit decays the feature weights that have not been updated for a long time. The version iteration unit generates a new model version when the cumulative error exceeds a threshold to continuously optimize system performance.

[0068] Preferably, the multi-source interface unit in the data acquisition module can use multiple communication protocols to connect to medical devices and databases to meet the data transmission requirements of different devices and systems. The preset data quality threshold of the dynamic cleaning unit can be adjusted according to the actual data situation to ensure the accuracy and effectiveness of data filtering. The distributed index storage structure of the real-time storage unit can adopt multiple indexing methods, such as time partitioning layer, feature encoding layer, and fast retrieval layer, to improve the efficiency of data storage and retrieval. The time series alignment unit in the data preprocessing module can adopt multiple interpolation methods, such as linear interpolation and spline interpolation, to meet the time alignment requirements of different data. The dynamic adjustment coefficient of the anomaly correction unit can be adjusted according to the data distribution and actual needs to improve the accuracy of anomaly data correction. The normalization unit can adopt multiple normalization methods, such as minimum-maximum normalization and Z-score normalization, to meet the normalization requirements of different data. The dynamic weight allocation algorithm in the feature extraction module can adopt multiple weight allocation strategies.

[0069] In some embodiments, the data acquisition module includes:

[0070] Multi-source interface units for connecting to electronic medical record systems, medical imaging equipment, and wearable sensors;

[0071] Dynamic cleaning unit, used to filter invalid data in real time based on preset data quality thresholds;

[0072] Real-time storage unit, used to establish a distributed index storage structure based on data type and timestamp.

[0073] It should be noted that the data acquisition module plays a crucial role in this system. It is responsible for acquiring structured and unstructured data from various medical data sources, providing a foundation for subsequent data analysis and decision-making. The multi-source interface unit is a key component of the data acquisition module. It connects to medical devices and systems such as electronic medical record systems, medical imaging equipment, and wearable sensors through various interface protocols to ensure stable data transmission and acquisition. The dynamic cleaning unit monitors and filters invalid data in real time based on preset data quality thresholds to ensure data accuracy and reliability. The real-time storage unit establishes a distributed index storage structure based on data type and timestamp, facilitating rapid data retrieval and management, and providing support for subsequent data preprocessing and analysis.

[0074] In some embodiments, the data acquisition module is configured to achieve stable integration with CT / MRI imaging devices, wearable monitoring devices, and electronic medical record databases through a medical device protocol adaptation layer that integrates DICOM 3.0, HL7, and FHIR protocol conversion interfaces. During the data acquisition process, a two-layer filtering mechanism based on a rule engine and LSTM anomaly detection is implemented to ensure high data quality.

[0075] The first level of filtering uses a Bayesian classifier to perform confidence assessment on unstructured text data. The unstructured text data here mainly includes doctors' handwritten notes, imaging diagnostic reports, etc. The Bayesian classifier calculates the probability that each piece of unstructured text data is valid data based on the prior probability and sample data, and sets a confidence threshold. Data below this threshold will be initially screened. For example, if the confidence threshold is set to 0.6, when the Bayesian classifier calculates that the confidence of a certain doctor's handwritten note is 0.5, then the note will enter the next step of processing or be directly discarded, depending on the rules set subsequently.

[0076] The second layer of filtering uses a time series similarity detection algorithm to identify abnormal device data streams. Based on technologies such as dynamic time warping (DTW), this algorithm analyzes collected device data in time series and compares it to the time series patterns of normal device data. If the similarity between a particular data stream and the normal pattern falls below a preset similarity threshold, the data stream is considered abnormal. For example, heart rate data from a wearable monitoring device should normally exhibit a certain fluctuation pattern. If the algorithm detects that the fluctuation pattern of a particular heart rate data segment differs significantly from the normal pattern and the similarity falls below 0.7 (a hypothetical threshold), the heart rate data stream is considered abnormal.

[0077] Specifically, the multi-source interface unit needs to have multiple interface protocols to adapt to the data transmission requirements of different medical devices and systems. For example, electronic medical record systems usually use standard protocols such as HL7 and DICOM for data transmission, while wearable sensors may use wireless communication protocols such as Bluetooth and Wi-Fi. The preset data quality threshold of the dynamic cleaning unit can be adjusted according to the actual data situation to ensure the accuracy and effectiveness of data filtering. For example, for heart rate data, a reasonable heart rate range can be set. Data outside this range will be regarded as invalid data and filtered. The distributed index storage structure of the real-time storage unit can be designed according to the characteristics of the data and query requirements. For example, for time series data, a time partition layer can be used for storage to improve data retrieval efficiency.

[0078] Preferably, the multi-source interface unit can adopt a modular design, flexibly configuring corresponding interface modules according to the interface requirements of different medical devices and systems to improve the compatibility and scalability of the system. The dynamic cleaning unit can be combined with machine learning algorithms to conduct real-time assessment and prediction of data quality, automatically adjusting data quality thresholds to adapt to changes in different data sources and data types. The real-time storage unit can use distributed database technologies such as Hadoop and Cassandra to achieve high data availability and high scalability. At the same time, it can combine data compression, caching and other technologies to optimize data storage and retrieval performance.

[0079] In some embodiments, the data preprocessing module includes:

[0080] A timing alignment unit is used to interpolate and align multi-source asynchronous data according to a unified time base;

[0081] The anomaly correction unit is used to dynamically correct data that exceeds the range of the current data window mean plus or minus three times the standard deviation. The corrected data is generated based on the relationship between the dynamic adjustment coefficient and the mean and standard deviation; the normalization unit is used to map the data to the range of zero to one.

[0082] Furthermore, the abnormality correction unit uses the following formula to correct the data that deviates from the normal range:

[0083]

[0084] in, is the corrected data, x is the original data, is the mean of the current data window, is the standard deviation, k is the dynamic adjustment coefficient;

[0085] It should be noted that the data preprocessing module plays a crucial role in this system. It is responsible for filtering noise, filling missing values, and standardizing the data output by the data acquisition module to ensure data quality and availability. The timing alignment unit interpolates and aligns multi-source asynchronous data according to a unified time base, ensuring temporal consistency of data from different sources. The anomaly correction unit uses specific formulas to correct data that deviates from the normal range to eliminate the impact of abnormal data on analysis results. The normalization unit maps data to the [0, 1] interval, eliminating dimensional differences between different features and providing standardized data for subsequent feature extraction and model training.

[0086] Specifically, the timing alignment unit uses interpolation methods to align data with different timestamps to a unified time base. For example, if heart rate data and blood pressure data have inconsistent timestamps, the timing alignment unit can use methods such as linear interpolation or spline interpolation to align them to a per-minute or per-second time base. In the formula for the anomaly correction unit, the mean and standard deviation of the current data window are used to determine whether the data is abnormal, and the dynamic adjustment coefficient is adjusted based on the actual situation. For example, for blood glucose data, if the mean of the current data window is 5.0 mmol / L and the standard deviation is 0.5 mmol / L, blood glucose values ​​exceeding 7.5 mmol / L or falling below 2.5 mmol / L are considered abnormal and corrected. The normalization unit maps data to the [0, 1] interval. For example, for age data, minimum-maximum normalization can be used to map age values ​​to the [0, 1] interval, i.e., the normalized age value = (age - minimum age) / (maximum age - minimum age).

[0087] Preferably, the timing alignment unit can adopt a variety of interpolation methods, such as linear interpolation, spline interpolation, nearest neighbor interpolation, etc., to adapt to the time alignment requirements of different data. The dynamic adjustment coefficient of the anomaly correction unit can be adjusted according to the distribution of the data and actual needs. For example, for certain key indicators, the dynamic adjustment coefficient can be appropriately increased to improve the accuracy of abnormal data correction. The normalization unit can adopt a variety of normalization methods, such as minimum-maximum normalization, Z-score normalization, decimal calibration normalization, etc., to meet the normalization requirements of different data. In addition, the data preprocessing module can also include a data cleaning unit for removing duplicate data, abnormal data and irrelevant data to further improve the quality of the data.

[0088] In some embodiments, the specific steps of the dynamic weight allocation algorithm include:

[0089] Step S31: Calculate the sum of the absolute values ​​of the correlation coefficients between each feature dimension and the clinical indicator, and divide it by the absolute value of the maximum correlation coefficient to obtain the feature importance score;

[0090] Step S32: assigning dynamic weights based on feature importance scores through exponential function normalization;

[0091] Step S33: Input the weighted features into the convolutional attention network for fusion.

[0092] Furthermore, the step S31 uses the following formula to obtain the feature importance score:

[0093]

[0094] in, is the correlation coefficient between the i-th feature and the j-th clinical indicator, is the importance score of the i-th feature, and n is the total number of clinical indicators;

[0095] The step S32 uses the following formula to normalize and allocate dynamic weights:

[0096]

[0097] in, is the dynamic weight of the i-th feature, is the importance score of the kth feature, and m is the total number of features.

[0098] It should be noted that the dynamic weight allocation algorithm in the feature extraction module is one of the key technologies of this system. It calculates the importance score of each feature dimension and dynamically allocates weights according to the score, thereby realizing the effective extraction and fusion of multi-dimensional features. Among them, the importance score of the feature dimension is determined by calculating the correlation coefficient between the i-th feature and the j-th clinical indicator. The larger the absolute value of the correlation coefficient, the stronger the correlation between the feature and the clinical indicator, and the higher its importance score. The dynamic weight is obtained after normalization based on the importance score of each feature, and is used to measure the relative importance of each feature in the model. By inputting the weighted features into the convolutional attention network for fusion, more representative and discriminative features can be further extracted, providing strong support for subsequent model training and decision generation.

[0099] Specifically, when calculating the importance score of each feature dimension, it is necessary to traverse the correlation coefficients of all features and all clinical indicators, sum the absolute values ​​of the correlation coefficients, and then divide them by the total number of clinical indicators to obtain the importance score of each feature. In terms of specific parameter settings, the total number of clinical indicators should be determined according to actual clinical needs and data conditions. For example, it can include multiple indicators such as heart rate, blood pressure, and blood sugar. When determining the dynamic weight, the importance score of each feature is normalized, that is, the importance score of each feature is divided by the sum of the importance scores of all features to obtain the dynamic weight of each feature. These weights reflect the relative importance of each feature in the model. The larger the weight value, the greater the contribution of the feature to the model. In addition, the convolutional attention network is a deep learning model that combines convolutional neural networks and attention mechanisms. It can effectively extract local and global information of features and further improve the expressive power of features.

[0100] Preferably, when calculating the correlation coefficient between features and clinical indicators, a variety of correlation measurement methods can be used, such as the Pearson correlation coefficient, the Spearman rank correlation coefficient, etc., to adapt to different data types and distribution characteristics. For data with obvious time series characteristics, it is also possible to consider using measurement methods such as dynamic time warping distance to better capture the temporal correlation between features and clinical indicators. When inputting weighted features into the convolutional attention network for fusion, a variety of feature fusion strategies can be used, such as element-by-element multiplication, element-by-element addition, splicing, etc., to achieve effective interaction and fusion between different features. In addition, the pre-trained convolutional attention network model can be applied to this system in combination with transfer learning technology to improve the training efficiency and generalization ability of the model.

[0101] In some embodiments, the parameter update process of the dynamic weight optimization algorithm includes:

[0102] Based on the adaptive learning rate, momentum factor and regularization coefficient, the model parameters are dynamically adjusted in combination with the historical parameter changes and the loss function gradient. The parameter update formula of the dynamic weight optimization algorithm is:

[0103]

[0104] in, is the model parameter of the t+1th iteration, is the model parameter of the tth iteration, is the model parameter of the t-1th iteration, t is the number of iterations, is the adaptive learning rate, is the momentum factor, is the regularization coefficient, L is the loss function, is the gradient of the loss function at the i-th iteration;

[0105] It should be noted that the dynamic weight optimization algorithm in the model training module is a key technology for improving model performance. Its core lies in continuously iteratively updating model parameters to minimize the loss function and improve the model's generalization ability. The algorithm's parameter update formula comprehensively considers factors such as the parameters of the previous iteration, the gradient of the loss function, the adaptive learning rate, the momentum factor, and the regularization coefficient. Among them, the number of iterations is used to record the number of rounds of the current iteration, and the adaptive learning rate is dynamically adjusted according to the iterative process to ensure that the model can converge quickly in the early stages of training, while fine-tuning is performed in the later stages. The momentum factor is used to accelerate the convergence rate during the gradient descent process, the regularization coefficient helps prevent model overfitting, and the loss function is used to measure the difference between the model's predicted value and the actual value. Through the reasonable setting and optimization of the above parameters, the model can more accurately learn complex patterns in the data, thereby improving the reliability of clinical decision-making.

[0106] Specifically, the various parameters in the parameter update formula need to be carefully set according to the actual data and model training situation. For example, the adaptive learning rate can be dynamically adjusted according to the changes in the loss function. The initial value is usually set to 0.01 or 0.001, and is gradually reduced according to the set attenuation strategy during training. The momentum factor generally takes a value between 0.5 and 0.9, which is used to accelerate convergence and reduce oscillations. The regularization coefficient is adjusted according to the complexity of the data and the scale of the model, and is usually selected between 0.0001 and 0.1. The loss function can take various forms such as mean square error and cross entropy. The specific choice depends on the task type and data characteristics of the model. In addition, the setting of the number of iterations needs to balance the training time and performance of the model. The optimal number of iterations is usually determined by the performance on the validation set to avoid overfitting or underfitting.

[0107] Preferably, when implementing the dynamic weight optimization algorithm, various strategies can be employed to further enhance the model's training performance. For example, second-order derivative information, such as RMSprop or the Adam optimizer, can be introduced to more effectively adjust the learning rate and improve the model's ability to navigate complex loss landscapes. Furthermore, an early stopping strategy can be employed to terminate training early when performance on the validation set has not significantly improved over several consecutive iterations, thereby conserving computing resources and preventing overfitting. Transfer learning can also be employed, using the parameters of a pre-trained model as the initial parameters and fine-tuning them to adapt to a specific medical dataset, thereby accelerating the model's convergence and improving its generalization performance.

[0108] In some embodiments, the multi-stage logical judgment includes:

[0109] Initial judgment stage: matching preset clinical pathway rules;

[0110] Confidence evaluation stage: Calculate the decision confidence by the average difference between the model prediction probability and the historical verification label;

[0111] Expert rule verification stage: When the confidence level is lower than the preset threshold, the expert knowledge base is called for secondary verification.

[0112] Furthermore, in the confidence evaluation stage, the decision confidence is calculated using the following formula:

[0113]

[0114] Among them, pk is the model's predicted probability for the k-th sample, yk is the historical verification label of the k-th sample, and k is the total number of historical verification samples;

[0115] Expert rule verification phase: When When the value is 0.85, the expert knowledge base is called for secondary verification.

[0116] It should be noted that the multi-stage logical judgment in the decision generation module is an important link to ensure the reliability of clinical decision recommendations. The process includes the initial judgment stage, the confidence assessment stage and the expert rule verification stage. In the initial judgment stage, the system matches the preset clinical pathway rules according to the decision category output by the model to ensure the consistency of the decision with the clinical standard process. The confidence assessment stage quantifies the model's confidence level in the decision by calculating the decision confidence. The confidence calculation is based on the difference between the model's predicted probability of the sample and the historical verification label. When the confidence is lower than the set threshold (such as 0.85), the system will enter the expert rule verification stage and call the expert knowledge base for secondary verification to ensure the accuracy and reliability of the decision.

[0117] Specifically, the clinical pathway rules in the initial judgment stage are a series of decision rules preset based on clinical guidelines and expert experience, which are used to guide the decision categories output by the model. These rules can include standard procedures for disease diagnosis, criteria for selecting treatment options, etc. In the calculation formula of the confidence assessment stage, the predicted probability reflects the model's decision confidence for each sample, the historical verification label is an accurate label based on previous data verification, and the total number of samples is used to calculate the average confidence. The trigger condition for the expert rule verification stage is that the confidence level is lower than 0.85. This threshold can be adjusted according to actual application requirements and model performance. The expert knowledge base contains rich clinical experience and rules to provide additional decision support when confidence is insufficient.

[0118] In some embodiments, the decision generation module implements a three-level verification mechanism:

[0119] Phase 1 Verification: A causal inference engine is used to check the logical consistency of the decision recommendation with the clinical data. Based on technologies such as Bayesian networks, the causal inference engine constructs a causal relationship model between clinical data. The decision recommendation is fed into this model to verify that it aligns with the causal logic within the clinical data. For example, if the decision recommendation is to administer a certain medication to a patient, but the patient's history of allergies indicates an allergy to that medication, the causal inference engine will determine that the decision recommendation is inconsistent with the clinical data.

[0120] Phase II validation: Evidence-based medicine compliance assessment is conducted based on a knowledge graph (including structured data from UpToDate and NCCN guidelines). The knowledge graph integrates authoritative medical guidelines and clinical research findings, matching and comparing decision recommendations with information in the knowledge graph. For example, for a treatment decision for a specific disease, the decision is checked for compliance with the treatment options recommended in UpToDate and NCCN guidelines. The similarity between the decision recommendation and the guideline content is calculated. If the similarity falls below a preset threshold (e.g., 0.8), the decision recommendation is considered to have insufficient evidence-based medicine compliance.

[0121] The third phase of verification involves invoking the expert review interface and manually confirming high-risk decisions using a double-blind voting mechanism. When a decision is determined to be high-risk (e.g., involving major surgery or the use of high-risk medications), the expert review process is initiated. Multiple medical experts are invited to evaluate the decision. These experts vote without knowing the specific source of the decision and the opinions of other experts. If a majority of experts (e.g., more than two-thirds of the experts) agree with the decision, the decision passes verification; otherwise, it requires further adjustment or reassessment.

[0122] Preferably, when implementing multi-stage logical judgment, the parameters and processes of each stage can be optimized. For example, in the initial judgment stage, clinical pathway rules can be regularly updated to incorporate the latest clinical research results and treatment guidelines. In the confidence assessment stage, in addition to calculating the overall confidence, the confidence of each sample can also be individually evaluated to identify potential misjudgment samples. In addition, more expert rule verification mechanisms can be introduced, such as machine learning-based anomaly detection models, to further improve the reliability of decision-making. When the confidence level is lower than the threshold, the system can also automatically record relevant samples and decision information for subsequent analysis and model optimization.

[0123] In some embodiments, the feedback optimization module performs the following operations:

[0124] Parameter correction unit: reversely adjusts the penalty term weight of the model loss function according to the decision execution results;

[0125] Weight decay unit: exponentially decays the feature weights that have not been updated for a long time according to the decay rate and the time interval of non-update;

[0126] Version iteration unit: Generates a new model version when the cumulative error exceeds a preset threshold.

[0127] Furthermore, the weight decay unit uses the following formula to decay the feature weights that have not been updated for a long time:

[0128]

[0129] in, is the updated weight of the i-th feature, is the weight of the i-th feature before updating, is the decay rate, and t is the time interval without update.

[0130] It should be noted that the feedback optimization module in this system is responsible for adjusting model parameters and weight distribution strategies based on actual application feedback to continuously optimize system performance. This module includes a parameter correction unit, a weight decay unit, and a version iteration unit. The parameter correction unit reversely adjusts the penalty term weight of the model loss function based on the decision execution results to correct the model's deviations in the decision-making process. The weight decay unit decays the weights of features that have not been updated for a long time to reduce the impact of these features on the model and prevent overfitting. The version iteration unit generates a new model version when the cumulative error exceeds a preset threshold to ensure continuous optimization and performance improvement of the system.

[0131] Specifically, the parameter correction unit adjusts the penalty item weight in the model loss function by analyzing the difference between the decision execution result and the expected result. For example, if a certain feature causes a large error in the decision-making process, the penalty item weight of the feature will be increased to reduce its negative impact on the model. The weight decay unit uses a specific decay formula to adjust the weight of features that have not been updated for a long time. The decay rate and the non-update time interval are key parameters of the unit and can be set according to actual application requirements. For example, for features that have not been updated within a month, the decay rate can be set to 0.1, that is, the weight decays by 10% every month. The version iteration unit monitors the cumulative error of the model and triggers the generation of a new model version when the error exceeds the preset threshold. The preset threshold can be adjusted according to the performance requirements and application scenarios of the model. For example, it can be set to 0.05 or 0.1.

[0132] Preferably, when implementing the feedback optimization module, the operating steps of each unit can be further refined or alternative solutions can be provided. For example, the parameter correction unit can adopt an adaptive adjustment strategy to dynamically adjust the weight of the penalty term according to the frequency and error size of the decision execution result, rather than just a simple reverse adjustment. The weight decay unit can perform differentiated attenuation based on the importance of the feature, using a lower decay rate for important features and a higher decay rate for secondary features. The version iteration unit can introduce a model performance evaluation mechanism to conduct a comprehensive evaluation of the current model before generating a new model version to ensure that the generation of the new version is necessary and effective. In addition, a model version rollback mechanism can be set up. When the performance of the new version is not as good as the old version, it can be rolled back to the old version in time to ensure the stability and reliability of the system.

[0133] In some embodiments, the spatial distribution analysis includes:

[0134] Identify data abnormal clustering areas through density clustering algorithm that dynamically adjusts neighborhood radius;

[0135] Generate a 3D visualization map of feature distribution density;

[0136] Establish a pathological feature transfer matrix between different spatial regions.

[0137] Specifically, the improved DBSCAN algorithm can be used to identify abnormal data clustering areas;

[0138] Generate a three-dimensional visualization map based on the feature distribution density; establish a pathological feature transfer matrix between different spatial regions.

[0139] It should be noted that the spatial distribution analysis within the multidimensional analysis module is a crucial component of this system. It aims to reveal the spatial distribution characteristics and underlying patterns of data through a variety of analytical methods. This analysis includes a density clustering unit, a heat map generation unit, and an association mapping unit. The density clustering unit uses an improved DBSCAN algorithm to identify areas of abnormal data clustering. The heat map generation unit generates a three-dimensional visualization based on the density of feature distributions. The association mapping unit establishes a pathological feature transfer matrix between different spatial regions, providing intuitive data support for clinical decision-making.

[0140] Specifically, the DBSCAN algorithm used by the density clustering unit is a density-based clustering algorithm that can effectively identify clusters of any shape. The improved DBSCAN algorithm better adapts to data distributions of different densities by dynamically adjusting the neighborhood radius. The heat map generation unit generates a three-dimensional visualization map based on the feature distribution density, where color or height represents data density, helping users to intuitively observe the data distribution. The transfer matrix established by the association mapping unit is used to describe the changes in pathological features between different spatial regions. The elements in the matrix represent the probability or intensity of feature transitions from one region to another.

[0141] Preferably, when implementing spatial distribution analysis, the operating steps of each unit can be further refined or alternatives can be provided. For example, the improved DBSCAN algorithm of the density clustering unit can be combined with other clustering algorithms, such as K-means or Gaussian Mixture Models, to improve the accuracy and robustness of clustering. The heat map generation unit can use different visualization techniques, such as two-dimensional heat maps, contour maps, or three-dimensional surface maps, to adapt to different data characteristics and user needs. The transfer matrix of the association mapping unit can be combined with time series information to construct a dynamic transfer matrix to capture the changing trend of pathological characteristics over time. In addition, machine learning algorithms, such as random forests or support vector machines, can be introduced to further analyze and predict the clustering results and transfer matrices to provide deeper insights and decision support.

[0142] In some embodiments, the density clustering algorithm includes:

[0143] Dynamically adjust the neighborhood radius based on the average distance, distance standard deviation, and number of data points:

[0144] A bidirectional linkage strategy was used to merge overlapping clusters;

[0145] Make probabilistic attribution judgment on boundary points.

[0146] Furthermore, the neighborhood radius is dynamically adjusted using the following formula:

[0147]

[0148] in, is the average distance, is the standard deviation of the distance, and N is the number of data points.

[0149] It should be noted that the improved DBSCAN algorithm is the core technology for spatial distribution analysis in the multidimensional analysis module, used to identify anomalous clusters in data. This algorithm improves clustering accuracy and robustness by dynamically adjusting the neighborhood radius, merging overlapping clusters using a bidirectional linkage strategy, and performing probabilistic attribution determination on boundary points. In the formula for dynamically adjusting the neighborhood radius, the average distance and standard deviation of the distance are used to measure the distribution density of data points, while the number of data points is used to adjust the size of the neighborhood radius to accommodate clusters of varying densities. The bidirectional linkage strategy effectively merges overlapping clusters by establishing bidirectional links during the clustering process, avoiding duplicate assignment of data points. Probabilistic attribution determination assigns a probability value to boundary points, indicating the probability of them belonging to a particular cluster, thereby improving the reliability of the clustering results.

[0150] Specifically, in the formula for dynamically adjusting the neighborhood radius, the average distance can be obtained by calculating the average Euclidean distance between all data points, and the standard deviation of the distance is obtained by calculating the standard deviation of the distance between all data points. The number of data points refers to the number of data points in the current clustering area. In the bidirectional link strategy, each data point will establish a link with other data points during the clustering process. When two clustering areas overlap, they are merged into one cluster through bidirectional links. In the probabilistic attribution judgment, the probability value of the boundary point can be determined by calculating its distance from the cluster center and the distance distribution of other data points in the cluster. The closer the probability value is to 1, the more likely the boundary point is to belong to the cluster.

[0151] Preferably, when implementing the improved DBSCAN algorithm, each step can be further refined or alternatives can be provided. For example, when dynamically adjusting the neighborhood radius, the local density characteristics of the data can be combined to adopt a local density estimation method to calculate the neighborhood radius to better adapt to clustering areas of different densities. In the bidirectional link strategy, a weight mechanism can be introduced to assign weights to links according to the similarity between data points, so as to better consider the similarity of data points when merging overlapping clusters. In the probabilistic attribution judgment, the Bayesian method can be used to calculate the posterior probability of a data point belonging to a certain cluster based on the prior probability and likelihood probability of the data point to improve the accuracy of the judgment. In addition, the results of the DBSCAN algorithm can be post-processed in combination with other clustering algorithms, such as K-means or Gaussian Mixture Models, to further improve the accuracy and robustness of the clustering.

[0152] In some embodiments, the real-time storage unit adopts the following index structure:

[0153] Time partition layer: divides data storage blocks into hourly granularity;

[0154] Feature encoding layer: adds a feature hash value to each piece of data;

[0155] Fast retrieval layer: Builds cross-partition joint query indexes based on B+ trees.

[0156] It should be noted that the real-time storage unit in this system is responsible for efficiently storing and managing multi-source medical data. The index structure it employs is crucial for rapid data retrieval and analysis. This index structure comprises a time partitioning layer, a feature encoding layer, and a fast retrieval layer. The time partitioning layer divides data storage blocks into hourly granularity to ensure orderly storage of time series data. The feature encoding layer appends a feature hash value to each piece of data to quickly identify and locate data with the same characteristics. The fast retrieval layer constructs a cross-partition joint query index based on a B+ tree, enabling efficient data retrieval and query operations.

[0157] Specifically, the hourly granularity of the time partitioning layer means that data will be stored and indexed according to hourly time periods. For example, all data from 2024-06-01 00:00:00 to 2024-06-01 00:59:59 will be stored in the same block. This partitioning method helps improve the retrieval efficiency of time series data, especially when performing time series analysis. The feature encoding layer encodes the data's feature information into a fixed-length string by generating a feature hash value for each piece of data. This not only facilitates rapid data identification but also reduces storage space usage when the data volume is large. The fast retrieval layer utilizes the B+ tree index structure to enable fast queries across multiple time partitions. Each node in the B+ tree contains multiple key values ​​and pointers to child nodes. This structure is particularly suitable for the storage and retrieval of large amounts of data.

[0158] Preferably, when implementing a real-time storage unit, the parameters and operating steps of each layer can be further refined or alternative solutions can be provided. For example, the granularity of the time partitioning layer can be adjusted according to the actual data volume and query requirements, such as dividing by day or minute. The hash function of the feature encoding layer can choose different algorithms, such as MD5, SHA-1 or SHA-256, to ensure the uniqueness and security of the hash value. The B+ tree structure of the fast retrieval layer can be optimized based on the distribution characteristics of the data, such as adjusting the order of the tree or the node size to improve retrieval performance. In addition, a caching mechanism can be introduced to cache frequently accessed data to further improve the speed of data retrieval.

[0159] The aforementioned embodiments of the present invention have the following beneficial effects: the system can effectively integrate multi-source heterogeneous medical data. Through dynamic weight allocation and optimization algorithms, it accurately extracts multidimensional features and trains adaptive decision models, improving the accuracy and reliability of medical data analysis. The system can also analyze spatial distribution, temporal evolution, and correlation, providing a comprehensive and scientific basis for clinical decision-making and effectively assisting medical staff in developing more precise treatment plans.

[0160] The system also verifies the reliability of its decisions through multi-stage logical judgment based on actual application feedback, dynamically adjusts model parameters and weight distribution strategies, and continuously optimizes system performance. Furthermore, the system boasts efficient data acquisition, preprocessing, and storage capabilities, enabling real-time filtering of invalid data, filling in missing values, performing standardization, and establishing a distributed index storage structure to ensure data integrity and availability, providing strong support for medical data analysis and decision-making.

[0161] Example 2: Please refer to Figure 2 , Figure 2 This is a flow chart of a multi-dimensional medical data analysis and decision-making assistance method provided by one embodiment of the present invention. Figure 2As shown, a specific embodiment of the present invention discloses a multi-dimensional medical data analysis and decision-making support method, including the following steps:

[0162] S100, which acquires structured and unstructured data from multiple medical devices and databases;

[0163] S200, performing noise filtering, missing value filling and standardization on the acquired structured and unstructured data;

[0164] S300, extracts multidimensional features from preprocessed data using a dynamic weight assignment algorithm;

[0165] S400, trains an adaptive decision model based on the extracted multi-dimensional features and uses a dynamic weight optimization algorithm to update the model parameters;

[0166] S500, performing spatial distribution analysis, temporal evolution analysis, and correlation analysis on the feature data;

[0167] S600 generates clinical decision recommendations based on multidimensional analysis results and verifies the reliability of the decisions through multi-stage logical judgment;

[0168] S700: Adjust model parameters and weight distribution strategy based on actual application feedback.

[0169] It is understandable that the steps and references recorded in the multi-dimensional medical data analysis and decision support method Figure 1 The modules described above correspond to the modules in the multidimensional medical data analysis and decision support system. Therefore, the modules, features, and beneficial effects described above for the multidimensional medical data analysis and decision support system are also applicable to the multidimensional medical data analysis and decision support method and the operations included therein, and will not be repeated here.

[0170] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, server, mobile phone, or tablet.

[0171] The above descriptions merely illustrate some preferred embodiments of the present invention and the underlying technical principles. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A multi-dimensional medical data analysis and decision support system, characterized by: Includes the following modules: Data acquisition module, used to obtain structured and unstructured data from multiple source medical devices and databases; A data preprocessing module is used to perform noise filtering, missing value filling and standardization on the data output by the data acquisition module; Feature extraction module, used to extract multidimensional features from preprocessed data through dynamic weight allocation algorithm; Model training module, used to train the adaptive decision model based on the extracted multi-dimensional features and update the model parameters using a dynamic weight optimization algorithm; Multidimensional analysis module, used to perform spatial distribution analysis, temporal evolution analysis, and correlation analysis on feature data; The decision generation module is used to generate clinical decision recommendations based on multidimensional analysis results and verify the reliability of the decision through multi-stage logical judgment; Feedback optimization module, used to adjust model parameters and weight distribution strategies based on actual application feedback; The feedback optimization module includes: The parameter correction unit reversely adjusts the penalty term weight of the model loss function according to the decision execution results; The weight decay unit exponentially decays the feature weights that have not been updated for a long time according to the decay rate and the time interval of non-update; Version iteration unit, which generates a new model version when the cumulative error exceeds a preset threshold; The multi-stage logical judgment includes: In the initial judgment stage, matching the preset clinical pathway rules; In the confidence evaluation phase, the decision confidence is calculated by the average difference between the model prediction probability and the historical verification label; In the expert rule verification stage, when the confidence level is lower than the preset threshold, the expert knowledge base is called for secondary verification.

2. The system according to claim 1, wherein: The data acquisition module includes: Multi-source interface units for connecting to electronic medical record systems, medical imaging equipment, and wearable sensors; Dynamic cleaning unit, used to filter invalid data in real time based on preset data quality thresholds; Real-time storage unit, used to establish a distributed index storage structure based on data type and timestamp.

3. The system according to claim 1, wherein: The data preprocessing module includes: A timing alignment unit is used to interpolate and align multi-source asynchronous data according to a unified time base; Anomaly correction unit, used to dynamically correct data that exceeds the range of the current data window mean plus or minus three times the standard deviation. The corrected data is generated based on the relationship between the dynamic adjustment coefficient and the mean and standard deviation; Normalization unit, used to map data to the range of zero to one.

4. The system according to claim 3, characterized in that The specific steps of the dynamic weight allocation algorithm include: The sum of the absolute values ​​of the correlation coefficients between each feature dimension and the clinical indicator was calculated and divided by the absolute value of the maximum correlation coefficient to obtain the feature importance score; According to the feature importance score, dynamic weights are assigned through exponential function normalization; The weighted features are input into the convolutional attention network for fusion.

5. The system according to claim 1, wherein: The parameter updating process of the dynamic weight optimization algorithm includes: Based on the adaptive learning rate, momentum factor and regularization coefficient, the model parameters are dynamically adjusted in combination with historical parameter changes and loss function gradient.

6. The system according to claim 1, wherein: The spatial distribution analysis includes: Identify data abnormal clustering areas through density clustering algorithm that dynamically adjusts neighborhood radius; Generate a 3D visualization map of feature distribution density; Establish a pathological feature transfer matrix between different spatial regions.

7. The system according to claim 6, characterized in that The density clustering algorithm includes: Dynamically adjust the neighborhood radius based on the average distance, distance standard deviation, and number of data points; A bidirectional linkage strategy was used to merge overlapping clusters; Make probabilistic attribution judgment on boundary points.

8. A multi-dimensional medical data analysis and decision-making support method, characterized in that: The following steps are involved: Acquire structured and unstructured data from multiple source medical devices and databases; Perform noise filtering, missing value filling and standardization on the acquired structured and unstructured data; Extract multidimensional features from preprocessed data through dynamic weight assignment algorithm; The adaptive decision model is trained based on the extracted multi-dimensional features, and the model parameters are updated using a dynamic weight optimization algorithm; Conduct spatial distribution analysis, temporal evolution analysis, and correlation analysis on feature data; Generate clinical decision recommendations based on multidimensional analysis results, and verify the reliability of the decisions through multi-stage logical judgment; Adjust model parameters and weight distribution strategies based on actual application feedback.

Images