Medical data dynamic management method based on multi-modal data fusion and deep reasoning

Abstract

The invention relates to a medical data dynamic management method based on multi-modal data fusion and deep reasoning, which comprises the following steps: establishing a three-layer dynamic management architecture, collecting various medical data of a patient in real time through a bottom layer, performing primary processing on the various medical data of the patient, receiving the primarily processed medical data through a middle management layer, and performing dynamic management on the medical data through the middle management layer. The double-flow Transform network architecture performs medical event prediction, the event-driven architecture triggers a data processing flow, medical data after primary processing is subjected to dynamic extraction, data conversion, data loading and multi-modal data fusion processing, a data processing result is transmitted to the application layer, the application layer performs dynamic field generation, and the event-driven architecture performs multi-modal data fusion processing. And data automatic dynamic management and real-time feedback calibration are carried out based on a data fusion analysis result. The limitation of a static field structure existing in an existing medical data management SEER database is solved, the data quality is improved, automatic management and real-time feedback calibration of follow-up visit data are achieved, and separation of a follow-up visit system and a data management platform is avoided.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and particularly relates to a dynamic management method for medical data based on multi-modal data fusion and deep reasoning. Background Art

[0002] The SEER database is a currently widely used medical data storage and analysis system, mainly used for cancer research. It stores patients' clinical data through a predefined static field structure, including pathological information, treatment plans, follow-up results, etc. Although this static field structure can meet the basic data recording requirements, when facing precision medicine research, its flexibility and scalability are insufficient, and it is difficult to adapt to the rapidly changing clinical research needs.

[0003] Traditional medical data governance mainly relies on manual review and regular data verification mechanisms. After data is collected, it usually undergoes centralized processing and storage, lacking a real-time logic verification mechanism, resulting in difficulties in timely discovery and correction of data errors and contradictions. The follow-up system is usually separated from the data governance platform, leading to information gaps. In clinical research, patients' follow-up data needs to be manually entered and integrated, increasing the risk of data loss and inconsistency.

[0004] The ICD coding system is an important tool for medical data standardization, but in actual applications, there is a semantic gap between ICD coding and clinical practice. Many clinical terms cannot be directly mapped to ICD codes, resulting in inaccurate data records and information loss. For example, some complex clinical symptoms or treatment plans may not be accurately described by the existing ICD coding system, affecting the integrity and availability of data.

[0005] Therefore, in the medical data management of the prior art, there are many technical problems. On the one hand, the static field structure of the SEER database used cannot meet the needs of precision medicine research. With the in-depth development of medical research, the requirements for data refinement and dynamization are getting higher and higher, and the static field structure is difficult to flexibly expand and adjust, unable to meet the needs of complex clinical research. On the other hand, traditional medical data governance lacks a real-time logic verification mechanism, and data errors and contradictions are difficult to discover and correct in a timely manner, resulting in low data quality and affecting the accuracy and reliability of clinical research. Moreover, the follow-up system is separated from the data governance platform, leading to information gaps. Patients' follow-up data needs to be manually entered and integrated, increasing the risk of data loss and inconsistency, affecting the integrity and continuity of clinical research. Finally, there is a semantic gap between ICD coding and clinical practice, and many clinical terms cannot be directly mapped to ICD codes, resulting in inaccurate data records and information loss, affecting the integrity and availability of data.

[0006] Therefore, how to improve the existing medical data management method, solve the limitations of the static field structure in the SEER database used in the existing medical data management, improve the data processing quality, and achieve the automated management and real-time feedback calibration of follow-up data, avoid the separation of the follow-up system and the data governance platform, and solve the semantic gap between ICD coding and clinical practice are the technical problems that need to be solved urgently at present. Summary of the Invention

[0007] The purpose of the present invention is to provide a dynamic management method for medical data based on multimodal data fusion and deep reasoning, which is used to improve the existing medical data management method, solve the limitations of the static field structure in the SEER database used in the existing medical data management, improve the data processing quality, and achieve the automated management and real-time feedback calibration of follow-up data, avoid the separation of the follow-up system and the data governance platform, and solve the semantic gap between ICD coding and clinical practice.

[0008] To solve the above technical problems, the technical solutions adopted by the present invention are as follows: A dynamic management method for medical data based on multimodal data fusion and deep reasoning includes the following steps: S1: Establish at least a three-layer dynamic governance architecture, including a bottom layer, an intermediate governance layer, and an application layer. The bottom layer is used to realize real-time medical data collection and preliminary processing; a dual-stream Transformer network architecture and an event-driven architecture are constructed in the intermediate governance layer to process the multimodal medical data after preliminary processing; an adaptive database is constructed in the application layer to perform multimodal data analysis based on the processing results of the multimodal medical data in the intermediate layer to achieve dynamic management. The bottom layer is edge computing nodes deployed at various key positions in the hospital, including positions such as outpatient clinics, wards, and operating rooms. The edge computing nodes are close to the data sources, such as medical devices, sensors, etc., and can process and analyze data in real time.

[0009] S2: Real-time collect various medical data of the patient through the bottom layer, perform preliminary processing on the various medical data of the patient. The preliminary processing includes data cleaning, normalization processing, abnormal data processing, and immediate verification processing, and transmit the preliminarily processed data to the intermediate layer; S3: After receiving the preliminarily processed medical data through the intermediate governance layer, the dual-stream Transformer network architecture predicts medical events based on the preliminarily processed medical data, and the event-driven architecture triggers a data processing flow based on the predicted medical events or the occurred medical events, performs dynamic extraction, data conversion, data loading, and multimodal data fusion processing on the preliminarily processed medical data, and transmits the data processing results to the application layer; S4: The application layer generates dynamic fields through an adaptive database based on the data processing results, and performs automated dynamic management and real-time feedback calibration of the data based on the data fusion analysis results.

[0010] Preferably, the dual-stream Transformer network architecture embedded in the intermediate governance layer includes a main stream layer, an auxiliary stream layer, an interaction layer, and an output layer; The main stream layer and the auxiliary stream layer are connected through the interaction layer. The interaction layer is provided with a cross-attention mechanism to realize the fusion of data between the main stream layer and the auxiliary stream layer and the mutual attention and complement between the main stream layer and the auxiliary stream layer. The main stream layer extracts the context information provided by the auxiliary stream layer through the interaction layer to enhance the understanding of the main medical data, and the auxiliary stream layer extracts the time series features of the medical data in the main stream layer through the interaction layer; The output layer is respectively connected to the main stream layer and the auxiliary stream layer. A fully connected layer is provided in the output layer, and the data processing results output by the main stream layer and the auxiliary stream layer are combined by weighted summation and splicing to obtain the final output data.

[0011] Preferably, the process of predicting medical events based on the dual-stream Transformer network architecture in step S3 for the preliminarily processed medical data includes the following specific processes: S31: Extract a specified amount of preliminarily processed medical data and divide it into a training set, a validation set, and a test set. After data annotation of the training set, input it into the model for training, verify the model based on the validation set, and test the model based on the test set; S32: Re-divide the remaining preliminarily processed medical data according to a specified ratio into main stream data and auxiliary stream data. The main stream data is time-series medical event data, and the time-series medical event data is divided into sequence segments of a fixed length as the input of the main stream layer. Feature extraction and encoding are performed on the auxiliary stream data to generate an input corresponding to the main stream layer input for the auxiliary stream layer; S33: The encoder in the main stream layer encodes the input sequence segments of a fixed length to extract the global features in the sequence segments of a fixed length; the encoder of the auxiliary stream layer encodes the auxiliary stream data to extract the context features related to the main stream data; When the main stream layer performs feature extraction, it obtains the context information provided by the auxiliary stream layer based on the interaction layer to enhance the understanding of the main data and realize the extraction of global features. The auxiliary stream layer obtains more accurate time series features of the main stream layer based on the interaction layer to realize the extraction of context features related to the main stream data, and realizes the information interaction and fusion of the global features extracted by the main stream layer and the context features extracted by the auxiliary stream layer; S34: Transmit the global features extracted from the main stream layer and the context features extracted from the auxiliary stream layer to the output layer. The output layer combines them through weighted summation and concatenation of the data processing results output by the main stream layer and the auxiliary stream layer through a fully connected layer, obtains the final medical event prediction result and outputs it.

[0012] Preferably, the preliminarily processed medical data is the time-series medical event data of the patient, including the key nodes in the clinical pathway, such as admission nodes, examination nodes, treatment nodes and follow-up nodes, and the event features corresponding to each key node, including physiological indicators and diagnostic results.

[0013] Preferably, in step S33, the encoder is set with a multi-head self-attention mechanism layer to encode the input time-series event sequence, and the long-distance dependencies in the medical event sequence are captured through the multi-head self-attention mechanism. In each layer of the encoder, an attention mechanism module with clinical pathway constraints is added. The attention mechanism module with clinical pathway constraints adjusts the attention weights according to the predefined clinical pathway rules, including for the standard treatment path of a specified disease, the clinical pathway constraint module will increase the attention weights of events related to the key nodes and reduce the weights of events unrelated to the path; Based on the output of the encoder, the attention mechanism layer further applies the attention mechanism with clinical pathway constraints. The attention mechanism dynamically adjusts the weights of medical events by calculating the matching degree between each medical event and the clinical pathway. The calculation of the matching degree is based on the type, time sequence of the medical event and the degree of association with the key nodes in the path.

[0014] Preferably, the implementation process of the attention mechanism module with clinical pathway constraints is as follows: S331: Collect the time-series medical event data of the patient, including the key nodes and event features in the clinical pathway, and clean, normalize and convert the data format to meet the model input requirements; S332: Initialize the parameters of the encoder in the dual-stream Transformer network architecture, including the weights of the multi-head self-attention mechanism and load the predefined clinical pathway rules for subsequent attention weight adjustment; S333: Send the input time-series event sequence into the encoder. The encoder encodes the events layer by layer. During each layer of encoding, the clinical pathway constraint module adjusts the attention weights according to the path rules and outputs the encoded event data; S334: Based on the output of the encoder, further apply the attention mechanism with clinical pathway constraints, calculate the matching degree between each event and the clinical pathway, and dynamically adjust the attention weights of the events according to the matching degree; S335: Feed the output of the encoder and the adjusted attention weights into the decoder. The decoder generates predictions for subsequent medical events based on its input data. During the decoding process, continue to apply clinical pathway constraints to ensure that the prediction results conform to the logic of the clinical pathway; S336: The prediction results generated by the decoder are converted into specific medical event predictions through the output layer, and the prediction results are applied to clinical decision support.

[0015] Preferably, the specific processes of dynamically extracting, data transforming, data loading, and multi-modal data fusion processing for the preliminarily processed medical data in step S3 are as follows: S35: Use a dynamic ETL engine to extract medical data from multiple heterogeneous data sources that have been preliminarily processed in real time, and perform data mapping and transformation according to the preset ETL specification guidelines during the extraction process. Dynamically adjust the transformation rules according to the type and content of medical events during the data transformation process; S36: Load the processed medical data into a preset database. During the data loading process, the dynamic ETL engine simultaneously performs data verification and conflict resolution to ensure data consistency; S37: Extract features from medical data of different modalities respectively, convert the features of the medical data of different modalities that have been extracted into corresponding feature vectors respectively, and merge the feature vectors of the medical data of different modalities according to specified rules to obtain a fused feature vector.

[0016] Preferably, the specific process of the application layer generating dynamic fields through an adaptive database based on the data processing results in step S4 is as follows: S41: Obtain the respective feature values based on the fused feature vector, perform polynomial expansion on the respective feature values to generate high-level features, and capture the non-linear relationships between the features; S42: Randomly combine different feature values to generate new feature values and capture the interaction relationships between the features; S43: Perform standardization transformation on the obtained respective feature values and the new feature values, and select the feature values greater than the specified information quantity threshold to form a new feature subset.

[0017] Preferably, the encoder uses an intelligent coding model with three-stage coding verification. The intelligent coding model includes an input layer, a knowledge graph embedding layer, a path encoder, and an output layer; The input of the input layer is clinical text data. Use a pre-trained medical text embedding model to encode the text and generate an initial embedding representation of the text; The knowledge graph embedding layer embeds the concepts and relationships in the knowledge graph into a vector space, encodes the knowledge graph using a graph neural network, and obtains the context representation of the concept nodes; The path encoder encodes the paths searched from the knowledge graph, generates the embedded representation of the paths, and includes a multi-head attention mechanism and a feed-forward network for processing the path sequence and integrating features; The output of the output layer is the embedded representation of the clinical concept after semantic parsing, and the embedded representation of the clinical concept will be used for subsequent logical consistency verification.

[0018] The beneficial effects of the present invention include: The medical data dynamic management method based on multi-modal data fusion and deep reasoning provided by the present invention establishes a three-layer dynamic governance architecture. By collecting various medical data of patients in real time at the bottom layer and performing preliminary processing on the various medical data of patients, after the middle governance layer receives the preliminarily processed medical data, the dual-stream Transformer network architecture is used for medical event prediction, and the event-driven architecture triggers the data processing flow, performs dynamic extraction, data conversion, data loading and multi-modal data fusion processing on the preliminarily processed medical data, transmits the data processing result to the application layer, the application layer generates dynamic fields, and performs automated dynamic management and real-time feedback calibration of the data based on the data fusion analysis result. It solves the limitations of the static field structure existing in the existing medical data management SEER database, improves the data quality, realizes the automated management and real-time feedback calibration of follow-up data, and avoids the separation of the follow-up system and the data governance platform.

[0019] First, a three-layer dynamic governance architecture is set up. By realizing real-time collection and real-time processing of data at the bottom layer, the accuracy and real-time nature of various medical data in the collection stage are ensured, providing a real-time and accurate data basis for subsequent medical data dynamic management and ensuring the timeliness of management. The event-driven architecture realizes asynchronous processing of data through a message queue and an event processing engine. The event-driven architecture allows the system to automatically trigger the data processing flow when receiving a medical event.

[0020] Second, the middle governance layer constructs a dual-stream Transformer network architecture and an event-driven architecture. By respectively obtaining the global time dependence relationship and context information through the main stream layer and the auxiliary stream layer of the dual-stream Transformer network architecture, it realizes the simultaneous capture of global features and context features, and combines the multi-head self-attention mechanism to process features in multiple subspaces in parallel, so as to more comprehensively understand the complex relationships in the data and realize the accurate processing of medical data, providing an efficient and accurate network architecture for the management of medical data.

[0021] Thirdly, an interaction layer is set between the mainstream layer and the auxiliary layer of the dual-stream Transformer network architecture to fuse the information of the main and auxiliary streams. The interaction layer adopts a cross-attention mechanism to enable the two streams to pay attention to and complement each other. Through the interaction layer, the mainstream can use the contextual information provided by the auxiliary stream to enhance its understanding of the main data, and the auxiliary stream can also obtain more accurate time series features from the mainstream, thereby enhancing the accuracy of feature extraction. The output layer merges the outputs of the two streams and generates the final output by weighted summation or splicing of the fully connected layer. For medical event prediction tasks, the output layer can output the probability distribution of the event, and can judge whether the event has occurred based on the probability value, thereby enhancing the flexibility of subsequent dynamic management.

[0022] Finally, the encoders in the mainstream layer and auxiliary stream layer use an intelligent coding system with three-stage coding verification. The intelligent coding system adopts a three-stage coding verification process, combined with the UMLS knowledge graph and SWRL rules, and optimizes the coding process through a reinforcement learning reward mechanism, thus solving the semantic gap problem between ICD coding in existing technology and clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flowchart of the medical data dynamic management method based on multimodal data fusion and deep reasoning of the present invention.

[0024] Figure 2 Schematic diagram of the dual-stream Transformer network architecture of the present invention. DETAILED DESCRIPTION

[0025] The following is combined with Figures 1 to 2 The present invention is further described in detail: Example 1 See attached Figure 1 As shown, the medical data dynamic management method based on multimodal data fusion and deep reasoning includes the following steps: S1: Establish at least three layers of dynamic governance architecture, including the bottom layer, the middle governance layer and the application layer. The bottom layer is used to realize real-time medical data collection and preliminary processing; a dual-stream Transformer network architecture and event-driven architecture are built in the middle governance layer to process the multimodal medical data after preliminary processing; an adaptive database is built in the application layer to perform multimodal data analysis based on the processing results of the multimodal medical data in the middle layer to achieve dynamic management. The bottom layer is the edge computing nodes deployed in various key locations in the hospital, including outpatient clinics, wards, operating rooms, etc. Edge computing nodes are close to data sources, such as medical equipment, sensors, etc., and can process and analyze data in real time.

[0026] The edge computing node collects data in real time from medical devices and information systems using sensors, monitors, or software tools. After the data is encapsulated, a timestamp is added, and a time synchronization mechanism is used to ensure the accuracy of the data. The best path is selected for data transmission according to the network status and optimization objectives.

[0027] S2: Collect various medical data of the patient in real time through the underlying layer, and perform preliminary processing on the various medical data of the patient. The preliminary processing includes data cleaning, normalization processing, abnormal data processing, and instant verification processing, and transmit the preliminarily processed data to the middle layer.

[0028] S3: After receiving the preliminarily processed medical data through the middle governance layer, the dual-stream Transformer network architecture performs medical event prediction based on the preliminarily processed medical data. The event-driven architecture triggers the data processing flow based on the predicted medical events or the occurred medical events, and performs dynamic extraction, data transformation, data loading, and multimodal data fusion processing on the preliminarily processed medical data, and transmits the data processing result to the application layer.

[0029] S4: The application layer generates dynamic fields through an adaptive database based on the data processing result, and performs automated dynamic management and real-time feedback calibration of the data based on the data fusion analysis result. The database adopts a distributed architecture for high-concurrency access and dynamic expansion of data. The medical data storage and dynamic management use a combination of relational databases and non-relational databases to store structured and unstructured data. Data dynamic sharding and partitioning are achieved through the database to ensure efficient storage and access of data. According to the research protocol and clinical needs, the adaptive database dynamically generates the required fields. For example, for breast cancer research, the RCB grading field is automatically generated, and the dynamic field generation system implements fine-grained data such as molecular typing based on the protocol-driven extension algorithm.

[0030] In this embodiment, the diagnosis and treatment data of the patient are collected in real time from medical devices and information systems. During the data collection process, the edge computing node performs preliminary processing on the data, such as data cleaning and formatting. The edge computing node deploys a real-time data analysis engine to perform instant verification on the collected data. For example, detect whether there is a conflict between the chemotherapy plan and the cardiac toxicity monitoring data. Data verification technology is used to ensure the integrity and accuracy of the data. When data anomalies are detected, the edge computing node will immediately trigger an alarm and send the abnormal data to the central server for further analysis. The central server processes the abnormal data according to the preset rules and the logic inference engine, and feeds back the processing result. Data synchronization: A real-time synchronization strategy is adopted between the edge computing node and the central server to ensure the consistency and integrity of the data. Mechanisms such as checksum, sequence number, and timestamp are used to handle delays, packet losses, or errors during the data synchronization process.

[0031] Build a dynamic ETL (Extract, Transform, Load) engine driven by medical events to perform dynamic extraction, transformation, and loading of data to adapt to different research needs. The event-driven architecture constructs an event-driven architecture, and realizes asynchronous processing of data through a message queue and an event processing engine. The event-driven architecture allows the system to automatically trigger the data processing flow when a medical event is received.

[0032] Embodiment 2 Based on Embodiment 1, the two-stream Transformer network architecture embedded in the intermediate governance layer includes a main stream layer, an auxiliary stream layer, an interaction layer, and an output layer. The input of the main stream layer is the original time-series medical event data, such as the continuous physiological monitoring data of patients. Auxiliary stream: The input is the data after preprocessing or feature extraction, such as the key feature sequence extracted from the original data, or the context information related to the main data stream, and customized adjustments can be made according to the characteristics of the auxiliary data, such as adjusting the weight distribution of the attention mechanism.

[0033] The main stream layer and the auxiliary stream layer are connected through the interaction layer. The interaction layer is set with a cross-attention mechanism to achieve data fusion between the main stream layer and the auxiliary stream layer and mutual attention and complementarity between the main stream layer and the auxiliary stream layer. The main stream layer extracts the context information provided by the auxiliary stream layer through the interaction layer to enhance the understanding of the main medical data, and the auxiliary stream layer extracts the time-series features of the medical data in the main stream layer through the interaction layer.

[0034] The output layer is respectively connected to the main stream layer and the auxiliary stream layer. A fully connected layer is provided in the output layer, and the data processing results output by the main stream layer and the auxiliary stream layer are combined by weighted summation and splicing to obtain the final output data. The output layer is adjusted according to specific task requirements. For example, for the medical event prediction task, the output layer can output the probability distribution of the event occurrence, and judge whether the event occurs according to the probability value.

[0035] The specific process of the two-stream Transformer network architecture performing medical event prediction based on the preliminarily processed medical data in step S3 is as follows: S31: Extract a specified amount of pre - processed medical data and divide it into a training set, a validation set, and a test set. After data annotation of the training set, input it into the model for training. Verify the model based on the validation set and test the model based on the test set. When using the training set data to train the dual - stream Transformer model, adjust the model parameters by optimizing the loss function. Use the validation set to verify the model and select the model parameters with the best performance. During the training process, use the early stopping strategy to prevent overfitting. Use the test set to evaluate the trained model. The evaluation metrics can include accuracy, recall, F1 - value, etc. Apply the model to the actual medical event prediction task to provide support for clinical decision - making.

[0036] S32: Re - divide the remaining pre - processed medical data according to a specified ratio into mainstream data and auxiliary - stream data. The mainstream data is time - series medical event data, and divide the time - series medical event data into sequence segments of a fixed length as the input of the mainstream layer. Extract features and encode the auxiliary - stream data to generate the input of the auxiliary - stream layer corresponding to the input of the mainstream layer. S33: The encoder in the mainstream layer encodes the input sequence segments of a fixed length to extract the global features in the sequence segments of a fixed length. The encoder in the auxiliary - stream layer encodes the auxiliary - stream data to extract the context features related to the mainstream data. When the mainstream layer extracts features, it obtains the context information provided by the auxiliary - stream layer based on the interaction layer to enhance the understanding of the main data and achieve the extraction of global features. The auxiliary - stream layer obtains more accurate time - series features of the mainstream layer based on the interaction layer to achieve the extraction of context features related to the mainstream data, realizing the information interaction and fusion of the global features extracted by the mainstream layer and the context features extracted by the auxiliary - stream layer. S34: Transmit the global features extracted by the mainstream layer and the context features extracted by the auxiliary - stream layer to the output layer. The output layer uses a fully - connected layer to combine the data processing results of the mainstream layer and the auxiliary - stream layer through weighted summation and concatenation to obtain the final medical event prediction result and output it.

[0037] The pre - processed medical data is the time - series medical event data of patients, including key nodes in the clinical pathway, such as admission nodes, examination nodes, treatment nodes, and follow - up nodes, as well as the event features corresponding to each key node, including physiological indicators and diagnostic results.

[0038] In step S33, the encoder sets a multi-head self-attention mechanism layer to encode the input time-series event sequence and capture long-range dependencies in the medical event sequence through the multi-head self-attention mechanism. In each layer of the encoder, an attention mechanism module with clinical pathway constraints is added. The attention mechanism module with clinical pathway constraints adjusts the attention weights according to predefined clinical pathway rules. For the standard treatment pathway of a specified disease, the clinical pathway constraint module will increase the attention weights of events related to key nodes and reduce the weights of events unrelated to the pathway.

[0039] Based on the output of the encoder, the attention mechanism layer further applies the attention mechanism with clinical pathway constraints. The attention mechanism dynamically adjusts the weights of medical events by calculating the matching degree between each medical event and the clinical pathway. The calculation of the matching degree is based on the type of medical event, the time sequence, and the degree of association with key nodes in the pathway. The network architecture of the attention mechanism with clinical pathway constraints includes an input layer that inputs the time-series medical event data of the patient, including the key nodes in the clinical pathway and the corresponding event features (including physiological indicators, diagnostic results). These data are converted into a format suitable for model processing, such as embedding time-series data into a high-dimensional space. The encoder layer also adopts a Transformer encoder structure to encode the input time-series event sequence. The long-range dependencies in the event sequence are captured through the multi-head self-attention mechanism in the encoder. In each layer of the encoder, a clinical pathway constraint module is introduced, which adjusts the attention weights according to predefined clinical pathway rules. For example, for the standard treatment pathway of a specific disease, the constraint module can increase the attention weights of events related to key treatment nodes and reduce the weights of events unrelated to the pathway.

[0040] Based on the output of the encoder, the attention mechanism layer further applies the attention mechanism with clinical pathway constraints. This attention mechanism dynamically adjusts the weights of events by calculating the matching degree between each event and the clinical pathway. The calculation of the matching degree can be based on the type of event, the time sequence, and the degree of association with key nodes in the pathway. For example, for a patient undergoing chemotherapy, events related to the chemotherapy cycle (including blood test and drug dose adjustment) will obtain higher weights. The decoder layer generates predictions for subsequent medical events based on the output of the encoder and the attention weights. The decoder also adopts a Transformer structure and can use the key information captured by the attention mechanism for event prediction. During the decoding process, the clinical pathway constraints continue to play a role to ensure that the prediction results conform to the logic of the clinical pathway. Output layer: The output layer converts the prediction results generated by the decoder into specific medical event predictions, such as predicting the type of event that the patient may experience at the next time point and its probability. The output results can be used for clinical decision support, such as reminding doctors to pay attention to key events or adjust treatment plans.

[0041] The implementation process of the attention mechanism module constrained by the clinical pathway is as follows: S331: Collect the temporal medical event data of the patient, including the key nodes and event features in the clinical pathway, clean, normalize, and convert the data to meet the model input requirements, such as embedding the time series data into a high-dimensional space.

[0042] S332: Initialize the encoder parameters in the two-stream Transformer network architecture, including the weights of the multi-head self-attention mechanism and load the predefined clinical pathway rules for subsequent attention weight adjustment; S333: Feed the input temporal event sequence into the encoder. The encoder encodes the events layer by layer. During each layer of encoding, the clinical pathway constraint module adjusts the attention weights according to the pathway rules and outputs the encoded event data. For example, for a specific treatment stage, the constraint module will increase the attention weights of the key events in that stage.

[0043] S334: Based on the output of the encoder, further apply the attention mechanism constrained by the clinical pathway, calculate the matching degree of each event with the clinical pathway, and dynamically adjust the attention weights of the events according to the matching degree; for example, for a patient undergoing a specific treatment, the events related to the treatment will obtain higher weights.

[0044] S335: Feed the output of the encoder and the adjusted attention weights into the decoder. The decoder generates predictions for subsequent medical events based on its input data. During the decoding process, continue to apply the clinical pathway constraints to ensure that the prediction results conform to the logic of the clinical pathway; S336: The prediction results generated by the decoder are converted into specific medical event predictions through the output layer, and the prediction results are applied to clinical decision support, such as reminding doctors to pay attention to key events or adjusting treatment plans.

[0045] Embodiment 3 Based on Embodiment 1 or Embodiment 2, the specific process of dynamically extracting, data converting, data loading, and multi-modal data fusion processing of the preliminarily processed medical data in step S3 is as follows: S35: Real-time extract the medical data of multiple heterogeneous data sources that have been preliminarily processed through a dynamic ETL engine, and perform data mapping and conversion according to the preset ETL specification guidelines during the extraction process. Dynamically adjust the conversion rules according to the type and content of medical events during the data conversion process. Use streaming ETL technology to perform real-time conversion of the data. The dynamic ETL engine supports rule-based conversion logic and can dynamically adjust the conversion rules according to the type and content of medical events.

[0046] S36: Load the processed medical data into a preset database. During the data loading process, the dynamic ETL engine simultaneously performs data verification and conflict resolution to ensure data consistency.

[0047] S37: Extract features from medical data of different modalities respectively, convert the features of the medical data of different modalities into corresponding feature vectors respectively, and merge the feature vectors of the medical data of different modalities according to specified rules to obtain a fused feature vector.

[0048] In this embodiment, the specific process of the application layer generating dynamic fields through the adaptive database based on the data processing result in step S4 is as follows: S41: Obtain the respective eigenvalue of the fused feature vector, perform polynomial expansion on the respective eigenvalue to generate high-level features, and capture the non-linear relationship between features; S42: Randomly combine different eigenvalues to generate new eigenvalues and capture the interaction relationship between features; S43: Perform normalization transformation on the obtained respective eigenvalue and the new eigenvalue, and select the eigenvalues greater than the specified information quantity threshold to form a new feature subset.

[0049] The encoder uses an intelligent coding model with three-stage coding verification. The intelligent coding model includes an input layer, a knowledge graph embedding layer, a path encoder, and an output layer. The input of the input layer is clinical text data, and a pre-trained medical text embedding model is used to encode the text to generate an initial embedding representation of the text. The knowledge graph embedding layer embeds the concepts and relationships in the knowledge graph into the vector space, and uses a graph neural network to encode the knowledge graph to obtain the context representation of the concept nodes. The path encoder encodes the paths searched from the knowledge graph to generate an embedding representation of the paths, which includes a multi-head attention mechanism and a feed-forward network for processing the path sequence and integrating features. The output of the output layer is the clinical concept embedding representation after semantic parsing, and the clinical concept embedding representation will be used for subsequent logical consistency verification.

[0050] Before encoding using the intelligent encoding model, the clinical text is first segmented, part-of-speech tagged, and named entity recognized to extract medical concepts. Then, knowledge graph matching is performed to match the extracted medical concepts with the concepts in the UMLS knowledge graph to find the corresponding nodes. Next, path search is carried out, and multi-hop path search is performed in the UMLS knowledge graph to find the paths related to the input concept. Path encoding uses the Transformer architecture to encode the paths and generate the embedded representations of the paths. Then, semantic fusion is performed to fuse the path embeddings with the original text embeddings to generate the final semantic representation, and SWRL rules are applied for logical consistency verification. The rule engine layer loads the SWRL rules, which define the logical relationships between clinical concepts. An inference engine is used to match and reason about the instances and rules in the knowledge base. Knowledge base layer: The embedded representations of the clinical concepts after semantic parsing are loaded into the knowledge base. The knowledge base contains OWL ontologies and SWRL rules to support logical reasoning. The output layer outputs the encoded results after logical consistency verification, including the verified instances and the derived new instances. Rule loading: Load the defined SWRL rules into the inference engine. Instance matching: According to the antecedents of the rules, find the instances that meet the conditions in the knowledge base. Inference execution: Apply the consequents of the rules to the found instances to generate new instances or modify existing instances. Result verification: Verify whether the inference results meet the requirements of logical consistency to ensure there are no logical conflicts. Error handling: For inconsistent results, perform error analysis and record them for subsequent correction.

[0051] During the three-stage encoding, dynamic encoding optimization of the reinforcement learning reward mechanism is also carried out. The network structure of the reinforcement learning reward mechanism includes a reinforcement learning agent layer: Define the reinforcement learning agent, whose goal is to optimize the accuracy and consistency of the encoding results. The agent adjusts the encoding strategy according to the environmental feedback (reward signal). Environment layer: The environment includes the encoded results after logical consistency verification and the true annotations of clinical data. The environment provides a reward signal to evaluate the behavior of the agent. Output layer: The output is the encoded results after dynamic optimization, which are improved in terms of accuracy and consistency. The implementation process includes initializing the agent: Initialize the reinforcement learning agent and define its policy network and value network. Environment interaction: The agent interacts with the environment and generates encoding results according to the current policy. Reward calculation: The environment calculates the reward signal based on the accuracy and consistency of the encoding results. Policy update: The agent updates its policy network according to the reward signal to optimize the encoding strategy. Iterative optimization: Repeat the environment interaction and policy update process until the encoding results reach satisfactory accuracy and consistency.

[0052] The dynamic management of medical data in the present invention uses a closed-loop follow-up system, including a four-dimensional follow-up decision-making model: Research dimension: Prediction of the achievement degree of research endpoints; The network structure includes a data input layer: The inputs include the baseline characteristics of patients and the dynamic data during the follow-up process (including treatment responses, laboratory test results, etc.). Feature extraction layer: The mixed effect model is used to extract the time-varying characteristics of key indicators (including LDH, hs-CRP, etc.) during the follow-up process. Prediction model layer: Based on the Cox proportional hazards model or the joint modeling method, combined with the baseline characteristics and dynamic characteristics, predict the probability of achieving the research endpoint. Output layer: The output is the predicted value of the achievement degree of the research endpoint for each patient at different time points, which is used to guide subsequent research strategies.

[0053] The implementation process includes data preparation: Collect the baseline data of patients and the dynamic data during the follow-up process. Feature extraction: Use the mixed effect model to fit the changes of key indicators over time. Based on the extracted features, fit the Cox model or the joint model to predict the research endpoint. Result output: Output the predicted value of the achievement degree of the research endpoint at each time point.

[0054] In the clinical dimension, the network structure for setting the early warning of adverse reaction risks is as follows: Data input layer: Input the clinical data of patients, including treatment plans, drug use records, physiological indicators, etc. Risk assessment model layer: Use machine learning or statistical methods to establish an adverse reaction risk assessment model. Early warning output layer: Output the early warning signals of adverse reactions, including the risk level and possible types of adverse reactions. The implementation process is data collection: Collect the treatment plans, drug use records and physiological indicators of patients. Model training: Use historical data to train the adverse reaction risk assessment model. Risk assessment: Evaluate the risk of patients having adverse reactions according to the input data. Early warning generation: Generate the early warning signals of adverse reaction risks according to the risk assessment results.

[0055] Patient dimension: Dynamic assessment of compliance. The network structure includes a data input layer: Input the data related to the treatment compliance of patients, such as medication records, implementation of treatment plans, etc. Compliance assessment model layer: Use the scale assessment method or the machine learning model to evaluate the compliance of patients. Dynamic assessment output layer: Output the compliance assessment results of patients, including the compliance level and possible compliance problems. The implementation process: Data collection: Collect the medication records and implementation of treatment plans of patients. Compliance assessment: Use the MMAS-8 scale or the machine learning model to evaluate the compliance of patients. Result output: Output the compliance assessment results of patients, including the compliance level and possible problems.

[0056] System Dimension: Data Quality Feedback Calibration; The network structure includes a data input layer: Input data quality indicators during the follow-up process, such as data integrity, accuracy, etc. Quality Assessment Model Layer: Use a data quality assessment model to evaluate the quality level of the data. Feedback Calibration Output Layer: Output the data quality assessment results and calibration suggestions for optimizing the data collection and processing processes. Implementation Process: Data Collection: Collect data quality indicators during the follow-up process. Quality Assessment: Use a data quality assessment model to evaluate the quality level of the data. Feedback Generation: Generate data quality feedback and calibration suggestions based on the assessment results. Process Optimization: Optimize the data collection and processing processes based on the feedback and suggestions.

[0057] Take the data governance of neoadjuvant treatment for breast cancer as an example: Data Capture: Integrate pathological reports, MRI images, and genomic data in real time Network Structure Setting Input Layer: Include multi-modal data such as pathological reports, MRI images, and genomic data. Preprocessing Layer: Perform preprocessing operations on MRI images such as size unification and brightness contrast adjustment. Standardize genomic data. Feature Extraction Layer: Use a convolutional neural network to extract features from MRI images. Use a graph neural network to extract features from genomic data. Fusion Layer: Fuse the features of different modal data to generate a comprehensive feature representation. The implementation process includes Data Collection: Obtain pathological reports, MRI images, and genomic data from the hospital information system in real time. Data Preprocessing: Perform operations such as cropping, scaling, and histogram equalization on MRI images. Standardize genomic data. Feature Extraction: Use CNN to extract features from MRI images. Use GNN to extract features from genomic data. Feature Fusion: Perform weighted average fusion on the features of different modalities. Data Integration: Store the integrated data in a database for subsequent processing. Conflict Detection: Early warning of conflicts between chemotherapy regimens and cardiac toxicity monitoring data. The network structure includes an Input Layer: Input chemotherapy regimens and cardiac toxicity monitoring data.

[0058] Rule Engine Layer: Use predefined rules to perform logical consistency checks on the input data. Early Warning Output Layer: Output conflict early warning signals to indicate potential contradictions. The implementation process includes Data Collection: Collect the chemotherapy regimens and cardiac toxicity monitoring data of patients. Rule Matching: According to the predefined rules, check whether there are logical contradictions between the chemotherapy regimens and the cardiac toxicity monitoring data. Conflict Detection: If contradictions are detected, generate conflict early warning signals. Early Warning Output: Send the conflict early warning signals to clinicians for timely adjustment of treatment plans. Field Generation: Automatically create dynamic fields for RCB grading. The network structure includes Input Layer: Input the pathological reports and treatment records of patients. Field Generation Layer: Use protocol-driven field derivation algorithms to automatically generate RCB grading fields according to the research protocol. Output Layer: Output the generated dynamic fields for RCB grading. The implementation process: Data Collection: Collect the pathological reports and treatment records of patients. Field Generation: Automatically create RCB grading fields according to the research protocol. Field Storage: Store the generated fields in the database for subsequent analysis.

[0059] Intelligent Coding: Synchronously generate ICD-11, CPT-4, and LOINC codes. Network Structure: Input Layer: Input the clinical data of patients. Coding Generation Layer: Use clinical semantic parsing based on the UMLS knowledge graph to generate preliminary codes. Optimize the codes using logical consistency checks. Further optimize the codes using dynamic coding optimization. Output Layer: Output the synchronously generated ICD-11, CPT-4, and LOINC codes. The implementation process: Data Collection: Collect the clinical data of patients. Semantic Parsing: Use the UMLS knowledge graph to perform semantic parsing on the clinical data. Logical Verification: Use SWRL rules to perform logical consistency checks on the parsing results.

[0060] Coding Optimization: Use the reinforcement learning reward mechanism to optimize the coding results. Coding Output: Output the synchronously generated ICD-11, CPT-4, and LOINC codes. Follow-up Trigger: Initiate imaging review follow-up based on the treatment response model. Network Structure: Input Layer: Input the treatment response data of patients. Model Evaluation Layer: Use deep learning models (such as 3D CNN) to evaluate the treatment response of patients.

[0061] Follow-up Decision Layer: According to the model evaluation results, decide whether to initiate imaging review follow-up. Output Layer: Output the follow-up decision results. The implementation process: Data Collection: Collect the treatment response data of patients. Model Evaluation: Use 3D CNN to evaluate the treatment response of patients. Follow-up Decision: According to the evaluation results, decide whether to initiate imaging review follow-up. Follow-up Trigger: If necessary, automatically trigger imaging review follow-up.

[0062] Second, taking immune therapy toxicity management as an example: Dynamic field: Multidimensional record structure of irAE (immune-related adverse events). Network structure: Input layer: Input the immunotherapy records of patients. Field generation layer: Generate a multidimensional record structure according to the grading criteria of irAE. Output layer: Output the multidimensional record structure of irAE. Implementation process: Data collection: Collect the immunotherapy records of patients. Field generation: Generate a multidimensional record structure according to the grading criteria of irAE. Field storage: Store the generated fields in the database.

[0063] Logical verification: Detection of contradictions between cortisol levels and the use of immunosuppressants. Network structure: Input layer: Input the cortisol levels and immunosuppressant use records of patients. Rule engine layer: Use predefined rules for logical verification. Output layer: Output the contradiction detection results. Implementation process: Data collection: Collect the cortisol levels and immunosuppressant use records of patients. Rule matching: Use predefined rules to check whether there are contradictions between cortisol levels and immunosuppressant use. Contradiction detection: If contradictions are detected, generate contradiction detection results. Result output: Output the contradiction detection results to clinicians.

[0064] Follow-up strategy: Personalized follow-up interval algorithm based on T cell subset analysis. Network structure: Input layer: Input the T cell subset analysis data of patients. Model evaluation layer: Use machine learning models to evaluate the follow-up intervals of patients. Output layer: Output personalized follow-up interval suggestions. Implementation process: Data collection: Collect the T cell subset analysis data of patients. Model evaluation: Use machine learning models to evaluate the follow-up intervals of patients. Follow-up suggestion: Generate personalized follow-up interval suggestions based on the evaluation results. Suggestion output: Output the follow-up interval suggestions to clinicians.

[0065] In summary, the medical data dynamic management method based on multimodal data fusion and deep reasoning provided by the present invention establishes a three-layer dynamic governance architecture. By collecting various medical data of patients in real time at the bottom layer and performing preliminary processing on the medical data of patients, after the middle governance layer receives the preliminarily processed medical data, a two-stream Transformer network architecture is used for medical event prediction, and the event-driven architecture triggers the data processing process, performing dynamic extraction, data transformation, data loading, and multimodal data fusion processing on the preliminarily processed medical data, transmitting the data processing results to the application layer, where the application layer generates dynamic fields and performs automated dynamic management and real-time feedback calibration of the data based on the data fusion analysis results. It solves the limitations of the static field structure existing in the existing medical data management SEER database, improves data quality, and realizes the automated management and real-time feedback calibration of follow-up data, avoiding the separation of the follow-up system and the data governance platform.

[0066] Set up a three-layer dynamic governance architecture. By realizing real-time data collection and real-time processing at the bottom layer, it ensures the accuracy and real-time nature of various medical data during the collection stage. The event-driven architecture realizes the asynchronous processing of data through a message queue and an event processing engine. The event-driven architecture allows the system to automatically trigger the data processing flow when a medical event is received. The middle governance layer constructs a dual-stream Transformer network architecture and an event-driven architecture. Through the main stream layer and the auxiliary stream layer of the dual-stream Transformer network architecture, it obtains global temporal dependencies and context information respectively, realizes the simultaneous capture of global features and context features, and combines the multi-head self-attention mechanism to process features in multiple subspaces in parallel, so as to more comprehensively understand the complex relationships in the data and achieve the precise processing of medical data, providing an efficient and accurate network architecture for the management of medical data.

[0067] An interaction layer is set between the main stream layer and the auxiliary stream layer of the dual-stream Transformer network architecture to fuse the information of the main and auxiliary streams. The interaction layer adopts a cross-attention mechanism, enabling the two streams to pay attention to and complement each other, enhancing the accuracy of feature extraction. The output layer merges the outputs of the two streams and generates the final output through weighted summation or concatenation of the fully connected layer. For the medical event prediction task, the output layer can output the probability distribution of the event occurrence, which can further enhance the flexibility of subsequent dynamic management. The encoders in the main stream layer and the auxiliary stream layer use an intelligent coding system with three-stage coding verification. The intelligent coding system adopts a three-stage coding verification process, combines the UMLS knowledge graph and SWRL rules, and optimizes the coding process through a reinforcement learning reward mechanism, solving the semantic gap problem between ICD coding and clinical practice in the existing technology.

Claims

1. A medical data dynamic management method based on multi-modal data fusion and deep reasoning, characterized in that It includes the following steps: S1: Establish at least a three-layer dynamic governance architecture, including a bottom layer, an intermediate governance layer, and an application layer. The bottom layer is used to realize real-time medical data collection and preliminary processing; a dual-stream Transformer network architecture and an event-driven architecture are constructed in the intermediate governance layer for processing the multimodal medical data after preliminary processing; an adaptive database is constructed in the application layer for performing multimodal data analysis based on the processing results of the multimodal medical data in the intermediate layer to achieve dynamic management; S2: Real-time collect various medical data of the patient through the bottom layer, and perform preliminary processing on the various medical data of the patient. The preliminary processing includes data cleaning, normalization processing, abnormal data processing, and immediate verification processing, and transmit the data after preliminary processing to the intermediate layer; S3: After receiving the medical data after preliminary processing through the intermediate governance layer, the dual-stream Transformer network architecture predicts medical events based on the medical data after preliminary processing, and the event-driven architecture triggers a data processing flow based on the predicted medical events or the occurred medical events, and performs dynamic extraction, data conversion, data loading, and multimodal data fusion processing on the medical data after preliminary processing, and transmits the data processing results to the application layer; S4: The application layer generates dynamic fields through the adaptive database based on the data processing results, and performs data automated dynamic management and real-time feedback calibration based on the data fusion analysis results.

2. The medical data dynamic management method based on multi-modal data fusion and deep reasoning according to claim 1, characterized in that The dual-stream Transformer network architecture embedded in the intermediate governance layer includes a main stream layer, an auxiliary stream layer, an interaction layer, and an output layer; The main stream layer and the auxiliary stream layer are connected through the interaction layer. The interaction layer is provided with a cross-attention mechanism to realize the fusion of data between the main stream layer and the auxiliary stream layer and the mutual attention and complement between the main stream layer and the auxiliary stream layer. The main stream layer extracts the context information provided by the auxiliary stream layer through the interaction layer to enhance the understanding of the main medical data, and the auxiliary stream layer extracts the time series features of the medical data in the main stream layer through the interaction layer; The output layer is respectively connected to the main stream layer and the auxiliary stream layer. A fully connected layer is provided in the output layer, and the output data processing results of the main stream layer and the auxiliary stream layer are combined in a way of weighted summation and splicing to obtain the final output data.

3. The medical data dynamic management method based on multi-modal data fusion and deep reasoning according to claim 2, characterized in that The process that the dual-stream Transformer network architecture in step S3 predicts medical events based on the medical data after preliminary processing includes the following specific processes: S31: Extract a specified amount of the medical data after preliminary processing and divide it into a training set, a validation set, and a test set. After data annotation of the training set, input it into the model to train the model, verify the model based on the validation set, and test the model based on the test set; S32: Divide the remaining medical data after preliminary processing again according to a specified ratio into main stream data and auxiliary stream data. The main stream data is time-series medical event data, and the time-series medical event data is divided into sequence segments of a fixed length as the input of the main stream layer. Feature extraction and encoding are performed on the auxiliary stream data to generate the input of the auxiliary stream layer corresponding to the input of the main stream layer; S33: The encoder in the main stream layer encodes the input fixed-length sequence segments and extracts the global features in the fixed-length sequence segments; The encoder in the auxiliary stream layer encodes the auxiliary stream data and extracts the context features related to the main stream data; When performing feature extraction, the main stream layer obtains the context information provided by the auxiliary stream layer based on the interaction layer to enhance the understanding of the main data and achieve the extraction of global features. The auxiliary stream layer obtains more accurate time series features of the main stream layer based on the interaction layer to achieve the extraction of context features related to the main stream data, realizing the information interaction and fusion of the global features extracted by the main stream layer and the context features extracted by the auxiliary stream layer; S34: Transmit the global features extracted by the main stream layer and the context features extracted by the auxiliary stream layer to the output layer. The output layer combines them in the way of weighted summation and concatenation through the fully connected layer for the data processing results output by the main stream layer and the auxiliary stream layer, and obtains the final medical event prediction result and outputs it.

4. The medical data dynamic management method based on multi-modal data fusion and deep reasoning according to claim 3, characterized in that, The preliminarily processed medical data is the time series medical event data of the patient, including the key nodes in the clinical pathway, such as the admission node, examination node, treatment node and follow-up node, and the event features corresponding to each key node, including physiological indicators and diagnosis results.

5. The medical data dynamic management method based on multi-modal data fusion and deep reasoning according to claim 3, wherein In step S33, the encoder is set with a multi-head self-attention mechanism layer to encode the input time series event sequence, and captures the long-range dependencies in the medical event sequence through the multi-head self-attention mechanism. In each layer of the encoder, an attention mechanism module with clinical pathway constraints is added. The attention mechanism module with clinical pathway constraints adjusts the attention weights according to the predefined clinical pathway rules, including for the standard treatment path of a specified disease, the clinical pathway constraint module will increase the attention weights of the events related to the key nodes and reduce the weights of the events unrelated to the path; Based on the output of the encoder, the attention mechanism layer further applies the attention mechanism with clinical pathway constraints. The attention mechanism dynamically adjusts the weights of medical events by calculating the matching degree between each medical event and the clinical pathway. The calculation of the matching degree is based on the type, time sequence of the medical event and the degree of association with the key nodes in the path.

6. The medical data dynamic management method based on multi-modal data fusion and deep reasoning according to claim 5, characterized in that The implementation process of the attention mechanism module with clinical pathway constraints is as follows: S331: Collect the time series medical event data of the patient, including the key nodes and event features in the clinical pathway, and clean, normalize and convert the data to meet the model input requirements; S332: Initialize the encoder parameters in the dual-stream Transformer network architecture, including the weights of the multi-head self-attention mechanism and load the predefined clinical pathway rules for subsequent attention weight adjustment; S333: Feed the input time series event sequence into the encoder. The encoder encodes the events layer by layer. During each layer of encoding, the clinical pathway constraint module adjusts the attention weights according to the path rules and outputs the encoded event data; S334: Based on the encoder output, further apply the attention mechanism constrained by the clinical pathway to calculate the matching degree between each event and the clinical pathway, and dynamically adjust the attention weights of the events according to the matching degree; S335: Feed the output of the encoder and the adjusted attention weights into the decoder. The decoder generates predictions for subsequent medical events based on its input data. During the decoding process, continue to apply the clinical pathway constraints to ensure that the prediction results conform to the logic of the clinical pathway; S336: The prediction results generated by the decoder are converted into specific medical event predictions through the output layer, and the prediction results are applied to clinical decision support.

7. The medical data dynamic management method based on multi-modal data fusion and deep reasoning according to claim 1, wherein, The specific processes of dynamically extracting, transforming, loading, and fusing multi-modal data for the pre-processed medical data in step S3 are as follows: S35: Use a dynamic ETL engine to extract the medical data from multiple heterogeneous data sources after preliminary processing in real time, and perform data mapping and transformation according to the preset ETL specification guidelines during the extraction process. Dynamically adjust the transformation rules according to the type and content of medical events during the data transformation process; S36: Load the processed medical data into a preset database. During the data loading process, the dynamic ETL engine simultaneously performs data verification and conflict resolution to ensure data consistency; S37: Extract features from medical data of different modalities respectively, convert the features of the extracted medical data of different modalities into corresponding feature vectors respectively, and merge the feature vectors of the medical data of different modalities according to specified rules to obtain a fused feature vector.

8. The medical data dynamic management method based on multi-modal data fusion and deep reasoning according to claim 7, characterized in that The specific process of the application layer generating dynamic fields through an adaptive database based on the data processing results in step S4 is as follows: S41: Obtain the respective feature values based on the fused feature vector, perform polynomial expansion on the respective feature values to generate high-level features and capture the non-linear relationships between features; S42: Randomly combine different feature values to generate new feature values and capture the interaction relationships between features; S43: Perform standardization transformation on the obtained respective feature values and the new feature values, and select the feature values greater than the specified information quantity threshold to form a new feature subset.

9. The medical data dynamic management method based on multi-modal data fusion and deep reasoning according to claim 5, characterized in that, The encoder uses an intelligent coding model with three-stage coding verification. The intelligent coding model includes an input layer, a knowledge graph embedding layer, a path encoder, and an output layer; The input of the input layer is clinical text data. Use a pre-trained medical text embedding model to encode the text to generate an initial embedding representation of the text; The knowledge graph embedding layer embeds the concepts and relationships in the knowledge graph into the vector space, and uses a graph neural network to encode the knowledge graph to obtain the context representation of the concept nodes; The path encoder encodes the paths searched from the knowledge graph to generate an embedding representation of the paths, which includes a multi-head attention mechanism and a feed-forward network for processing path sequences and integrating features; The output of the output layer is the clinical concept embedding representation after semantic parsing, and the clinical concept embedding representation will be used for subsequent logical consistency verification.

Images