Medical strategy generation method and system combined with data flywheel reinforcement learning

Abstract

The invention provides a medical strategy generation method and system combined with data flywheel reinforcement learning in the field of artificial intelligence and clinical medical cross technology. The method comprises the following steps: step S1, creating a multimode data annotation model, and setting an annotation loss function and an annotation reward function; s2, acquiring a large amount of historical medical data to construct a data set; s3, carrying out training and compression on the multi-mode data labeling model; s4, creating a medical decision model, and setting a decision loss function and a decision reward function; s5, training and compressing the medical decision model; step S6, carrying out deployment on the trained model; step S7, labeling the real-time medical data through the multi-mode data labeling model, and performing incremental training on the medical decision model through the labeled real-time medical data; and S8, performing medical question and answer through the deployed medical decision model. The method has the advantages that the accuracy, timeliness and safety of medical decision making are greatly improved.

Description

Technical Field

[0001] The present invention relates to the cross - technical field of artificial intelligence and clinical medicine, and particularly to a medical strategy generation method and system combining data - flywheel reinforcement learning. Background Art

[0002] As a core component of the modern medical informatization system, the Clinical Decision Support System (CDSS) has evolved in tandem with the development of medical informatics. Since the MYCIN system pioneered the rule - based expert system in the 1970s, it has gone through four generations of technological changes: The first - generation systems (1970 - 1985) were based on symbolic logic inference engines and used IF - THEN rule - bases to implement disease diagnosis reasoning. Typical cases include the INTERNIST - 1 internal medicine diagnosis system, whose knowledge base contains 572 diseases and 4,500 symptom - related rules; The second - generation systems (1986 - 2000) introduced probabilistic graphical models such as Bayesian networks and hidden Markov models, and implemented the calculation of the probability of multiple concurrent diseases in the QMR - DT system at the University of Washington, with the diagnostic accuracy rate increased to 78%; The third - generation systems (2001 - 2015) integrated supervised machine - learning algorithms. The DeepDx system at Stanford University used SVM and random forest models to achieve an AUC value of 91% in the classification task of diabetic retinopathy; The fourth - generation systems (since 2016) began to explore the application of deep - learning technologies. The fundus image analysis system of Google Health achieved the automatic grading of diabetic macular edema through the Inception - v3 network, with a sensitivity of 94.5%.

[0003] Currently, the mainstream clinical decision support systems mostly adopt a hybrid architecture, integrating multi - source heterogeneous data such as the Electronic Medical Record System (EMR), Laboratory Information System (LIS), and Picture Archiving and Communication System (PACS). A typical system such as the Cognitive Computing Platform of Epic Company uses natural language processing technology to parse unstructured medical record texts, and combines predefined clinical pathways and diagnostic rule - bases (including more than 5,000 clinical pathways such as NCCN guidelines) to provide medical staff with services such as diagnostic suggestions, treatment plan recommendations, and medication safety warnings. However, the current clinical decision support systems face a vicious cycle of "data islands - model rigidity - application limitations", which is specifically manifested in the following dimensions: 1. Lack of feedback on misdiagnosis cases: Traditionally, misdiagnosis cases need to be manually labeled for retraining, which is not only time-consuming and labor-intensive, but also involves issues with defining responsibilities for labeling errors. Therefore, most misdiagnosis cases in clinical practice cannot be effectively fed back to the training system. Statistics from the Mayo Clinic show that only 12% of misdiagnosis reports are entered into the system in a structured manner, resulting in serious sample bias in continuous model training.

[0004] 2. Dilemma of multimodal data fusion: Medical data includes multimodal data such as images (CT / MRI), waveforms (ECG / EEG), text (electronic medical records), and gene sequencing. Existing fusion methods mostly use early fusion (such as feature-level splicing in the MITK framework) or late fusion (such as decision-level weighted voting), which leads to the loss of cross-modal semantic associations.

[0005] 3. Reward mechanism clinical mismatch: Existing reinforcement learning uses a single diagnostic accuracy as a reward function (such as DeepMind Health's retinopathy system), but ignores the evaluation of the rationality of the clinical pathway; according to NEJM research, this design results in 15% of recommended regimens having drug interaction risks or violating treatment guidelines.

[0006] 4. Knowledge update hysteresis effect: The system update cycle generally exceeds 6 months, which forms a scissors gap with the iteration speed of medical knowledge; taking tumor targeted therapy as an example, of the 8 new drugs approved by the FDA in 2023, only 2 were integrated into the mainstream CDSS knowledge base within 6 months.

[0007] 5. Insufficient response timeliness: Due to multiple factors such as limited computing resources of the deployed terminals, large model size, and pre-processing of multi-source heterogeneous data, the response timeliness is poor. For example, the decision response time in emergency scenarios exceeds 5 seconds, which cannot meet the golden treatment window for diseases such as myocardial infarction.

[0008] 6. Security risks of patient privacy data: In order to facilitate later traceability, reflux training, medical quality monitoring, etc., the session data generated during the question-and-answer process will be stored. Traditionally, a centralized storage architecture is adopted, and basic keys such as AES-128 are used. Once a hacker attack or key leakage occurs, it will trigger a large-scale data leakage incident, posing a huge security risk.

[0009] Therefore, how to provide a medical strategy generation method and system that combines data flywheel reinforcement learning to improve the accuracy, timeliness and safety of medical decision-making has become a technical problem that needs to be solved urgently. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide a medical strategy generation method and system combined with data flywheel reinforcement learning to improve the accuracy, timeliness and safety of medical decision-making.

[0011] In a first aspect, the present invention provides a medical policy generation method combining data flywheel reinforcement learning, comprising the following steps: Step S1: Create a multi-modal data annotation model based on a feature extraction layer, a modality fusion layer, a feature enhancement layer, an annotation prediction layer, and a feedback optimization layer, and set an annotation loss function and an annotation reward function for the multi-modal data annotation model; The feature extraction layer is constructed based on an image feature extraction module, a waveform feature extraction module, a case feature extraction module, and a gene feature extraction module; the image feature extraction module is used to extract image features from medical image data; the waveform feature extraction module is used to extract waveform features from detection waveform data; the case feature extraction module is used to extract case features from electronic case data; the gene feature extraction module is used to extract gene features from gene sequencing data; The modality fusion layer is constructed based on a spatio-temporal alignment module and a multi-scale feature fusion module; the spatio-temporal alignment module is used to perform spatio-temporal alignment operations on the image features, waveform features, case features, and gene features; the multi-scale feature fusion module is used to fuse the image features, waveform features, case features, and gene features after spatio-temporal alignment to obtain fused features; The feature enhancement layer is constructed based on a self-supervised contrast learning module and a knowledge graph enhancement module; the self-supervised contrast learning module is used to construct positive and negative sample pairs to perform enhancement operations on the fused features to obtain first-level enhanced features; the knowledge graph enhancement module is used to perform enhancement operations on the first-level enhanced features through a dynamic medical knowledge graph to obtain second-level enhanced features; The annotation prediction layer is constructed based on a multi-task annotation prediction module and an uncertainty estimation module; the multi-task annotation prediction module is used to respectively predict the annotation results of the medical image data, detection waveform data, electronic case data, and gene sequencing data according to the second-level enhanced features; the uncertainty estimation module is used to predict the uncertainty of each of the annotation results, screen the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty; The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model; Step S2: Obtain a large amount of historical medical data including medical image data, detection waveform data, electronic case data, and gene sequencing data, desensitize, preprocess, and annotate each piece of the historical medical data, and then construct a data set; Step S3: Train the multi-modal data annotation model through the data set, the annotation loss function, and the annotation reward function, and compress the multi-modal data annotation model during the training process; Step S4: Create a medical decision-making model based on the multi-modal data fusion layer, feature interaction layer, and diagnostic decision-making layer, and set the decision loss function and decision reward function of the medical decision-making model; The multi-modal data fusion layer is used to extract joint representations from medical image data, detection waveform data, electronic medical record data, and gene sequencing data; the feature interaction layer is used to perform feature interaction operations on the joint representations to obtain fused features; the diagnostic decision-making layer is used to output medical strategy suggestions carrying disease classification, prognosis assessment, and treatment recommendations based on the fused features; Step S5: Train the medical decision-making model using the dataset, decision loss function, and decision reward function, and compress the medical decision-making model during the training process; Step S6: Deploy the trained multi-modal data annotation model and medical decision-making model, and perform model drift training on the deployed multi-modal data annotation model and medical decision-making model; Step S7: Collect real-time medical data and the latest medical knowledge, automatically annotate the real-time medical data and the latest medical knowledge using the deployed multi-modal data annotation model, and perform incremental training on the medical decision-making model using the annotated real-time medical data and the latest medical knowledge; Step S8: Conduct medical Q&A using the deployed medical decision-making model, and record the Q&A log including at least medical questions, medical strategy suggestions, Q&A time, questioner account, and misdiagnosis feedback in real time; Step S9: Annotate the Q&A log using the deployed multi-modal data annotation model, perform incremental training on the medical decision-making model using the annotated Q&A log, and encrypt and store the Q&A log in a distributed manner.

[0012] In a second aspect, the present invention provides a medical strategy generation system combining data flywheel reinforcement learning, including the following modules: A multi-modal data annotation model creation module, used to create a multi-modal data annotation model based on a feature extraction layer, a modal fusion layer, a feature enhancement layer, an annotation prediction layer, and a feedback optimization layer, and set the annotation loss function and annotation reward function of the multi-modal data annotation model; The feature extraction layer is constructed based on an image feature extraction module, a waveform feature extraction module, a case feature extraction module, and a gene feature extraction module; the image feature extraction module is used to extract image features from medical image data; the waveform feature extraction module is used to extract waveform features from detection waveform data; the case feature extraction module is used to extract case features from electronic medical record data; the gene feature extraction module is used to extract gene features from gene sequencing data; The modality fusion layer is constructed based on a spatio-temporal alignment module and a multi-scale feature fusion module; the spatio-temporal alignment module is used to perform spatio-temporal alignment operations on image features, waveform features, case features, and gene features; the multi-scale feature fusion module is used to fuse the image features, waveform features, case features, and gene features after spatio-temporal alignment to obtain fused features; The feature enhancement layer is constructed based on a self-supervised contrast learning module and a knowledge graph enhancement module; the self-supervised contrast learning module is used to construct positive and negative sample pairs to perform enhancement operations on the fused features to obtain first-level enhanced features; the knowledge graph enhancement module is used to perform enhancement operations on the first-level enhanced features through a dynamic medical knowledge graph to obtain second-level enhanced features; The annotation prediction layer is constructed based on a multi-task annotation prediction module and an uncertainty estimation module; the multi-task annotation prediction module is used to predict the annotation results of medical image data, detection waveform data, electronic case data, and gene sequencing data respectively according to the second-level enhanced features; the uncertainty estimation module is used to predict the uncertainty of each annotation result, screen the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty; The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model; The dataset construction module is used to obtain a large amount of historical medical data including medical image data, detection waveform data, electronic case data, and gene sequencing data, desensitize, preprocess, and annotate each piece of historical medical data to construct a dataset; The multi-modal data annotation model training module is used to train the multi-modal data annotation model through the dataset, annotation loss function, and annotation reward function, and compress the multi-modal data annotation model during the training process; The medical decision-making model creation module is used to create a medical decision-making model based on the multi-modal data fusion layer, feature interaction layer, and diagnostic decision-making layer, and set the decision loss function and decision reward function of the medical decision-making model; The multi-modal data fusion layer is used to extract joint representations from medical image data, detection waveform data, electronic case data, and gene sequencing data; the feature interaction layer is used to perform feature interaction operations on the joint representations to obtain fused features; the diagnostic decision-making layer is used to output medical strategy suggestions carrying disease classification, prognosis assessment, and treatment suggestions based on the fused features; The medical decision-making model training module is used to train the medical decision-making model through the dataset, decision loss function, and decision reward function, and compress the medical decision-making model during the training process; A model deployment module, configured to deploy the trained multi-modal data annotation model and the medical decision-making model, and perform model drift training on the deployed multi-modal data annotation model and medical decision-making model; An incremental knowledge training module, configured to collect real-time medical data and the latest medical knowledge, automatically annotate the real-time medical data and the latest medical knowledge through the deployed multi-modal data annotation model, and perform incremental training on the medical decision-making model through the annotated real-time medical data and the latest medical knowledge; A medical Q&A module, configured to perform medical Q&A through the deployed medical decision-making model, and record in real time a Q&A log including at least medical questions, medical strategy suggestions, Q&A time, question-asking accounts, and misdiagnosis feedback; A Q&A log management module, configured to annotate the Q&A log through the deployed multi-modal data annotation model, perform incremental training on the medical decision-making model through the annotated Q&A log, and encrypt and distribute the storage of the Q&A log.

[0013] The advantages of the present invention are as follows: 1. Create a multi-modal data annotation model through a feature extraction layer, a modality fusion layer, a feature enhancement layer, a label prediction layer, and a feedback optimization layer, and set the annotation loss function and annotation reward function of the multi-modal data annotation model; then obtain a large amount of historical medical data including medical image data, detection waveform data, electronic medical record data, and gene sequencing data, desensitize, preprocess, and annotate each piece of historical medical data to construct a data set; train the multi-modal data annotation model through the data set, the annotation loss function, and the annotation reward function, and compress the multi-modal data annotation model during the training process; then create a medical decision-making model based on a multi-modal data fusion layer, a feature interaction layer, and a diagnostic decision-making layer, and set the decision loss function and decision reward function of the medical decision-making model; train the medical decision-making model through the data set, the decision loss function, and the decision reward function, and compress the medical decision-making model during the training process; deploy the trained multi-modal data annotation model and medical decision-making model, and perform model drift training on the deployed multi-modal data annotation model and medical decision-making model; then collect real-time medical data and the latest medical knowledge, automatically annotate the real-time medical data and the latest medical knowledge through the deployed multi-modal data annotation model, and perform incremental training on the medical decision-making model through the annotated real-time medical data and the latest medical knowledge; then perform medical Q&A through the deployed medical decision-making model, and record the Q&A log including at least medical questions, medical strategy suggestions, Q&A time, questioner accounts, and misdiagnosis feedback in real time; annotate the Q&A log through the deployed multi-modal data annotation model, perform incremental training on the medical decision-making model through the annotated Q&A log, and encrypt and distribute the storage of the Q&A log; that is, perform medical Q&A through the pre-trained medical decision-making model, record the Q&A log containing misdiagnosis feedback in real time, automatically annotate the Q&A log through the pre-trained multi-modal data annotation model and perform incremental training on the medical decision-making model (flow back to the training system), effectively reducing sample bias; the medical decision-making model effectively improves the feature extraction ability of multi-modal data through the multi-modal data fusion layer and the feature interaction layer; by setting the decision reward function of the medical decision-making model to carry path rationality rewards and risk aversion rewards, the risk of drug interactions or violations of treatment guidelines can be effectively prevented; by collecting real-time medical data and the latest medical knowledge for incremental training, the lag in knowledge update can be avoided; by compressing the multi-modal data annotation model and the medical decision-making model during the training phase, combining with a streaming computing engine in the data preprocessing phase, and combining with hardware acceleration and memory caching in the inference phase, the response time can be effectively improved; by performing multi-level encryption and distributed storage on the Q&A log, the security of the relevant data storage is guaranteed, and the data flywheel (model drift training, real-time data automatic annotation and incremental training, Q&A log feedback, knowledge update mechanism) and reinforcement learning (double reward function, policy optimization mechanism, uncertainty-driven exploration) are integrated as a whole, ultimately greatly improving the accuracy, timeliness, and security of medical decision-making.

[0014] 2. By integrating medical imaging data, detection waveform data, electronic case data, and gene sequencing data, comprehensively capture patient information to avoid the limitations of a single modality; by setting imaging features, combine 3D CNN and Transformer to take into account local details and global context, set waveform features to process temporal dependencies through 1D temporal convolution and causal convolution, incorporate case features into pre-trained medical term vectors to enhance semantic understanding, combine gene features with graph neural networks and knowledge graphs to analyze gene interactions and biological significance, and then through cross-modal contrast learning and dynamic weighted fusion, strengthen the multi-modal synergy effect and effectively improve the feature extraction ability.

[0015] 3. Through spatio-temporal alignment and multi-scale fusion, that is, solve the spatio-temporal inconsistency problem through deformable convolution and time warping alignment, and dynamically adjust the weights through the cross-attention mechanism to effectively improve the quality of the fused features, and then effectively improve the annotation quality.

[0016] 4. By fusing contrastive loss and focal loss in the annotation loss function, effectively solve class imbalance and modality consistency; by fusing KL divergence and temporal consistency loss in the decision loss function, effectively constrain the rationality of the clinical path and the coherence of the decision; by introducing indicators such as user recognition and diagnosis and treatment efficiency into the reward function, and combining reinforcement learning to achieve end-to-end optimization.

[0017] 5. By combining weight pruning, weight quantization, and parameter sharing during the model training stage, effectively compress the model volume, reduce the consumption of computing resources, and adapt to edge device deployment.

[0018] 6. During the process of generating encrypted logs, by combining two SM3 hash calculations, and respectively adopting AES-GCM algorithm, SM9 algorithm, ECIES algorithm, and Paillier algorithm for different data types, and finally using SM4 algorithm and Serpent algorithm for sandwich encryption, and uploading the data fingerprint of the encrypted logs to the blockchain, realize multi-level encryption, integrity verification, and anti-tampering of the Q&A logs, greatly improving the security of Q&A log storage.

[0019] 7. By automatically annotating real-time medical data and Q&A logs and then feeding them back to the model to form a "data-model" closed-loop optimization; through model drift training combined with indicators such as the ROC curve for verification, ensure long-term performance stability.

[0020] 8. By dividing the training set, validation set, and test set in a ratio of 8:1:1, avoid overfitting; automatically expand the training set when the validation fails, effectively improve the generalization ability; verify the model performance from multiple angles such as accuracy, recall rate, F1 value, and ROC curve to ensure clinical reliability.

[0021] 9. By setting a modular structure for each layer such as feature extraction, fusion, and enhancement, it is convenient to add new data modalities (such as voice records, sensor data) or task types (such as drug recommendation) without reconstructing the overall framework; by supporting the dynamic integration of the latest medical knowledge and user feedback (misdiagnosis feedback), it adapts to the characteristics of rapid iteration in the medical field.

[0022] 10. By predicting the uncertainty of the annotation results, automatically screening high-confidence data, and reducing the negative impact of noisy annotations on the model; by adopting multi-expert consensus for historical data annotation, reducing the bias of single annotators, and improving the reliability of training data.

[0023] 11. By using a multi-modal data annotation model to automatically annotate Q&A logs with misdiagnosis feedback, effectively improving the annotation efficiency; by generating data fingerprints through the HMAC algorithm and uploading them to the chain, ensuring the traceability and anti-tampering of misdiagnosis cases; by setting a quantity threshold for the incremental training trigger mechanism, when the misdiagnosis cases accumulate to the clinical significance level, the model update is automatically triggered, effectively solving the problem of missing feedback, and then constructing a closed loop of "misdiagnosis cases - annotation feedback - model iteration".

[0024] 12. Through a multi-modal data annotation model designed in a progressive manner of "professional feature extraction → spatio-temporal semantic alignment → knowledge enhancement fusion → credibility-aware annotation", breakthroughs have been achieved in three dimensions: accuracy, security, and adaptability of medical multi-modal data processing, constructing an autonomously evolving medical AI annotation infrastructure, and providing a high-quality, interpretable, and low-latency data foundation for subsequent clinical decisions.

[0025] 13. By setting a multi-modal deep collaborative architecture for the medical decision-making model, high-precision and comprehensiveness of medical decision-making are achieved: First, the 3D ResNet-50, 1D CNN-BiLSTM, BERT, and GAT networks are used to respectively adapt to the feature extraction of medical images, waveforms, electronic medical records, and gene data, specifically solving the problem of multi-source heterogeneous data representation; Second, through cross-modal contrast learning to align the semantic of different modal features, and combining the cross-attention mechanism and dynamic graph neural network to achieve adaptive fusion of feature interaction, effectively capturing complex medical associations and compressing redundant information; Further, based on the multi-task decision-making layer, disease classification, prognosis assessment, and treatment recommendations are synchronously output, integrating the loss-reward function optimization mechanisms of supervised learning and reinforcement learning to ensure a balance between the diagnostic accuracy and the rationality of clinical decisions. Its innovation is reflected in the deep fusion of multi-modal data, the optimization of dynamic feature interaction, and the integrated design of end-to-end diagnosis and treatment decision-making, significantly improving the reliability, efficiency, and interpretability of medical decision-making, and providing core technical support for precision medicine.

[0026] 14. By innovatively constructing an intelligent decision-making system for deep integration and dynamic optimization of multimodal medical data, integrating multi-source data such as medical image data, detection waveform data, electronic case data, and gene sequencing data, and adopting advanced technologies such as 3D convolution, temporal causal convolution, and knowledge graph enhancement to achieve accurate feature extraction and cross-modal collaborative expression. Combining the data flywheel mechanism of reinforcement learning and the dual loss-reward function to optimize the model performance, while ensuring data security (full-process desensitization, multi-level national secret encryption, and blockchain evidence storage), it has the capabilities of real-time reasoning, incremental learning, and clinical path alignment. Its modular design supports flexible expansion, and significantly improves the diagnostic accuracy and generalization ability through self-supervised contrast learning, uncertainty estimation, and multi-task joint prediction, forming a closed loop of "data annotation - decision optimization - feedback iteration", providing a highly reliable, highly interpretable, and continuously evolving full-stack solution for medical AI. Brief Description of the Drawings

[0027] The present invention will be further described below with reference to the accompanying drawings in conjunction with embodiments.

[0028] Figure 1 It is a flowchart of a medical policy generation method combining data flywheel reinforcement learning of the present invention.

[0029] Figure 2 It is a schematic structural diagram of a medical policy generation system combining data flywheel reinforcement learning of the present invention. Detailed Embodiments

[0030] The overall idea of the technical solution in the embodiments of the present application is as follows: Conduct medical Q&A through a pre-trained medical decision-making model, and record the Q&A logs containing misdiagnosis feedback in real time. Automatically annotate the Q&A logs through a pre-trained multi-modal data annotation model and perform incremental training on the medical decision-making model (flow back to the training system) to effectively reduce sample bias; the medical decision-making model effectively improves the feature extraction ability of multi-modal data through the multi-modal data fusion layer and the feature interaction layer; by setting the decision reward function of the medical decision-making model to carry path rationality rewards and risk aversion rewards, the risk of drug interactions or violations of treatment guidelines can be effectively prevented; by collecting real-time medical data and the latest medical knowledge for incremental training, the lag in knowledge update can be avoided; by compressing the multi-modal data annotation model and the medical decision-making model during the training stage, combining a streaming computing engine in the data preprocessing stage, and combining hardware acceleration and memory caching in the inference stage, the response time can be effectively improved; by performing multi-level encryption and distributed storage on the Q&A logs, the security of relevant data storage is guaranteed. Overall, it combines the data flywheel (model drift training, real-time data automatic annotation and incremental training, Q&A log feedback, knowledge update mechanism) and reinforcement learning (dual reward function, policy optimization mechanism, uncertainty-driven exploration), thereby improving the accuracy, timeliness, and security of medical decision-making.

[0031] Please refer to Figures 1 to 2 as shown in the following preferred embodiment of a medical policy generation method combining data flywheel and reinforcement learning, which includes the following steps: Step S1: Create a multi-modal data annotation model based on a feature extraction layer, a modality fusion layer, a feature enhancement layer, a label prediction layer, and a feedback optimization layer, and set the annotation loss function and annotation reward function of the multi-modal data annotation model; By setting each layer such as feature extraction, fusion, and enhancement to adopt a modular structure, it is convenient to add new data modalities (such as voice records, sensor data) or task types (such as drug recommendations) without reconstructing the overall framework; by supporting dynamic integration of the latest medical knowledge and user feedback (misdiagnosis feedback), it adapts to the characteristics of rapid iteration in the medical field.

[0032] The feature extraction layer is constructed based on an image feature extraction module, a waveform feature extraction module, a case feature extraction module, and a gene feature extraction module; the image feature extraction module is used to extract image features from medical image data; the waveform feature extraction module is used to extract waveform features from detection waveform data; the case feature extraction module is used to extract case features from electronic case data; the gene feature extraction module is used to extract gene features from gene sequencing data; The modality fusion layer is constructed based on a spatio-temporal alignment module and a multi-scale feature fusion module; the spatio-temporal alignment module is used to perform spatio-temporal alignment operations on image features, waveform features, case features, and gene features; the multi-scale feature fusion module is used to fuse the image features, waveform features, case features, and gene features after spatio-temporal alignment to obtain fused features; The feature enhancement layer is constructed based on a self-supervised contrast learning module and a knowledge graph enhancement module; the self-supervised contrast learning module is used to construct positive and negative sample pairs to perform enhancement operations on the fused features to obtain first-level enhanced features; the knowledge graph enhancement module is used to perform enhancement operations on the first-level enhanced features through a dynamic medical knowledge graph to obtain second-level enhanced features; The self-supervised contrastive learning module works a bit like a "spot the difference" game. It extracts two sets of samples from the data: one is the "positive sample pair", which are essentially similar (such as the data of two patients with the same disease); the other is the "negative sample pair", which are essentially different (such as the data of a diseased patient and a healthy patient). By comparing these two sets of samples, it can learn which features are important, thereby enhancing the features of the data and obtaining the "primary enhanced features". The knowledge graph enhancement module can be understood as a "medical knowledge base", which contains various medical knowledge (such as the relationships between diseases, the associations between symptoms and diseases, etc.). This module will use the dynamic medical knowledge graph to further enhance the previously obtained "primary enhanced features". By comparing and matching the data with the knowledge in the dynamic medical knowledge graph, it can better understand the meaning of the data, thereby obtaining the "secondary enhanced features".

[0033] In specific implementation, the dynamic medical knowledge graph can be automatically updated based on the real-time update mechanism of the data stream, through the automatic extraction technology of natural language processing (NLP) to automatically update the dynamic medical knowledge graph, and through the incremental update mechanism of machine learning to automatically update the dynamic medical knowledge graph.

[0034] The annotation prediction layer is constructed based on a multi-task annotation prediction module and an uncertainty estimation module; the multi-task annotation prediction module is used to predict the annotation results of medical image data, detection waveform data, electronic case data, and gene sequencing data respectively according to the secondary enhanced features; the uncertainty estimation module is used to predict the uncertainty of each annotation result, screen the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty. The multi-task annotation prediction module is constructed based on a multi-branch classification / regression network sharing underlying features, and is used to simultaneously predict multiple annotation types (such as lesion location, disease type, etc.), and adopts a modular design to support flexible expansion of new annotation tasks. The uncertainty estimation module is constructed based on a Bayesian neural network or a deep ensemble method, and is used to quantify prediction uncertainty and assist in clinical decision-making. By predicting the uncertainty of the annotation results, high-confidence data is automatically screened to reduce the negative impact of noisy annotations on the model; by adopting multi-expert consensus for historical data annotation, the bias of a single annotator is reduced, and the reliability of the training data is improved.

[0035] The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model. The feedback optimization layer is constructed based on an active learning module and a continuous learning module; the active learning module designs a sampling strategy based on uncertainty sampling and domain adaptation strategies, combined with medical expert knowledge, and is used to intelligently select the most valuable data for annotation; the continuous learning module is based on elastic weight consolidation and knowledge distillation, combined with a robust processing mechanism for changes in medical data distribution, and supports incremental learning of the model without forgetting existing knowledge; The core goal of the active learning module is to enable the model to automatically select the most valuable data for human experts to annotate, avoiding wasting energy on unimportant data. Uncertainty sampling (prioritize learning where the model is "uncertain"): When the model makes predictions, it assigns a "confidence score" to each sample. The active learning module will preferentially select these "low-confidence" samples for doctors or experts to annotate because this data may be the "knowledge blind spot" of the model; Diversity sampling (avoid repeated learning of similar data): If the model has learned some common cases (such as pneumonia), it will actively look for uncommon cases (such as rare diseases) for experts to annotate, avoiding continuous learning of repetitive things; Domain adaptation (optimize for the particularity of medicine): Medical data may be noisy (such as blurred images, incorrect medical records). The active learning module will identify these "difficult" samples and preferentially let experts confirm them to prevent the model from being misled.

[0036] The core goal of the continuous learning module is to enable the model to continuously learn new data without forgetting the knowledge learned previously. Elastic weight consolidation (EWC) (protect old knowledge): The model will add a "protective cover" to the "important knowledge points" learned previously so that they will not be easily overwritten during the training of new data; for example, if the model has learned the "characteristics of pneumonia", when a new batch of "heart disease" data comes in, it will not completely forget the diagnostic method of pneumonia; Knowledge distillation (the old model teaches the new model): When training a new model, let the old model act as a "teacher" to supervise the predictions of the new model to ensure that the new model does not deviate from the original correct direction; for example, after the new model has learned the new characteristics of "diabetes", it still needs to be checked by the old model to ensure that it can still correctly diagnose "hypertension"; Incremental learning (update gradually): Instead of retraining the model with all new data at once, update it batch by batch and gradually to avoid "collapse during learning"; for example, first let the model learn 100 new cases, test the effect, and then learn the next batch, which is more stable.

[0037] That is, the active learning module is used to let the model itself select "difficult" data for experts to label, thereby optimizing the model parameters, improving efficiency, and reducing repetitive labor. The continuous learning module is used to enable the model to continuously learn new things without forgetting old knowledge, ensuring long-term stability and reliability.

[0038] When only one modality of data is input in practical applications, the multi-modal medical data annotation model can still perform automatic annotation because each modality has its own dedicated feature extraction module.

[0039] Step S2: Obtain a large amount of historical medical data including medical image data, detection waveform data, electronic medical record data, and gene sequencing data, desensitize, preprocess, and label each piece of the historical medical data, and then construct a dataset; Step S3: Train a multi-modal data annotation model with the dataset, annotation loss function, and annotation reward function, and compress the multi-modal data annotation model during the training process; Step S4: Create a medical decision-making model based on a multi-modal data fusion layer, a feature interaction layer, and a diagnostic decision-making layer, and set a decision loss function and a decision reward function for the medical decision-making model; The multi-modal data fusion layer is used to extract a joint representation from medical image data, detection waveform data, electronic medical record data, and gene sequencing data; the feature interaction layer is used to perform feature interaction operations on the joint representation to obtain a fused feature; the diagnostic decision-making layer is used to output a medical strategy recommendation carrying disease classification, prognosis assessment, and treatment suggestions based on the fused feature; prognosis assessment refers to the prediction of the future development trend of a patient's disease, which can help doctors and patients understand the possible direction of the disease and formulate reasonable treatment plans and care plans.

[0040] Step S5: Train the medical decision-making model with the dataset, decision loss function, and decision reward function, and compress the medical decision-making model during the training process; Step S6: Deploy the trained multi-modal data annotation model and medical decision-making model, and perform model drift training on the deployed multi-modal data annotation model and medical decision-making model; Step S7: Collect real-time medical data and the latest medical knowledge, automatically annotate the real-time medical data and the latest medical knowledge with the deployed multi-modal data annotation model, and perform incremental training on the medical decision-making model with the annotated real-time medical data and the latest medical knowledge; Step S8: Conduct medical Q&A with the deployed medical decision-making model, and record a Q&A log including at least medical questions, medical strategy recommendations, Q&A time, questioner accounts, and misdiagnosis feedback in real time; Step S9: Annotate the Q&A log with the deployed multi-modal data annotation model, perform incremental training on the medical decision-making model with the annotated Q&A log, and encrypt and store the Q&A log in a distributed manner.

[0041] In the step S1, the image feature extraction module extracts initial image features from medical image data through a 3D convolutional neural network, integrates global context information in the initial image features through a Transformer to obtain context features, and filters the context features through an attention gating mechanism to obtain image features; That is, the 3D CNN is the "basic worker" responsible for extracting raw features; the Transformer is the "integration expert" responsible for stitching features into a more meaningful whole; the attention gating mechanism is the "quality inspector" responsible for screening out the most important features. Through this collaborative work, the medical image data feature extraction module can extract key information in the image more efficiently and accurately, providing strong support for subsequent diagnosis or analysis.

[0042] The waveform feature extraction module extracts waveform local features from the detection waveform data through a 1D temporal convolutional network, regularizes the order of each local feature through causal convolution, and captures temporal dependencies from each locally regularized feature through a recurrent neural network to obtain waveform features; That is, the 1D CNN is the "starter" responsible for extracting local features in the waveform data; causal convolution is the "relayer" ensuring that no "cheating" occurs during the feature extraction process (i.e., future information is not used); the RNN is the "sprinter" responsible for integrating time series information and capturing the dynamic changes of the waveform. Through this collaborative work, the detection waveform data feature extraction module can extract key features in the waveform more efficiently and accurately, providing strong support for subsequent analysis or detection.

[0043] The case feature extraction module extracts case features from the electronic case data through a Transformer model incorporating pre-trained medical term vectors; That is, the Transformer model is the "basic translator" responsible for converting electronic case data into feature vectors that can be understood by a computer; pre-training of medical term embedding is the "professional translator" helping the model better understand professional terms in the medical field. Through this collaborative work, the electronic medical record data feature extraction module can extract key features in the medical record more efficiently and accurately, providing strong support for subsequent diagnosis, treatment, or research.

[0044] The gene feature extraction module extracts key gene features from gene sequencing data through a convolutional autoencoder, learns the interactions between each key gene feature through a graph neural network to obtain gene enhanced features, and integrates semantic information for each gene enhanced feature through a gene ontology knowledge graph to obtain gene features; That is, the convolutional autoencoder is the "basic detective" responsible for extracting preliminary features from gene sequencing data; the graph neural network is the "relationship expert" that further improves the feature representation of genes by analyzing the interactions between genes; the Gene Ontology knowledge graph is the "knowledge base expert" that provides rich semantic information for genes and helps us better understand the biological significance of genes. Through this collaborative work, the gene sequencing data feature extraction module can extract gene features more efficiently and accurately, providing strong support for subsequent gene analysis, disease diagnosis, or drug research and development.

[0045] By integrating medical image data, detection waveform data, electronic case data, and gene sequencing data, patient information is comprehensively captured to avoid the limitations of a single modality; by setting the image features to combine 3D CNN and Transformer, both local details and global context are taken into account. The waveform features are processed by 1D temporal convolution and causal convolution to handle temporal dependencies. The case features are incorporated into pre-trained medical term vectors to enhance semantic understanding. The gene features are combined with graph neural networks and knowledge graphs to analyze the interactions between genes and their biological significance. Then, through cross-modal contrast learning and dynamic weighted fusion, the multi-modal collaborative effect is strengthened, effectively improving the feature extraction ability.

[0046] The spatio-temporal alignment module performs spatial alignment operations on the image features, waveform features, case features, and gene features through deformable convolution, and performs temporal alignment operations on the spatially aligned image features, waveform features, case features, and gene features through the time warping alignment network to complete the spatio-temporal alignment operation. That is, the deformable convolution is the "spatial adjuster" responsible for solving the spatial alignment problem of different modality features; the time warping alignment network is the "time adjuster" responsible for solving the temporal alignment problem of different modality features. Through this collaborative work, the spatio-temporal alignment module can align multi-modal features more efficiently and accurately, providing more consistent feature representations for subsequent tasks (such as multi-modal fusion, video understanding, or behavior recognition).

[0047] The multi-scale feature fusion module extracts correlation features from the spatio-temporally aligned image features, waveform features, case features, and gene features through the cross-attention mechanism, dynamically adjusts the weights of the image features, waveform features, case features, and gene features according to the correlation features through the gated fusion unit to fuse and obtain the fused features, and enhances the complementary feature expression of the fused features through inter-modal contrast learning. ​That is, the cross-attention mechanism is the "communicator" responsible for enabling mutual understanding between different features; the gated fusion unit is the "regulator" responsible for adjusting the degree of feature fusion according to importance; the cross-modal contrastive learning is the "optimizer" responsible for making the fused features more consistent and effective. Through this collaborative work, the multi-scale feature fusion module can fuse features from different sources more efficiently and accurately, providing a more powerful feature representation for subsequent tasks (such as classification, detection, or generation).

[0048] Through spatio-temporal alignment and multi-scale fusion, that is, by using deformable convolution and time warping alignment to solve the spatio-temporal inconsistency problem, and dynamically adjusting the weights through the cross-attention mechanism, the quality of the fused features is effectively improved, and thus the annotation quality is effectively improved.

[0049] The formula of the annotation loss function is as follows: L1 = λ1 * L vis + λ2 * L text + λ3 * L cons + λ4 * L fusionLtotal ; Among them, L1 represents the loss value of the annotation loss function; L vis represents the contrastive loss of medical image data, which is used to constrain the correlation between key image features and labels; L text represents the focal loss of electronic case data, which is used to alleviate the class imbalance problem; L cons represents the cross-modal consistency loss, which is used to constrain the consistency of different modal prediction distributions using KL divergence; L fusionLtotal represents the fusion enhancement loss, which is used to strengthen the multi-modal synergy effect using dynamic weighted cross-entropy; λ1, λ2, λ3, and λ4 all represent weight coefficients; The formula of the annotation reward function is as follows: R1 = α1 * R acc1 + α2 * R comp + α3 * R eff1 + α4 * R userRtotal ; Among them, R1 represents the reward value of the annotation reward function; R acc1 represents the annotation accuracy; R comp represents the modal complementarity; R eff1 represents the annotation efficiency; R userRtotal represents the user recognition; α1, α2, α3, and α4 all represent weight coefficients.

[0050] Step S2 is specifically as follows: Obtain a large amount of historical medical data including medical image data, detection waveform data, electronic case data, and gene sequencing data; Perform a deletion operation on the names in the metadata carried by each of the medical image data, and perform a blurring operation on the facial regions in each of the medical image data to complete the desensitization of each of the medical image data. Perform preprocessing such as noise reduction, grayscale transformation, image enhancement, image normalization, and image segmentation on each of the desensitized medical image data. Based on the preset image annotation specification, use the multi-annotator fusion method to annotate the patient region, disease classification, prognosis assessment, and treatment suggestions for each of the preprocessed medical image data, and generate corresponding labels; Prognosis assessment refers to the prediction of the future development trend of a patient's disease, which can help doctors and patients understand the possible course of the disease and formulate reasonable treatment plans and care plans.

[0051] Perform a deletion operation on the patient identifiers carried by each of the detection waveform data to complete desensitization. Perform preprocessing such as noise reduction, baseline correction, mean removal, linear trend removal, waveform clipping, and data segmentation on each of the desensitized detection waveform data. Based on the preset waveform annotation specification, use the multi-annotator fusion method to annotate the patient waveform, disease classification, prognosis assessment, and treatment suggestions for each of the preprocessed detection waveform data, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each of the electronic case data to complete desensitization. Perform preprocessing such as irrelevant content deletion, missing value filling, error data correction, format unification, and part-of-speech tagging on each of the desensitized electronic case data. Based on the preset case annotation specification, use the multi-annotator fusion method to annotate the disease classification, prognosis assessment, and treatment suggestions for each of the preprocessed electronic case data, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each of the gene sequencing data to complete desensitization. Perform preprocessing such as low-quality sequence removal, repetitive sequence removal, adapter sequence removal, and error correction on each of the desensitized gene sequencing data. Based on the preset gene annotation specification, use the multi-annotator fusion method to annotate the abnormal sequences, disease classification, prognosis assessment, and treatment suggestions for each of the preprocessed gene sequencing data, and generate corresponding labels; Construct a data set based on each of the annotated historical medical data; The specific step S3 is as follows: Divide the data set into a first training set, a first validation set, and a first test set according to a ratio of 8:1:1 by the stratified sampling method. Use the first training set to train the multi-modal data annotation model. In each round of training, alternately use the annotation loss function and the annotation reward function to optimize the multi-modal data annotation model. The training process combines weight pruning, weight quantization, and parameter sharing to compress the multi-modal data annotation model until the preset first early stopping condition is met; Calculate the annotation accuracy of the multi-modal data annotation model through the first validation set to verify the multi-modal data annotation model. If the verification fails, expand the first training set and continue training; if the verification passes, then: Calculate the annotation confidence of the multi-modal data annotation model through the first test set to test the multi-modal data annotation model. If the test fails, expand the first training set and continue training; if the test passes, end the training.

[0052] In step S4, the multi-modal data fusion layer is constructed based on a medical image branch, a waveform processing branch, a medical record coding branch, a gene analysis branch, and a cross-modal alignment module; the medical image branch is used to extract image features from medical image data through a 3D ResNet-50 network; the waveform processing branch is used to extract periodic features from detected waveform data through a 1D CNN-BiLSTM network; the medical record coding branch is used to extract temporal features from electronic medical record data through a BERT model; the gene analysis branch is used to extract genetic features from gene sequencing data through a GAT network; the cross-modal alignment module is used to align and fuse image features, periodic features, temporal features, and genetic features through a contrastive learning loss to obtain a joint representation; The feature interaction layer performs feature interaction on the joint representation through a cross-attention mechanism to obtain interaction features, fuses the interaction features through a dynamic graph neural network to obtain fusion features, and compresses redundant features in the fusion features through a feature distillation module; The feature interaction layer is an indispensable part of modern deep learning models. It promotes information exchange and fusion between different features, improving the model's overall understanding ability and prediction performance of data; in the fields of multi-modal learning, multi-task learning, recommendation systems, and natural language processing, the feature interaction layer plays an important role.

[0053] The formula for the decision loss function is: L2 = β1 * L diag + β2 * L path + β3 * L time ; where L2 represents the loss value of the decision loss function; L diag represents the diagnostic cross-entropy loss, which is used to measure the difference between the medical policy recommendation and the true label; L path represents the clinical pathway regularization loss, which uses KL divergence to measure the deviation of the model's decision path from the standard clinical pathway; L time represents the temporal consistency loss, which is used to ensure the rationality between consecutive decision steps; β1, β2, and β3 all represent weight coefficients; The formula for the decision reward function is: R2 = γ1 * Racc2 +γ2*R path +γ3*R risk +γ4*R eff2 ; wherein, R2 represents the reward value of the decision reward function; R acc2 represents the diagnostic accuracy reward, and the weighted F1 score is adopted; R path represents the path rationality reward, and the cosine similarity is adopted; R risk represents the risk aversion reward; R eff2 represents the diagnosis and treatment efficiency reward; γ1, γ2, γ3, and γ4 all represent weight coefficients; By fusing the contrastive loss and the focal loss through the annotation loss function, the class imbalance and the modality consistency are effectively solved; by fusing the KL divergence and the temporal consistency loss through the decision loss function, the rationality of the clinical path and the decision coherence are effectively constrained; by introducing indicators such as user recognition and diagnosis and treatment efficiency through the reward function, end-to-end optimization is achieved in combination with reinforcement learning.

[0054] The specific steps of step S5 are as follows: The data set is divided into a second training set, a second validation set, and a second test set according to a ratio of 8:1:1 by the stratified sampling method. The medical decision-making model is trained by the second training set. In each round of training, the decision loss function and the decision reward function are alternately used to optimize the medical decision-making model. The weight pruning, weight quantization, and parameter sharing are combined in the training process to compress the medical decision-making model until the preset second early stopping condition is met; The decision accuracy of the medical decision-making model is calculated through the second validation set to verify the medical decision-making model. If the verification fails, the second training set is expanded and training continues; if the verification passes, then: The decision confidence of the medical decision-making model is calculated through the second test set to test the medical decision-making model. If the test fails, the second training set is expanded and training continues; if the test passes, the training ends; By combining weight pruning, weight quantization, and parameter sharing in the model training stage, the model volume is effectively compressed, the consumption of computing resources is reduced, and it is adapted to the deployment of edge devices.

[0055] The specific steps of step S6 are as follows: The trained multi-modal data annotation model and medical decision-making model are locally deployed through containerization technology. A large amount of actual medical data is collected, and each piece of the actual medical data is desensitized, preprocessed, and annotated. Based on each piece of the annotated actual medical data, model drift training is performed on the deployed multi-modal data annotation model and medical decision-making model. The performance of the multi-modal data annotation model and medical decision-making model after model drift training is verified through accuracy, recall rate, F1 value, and ROC curve.

[0056] By dividing the training set, validation set, and test set in the ratio of 8:1:1, overfitting is avoided; when the verification fails, the training set is automatically expanded to effectively improve the generalization ability; the performance of the model is verified from multiple angles through accuracy, recall rate, F1 value, and ROC curve to ensure clinical reliability.

[0057] The specific content of step S8 is as follows: Through a visual interface, a medical problem carrying medical image data, detection waveform data, electronic medical record data, or gene sequencing data is obtained. After preprocessing the medical problem through a streaming computing engine, it is input into the deployed medical decision-making model. The medical decision-making model combines GPU acceleration technology and memory caching technology to infer a medical strategy recommendation for the medical problem, and feeds back the medical strategy recommendation through the visual interface for medical Q&A. A Q&A log is recorded in real time, including at least the medical problem, medical strategy recommendation, Q&A time, questioner account, and misdiagnosis feedback. The specific content of step S9 is as follows: The deployed multi-modal data annotation model automatically annotates the Q&A log. When the number of the annotated Q&A logs reaches a preset quantity threshold, the medical decision-making model is automatically incrementally trained through the annotated Q&A logs within a preset time period. Through real-time medical data and the Q&A log automatically annotated and fed back to the model, a "data - model" closed-loop optimization is formed; through model drift training combined with indicators such as the ROC curve for verification, long-term performance stability is ensured.

[0058] The Q&A log is encrypted to obtain an encrypted log, and the encrypted log is stored distributively. The data fingerprint of the encrypted log is calculated through the HMAC algorithm, and the data fingerprint is uploaded to the blockchain. By using a multi-modal data annotation model to automatically annotate the Q&A log carrying misdiagnosis feedback, the annotation efficiency is effectively improved; by generating a data fingerprint through the HMAC algorithm and uploading it to the chain, the traceability and anti-tampering of misdiagnosis cases are ensured; through an incremental training trigger mechanism, a quantity threshold is set, and when misdiagnosis cases accumulate to a clinically significant level, the model is automatically updated, effectively solving the problem of missing feedback, and then constructing a closed loop of "misdiagnosis cases - annotation feedback - model iteration".

[0059] The specific process of encrypting the Q&A log to obtain an encrypted log is as follows: Perform DCT transformation on the medical image data in the Q&A log to extract frequency domain coefficients, perform Zigzag scanning based on the frequency domain coefficients to convert the medical image data into a one-dimensional sequence; after encoding the text data including electronic medical record data, medical strategy suggestions, and misdiagnosis feedback using UTF-8 encoding, compress it through Huffman coding to obtain compressed data; perform wavelet packet decomposition on the detected waveform data to extract feature coefficients; convert the gene sequencing data into FASTQ format and perform base sequence encoding to obtain encoded data; Perform SM3 hash calculation on the one-dimensional sequence, compressed data, feature coefficients, and encoded data respectively to obtain corresponding sub-check codes, and perform SM3 hash calculation on each sub-check code to obtain a main check code. Encrypt the one-dimensional sequence into the first encrypted data through the AES-GCM algorithm, encrypt the compressed data into the second encrypted data through the SM9 algorithm, encrypt the feature coefficients into the third encrypted data through the ECIES algorithm, and encrypt the encoded data into the fourth encrypted data through the Paillier algorithm; Use the SM4 algorithm and the Serpent algorithm to perform sandwich encryption on the first encrypted data, second encrypted data, third encrypted data, fourth encrypted data, and the main check code to obtain an encrypted log.

[0060] By combining two SM3 hash calculations during the generation of the encrypted log, adopting the AES-GCM algorithm, SM9 algorithm, ECIES algorithm, and Paillier algorithm for different data types respectively, and finally using the SM4 algorithm and the Serpent algorithm for sandwich encryption, and uploading the data fingerprint of the encrypted log to the blockchain, multi-level encryption, integrity verification, and anti-tampering of the Q&A log are realized, greatly improving the security of Q&A log storage.

[0061] A preferred embodiment of a medical strategy generation system combining data flywheel reinforcement learning according to the present invention includes the following modules: A multi-modal data annotation model creation module, used to create a multi-modal data annotation model based on a feature extraction layer, a modality fusion layer, a feature enhancement layer, an annotation prediction layer, and a feedback optimization layer, and set the annotation loss function and annotation reward function of the multi-modal data annotation model; By setting each layer such as feature extraction, fusion, and enhancement to adopt a modular structure, it is convenient to add new data modalities (such as voice records, sensor data) or task types (such as drug recommendations) without reconstructing the overall framework; by supporting dynamic integration of the latest medical knowledge and user feedback (misdiagnosis feedback), it adapts to the characteristics of rapid iteration in the medical field.

[0062] The feature extraction layer is constructed based on an image feature extraction module, a waveform feature extraction module, a case feature extraction module, and a gene feature extraction module; the image feature extraction module is used to extract image features from medical image data; the waveform feature extraction module is used to extract waveform features from detection waveform data; the case feature extraction module is used to extract case features from electronic case data; the gene feature extraction module is used to extract gene features from gene sequencing data; The modality fusion layer is constructed based on a spatio-temporal alignment module and a multi-scale feature fusion module; the spatio-temporal alignment module is used to perform spatio-temporal alignment operations on the image features, waveform features, case features, and gene features; the multi-scale feature fusion module is used to fuse the image features, waveform features, case features, and gene features after spatio-temporal alignment to obtain fused features; The feature enhancement layer is constructed based on a self-supervised contrast learning module and a knowledge graph enhancement module; the self-supervised contrast learning module is used to construct positive sample pairs and negative sample pairs to perform enhancement operations on the fused features to obtain first-level enhanced features; the knowledge graph enhancement module is used to perform enhancement operations on the first-level enhanced features through a dynamic medical knowledge graph to obtain second-level enhanced features; The working mode of the self-supervised contrast learning module is a bit like a "spot the difference" game. It will find two sets of samples from the data: one set is "positive sample pairs", which are essentially similar (such as data of two patients with the same disease); the other set is "negative sample pairs", which are essentially different (such as data of a diseased patient and a healthy patient). By comparing these two sets of samples, it can learn which features are important, thereby enhancing the features of the data to obtain "first-level enhanced features". The knowledge graph enhancement module can be understood as a "medical knowledge base", which contains various medical knowledge (such as the relationships between diseases, the associations between symptoms and diseases, etc.). This module will use the dynamic medical knowledge graph to further enhance the previously obtained "first-level enhanced features". By comparing and matching the data with the knowledge in the dynamic medical knowledge graph, it can better understand the meaning of the data, thereby obtaining "second-level enhanced features".

[0063] In specific implementation, the dynamic medical knowledge graph can be automatically updated based on the real-time update mechanism of the data stream, through the automatic extraction technology of natural language processing (NLP) to automatically update the dynamic medical knowledge graph, and through the incremental update mechanism of machine learning to automatically update the dynamic medical knowledge graph.

[0064] The annotation prediction layer is constructed based on a multi-task annotation prediction module and an uncertainty estimation module; the multi-task annotation prediction module is used to predict the annotation results of medical image data, detection waveform data, electronic case data, and gene sequencing data respectively according to the secondary enhanced features; the uncertainty estimation module is used to predict the uncertainty of each of the annotation results, screen the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty; The multi-task annotation prediction module is constructed based on a multi-branch classification / regression network sharing underlying features, and is used to simultaneously predict multiple annotation types (such as lesion location, disease type, etc.), and adopts a modular design to support flexible expansion of new annotation tasks; The uncertainty estimation module is constructed based on a Bayesian neural network or a deep ensemble method, and is used to quantify prediction uncertainty and assist in clinical decision-making; By predicting the uncertainty of the annotation results, high-confidence data is automatically screened to reduce the negative impact of noisy annotations on the model; through multi-expert consensus in historical data annotation, single-annotator bias is reduced, and the reliability of training data is improved.

[0065] The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model; The feedback optimization layer is constructed based on an active learning module and a continual learning module; the active learning module designs a sampling strategy based on uncertainty sampling and domain adaptation strategies, combined with medical expert knowledge, and is used to intelligently select the most valuable data for annotation; the continual learning module is based on elastic weight consolidation and knowledge distillation, combined with a robust processing mechanism for changes in the distribution of medical data, and supports incremental learning of the model without forgetting existing knowledge; The core goal of the active learning module is to let the model automatically select the most valuable data for human experts to annotate, avoiding wasting energy on unimportant data. Uncertainty sampling (prioritize learning where the model is "uncertain"): When the model makes predictions, it assigns a "confidence score" to each sample. The active learning module will preferentially select these "low-confidence" samples for doctors or experts to annotate because these data may be the "knowledge blind spots" of the model; Diversity sampling (avoid repeating learning similar data): If the model has learned some common cases (such as pneumonia), it will actively look for uncommon cases (such as rare diseases) for experts to annotate, avoiding always learning repetitive things; Domain adaptation (optimize for the particularity of medicine): Medical data may be noisy (such as blurred images, incorrect medical records). The active learning module will identify these "difficult" samples and preferentially let experts confirm them to avoid the model being misled.

[0066] The core goal of the continuous learning module is to enable the model to continuously learn new data without forgetting the knowledge learned previously. Elastic Weight Consolidation (EWC) (protecting old knowledge): The model will add a "protective cover" to the "important knowledge points" learned previously, so that they will not be easily overwritten during the training with new data; for example, if the model has learned the "characteristics of pneumonia", when a new batch of "heart disease" data comes in, it will not completely forget the diagnostic method of pneumonia; Knowledge Distillation (the old model teaches the new model): When training the new model, let the old model act as the "teacher" to supervise the prediction of the new model to ensure that the new model does not deviate from the original correct direction; for example, after the new model has learned the new characteristics of "diabetes", it still needs to pass the inspection of the old model to ensure that it can still correctly diagnose "hypertension"; Incremental Learning (updating step by step): Instead of retraining the model with all the new data at once, update it batch by batch and step by step to avoid "collapsing during learning"; for example, first let the model learn 100 new cases, test the effect, and then learn the next batch, which is more stable.

[0067] That is, the active learning module is used to let the model select the "difficult" data for the expert to label by itself, thereby optimizing the model parameters, improving efficiency, and reducing repetitive labor. The continuous learning module is used to let the model continuously learn new things without forgetting old knowledge, ensuring long-term stability and reliability.

[0068] When only one modality of data is input in practical applications, the multi-modal medical data annotation model can still perform automatic annotation because each modality has its own dedicated feature extraction module.

[0069] The dataset construction module is used to obtain a large amount of historical medical data including medical image data, detection waveform data, electronic case data, and gene sequencing data, desensitize, preprocess, and annotate each piece of the historical medical data, and then construct a dataset; The multi-modal data annotation model training module is used to train the multi-modal data annotation model through the dataset, the annotation loss function, and the annotation reward function, and compress the multi-modal data annotation model during the training process; The medical decision-making model creation module is used to create a medical decision-making model based on the multi-modal data fusion layer, the feature interaction layer, and the diagnostic decision-making layer, and set the decision loss function and the decision reward function of the medical decision-making model; The multi-modal data fusion layer is used to extract a joint representation from medical image data, detection waveform data, electronic case data, and gene sequencing data; the feature interaction layer is used to perform feature interaction operations on the joint representation to obtain fused features; the diagnostic decision-making layer is used to output medical strategy suggestions carrying disease classification, prognosis assessment, and treatment recommendations based on the fused features; A medical decision-making model training module for training a medical decision-making model through the dataset, a decision loss function, and a decision reward function, and compressing the medical decision-making model during the training process; A model deployment module for deploying the trained multi-modal data annotation model and medical decision-making model, and performing model drift training on the deployed multi-modal data annotation model and medical decision-making model; An incremental knowledge training module for collecting real-time medical data and the latest medical knowledge, automatically annotating the real-time medical data and the latest medical knowledge through the deployed multi-modal data annotation model, and performing incremental training on the medical decision-making model through the annotated real-time medical data and the latest medical knowledge; A medical Q&A module for performing medical Q&A through the deployed medical decision-making model, and real-time recording a Q&A log including at least a medical question, a medical strategy suggestion, a Q&A time, a question-asking account, and a misdiagnosis feedback; A Q&A log management module for annotating the Q&A log through the deployed multi-modal data annotation model, performing incremental training on the medical decision-making model through the annotated Q&A log, and encrypting and storing the Q&A log in a distributed manner.

[0070] In the multi-modal data annotation model creation module, the image feature extraction module extracts initial image features from medical image data through a 3D convolutional neural network, integrates global context information in the initial image features through a Transformer to obtain context features, and filters the context features through an attention gating mechanism to obtain image features; That is, 3D CNN is the "basic worker" responsible for extracting raw features; Transformer is the "integration expert" responsible for stitching features into a more meaningful whole; the attention gating mechanism is the "quality inspector" responsible for screening out the most important features. Through this collaborative work, the medical image data feature extraction module can extract key information in the image more efficiently and accurately, providing strong support for subsequent diagnosis or analysis.

[0071] The waveform feature extraction module extracts waveform local features from detection waveform data through a 1D temporal convolutional network, regularizes the order of each local feature through causal convolution, and captures temporal dependencies from each local feature after order regularization through a recurrent neural network to obtain waveform features; That is, the 1D CNN is the "starter", responsible for extracting local features from waveform data; the causal convolution is the "relay runner", ensuring that there is no "cheating" in the feature extraction process (that is, future information will not be used); the RNN is the "sprinter", responsible for integrating time series information and capturing the dynamic changes of the waveform. Through this collaborative work, the waveform data feature extraction module can extract key features in the waveform more efficiently and accurately, providing strong support for subsequent analysis or detection.

[0072] The case feature extraction module extracts case features from electronic case data through a Transformer model incorporating pre-trained medical term vectors; That is, the Transformer model is the "basic translator", responsible for converting electronic case data into feature vectors that can be understood by a computer; the pre-training of medical term embedding is the "professional translator", helping the model better understand professional terms in the medical field. Through this collaborative work, the electronic medical record data feature extraction module can extract key features in the medical record more efficiently and accurately, providing strong support for subsequent diagnosis, treatment, or research.

[0073] The gene feature extraction module extracts key gene features from gene sequencing data through a convolutional autoencoder, learns the interactions between the key gene features through a graph neural network to obtain gene enhanced features, and integrates semantic information for each of the gene enhanced features through a gene ontology knowledge graph to obtain gene features; That is, the convolutional autoencoder is the "basic detective", responsible for extracting preliminary features from gene sequencing data; the graph neural network is the "relationship expert", further improving the feature representation of genes by analyzing the interactions between genes; the gene ontology knowledge graph is the "knowledge base expert", providing rich semantic information for genes and helping us better understand the biological significance of genes. Through this collaborative work, the gene sequencing data feature extraction module can extract gene features more efficiently and accurately, providing strong support for subsequent gene analysis, disease diagnosis, or drug research and development.

[0074] By integrating medical image data, detection waveform data, electronic case data, and gene sequencing data, patient information is comprehensively captured, avoiding the limitations of a single modality; by setting the image features to combine 3D CNN and Transformer, taking into account local details and global context, setting the waveform features to process temporal dependencies through 1D temporal convolution and causal convolution, integrating pre-trained medical term vectors into the case features to enhance semantic understanding, and combining the gene features with a graph neural network and a knowledge graph to analyze the interactions between genes and their biological significance, and then through cross-modal contrast learning and dynamic weighted fusion, the multi-modal synergistic effect is strengthened, effectively improving the feature extraction ability.

[0075] The spatio-temporal alignment module performs spatial alignment operations on image features, waveform features, case features, and gene features through deformable convolution, and performs temporal alignment operations on the spatially aligned image features, waveform features, case features, and gene features through a temporal warping alignment network to complete the spatio-temporal alignment operation; That is, the deformable convolution is the "spatial adjuster" responsible for solving the spatial alignment problem of different modality features; the temporal warping alignment network is the "temporal adjuster" responsible for solving the temporal alignment problem of different modality features. Through this collaborative work, the spatio-temporal alignment module can align multi-modal features more efficiently and accurately, providing a more consistent feature representation for subsequent tasks (such as multi-modal fusion, video understanding, or behavior recognition).

[0076] The multi-scale feature fusion module extracts correlation features from the spatio-temporally aligned image features, waveform features, case features, and gene features through a cross-attention mechanism, dynamically adjusts the weights of the image features, waveform features, case features, and gene features according to the correlation features through a gated fusion unit to fuse and obtain fused features, and enhances the complementary feature expression of the fused features through inter-modal contrast learning; That is, the cross-attention mechanism is the "communicator" responsible for enabling different features to understand each other; the gated fusion unit is the "regulator" responsible for adjusting the fusion degree of features according to importance; the inter-modal contrast learning is the "optimizer" responsible for making the fused features more consistent and effective. Through this collaborative work, the multi-scale feature fusion module can fuse features from different sources more efficiently and accurately, providing a more powerful feature representation for subsequent tasks (such as classification, detection, or generation).

[0077] Through spatio-temporal alignment and multi-scale fusion, that is, by using deformable convolution and temporal warping alignment to solve the spatio-temporal inconsistency problem, and dynamically adjusting weights through the cross-attention mechanism, the quality of the fused features is effectively improved, and thus the annotation quality is effectively improved.

[0078] The formula of the annotation loss function is: L1 = λ1 * L vis + λ2 * L text + λ3 * L cons + λ4 * L fusionLtotal ; where L1 represents the loss value of the annotation loss function; L vis represents the contrast loss of medical image data, which is used to constrain the correlation between key image features and labels; L text represents the focal loss of electronic case data, which is used to alleviate the class imbalance problem; L cons represents the cross-modal consistency loss, which is used to constrain the consistency of different modality prediction distributions using KL divergence; L fusionLtotalDenote the fusion enhancement loss, which is used to strengthen the multimodal synergy using dynamic weighted cross-entropy; λ1, λ2, λ3, and λ4 all denote weight coefficients; The formula for the described annotation reward function is: R1 = α1 * R acc1 + α2 * R comp + α3 * R eff1 + α4 * R userRtotal ; Among them, R1 represents the reward value of the annotation reward function; R acc1 represents the annotation accuracy; R comp represents the modality complementarity; R eff1 represents the annotation efficiency; R userRtotal represents the user recognition; α1, α2, α3, and α4 all denote weight coefficients.

[0079] The described dataset construction module is specifically used for: Obtain a large amount of historical medical data including medical image data, detection waveform data, electronic case data, and gene sequencing data; Perform a deletion operation on the names in the metadata carried by each of the medical image data, and perform a blurring operation on the facial regions in each of the medical image data to complete the desensitization of each of the medical image data. Perform preprocessing such as noise reduction, grayscale transformation, image enhancement, image normalization, and image segmentation on the desensitized medical image data. Based on the preset image annotation specification, perform annotation of the patient region, disease classification, prognosis assessment, and treatment suggestions on each of the preprocessed medical image data using the multi-annotator fusion method, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each of the detection waveform data to complete desensitization, and perform preprocessing such as noise reduction, baseline correction, mean removal, linear trend removal, waveform clipping, and data segmentation on the desensitized detection waveform data. Based on the preset waveform annotation specification, perform annotation of the patient waveform, disease classification, prognosis assessment, and treatment suggestions on each of the preprocessed detection waveform data using the multi-annotator fusion method, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each of the electronic case data to complete desensitization, and perform preprocessing such as irrelevant content deletion, missing value filling, error data correction, format unification, and part-of-speech tagging on the desensitized electronic case data. Based on the preset case annotation specification, perform annotation of the disease classification, prognosis assessment, and treatment suggestions on each of the preprocessed electronic case data using the multi-annotator fusion method, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried in each of the gene sequencing data to complete desensitization, and perform preprocessing on the desensitized gene sequencing data, including removing low-quality sequences, duplicate sequences, adapter sequences, and error correction. Based on a preset gene annotation specification, use a multi-annotator fusion method to annotate the preprocessed gene sequencing data for abnormal sequences, disease classification, prognosis assessment, and treatment recommendations, and generate corresponding tags; Construct a dataset based on each of the annotated historical medical data; The multi-modal data annotation model training module is specifically used for: Divide the dataset into a first training set, a first validation set, and a first test set according to a ratio of 8:1:1 by the stratified sampling method. Use the first training set to train the multi-modal data annotation model. In each round of training, alternately use the annotation loss function and the annotation reward function to optimize the multi-modal data annotation model. The training process combines weight pruning, weight quantization, and parameter sharing to compress the multi-modal data annotation model until the preset first early stopping condition is met; Calculate the annotation accuracy of the multi-modal data annotation model through the first validation set to verify the multi-modal data annotation model. If the verification fails, expand the first training set and continue training; if the verification passes, then: Calculate the annotation confidence of the multi-modal data annotation model through the first test set to test the multi-modal data annotation model. If the test fails, expand the first training set and continue training; if the test passes, end the training.

[0080] In the medical decision-making model creation module, the multi-modal data fusion layer is constructed based on a medical image branch, a waveform processing branch, a medical record coding branch, a gene analysis branch, and a cross-modal alignment module; the medical image branch is used to extract image features from medical image data through a 3D ResNet-50 network; the waveform processing branch is used to extract periodic features from detection waveform data through a 1DCNN-BiLSTM network; the medical record coding branch is used to extract temporal features from electronic medical record data through a BERT model; the gene analysis branch is used to extract genetic features from gene sequencing data through a GAT network; the cross-modal alignment module is used to align and fuse image features, periodic features, temporal features, and genetic features through a contrastive learning loss to obtain a joint representation; The feature interaction layer performs feature interaction on the joint representation through a cross-attention mechanism to obtain interaction features, fuses the interaction features through a dynamic graph neural network to obtain fused features, and compresses the redundant features in the fused features through a feature distillation module; The feature interaction layer is an indispensable part of modern deep learning models. By promoting information exchange and fusion between different features, it enhances the model's overall understanding ability and prediction performance of data. The feature interaction layer plays an important role in fields such as multi-modal learning, multi-task learning, recommendation systems, and natural language processing.

[0081] The formula for the decision loss function is as follows: L2 = β1 * L diag + β2 * L path + β3 * L time ; where L2 represents the loss value of the decision loss function; L diag represents the diagnostic cross-entropy loss, which is used to measure the difference between the medical strategy recommendation and the true label; L path represents the clinical pathway regularization loss, and the KL divergence is used to measure the deviation between the model's decision path and the standard clinical pathway; L time represents the temporal consistency loss, which is used to ensure the rationality between consecutive decision steps; β1, β2, and β3 all represent weight coefficients; The formula for the decision reward function is as follows: R2 = γ1 * R acc2 + γ2 * R path + γ3 * R risk + γ4 * R eff2 ; where R2 represents the reward value of the decision reward function; R acc2 represents the diagnostic accuracy reward, and the weighted F1 score is used; R path represents the path rationality reward, and the cosine similarity is used; R risk represents the risk aversion reward; R eff2 represents the diagnosis and treatment efficiency reward; γ1, γ2, γ3, and γ4 all represent weight coefficients; By fusing the contrast loss and the focal loss in the annotation loss function, the class imbalance and modality consistency are effectively solved. By fusing the KL divergence and the temporal consistency loss in the decision loss function, the rationality of the clinical pathway and the decision coherence are effectively constrained. By introducing indicators such as user recognition and diagnosis and treatment efficiency in the reward function and combining reinforcement learning, end-to-end optimization is achieved.

[0082] The medical decision model training module is specifically used for: The dataset is divided into a second training set, a second validation set, and a second test set according to a ratio of 8:1:1 by stratified sampling. The medical decision-making model is trained using the second training set. In each round of training, the decision loss function and the decision reward function are alternately used to optimize the medical decision-making model. The training process combines weight pruning, weight quantization, and parameter sharing to compress the medical decision-making model until a preset second early stopping condition is met; The decision accuracy of the medical decision-making model is calculated using the second validation set to verify the medical decision-making model. If the verification fails, the second training set is expanded and training continues; if the verification passes, then: The decision confidence of the medical decision-making model is calculated using the second test set to test the medical decision-making model. If the test fails, the second training set is expanded and training continues; if the test passes, training ends; By combining weight pruning, weight quantization, and parameter sharing during the model training stage, the model volume is effectively compressed, the consumption of computing resources is reduced, and it is adapted to edge device deployment.

[0083] The model deployment module is specifically used for: The trained multi-modal data annotation model and medical decision-making model are locally deployed through containerization technology. A large amount of actual medical data is collected, and each of the actual medical data is desensitized, preprocessed, and annotated. Based on each of the annotated actual medical data, the deployed multi-modal data annotation model and medical decision-making model are subjected to model drift training, and the performance of the multi-modal data annotation model and medical decision-making model after model drift training is verified through accuracy, recall rate, F1 value, and ROC curve.

[0084] By dividing the training set, validation set, and test set according to 8:1:1, overfitting is avoided; when the verification fails, the training set is automatically expanded, effectively improving the generalization ability; the model performance is verified from multiple angles through accuracy, recall rate, F1 value, and ROC curve to ensure clinical reliability.

[0085] The medical Q&A module is specifically used for: Through a visual interface, a medical question carrying medical image data, detection waveform data, electronic case data, or gene sequencing data is obtained. After preprocessing the medical question through a streaming computing engine, it is input into the deployed medical decision-making model. The medical decision-making model combines GPU acceleration technology and memory caching technology to reason about the medical question to obtain medical strategy suggestions, and the medical strategy suggestions are fed back through the visual interface for medical Q&A. A Q&A log including at least the medical question, medical strategy suggestions, Q&A time, questioner account, and misdiagnosis feedback is recorded in real time; The Q&A log management module is specifically used for: Automatically annotate the Q&A logs through the deployed multi-modal data annotation model. When the number of the annotated Q&A logs reaches a preset quantity threshold, incrementally train the medical decision-making model automatically with the annotated Q&A logs within a preset time period; Feed back the real-time medical data and the automatically annotated Q&A logs to the model to form a "data-model" closed-loop optimization; verify through model drift training combined with indicators such as the ROC curve to ensure long-term performance stability.

[0086] Encrypt the Q&A logs to obtain encrypted logs, store the encrypted logs distributively, calculate the data fingerprint of the encrypted logs through the HMAC algorithm, and upload the data fingerprint to the blockchain; Automatically annotate the Q&A logs carrying misdiagnosis feedback by adopting a multi-modal data annotation model, effectively improving the annotation efficiency; generate a data fingerprint through the HMAC algorithm and upload it to the chain to ensure the traceability and anti-tampering of misdiagnosis cases; set a quantity threshold through the incremental training trigger mechanism, and automatically trigger model update when the misdiagnosis cases accumulate to a clinically significant level, effectively solving the problem of missing feedback, and then constructing a closed loop of "misdiagnosis cases-annotation feedback-model iteration".

[0087] Specifically, encrypting the Q&A logs to obtain encrypted logs is as follows: Perform DCT transformation on the medical image data in the Q&A logs to extract frequency domain coefficients, perform Zigzag scanning based on the frequency domain coefficients to convert the medical image data into a one-dimensional sequence; after encoding the text data including electronic case data, medical strategy suggestions, and misdiagnosis feedback using UTF-8 encoding, compress it through Huffman coding to obtain compressed data; perform wavelet packet decomposition on the detection waveform data to extract feature coefficients; convert the gene sequencing data into the FASTQ format and perform base sequence encoding to obtain encoded data; Perform SM3 hash calculation on the one-dimensional sequence, compressed data, feature coefficients, and encoded data respectively to obtain corresponding sub-check codes, and perform SM3 hash calculation on each sub-check code to obtain a main check code, Encrypt the one-dimensional sequence into the first encrypted data through the AES-GCM algorithm, encrypt the compressed data into the second encrypted data through the SM9 algorithm, encrypt the feature coefficients into the third encrypted data through the ECIES algorithm, and encrypt the encoded data into the fourth encrypted data through the Paillier algorithm; Use the SM4 algorithm and the Serpent algorithm to perform sandwich encryption on the first encrypted data, second encrypted data, third encrypted data, fourth encrypted data, and the main check code to obtain encrypted logs.

[0088] By combining two SM3 hash calculations during the generation of encrypted logs, adopting the AES-GCM algorithm, SM9 algorithm, ECIES algorithm, and Paillier algorithm for different data types respectively, and finally using the SM4 algorithm and Serpent algorithm for sandwich encryption, and uploading the data fingerprint of the encrypted logs to the blockchain, multi-level encryption, integrity verification, and anti-tampering of the Q&A logs are achieved, greatly improving the security of Q&A log storage.

[0089] Although the specific implementation manners of the present invention are described above, those skilled in the art should understand that the specific embodiments we described are illustrative rather than used to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope protected by the claims of the present invention.

Claims

1. A medical policy generation method combining data flywheel and reinforcement learning, characterized in that: It includes the following steps: Step S1: Create a multi-modal data annotation model based on a feature extraction layer, a modality fusion layer, a feature enhancement layer, a label prediction layer, and a feedback optimization layer, and set the annotation loss function and annotation reward function of the multi-modal data annotation model; Step S2: Obtain a large amount of historical medical data including medical image data, detection waveform data, electronic case data, and gene sequencing data, desensitize, preprocess, and annotate each piece of the historical medical data, and then construct a data set; Step S3: Train the multi-modal data annotation model with the data set, annotation loss function, and annotation reward function, and compress the multi-modal data annotation model during the training process; Step S4: Create a medical decision-making model based on a multi-modal data fusion layer, a feature interaction layer, and a diagnostic decision layer, and set the decision loss function and decision reward function of the medical decision-making model; The multi-modal data fusion layer is used to extract joint representations from medical image data, detection waveform data, electronic case data, and gene sequencing data; The feature interaction layer is used to perform feature interaction operations on the joint representations to obtain fused features; the diagnostic decision layer is used to output medical strategy suggestions carrying disease classification, prognosis assessment, and treatment recommendations based on the fused features; Step S7: Train the medical decision-making model with the data set, decision loss function, and decision reward function, and compress the medical decision-making model during the training process; Step S8: Deploy the trained multi-modal data annotation model and medical decision-making model, and perform model drift training on the deployed multi-modal data annotation model and medical decision-making model; Step S9: Collect real-time medical data and the latest medical knowledge, automatically annotate the real-time medical data and the latest medical knowledge with the deployed multi-modal data annotation model, and perform incremental training on the medical decision-making model with the annotated real-time medical data and the latest medical knowledge; Step S10: Conduct medical Q&A through the deployed medical decision-making model, and record a Q&A log including at least medical questions, medical strategy suggestions, Q&A time, questioner accounts, and misdiagnosis feedback in real time; Step S11: Annotate the Q&A log with the deployed multi-modal data annotation model, perform incremental training on the medical decision-making model with the annotated Q&A log, and encrypt and store the Q&A log in a distributed manner.

2. The medical policy generation method combining data flywheel and reinforcement learning according to claim 1, wherein: In step S1, the feature extraction layer is constructed based on an image feature extraction module, a waveform feature extraction module, a case feature extraction module, and a gene feature extraction module; the image feature extraction module is used to extract image features from medical image data; the waveform feature extraction module is used to extract waveform features from detection waveform data; the case feature extraction module is used to extract case features from electronic case data; the gene feature extraction module is used to extract gene features from gene sequencing data; The modality fusion layer is constructed based on a spatio-temporal alignment module and a multi-scale feature fusion module; the spatio-temporal alignment module is used to perform spatio-temporal alignment operations on image features, waveform features, case features, and gene features; the multi-scale feature fusion module is used to fuse the image features, waveform features, case features, and gene features after spatio-temporal alignment to obtain fused features; The feature enhancement layer is constructed based on a self-supervised contrastive learning module and a knowledge graph enhancement module; the self-supervised contrastive learning module is used to construct positive and negative sample pairs to perform enhancement operations on the fused features to obtain first-level enhanced features; The knowledge graph enhancement module is used to perform enhancement operations on the first-level enhanced features through a dynamic medical knowledge graph to obtain second-level enhanced features; The annotation prediction layer is constructed based on a multi-task annotation prediction module and an uncertainty estimation module; the multi-task annotation prediction module is used to predict the annotation results of medical image data, detection waveform data, electronic case data, and gene sequencing data respectively according to the second-level enhanced features; the uncertainty estimation module is used to predict the uncertainty of each of the annotation results, screen the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty; The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model; The image feature extraction module extracts initial image features from medical image data through a 3D convolutional neural network, integrates global context information in the initial image features through a Transformer, and screens the context features through an attention gating mechanism to obtain image features; The waveform feature extraction module extracts waveform local features from detection waveform data through a 1D temporal convolutional network, regularizes the order of each of the local features through causal convolution, and captures temporal dependencies from each of the local features after order regularization through a recurrent neural network to obtain waveform features; The case feature extraction module extracts case features from electronic case data through a Transformer model incorporating pre-trained medical term vectors; The gene feature extraction module extracts gene key features from gene sequencing data through a convolutional autoencoder, learns the interactions between each of the gene key features through a graph neural network to obtain gene enhanced features, and integrates semantic information for each of the gene enhanced features through a gene ontology knowledge graph to obtain gene features; The spatio-temporal alignment module performs spatial alignment operations on image features, waveform features, case features, and gene features through deformable convolution, and performs temporal alignment operations on the image features, waveform features, case features, and gene features after spatial alignment through a time warping alignment network to complete the spatio-temporal alignment operation; The multi-scale feature fusion module extracts correlation features from the spatio-temporally aligned image features, waveform features, case features, and gene features through a cross-attention mechanism, dynamically adjusts the weights of the image features, waveform features, case features, and gene features according to the correlation features through a gated fusion unit to fuse and obtain fused features, and enhances the complementary feature expression of the fused features through cross-modal contrast learning; The formula of the annotation loss function is: L1 = λ1 * L vis + λ2 * L text + λ3 * L cons + λ4 * L fusionLtotal ; Among them, L1 represents the loss value of the annotation loss function; L vis represents the contrastive loss of medical image data, which is used to constrain the correlation between key image features and labels; L text represents the focal loss of electronic case data, which is used to alleviate the problem of class imbalance; L cons represents the cross-modal consistency loss, which is used to constrain the consistency of different modal prediction distributions using KL divergence; L fusionLtotal represents the fusion enhancement loss, which is used to strengthen the multi-modal synergy effect using dynamic weighted cross-entropy; λ1, λ2, λ3, and λ4 all represent weight coefficients; The formula of the annotation reward function is: R1 = α1 * R acc1 + α2 * R comp + α3 * R eff1 + α4 * R userRtotal ; Among them, R1 represents the reward value of the labeled reward function; R acc1 represents the labeling accuracy; R comp represents the modality complementarity; R eff1 represents the labeling efficiency; R userRtotal represents the user acceptance; α1, α2, α3, and α4 all represent weight coefficients.

3. The medical policy generation method combining data flywheel and reinforcement learning according to claim 1, wherein: The specific steps of step S2 are as follows: Obtain a large amount of historical medical data including medical image data, detection waveform data, electronic case data, and gene sequencing data; Perform a deletion operation on the names in the metadata carried by each medical image data, and perform a blurring operation on the facial regions in each medical image data to complete the desensitization of each medical image data. Perform preprocessing such as noise reduction, grayscale transformation, image enhancement, image normalization, and image segmentation on the desensitized medical image data. Based on a multi-annotator fusion method and a preset image annotation specification, perform annotations on the preprocessed medical image data for patient regions, disease classification, prognosis assessment, and treatment recommendations, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each detection waveform data to complete desensitization. Perform preprocessing such as noise reduction, baseline correction, mean removal, linear trend removal, waveform spike extinction, and data segmentation on the desensitized detection waveform data. Based on a multi-annotator fusion method and a preset waveform annotation specification, perform annotations on the preprocessed detection waveform data for patient waveforms, disease classification, prognosis assessment, and treatment recommendations, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each electronic case data to complete desensitization. Perform preprocessing such as irrelevant content deletion, missing value filling, error data correction, format unification, and part-of-speech tagging on the desensitized electronic case data. Based on a multi-annotator fusion method and a preset case annotation specification, perform annotations on the preprocessed electronic case data for disease classification, prognosis assessment, and treatment recommendations, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each gene sequencing data to complete desensitization. Perform preprocessing such as low-quality sequence removal, repetitive sequence removal, adapter sequence removal, and error correction on the desensitized gene sequencing data. Based on a multi-annotator fusion method and a preset gene annotation specification, perform annotations on the preprocessed gene sequencing data for abnormal sequences, disease classification, prognosis assessment, and treatment recommendations, and generate corresponding labels; Construct a data set based on each of the annotated historical medical data; The specific steps of step S3 are as follows: The dataset is divided into a first training set, a first validation set, and a first test set according to a ratio of 8:1:1 by the stratified sampling method. The multi-modal data annotation model is trained using the first training set. In each round of training, the annotation loss function and the annotation reward function are alternately used to optimize the multi-modal data annotation model. The training process combines weight pruning, weight quantization, and parameter sharing to compress the multi-modal data annotation model until a preset first early stopping condition is met; The annotation accuracy of the multi-modal data annotation model is calculated using the first validation set to verify the multi-modal data annotation model. If the verification fails, the first training set is expanded and training continues; if the verification passes, then: The annotation confidence of the multi-modal data annotation model is calculated using the first test set to test the multi-modal data annotation model. If the test fails, the first training set is expanded and training continues; if the test passes, training ends.

4. The medical policy generation method combining data flywheel and reinforcement learning according to claim 1, characterized in that: In step S4, the multi-modal data fusion layer is constructed based on a medical image branch, a waveform processing branch, a medical record coding branch, a gene analysis branch, and a cross-modal alignment module; the medical image branch is used to extract image features from medical image data through a 3D ResNet-50 network; the waveform processing branch is used to extract periodic features from detection waveform data through a 1D CNN-BiLSTM network; the medical record coding branch is used to extract temporal features from electronic medical record data through a BERT model; the gene analysis branch is used to extract genetic features from gene sequencing data through a GAT network; the cross-modal alignment module is used to align and fuse the image features, periodic features, temporal features, and genetic features through a contrastive learning loss to obtain a joint representation; The feature interaction layer performs feature interaction on the joint representation through a cross-attention mechanism to obtain interaction features, fuses the interaction features through a dynamic graph neural network to obtain fusion features, and compresses the redundant features in the fusion features through a feature distillation module; The formula for the decision loss function is: L2 = β1 * L diag + β2 * L path + β3 * L time ; Among them, L2 represents the loss value of the decision loss function; L diag represents the diagnostic cross-entropy loss, which is used to measure the difference between the medical strategy recommendation and the true label; L path represents the clinical pathway regularization loss, and the KL divergence is used to measure the deviation between the model decision path and the standard clinical pathway; L time represents the temporal consistency loss, which is used to ensure the rationality between consecutive decision steps; β1, β2, and β3 all represent weight coefficients; The formula for the decision reward function is: R2 = γ1 * R acc2 + γ2 * R path + γ3 * R risk + γ4 * R eff2 ; Among them, R2 represents the reward value of the decision reward function; R acc2 represents the diagnostic accuracy reward, and the weighted F1 score is adopted; R path represents the path rationality reward, and the cosine similarity is adopted; R risk represents the risk aversion reward; R eff2 represents the diagnosis and treatment efficiency reward; γ1, γ2, γ3, and γ4 all represent weight coefficients; Step S5 is specifically as follows: The dataset is divided into a second training set, a second validation set, and a second test set according to a ratio of 8:1:1 by the stratified sampling method. The medical decision-making model is trained using the second training set. In each round of training, the decision loss function and the decision reward function are alternately used to optimize the medical decision-making model. The training process combines weight pruning, weight quantization, and parameter sharing to compress the medical decision-making model until a preset second early stopping condition is met; The decision accuracy of the medical decision-making model is calculated using the second validation set to verify the medical decision-making model. If the verification fails, the second training set is expanded and training continues; if the verification passes, then: The decision confidence of the medical decision-making model is calculated using the second test set to test the medical decision-making model. If the test fails, the second training set is expanded and training continues; if the test passes, training ends; Step S6 is specifically as follows: The trained multi-modal data annotation model and medical decision-making model are locally deployed through containerization technology, a large amount of actual medical data is collected, and each piece of the actual medical data is desensitized, preprocessed, and annotated. Based on each piece of the annotated actual medical data, model drift training is performed on the deployed multi-modal data annotation model and medical decision-making model, and the performance of the multi-modal data annotation model and medical decision-making model after model drift training is verified through accuracy, recall rate, F1 value, and ROC curve.

5. The medical policy generation method combining data flywheel and reinforcement learning according to claim 1, characterized in that: The specific steps of step S8 are as follows: Through a visual interface, obtain a medical problem carrying medical image data, detection waveform data, electronic medical record data, or gene sequencing data. After preprocessing the medical problem through a streaming computing engine, input it into the deployed medical decision-making model. The medical decision-making model combines GPU acceleration technology and memory caching technology to infer a medical strategy recommendation for the medical problem, and feedback the medical strategy recommendation through the visual interface for medical Q&A, and record a Q&A log including at least the medical problem, medical strategy recommendation, Q&A time, questioner account, and misdiagnosis feedback in real time; The specific steps of step S9 are as follows: Automatically annotate the Q&A log through the deployed multi-modal data annotation model. When the number of the annotated Q&A logs reaches a preset quantity threshold, automatically perform incremental training on the medical decision-making model through the annotated Q&A logs within a preset time period; Encrypt the Q&A log to obtain an encrypted log, perform distributed storage on the encrypted log, calculate the data fingerprint of the encrypted log through the HMAC algorithm, and upload the data fingerprint to the blockchain; The specific process of encrypting the Q&A log to obtain an encrypted log is as follows: Perform DCT transformation on the medical image data in the Q&A log to extract frequency domain coefficients, perform Zigzag scanning based on the frequency domain coefficients to convert the medical image data into a one-dimensional sequence; after encoding the text data including electronic medical record data, medical strategy recommendation, and misdiagnosis feedback using UTF-8 encoding, compress it through Huffman encoding to obtain compressed data; Perform wavelet packet decomposition on the detection waveform data to extract feature coefficients; Convert the gene sequencing data into FASTQ format and perform base sequence encoding to obtain encoded data; Perform SM3 hash calculation on the one-dimensional sequence, compressed data, feature coefficients, and encoded data respectively to obtain corresponding sub-check codes, and perform SM3 hash calculation on each sub-check code to obtain a main check code; Encrypt the one-dimensional sequence into first encrypted data through the AES-GCM algorithm, encrypt the compressed data into second encrypted data through the SM9 algorithm, encrypt the feature coefficients into third encrypted data through the ECIES algorithm, and encrypt the encoded data into fourth encrypted data through the Paillier algorithm; Use the SM4 algorithm and the Serpent algorithm to perform sandwich encryption on the first encrypted data, second encrypted data, third encrypted data, fourth encrypted data, and the main check code to obtain an encrypted log.

6. A medical policy generation system that combines data flywheel reinforcement learning, characterized in that: It includes the following modules: A multi-modal data annotation model creation module, which is used to create a multi-modal data annotation model based on a feature extraction layer, a modality fusion layer, a feature enhancement layer, an annotation prediction layer, and a feedback optimization layer, and set an annotation loss function and an annotation reward function of the multi-modal data annotation model; A dataset construction module, which is used to obtain a large amount of historical medical data including medical image data, detection waveform data, electronic case data, and gene sequencing data, desensitize, preprocess, and annotate each piece of the historical medical data, and then construct a dataset; A multi-modal data annotation model training module, which is used to train the multi-modal data annotation model through the dataset, the annotation loss function, and the annotation reward function, and compress the multi-modal data annotation model during the training process; A medical decision-making model creation module, which is used to create a medical decision-making model based on a multi-modal data fusion layer, a feature interaction layer, and a diagnostic decision-making layer, and set a decision loss function and a decision reward function of the medical decision-making model; The multi-modal data fusion layer is used to extract a joint representation from medical image data, detection waveform data, electronic case data, and gene sequencing data; The feature interaction layer is used to perform a feature interaction operation on the joint representation to obtain a fused feature; the diagnostic decision-making layer is used to output a medical strategy recommendation carrying disease classification, prognosis assessment, and treatment suggestions based on the fused feature; A medical decision-making model training module, which is used to train the medical decision-making model through the dataset, the decision loss function, and the decision reward function, and compress the medical decision-making model during the training process; A model deployment module, which is used to deploy the trained multi-modal data annotation model and medical decision-making model, and perform model drift training on the deployed multi-modal data annotation model and medical decision-making model; An incremental knowledge training module, which is used to collect real-time medical data and the latest medical knowledge, automatically annotate the real-time medical data and the latest medical knowledge through the deployed multi-modal data annotation model, and perform incremental training on the medical decision-making model through the annotated real-time medical data and the latest medical knowledge; A medical Q&A module, which is used to perform medical Q&A through the deployed medical decision-making model, and record a Q&A log including at least medical questions, medical strategy recommendations, Q&A time, question-asking accounts, and misdiagnosis feedback in real time; A Q&A log management module, which is used to annotate the Q&A log through the deployed multi-modal data annotation model, perform incremental training on the medical decision-making model through the annotated Q&A log, and encrypt and store the Q&A log distributively.

7. The medical policy generation system combining data flywheel and reinforcement learning according to claim 6, characterized in that: In the multi-modal data annotation model creation module, the feature extraction layer is constructed based on an image feature extraction module, a waveform feature extraction module, a case feature extraction module, and a gene feature extraction module; the image feature extraction module is used to extract image features from medical image data; the waveform feature extraction module is used to extract waveform features from detection waveform data; the case feature extraction module is used to extract case features from electronic case data; the gene feature extraction module is used to extract gene features from gene sequencing data; The modality fusion layer is constructed based on a spatio-temporal alignment module and a multi-scale feature fusion module; the spatio-temporal alignment module is used to perform spatio-temporal alignment operations on image features, waveform features, case features, and gene features; the multi-scale feature fusion module is used to fuse the image features, waveform features, case features, and gene features after spatio-temporal alignment to obtain fused features; The feature enhancement layer is constructed based on a self-supervised contrast learning module and a knowledge graph enhancement module; the self-supervised contrast learning module is used to construct positive and negative sample pairs to perform enhancement operations on the fused features to obtain first-level enhanced features; The knowledge graph enhancement module is used to perform enhancement operations on the first-level enhanced features through a dynamic medical knowledge graph to obtain second-level enhanced features; The annotation prediction layer is constructed based on a multi-task annotation prediction module and an uncertainty estimation module; the multi-task annotation prediction module is used to predict the annotation results of medical image data, detection waveform data, electronic case data, and gene sequencing data respectively according to the second-level enhanced features; the uncertainty estimation module is used to predict the uncertainty of each of the annotation results, screen the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty; The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model; The image feature extraction module extracts initial image features from medical image data through a 3D convolutional neural network, integrates global context information in the initial image features through a Transformer, and screens the context features through an attention gating mechanism to obtain image features; The waveform feature extraction module extracts waveform local features from detection waveform data through a 1D temporal convolutional network, regularizes the order of each of the local features through causal convolution, and captures temporal dependencies from each of the local features after order regularization through a recurrent neural network to obtain waveform features; The case feature extraction module extracts case features from electronic case data through a Transformer model incorporating pre-trained medical term vectors; The gene feature extraction module extracts gene key features from gene sequencing data through a convolutional autoencoder, learns the interactions between each of the gene key features through a graph neural network to obtain gene enhanced features, and integrates semantic information for each of the gene enhanced features through a gene ontology knowledge graph to obtain gene features; The spatio-temporal alignment module performs spatial alignment operations on image features, waveform features, case features, and gene features through deformable convolution, and performs temporal alignment operations on the image features, waveform features, case features, and gene features after spatial alignment through a time warping alignment network to complete the spatio-temporal alignment operation; The multi-scale feature fusion module extracts correlation features from the spatio-temporally aligned image features, waveform features, case features, and gene features through a cross-attention mechanism, dynamically adjusts the weights of the image features, waveform features, case features, and gene features according to the correlation features through a gated fusion unit to fuse and obtain fusion features, and enhances the complementary feature expression of the fusion features through inter-modal contrast learning; The formula of the annotation loss function is: L1 = λ1 * L vis + λ2 * L text + λ3 * L cons + λ4 * L fusionLtotal ; Among them, L1 represents the loss value of the annotation loss function; L vis represents the contrastive loss of medical image data, which is used to constrain the correlation between key image features and labels; L text represents the focal loss of electronic case data, which is used to alleviate the class imbalance problem; L cons represents the cross-modal consistency loss, which is used to constrain the consistency of different modal prediction distributions using KL divergence; L fusionLtotal represents the fusion enhancement loss, which is used to strengthen the multi-modal synergistic effect using dynamic weighted cross-entropy; λ1, λ2, λ3, and λ4 all represent weight coefficients; The formula of the annotation reward function is: R1 = α1 * R acc1 + α2 * R comp + α3 * R eff1 + α4 * R userRtotal ; Among them, R1 represents the reward value of the labeled reward function; R acc1 represents the labeling accuracy; R comp represents the modality complementarity; R eff1 represents the labeling efficiency; R userRtotal represents the user acceptance; α1, α2, α3, and α4 all represent weight coefficients.

8. The medical policy generation system combining data flywheel and reinforcement learning according to claim 6, wherein: The dataset construction module is specifically used for: Obtain a large amount of historical medical data including medical image data, detection waveform data, electronic case data, and gene sequencing data; Perform a deletion operation on the names in the metadata carried by each medical image data, and perform a blurring operation on the facial regions in each medical image data to complete the desensitization of each medical image data. Perform preprocessing such as noise reduction, grayscale transformation, image enhancement, image normalization, and image segmentation on the desensitized medical image data. Based on a multi-annotator fusion method and a preset image annotation specification, perform annotations on the preprocessed medical image data for patient regions, disease classification, prognosis assessment, and treatment recommendations, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each detection waveform data to complete desensitization, and perform preprocessing such as noise reduction, baseline correction, mean removal, linear trend removal, waveform spike extinction, and data segmentation on the desensitized detection waveform data. Based on a multi-annotator fusion method and a preset waveform annotation specification, perform annotations on the preprocessed detection waveform data for patient waveforms, disease classification, prognosis assessment, and treatment recommendations, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each electronic case data to complete desensitization, and perform preprocessing such as irrelevant content deletion, missing value filling, error data correction, format unification, and part-of-speech tagging on the desensitized electronic case data. Based on a multi-annotator fusion method and a preset case annotation specification, perform annotations on the preprocessed electronic case data for disease classification, prognosis assessment, and treatment recommendations, and generate corresponding labels; Perform a deletion operation on the patient identifiers carried by each gene sequencing data to complete desensitization, and perform preprocessing such as low-quality sequence removal, repetitive sequence removal, adapter sequence removal, and error correction on the desensitized gene sequencing data. Based on a multi-annotator fusion method and a preset gene annotation specification, perform annotations on the preprocessed gene sequencing data for abnormal sequences, disease classification, prognosis assessment, and treatment recommendations, and generate corresponding labels; Construct a dataset based on the annotated historical medical data; The multi-modal data annotation model training module is specifically used for: The dataset is divided into a first training set, a first validation set, and a first test set according to a ratio of 8:1:1 by the stratified sampling method. The multi-modal data annotation model is trained using the first training set. In each round of training, the annotation loss function and the annotation reward function are alternately used to optimize the multi-modal data annotation model. During the training process, weight pruning, weight quantization, and parameter sharing are combined to compress the multi-modal data annotation model until a preset first early stopping condition is met; The annotation accuracy of the multi-modal data annotation model is calculated using the first validation set to verify the multi-modal data annotation model. If the verification fails, the first training set is expanded and training continues; if the verification passes, then: The annotation confidence of the multi-modal data annotation model is calculated using the first test set to test the multi-modal data annotation model. If the test fails, the first training set is expanded and training continues; if the test passes, training ends.

9. The medical policy generation system combining data flywheel and reinforcement learning according to claim 6, characterized in that: In the medical decision-making model creation module, the multi-modal data fusion layer is constructed based on a medical image branch, a waveform processing branch, a medical record coding branch, a gene analysis branch, and a cross-modal alignment module; the medical image branch is used to extract image features from medical image data through a 3D ResNet-50 network; the waveform processing branch is used to extract periodic features from detection waveform data through a 1D CNN-BiLSTM network; the medical record coding branch is used to extract temporal features from electronic medical record data through a BERT model; the gene analysis branch is used to extract genetic features from gene sequencing data through a GAT network; the cross-modal alignment module is used to align and fuse the image features, periodic features, temporal features, and genetic features through a contrastive learning loss to obtain a joint representation; The feature interaction layer performs feature interaction on the joint representation through a cross-attention mechanism to obtain interaction features, fuses the interaction features through a dynamic graph neural network to obtain fusion features, and compresses the redundant features in the fusion features through a feature distillation module; The formula for the decision loss function is: L2 = β1 * L diag + β2 * L path + β3 * L time ; Among them, L2 represents the loss value of the decision loss function; L diag represents the diagnostic cross-entropy loss, which is used to measure the difference between the medical strategy recommendation and the true label; L path represents the clinical pathway regularization loss, and the KL divergence is used to measure the deviation between the model decision path and the standard clinical pathway; L time represents the temporal consistency loss, which is used to ensure the rationality between consecutive decision steps; β1, β2, and β3 all represent weight coefficients; The formula for the decision reward function is: R2 = γ1 * R acc2 + γ2 * R path + γ3 * R risk + γ4 * R eff2 ; Among them, R2 represents the reward value of the decision reward function; R acc2 represents the diagnostic accuracy reward, using the weighted F1 score; R path represents the path rationality reward, using the cosine similarity; R risk represents the risk aversion reward; R eff2 represents the diagnosis and treatment efficiency reward; γ1, γ2, γ3, γ4 all represent weight coefficients; The medical decision-making model training module is specifically used for: The dataset is divided into a second training set, a second validation set, and a second test set according to a ratio of 8:1:1 by the stratified sampling method. The medical decision-making model is trained using the second training set. In each round of training, the decision loss function and the decision reward function are alternately used to optimize the medical decision-making model. During the training process, weight pruning, weight quantization, and parameter sharing are combined to compress the medical decision-making model until a preset second early stopping condition is met; The decision accuracy of the medical decision-making model is calculated using the second validation set to verify the medical decision-making model. If the verification fails, the second training set is expanded and training continues; if the verification passes, then: The decision confidence of the medical decision-making model is calculated using the second test set to test the medical decision-making model. If the test fails, the second training set is expanded and training continues; if the test passes, training ends; The model deployment module is specifically used for: Locally deploying the trained multi-modal data annotation model and medical decision-making model through containerization technology, collecting a large amount of actual medical data, desensitizing, preprocessing, and annotating each piece of the actual medical data, performing model drift training on the deployed multi-modal data annotation model and medical decision-making model based on each piece of the annotated actual medical data, and verifying the performance of the multi-modal data annotation model and medical decision-making model after model drift training through accuracy, recall rate, F1 value, and ROC curve.

10. A medical policy generation system integrating data flywheel and reinforcement learning, characterized in that: The medical Q&A module is specifically used for: Obtaining, through a visualization interface, a medical question carrying medical image data, detection waveform data, electronic medical record data, or gene sequencing data, preprocessing the medical question through a stream computing engine and then inputting it into the deployed medical decision-making model. The medical decision-making model combines GPU acceleration technology and memory caching technology to infer a medical strategy recommendation for the medical question, and feeds back the medical strategy recommendation through the visualization interface for medical Q&A, and real-time records a Q&A log including at least the medical question, medical strategy recommendation, Q&A time, questioner account, and misdiagnosis feedback; The Q&A log management module is specifically used for: Automatically annotating the Q&A log through the deployed multi-modal data annotation model, and when the number of the annotated Q&A logs reaches a preset quantity threshold, automatically performing incremental training on the medical decision-making model through the annotated Q&A logs within a preset time period; Encrypting the Q&A log to obtain an encrypted log, storing the encrypted log distributively, calculating the data fingerprint of the encrypted log through the HMAC algorithm, and uploading the data fingerprint to the blockchain; The encrypting the Q&A log to obtain an encrypted log is specifically: Performing DCT transformation on the medical image data in the Q&A log to extract frequency domain coefficients, performing Zigzag scanning based on the frequency domain coefficients to convert the medical image data into a one-dimensional sequence; encoding the text data including electronic medical record data, medical strategy recommendation, and misdiagnosis feedback using UTF-8 encoding, and then compressing it through Huffman coding to obtain compressed data; Performing wavelet packet decomposition on the detection waveform data to extract feature coefficients; Converting the gene sequencing data into FASTQ format and performing base sequence encoding to obtain encoded data; Performing SM3 hashing calculation on the one-dimensional sequence, compressed data, feature coefficients, and encoded data respectively to obtain corresponding sub-check codes, and performing SM3 hashing calculation on each of the sub-check codes to obtain a main check code; Encrypting the one-dimensional sequence into first encrypted data through the AES-GCM algorithm, encrypting the compressed data into second encrypted data through the SM9 algorithm, encrypting the feature coefficients into third encrypted data through the ECIES algorithm, and encrypting the encoded data into fourth encrypted data through the Paillier algorithm; Using the SM4 algorithm and the Serpent algorithm, perform sandwich encryption on the first encrypted data, the second encrypted data, the third encrypted data, the fourth encrypted data, and the main check code to obtain an encrypted log.

Images