A medical strategy generation method and system combining data flywheel reinforcement learning

Abstract

The application provides an artificial intelligence and clinical medical cross technical field medical strategy generation method and system combined with data flywheel reinforcement learning, which comprises the following steps: S1, creating a multi-modal data labeling model, setting a labeling loss function and a labeling reward function; S2, obtaining a large amount of historical medical data to construct a data set; S3, training and compressing the multi-modal data labeling model; S4, creating a medical decision model, setting a decision loss function and a decision reward function; S5, training and compressing the medical decision model; S6, deploying the trained model; S7, labeling real-time medical data through the multi-modal data labeling model, and incrementally training the medical decision model through the labeled real-time medical data; and S8, performing medical question answering through the deployed medical decision model. The application has the advantages that the accuracy, timeliness and safety of medical decision are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and clinical medicine, and in particular to a method and system for generating medical strategies by combining data flywheel reinforcement learning. Background Technology

[0002] Clinical Decision Support Systems (CDSS), as a core component of modern medical information systems, have evolved in tandem with the development of medical informatics. Since the MYCIN system pioneered rule-based expert systems in the 1970s, it has undergone four generations of technological transformation: the first generation (1970-1985) was based on a symbolic logic reasoning engine, using an IF-THEN rule base for disease diagnosis reasoning. A typical example is the INTERNIST-1 internal medicine diagnostic system, whose knowledge base contained association rules for 572 diseases and 4,500 symptoms; the second generation (1986-2000) introduced probabilistic graphical models such as Bayesian networks and Hidden Markov Models, exemplified by the QMR-DT system at the University of Washington. The first generation of systems achieved the calculation of the probability of multiple diseases occurring simultaneously, improving the diagnostic accuracy to 78%. The third generation (2001-2015) integrated supervised machine learning algorithms. Stanford University's DeepDx system, using SVM and random forest models, achieved an AUC value of 91% in the diabetic retinopathy classification task. The fourth generation (2016-present) began to explore the application of deep learning technology. Google Health's fundus image analysis system uses the Inception-v3 network to achieve automatic grading of diabetic macular edema, with a sensitivity of 94.5%.

[0003] Current mainstream clinical decision support systems mostly adopt a hybrid architecture, integrating heterogeneous data from multiple sources such as electronic medical record (EMR) systems, laboratory information systems (LIS), and picture archiving systems (PACS). A typical system, such as Epic's CognitiveComputing Platform, uses natural language processing technology to parse unstructured medical record text and combines it with predefined clinical pathways and diagnostic rule bases (including over 5,000 clinical pathways such as the NCCN guidelines) to provide medical staff with services such as diagnostic suggestions, treatment recommendations, and medication safety warnings. However, current clinical decision support systems face a vicious cycle of "data silos - rigid models - limited applications," specifically manifested in the following dimensions:

[0004] 1. Lack of feedback on misdiagnosed cases: Traditionally, retraining misdiagnosed cases requires manual annotation, which is not only time-consuming and labor-intensive, but also raises issues of responsibility for annotation errors. Therefore, most misdiagnosed cases in clinical practice cannot be effectively fed back into the training system. Mayo Clinic statistics show that only 12% of misdiagnosed reports are structured and entered into the system, resulting in serious sample bias in the continuous training of the model.

[0005] 2. Challenges in 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 often employ early fusion (such as feature-level splicing in the MITK framework) or late fusion (such as decision-level weighted voting), resulting in the loss of cross-modal semantic associations.

[0006] 3. Clinical mismatch in reward mechanism: Current reinforcement learning uses a single diagnostic accuracy as the reward function (such as DeepMind Health's retinopathy system), but ignores the assessment of the rationality of clinical pathways; according to the NEJM study, this design leads to 15% of recommended protocols having the risk of drug interactions or violating treatment guidelines.

[0007] 4. Knowledge update lag effect: The system update cycle generally exceeds 6 months, which creates a gap with the iteration speed of medical knowledge; taking tumor targeted therapy as an example, of the 8 drugs newly approved by the FDA in 2023, only 2 were integrated into the mainstream CDSS knowledge base within 6 months.

[0008] 5. Insufficient response time: Due to factors such as limited computing power resources of the deployment terminal, large model size, and preprocessing of multi-source heterogeneous data, the response time is not good; 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.

[0009] 6. Security risks of patient privacy data: For the purpose of facilitating later traceability, retraining, and medical quality monitoring, the conversation data generated during the question-and-answer process is 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 cause a large-scale data leakage incident, which poses a great security risk.

[0010] 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 decisions has become an urgent technical problem to be solved. Summary of the Invention

[0011] The technical problem to be solved by this invention is to provide a medical strategy generation method and system that combines data flywheel reinforcement learning, so as to improve the accuracy, timeliness and safety of medical decision-making.

[0012] In a first aspect, the present invention provides a method for generating medical strategies by combining data flywheel reinforcement learning, comprising the following steps:

[0013] Step S1: Create a multi-modal data annotation model based on the feature extraction layer, modality fusion layer, feature enhancement layer, annotation prediction layer, and feedback optimization layer; and set the annotation loss function and annotation reward function of the multi-modal data annotation model.

[0014] 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; and the gene feature extraction module is used to extract gene features from gene sequencing data.

[0015] The modality fusion layer is constructed based on a spatiotemporal alignment module and a multi-scale feature fusion module. The spatiotemporal alignment module is used to perform spatiotemporal alignment operations on image features, waveform features, case features, and gene features. The multi-scale feature fusion module is used to fuse the spatiotemporally aligned image features, waveform features, case features, and gene features to obtain fused features.

[0016] 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.

[0017] 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 medical record data and gene sequencing data respectively based on the secondary enhancement features. The uncertainty estimation module is used to predict the uncertainty of each annotation result, filter the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty.

[0018] The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model;

[0019] Step S2: Acquire a large amount of historical medical data, including medical imaging data, detection waveform data, electronic medical record data, and gene sequencing data; desensitize, preprocess, and label each of the historical medical data to construct a dataset.

[0020] Step S3: Train the multi-modal data labeling model using the dataset, labeling loss function, and labeling reward function, and compress the multi-modal data labeling model during the training process;

[0021] Step S4: Create a medical decision-making model based on the multimodal data fusion layer, feature interaction layer, and diagnostic decision layer, and set the decision loss function and decision reward function of the medical decision-making model;

[0022] The multimodal data fusion layer is used to extract joint representations from medical imaging 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 layer is used to output medical strategy recommendations carrying disease classification, prognostic assessment, and treatment suggestions based on the fused features.

[0023] 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;

[0024] 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;

[0025] Step S7: Collect real-time medical data and the latest medical knowledge, automatically label the real-time medical data and the latest medical knowledge using the deployed multi-modal data labeling model, and incrementally train the medical decision-making model using the labeled real-time medical data and the latest medical knowledge.

[0026] Step S8: Perform medical question and answer through the deployed medical decision model, and record question and answer logs in real time, including at least the medical question, medical strategy suggestions, question and answer time, question account, and misdiagnosis feedback;

[0027] Step S9: Annotate the question-and-answer logs using the deployed multi-modal data annotation model, incrementally train the medical decision-making model using the annotated question-and-answer logs, and encrypt and distribute the question-and-answer logs.

[0028] Secondly, the present invention provides a medical strategy generation system that combines data flywheel reinforcement learning, comprising the following modules:

[0029] The multimodal data annotation model creation module is used to create a multimodal 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 to set the annotation loss function and annotation reward function of the multimodal data annotation model.

[0030] 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; and the gene feature extraction module is used to extract gene features from gene sequencing data.

[0031] The modality fusion layer is constructed based on a spatiotemporal alignment module and a multi-scale feature fusion module. The spatiotemporal alignment module is used to perform spatiotemporal alignment operations on image features, waveform features, case features, and gene features. The multi-scale feature fusion module is used to fuse the spatiotemporally aligned image features, waveform features, case features, and gene features to obtain fused features.

[0032] 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.

[0033] 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 medical record data and gene sequencing data respectively based on the secondary enhancement features. The uncertainty estimation module is used to predict the uncertainty of each annotation result, filter the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty.

[0034] The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model;

[0035] The dataset construction module is used to acquire a large amount of historical medical data, including medical imaging data, detection waveform data, electronic medical record data, and gene sequencing data. The dataset is constructed after the historical medical data is desensitized, preprocessed, and labeled.

[0036] The multi-modal data annotation model training module is used to train the multi-modal data annotation model using the dataset, annotation loss function, and annotation reward function, and to compress the multi-modal data annotation model during the training process;

[0037] The medical decision model creation module is used to create a medical decision model based on a multimodal data fusion layer, a feature interaction layer, and a diagnostic decision layer, and to set the decision loss function and decision reward function of the medical decision model.

[0038] The multimodal data fusion layer is used to extract joint representations from medical imaging 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 layer is used to output medical strategy recommendations carrying disease classification, prognostic assessment, and treatment suggestions based on the fused features.

[0039] The medical decision-making model training module is used to train the medical decision-making model using the dataset, decision loss function, and decision reward function, and to compress the medical decision-making model during the training process.

[0040] The model deployment module is used to deploy the trained multi-modal data annotation model and medical decision-making model, and to perform model drift training on the deployed multi-modal data annotation model and medical decision-making model.

[0041] The incremental knowledge training module 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 incrementally train the medical decision-making model using the annotated real-time medical data and the latest medical knowledge.

[0042] The medical question-and-answer module is used to conduct medical question-and-answer through the deployed medical decision-making model, and to record question-and-answer logs in real time, including at least the medical question, medical strategy suggestions, question-and-answer time, question account, and misdiagnosis feedback.

[0043] The question-and-answer log management module is used to annotate the question-and-answer logs using the deployed multi-modal data annotation model, incrementally train the medical decision-making model using the annotated question-and-answer logs, and encrypt and distribute the question-and-answer logs.

[0044] The advantages of this invention are:

[0045] 1. A multi-modal data annotation model is created through a feature extraction layer, modality fusion layer, feature enhancement layer, annotation prediction layer, and feedback optimization layer. The annotation loss function and annotation reward function of the multi-modal data annotation model are set. Then, a large amount of historical medical data, including medical image data, detection waveform data, electronic medical record data, and gene sequencing data, is acquired. The historical medical data is anonymized, preprocessed, and annotated to construct a dataset. The multi-modal data annotation model is trained using the dataset, the annotation loss function, and the annotation reward function, and the model is compressed during training. Finally, a medical decision-making model is created based on the multi-modal data fusion layer, feature interaction layer, and diagnostic decision-making layer. The decision loss function and decision reward function of the medical decision-making model are set. The process involves: training a medical decision-making model using a dataset, decision loss function, and decision reward function, compressing the model during training; deploying the trained multi-model data annotation model and the medical decision-making model, and performing model drift training on the deployed model; collecting real-time medical data and the latest medical knowledge, automatically annotating the data using the deployed multi-model data annotation model, and incrementally training the medical decision-making model using the annotated data; and finally, using the deployed model to perform medical question-and-answer sessions, recording in real-time at least the medical question, suggested medical strategy, question-and-answer time, questioning account, and misdiagnosis. The system generates question-and-answer logs; these logs are annotated using a deployed multimodal data annotation model, and incrementally trained on the medical decision-making model using these annotated logs. The logs are also encrypted and stored in a distributed manner. Specifically, the pre-trained medical decision-making model performs medical question-and-answer sessions, recording real-time logs including misdiagnosis feedback. The pre-trained multimodal data annotation model automatically annotates these logs and incrementally trains the medical decision-making model (feedback to the training system), effectively reducing sample bias. The medical decision-making model, through a multimodal data fusion layer and a feature interaction layer, effectively enhances the feature extraction capability of multimodal data. By setting the decision reward function of the medical decision-making model to carry path rationality rewards and risk avoidance rewards, drug interactions can be effectively prevented. The system mitigates risks and may violate treatment guidelines. It employs incremental training by collecting real-time medical data and the latest medical knowledge to avoid knowledge update delays. During the training phase, it compresses multi-modal data annotation models and medical decision-making models; the data preprocessing phase combines a streaming computing engine; and the inference phase utilizes hardware acceleration and memory caching to effectively improve response time. Multi-level encryption and distributed storage of question-and-answer logs ensure the security of related data storage. The overall approach combines a data flywheel (model drift training, real-time automatic data annotation and incremental training, question-and-answer log feedback, and knowledge update mechanism) and reinforcement learning (dual reward function, policy optimization mechanism, and uncertainty-driven exploration), ultimately greatly improving the accuracy, timeliness, and safety of medical decisions.

[0046] 2. By integrating medical imaging data, detection waveform data, electronic medical record data, and gene sequencing data, we comprehensively capture patient information and avoid the limitations of a single modality. By setting image features using a combination of 3D CNN and Transformer, we take into account both local details and global context. By setting waveform features, we handle temporal dependencies through 1D temporal convolution and causal convolution. Medical record features are integrated with pre-trained medical terminology vectors to improve semantic understanding. Gene features are combined with graph neural networks and knowledge graphs to analyze the interactions and biological significance between genes. Furthermore, through cross-modal comparative learning and dynamic weighted fusion, we strengthen the multimodal synergistic effect and effectively improve feature extraction capabilities.

[0047] 3. By using spatiotemporal alignment and multi-scale fusion, the spatiotemporal inconsistency problem is solved by deformable convolution and temporal warp alignment, and the weights are dynamically adjusted by cross-attention mechanism, which effectively improves the quality of fused features and thus effectively improves the annotation quality.

[0048] 4. By fusing contrast loss and focus loss through the labeling loss function, class imbalance and modality consistency can be effectively addressed; by fusing KL divergence and temporal consistency loss through the decision loss function, the rationality of clinical pathways and decision coherence can be effectively constrained; by introducing indicators such as user satisfaction and treatment efficiency through the reward function, end-to-end optimization can be achieved in combination with reinforcement learning.

[0049] 5. By combining weight pruning, weight quantization, and parameter sharing during the model training phase, the model size is effectively compressed, the consumption of computing resources is reduced, and it is adapted to edge device deployment.

[0050] 6. By combining two SM3 hash calculations during the encrypted log generation process, and adopting AES-GCM, SM9, ECIES, and Paillier algorithms for different data types, and finally using SM4 and Serpent algorithms for sandwich encryption, and uploading the data fingerprint of the encrypted log to the blockchain, multi-level encryption, integrity verification, and tamper-proofing of the question-and-answer log are achieved, greatly improving the security of question-and-answer log storage.

[0051] 7. Real-time medical data and question-and-answer logs are automatically labeled and fed back to the model to form a closed-loop optimization of "data-model"; long-term performance stability is ensured by model drift training combined with ROC curves and other indicators.

[0052] 8. By dividing the training set, validation set, and test set into an 8:1:1 ratio, overfitting is avoided; the training set is automatically expanded when validation fails, effectively improving generalization ability; the model performance is verified from multiple perspectives, including accuracy, recall, F1 score, and ROC curve, to ensure clinical reliability.

[0053] 9. By setting up modular structures for each layer such as feature extraction, fusion, and enhancement, it is easy to add new data modalities (such as voice recordings and sensor data) or task types (such as drug recommendations) without reconstructing the overall framework; by supporting the dynamic integration of the latest medical knowledge and user feedback (misdiagnosis feedback), it adapts to the rapid iteration characteristics of the medical field.

[0054] 10. By predicting the uncertainty of the annotation results, high-confidence data is automatically filtered to reduce the negative impact of noisy annotations on the model; by using multi-expert consensus for historical data annotation, the bias of a single annotator is reduced and the reliability of training data is improved.

[0055] 11. By adopting a multi-modal data annotation model to automatically annotate question-and-answer logs carrying misdiagnosis feedback, the annotation efficiency is effectively improved; by generating data fingerprints through the HMAC algorithm and uploading them to the blockchain, the traceability and tamper-proof nature of misdiagnosis cases are ensured; by setting a quantity threshold through an incremental training trigger mechanism, when the number of misdiagnosis cases accumulates to a clinically significant level, the model update is automatically triggered, effectively solving the problem of missing feedback, and thus constructing a closed loop of "misdiagnosis cases - annotation feedback - model iteration".

[0056] 12. Through a multimodal data annotation model designed in a progressive manner of "professional feature extraction → spatiotemporal semantic alignment → knowledge enhancement fusion → credibility-aware annotation", breakthroughs have been achieved in the three dimensions of accuracy, security and adaptability of medical multimodal data processing. A medical AI annotation infrastructure that can evolve autonomously has been built, providing a high-quality, interpretable and low-latency data foundation for subsequent clinical decision-making.

[0057] 13. By setting up a multimodal deep collaborative architecture for the medical decision-making model, high accuracy and comprehensiveness of medical decision-making are achieved: First, 3D ResNet-50, 1D CNN-BiLSTM, BERT, and GAT networks are used to adapt feature extraction from medical images, waveforms, electronic medical records, and genetic data, respectively, specifically addressing the challenge of representing multi-source heterogeneous data; second, cross-modal contrastive learning is used to align the semantics of features from different modalities, and a cross-attention mechanism combined with a dynamic graph neural network is used to achieve adaptive fusion of feature interactions, effectively capturing complex medical correlations and compressing redundant information; furthermore, based on the simultaneous output of disease classification, prognostic assessment, and treatment suggestions from a multi-task decision layer, a loss-reward function optimization mechanism integrating supervised learning and reinforcement learning is integrated to ensure that the model achieves a balance between diagnostic accuracy and the rationality of clinical decisions. Its innovation lies in the deep fusion of multimodal data, dynamic feature interaction optimization, and end-to-end integrated design of diagnosis and treatment decision-making, significantly improving the reliability, efficiency, and interpretability of medical decisions, and providing core technical support for precision medicine.

[0058] 14. By innovatively constructing an intelligent decision-making system that deeply integrates and dynamically optimizes multimodal medical data, this system integrates multi-source data such as medical imaging data, detection waveform data, electronic medical record data, and gene sequencing data. It employs advanced technologies such as 3D convolution, temporal causal convolution, and knowledge graph enhancement to achieve accurate feature extraction and cross-modal collaborative expression. Combined with the data flywheel mechanism of reinforcement learning and the dual loss-reward function, it optimizes model performance. While ensuring data security (full-process desensitization, multi-level national cryptographic encryption, and blockchain evidence storage), it possesses real-time inference, incremental learning, and clinical pathway alignment capabilities. Its modular design supports flexible expansion. Through self-supervised comparative learning, uncertainty estimation, and multi-task joint prediction, it significantly improves diagnostic accuracy and generalization ability, 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. Attached Figure Description

[0059] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0060] Figure 1 This is a flowchart of a medical strategy generation method combining data flywheel reinforcement learning according to the present invention.

[0061] Figure 2 This is a schematic diagram of the structure of a medical strategy generation system that combines data flywheel reinforcement learning according to the present invention. Detailed Implementation

[0062] The overall approach of the technical solution in this application is as follows: A pre-trained medical decision-making model is used for medical question-and-answer sessions, recording question-and-answer logs including misdiagnosis feedback in real time. The logs are automatically labeled using a pre-trained multi-modal data annotation model, and the medical decision-making model is incrementally trained (backflowed to the training system), effectively reducing sample bias. The medical decision-making model effectively improves the feature extraction capability of multi-modal data through a multi-modal data fusion layer and a feature interaction layer. By setting the decision reward function of the medical decision-making model to carry path rationality rewards and risk avoidance rewards, the risk of drug interactions or violations of treatment guidelines can be effectively prevented. Real-time medical data is collected, and... The system uses incremental training with the latest medical knowledge to avoid knowledge update delays. During the training phase, it compresses multi-modal data annotation models and medical decision-making models. The data preprocessing phase incorporates a streaming computing engine, and the inference phase utilizes hardware acceleration and memory caching to effectively improve response time. Multi-level encryption and distributed storage of question-and-answer logs ensure the security of related data storage. The system combines a data flywheel (model drift training, real-time automatic data annotation and incremental training, question-and-answer log feedback, and knowledge update mechanism) with reinforcement learning (dual reward function, policy optimization mechanism, and uncertainty-driven exploration) to improve the accuracy, timeliness, and safety of medical decisions.

[0063] Please refer to Figures 1 to 2 As shown, a preferred embodiment of the medical strategy generation method combining data flywheel reinforcement learning of the present invention includes the following steps:

[0064] Step S1: Create a multi-modal data annotation model based on the feature extraction layer, modality fusion layer, feature enhancement layer, annotation prediction layer, and feedback optimization layer; and set the annotation loss function and annotation reward function of the multi-modal data annotation model.

[0065] By adopting a modular structure for each layer, such as feature extraction, fusion, and enhancement, it is easy to add new data modalities (such as voice recordings and sensor data) or task types (such as drug recommendations) without reconstructing the overall framework; by supporting the dynamic integration of the latest medical knowledge and user feedback (misdiagnosis feedback), it adapts to the rapid iteration characteristics of the medical field.

[0066] 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; and the gene feature extraction module is used to extract gene features from gene sequencing data.

[0067] The modality fusion layer is constructed based on a spatiotemporal alignment module and a multi-scale feature fusion module. The spatiotemporal alignment module is used to perform spatiotemporal alignment operations on image features, waveform features, case features, and gene features. The multi-scale feature fusion module is used to fuse the spatiotemporally aligned image features, waveform features, case features, and gene features to obtain fused features.

[0068] 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.

[0069] The self-supervised contrastive learning module works somewhat like a "spot the difference" game. It identifies two sets of samples from the data: one set is "positive sample pairs," which are essentially similar (e.g., data from two patients with the same disease); the other set is "negative sample pairs," which are essentially different (e.g., data from a sick patient and a healthy patient). By comparing these two sets of samples, it learns which features are important, thereby enhancing the data's features and obtaining "first-level enhanced features." The knowledge graph enhancement module can be understood as a "medical knowledge base," containing various medical knowledge (e.g., relationships between diseases, associations between symptoms and diseases). This module utilizes a 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, thus obtaining "second-level enhanced features."

[0070] In practice, the dynamic medical knowledge graph can be automatically updated based on the real-time update mechanism of data flow, automatically updated through the automatic extraction technology of natural language processing (NLP), or automatically updated through the incremental update mechanism of machine learning.

[0071] 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 medical record data and gene sequencing data respectively based on the secondary enhancement features. The uncertainty estimation module is used to predict the uncertainty of each annotation result, filter the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty.

[0072] The multi-task annotation prediction module is built on a multi-branch classification / regression network with shared underlying features. It is used to predict multiple annotation types (such as lesion location, disease type, etc.) at the same time, and adopts a modular design to support flexible expansion of new annotation tasks.

[0073] The uncertainty estimation module is built based on Bayesian neural networks or deep ensemble methods and is used to quantify and predict uncertainty to assist clinical decision-making.

[0074] By predicting the uncertainty of the annotation results, high-confidence data is automatically filtered to reduce the negative impact of noisy annotations on the model; by using multi-expert consensus for historical data annotation, the bias of a single annotator is reduced and the reliability of training data is improved.

[0075] The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model;

[0076] The feedback optimization layer is constructed based on an active learning module and a continuous learning module. The active learning module is based on uncertainty sampling and domain adaptation strategies, and combines medical expert knowledge to design sampling strategies for intelligently selecting the most valuable data for annotation. The continuous learning module is based on elastic weight solidification and knowledge distillation, and combines a robust processing mechanism for changes in the distribution of medical data to support incremental learning of the model without forgetting existing knowledge.

[0077] The core objective of the active learning module is to enable the model to automatically select the most valuable data for human experts to annotate, avoiding wasting effort on unimportant data. Uncertainty sampling (prioritizing learning from areas of model uncertainty): When the model predicts, it assigns a "confidence score" to each sample. The active learning module prioritizes these "low-confidence" samples for doctors or experts to annotate, as these data may represent the model's "knowledge blind spots." Diversity sampling (avoiding repetitive learning of similar data): If the model has already learned some common cases (such as pneumonia), it will proactively seek out uncommon cases (such as rare diseases) for expert annotation, avoiding continuously learning repetitive information. Domain adaptation (optimization for the specificities of medicine): Medical data may contain noise (such as blurry images or incorrect medical records). The active learning module identifies these "difficult" samples and prioritizes expert confirmation, preventing the model from being misled.

[0078] The core objective of the continuous learning module is to enable the model to continuously learn new data without forgetting previously learned knowledge. Elastic Weight Consolidation (EWC) (protecting old knowledge): The model adds a "protective shield" to previously learned "important knowledge points," preventing them from being easily overwritten during training with new data; for example, if the model has already learned the "features of pneumonia," it won't completely forget the diagnostic methods for pneumonia when a new batch of "heart disease" data arrives. Knowledge Distillation (old model teaching new model): When training a new model, the old model acts as the "teacher," supervising the new model's predictions and ensuring that the new model doesn't deviate from the correct direction; for example, after the new model learns new features of "diabetes," it still needs to be checked by the old model to ensure it can still correctly diagnose "hypertension." Incremental Learning (gradual updates): Instead of retraining the model with all new data at once, it updates in batches and gradually to avoid "learning collapse"; for example, let the model learn 100 new cases first to test the effect before learning the next batch, which is more stable.

[0079] The active learning module allows the model to select challenging data for expert labeling, thereby optimizing model parameters, improving efficiency, and reducing repetitive work. The continuous learning module enables the model to continuously learn new things without forgetting old knowledge, ensuring long-term stability and reliability.

[0080] When only one modality of data is input in a practical application, this multimodal medical data annotation model can still perform automatic annotation because each modality has its own dedicated feature extraction module.

[0081] Step S2: Acquire a large amount of historical medical data, including medical imaging data, detection waveform data, electronic medical record data, and gene sequencing data; desensitize, preprocess, and label each of the historical medical data to construct a dataset.

[0082] Step S3: Train the multi-modal data labeling model using the dataset, labeling loss function, and labeling reward function, and compress the multi-modal data labeling model during the training process;

[0083] Step S4: Create a medical decision-making model based on the multimodal data fusion layer, feature interaction layer, and diagnostic decision layer, and set the decision loss function and decision reward function of the medical decision-making model;

[0084] The multimodal data fusion layer is used to extract joint representations from medical imaging 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 layer is used to output medical strategy recommendations carrying disease classification, prognostic assessment, and treatment suggestions based on the fused features; prognostic 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.

[0085] 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;

[0086] 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;

[0087] Step S7: Collect real-time medical data and the latest medical knowledge, automatically label the real-time medical data and the latest medical knowledge using the deployed multi-modal data labeling model, and incrementally train the medical decision-making model using the labeled real-time medical data and the latest medical knowledge.

[0088] Step S8: Perform medical question and answer through the deployed medical decision model, and record question and answer logs in real time, including at least the medical question, medical strategy suggestions, question and answer time, question account, and misdiagnosis feedback;

[0089] Step S9: Annotate the question-and-answer logs using the deployed multi-modal data annotation model, incrementally train the medical decision-making model using the annotated question-and-answer logs, and encrypt and distribute the question-and-answer logs.

[0090] In step S1, the image feature extraction module extracts initial image features from medical image data through a 3D convolutional neural network, integrates global contextual information into the initial image features through a Transformer to obtain contextual features, and filters the contextual features through an attention gating mechanism to obtain image features.

[0091] In essence, 3D CNN acts as the "basic worker," responsible for extracting raw features; Transformer acts as the "integration expert," responsible for piecing together features into a more meaningful whole; and attention gating mechanism acts as the "quality inspector," responsible for filtering out the most important features. Through this collaborative work, the medical image data feature extraction module can extract key information from images more efficiently and accurately, providing strong support for subsequent diagnosis or analysis.

[0092] The waveform feature extraction module extracts local waveform features from the detected waveform data through a 1D temporal convolutional network, performs sequential regularization on each of the local features through causal convolution, and captures the temporal dependencies from each of the sequentially regularized local features through a recurrent neural network to obtain waveform features.

[0093] In essence, 1D CNN acts as the "starter," responsible for extracting local features from waveform data; causal convolution is the "relay runner," ensuring that the feature extraction process doesn't "cheat" (i.e., it doesn't use future information); and RNN is the "sprinter," responsible for integrating time-series information and capturing the dynamic changes in the waveform. Through this collaborative work, the waveform data feature extraction module can extract key features from the waveform more efficiently and accurately, providing strong support for subsequent analysis or detection.

[0094] The case feature extraction module extracts case features from electronic medical record data using a Transformer model that incorporates pre-trained medical term vectors.

[0095] In essence, the Transformer model acts as the "basic translator," responsible for converting electronic medical record data into feature vectors that computers can understand; the medical terminology embedding pre-training is the "professional translator," helping the model better understand medical terminology. Through this collaborative work, the electronic medical record data feature extraction module can extract key features from medical records more efficiently and accurately, providing strong support for subsequent diagnosis, treatment, or research.

[0096] The gene feature extraction module extracts key gene features from gene sequencing data through a convolutional autoencoder, learns the interaction between the key gene features through a graph neural network to obtain gene enhancement features, and integrates semantic information for each gene enhancement feature through a gene ontology knowledge graph to obtain gene features.

[0097] In essence, the convolutional autoencoder acts as the "basic detective," responsible for extracting preliminary features from gene sequencing data; the graph neural network is the "relationship expert," further refining gene feature representations by analyzing interactions between genes; and the gene ontology knowledge graph is the "knowledge base expert," providing rich semantic information for genes to help us better understand their biological significance. 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 development.

[0098] By integrating medical imaging data, waveform data, electronic medical record data, and gene sequencing data, this approach comprehensively captures patient information, avoiding the limitations of a single modality. Image features are designed using a combination of 3D CNN and Transformer to balance local details and global context. Waveform features are designed using 1D temporal and causal convolutions to handle temporal dependencies. Medical record features are integrated with pre-trained medical terminology vectors to enhance semantic understanding. Gene features are combined with graph neural networks and knowledge graphs to analyze gene interactions and biological significance. Finally, cross-modal comparative learning and dynamic weighted fusion are used to strengthen multimodal synergy and effectively improve feature extraction capabilities.

[0099] The spatiotemporal 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 warp alignment network to complete the spatiotemporal alignment operation.

[0100] Deformable convolutions act as "spatial adjusters," resolving the spatial alignment of different modal features; temporal warp alignment networks act as "temporal adjusters," resolving the temporal alignment of different modal features. Through this collaborative work, the spatiotemporal alignment module can more efficiently and accurately align multimodal features, providing a foundation for subsequent tasks (such as multimodal fusion, video understanding, or action recognition).

[0101] Provide a more consistent representation of features.

[0102] The multi-scale feature fusion module extracts relevant features from spatiotemporally aligned image features, waveform features, case features, and gene features through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of the image features, waveform features, case features, and gene features based on the relevant features to obtain fused features. The complementary feature expression of the fused features is enhanced through intermodal contrastive learning.

[0103] In essence, the cross-attention mechanism acts as the "communicator," enabling different features to understand each other; the gated fusion unit acts as the "regulator," adjusting the degree of feature fusion based on importance; and the intermodal contrastive learning acts as the "optimizer," ensuring that the fused features are more consistent and effective. Through this collaborative work, the multi-scale feature fusion module can more efficiently and accurately fuse features from different sources, providing more powerful feature representations for subsequent tasks (such as classification, detection, or generation).

[0104] By combining spatiotemporal alignment and multi-scale fusion, specifically by using deformable convolution and temporal warping alignment to address spatiotemporal inconsistency, and by dynamically adjusting weights through a cross-attention mechanism, the quality of fused features is effectively improved, thereby enhancing the quality of annotation.

[0105] The formula for the labeling loss function is:

[0106] L1=λ1*L vis +λ2*L text +λ3*L cons +λ4*L fusionLtotal ;

[0107] Where L1 represents the loss value of the labeled loss function; L vis This represents the contrast loss in medical imaging data, used to constrain the correlation between key image features and labels; L text This represents the focus loss in electronic medical record data, used to mitigate the class imbalance problem; L cons L represents the cross-modal consistency loss, used to constrain the consistency of prediction distributions across different modes using KL divergence constraints; fusionLtotal λ1, λ2, λ3, and λ4 represent the fusion enhancement loss, used to enhance multimodal synergistic effects using dynamic weighted cross-entropy; λ1, λ2, λ3, and λ4 all represent weight coefficients.

[0108] The formula for the labeled reward function is:

[0109] R1=α1*R acc1 +α2*R comp +α3*R eff1 +α4*R userRtotal ;

[0110] Where R1 represents the reward value of the labeled reward function; R acc1Indicates the accuracy of the annotation; R comp Represents modal complementarity; R eff1 Indicates annotation efficiency; R userRtotal α1, α2, α3, and α4 represent user approval; α1, α2, α3, and α4 all represent weighting coefficients.

[0111] Step S2 specifically involves:

[0112] Acquire a large amount of historical medical data, including medical imaging data, test waveform data, electronic medical record data, and gene sequencing data;

[0113] The names in the metadata of each medical image data are deleted, and the facial regions in each medical image data are blurred to desensitize the medical image data. The desensitized medical image data undergoes preprocessing including noise reduction, grayscale transformation, image enhancement, image normalization, and image segmentation. Based on a preset image annotation standard and a multi-annotator fusion method, the preprocessed medical image data is annotated with patient regions, disease classifications, prognostic assessments, and treatment recommendations, generating corresponding labels. Prognostic assessment refers to the prediction of the future development trend of a patient's disease; it can help doctors and patients understand the possible course of the disease and formulate reasonable treatment plans and care plans.

[0114] The patient identifiers carried in each of the detection waveform data are deleted to complete the desensitization. The desensitized detection waveform data are then preprocessed by noise reduction, baseline correction, mean removal, linear trend removal, waveform peaking, and data segmentation. Based on the preset waveform annotation specifications and the multi-annotator fusion method, the preprocessed detection waveform data are annotated with patient waveforms, disease classifications, prognostic assessments, and treatment suggestions to generate corresponding labels.

[0115] The patient identifiers carried in each of the electronic medical records are deleted to complete the desensitization. The desensitized electronic medical records are preprocessed by deleting irrelevant content, filling missing values, correcting erroneous data, unifying format, and tagging parts of speech. Based on the preset medical record tagging specifications and the multi-annotator fusion method, the preprocessed electronic medical records are annotated with disease classification, prognosis assessment and treatment suggestions to generate corresponding tags.

[0116] The patient identifiers carried in each of the gene sequencing data are deleted to complete the desensitization. The desensitized gene sequencing data are preprocessed by removing low-quality sequences, repetitive sequences, adapter sequences and performing error correction. Based on the preset gene annotation specifications and the multi-annotator fusion method, the preprocessed gene sequencing data are annotated with abnormal sequences, disease classification, prognostic assessment and treatment suggestions to generate corresponding tags.

[0117] A dataset is constructed based on the labeled historical medical data.

[0118] Step S3 specifically involves:

[0119] The dataset is divided into a first training set, a first validation set, and a first test set in a ratio of 8:1:1 using stratified sampling. The multi-modal data labeling model is trained using the first training set. In each training round, the labeling loss function and the labeling reward function are used alternately to optimize the multi-modal data labeling model. The training process combines weight pruning, weight quantization, and parameter sharing to compress the multi-modal data labeling model until the preset first early stopping condition is met.

[0120] The annotation accuracy of the multi-modal data annotation model is calculated using the first validation set to validate the model. If the validation fails, the first training set is expanded and training continues; if the validation passes, then:

[0121] The label confidence of the multi-modal data labeling model is calculated using the first test set to test the multi-modal data labeling model. If the test fails, the first training set is expanded and training continues; if the test passes, training ends.

[0122] In step S4, the multimodal data fusion layer is constructed based on a medical imaging branch, a waveform processing branch, a medical record coding branch, a gene analysis branch, and a cross-modal alignment module. The medical imaging branch is used to extract image features from medical image data using a 3D ResNet-50 network. The waveform processing branch is used to extract periodic features from detected waveform data using a 1D CNN-BiLSTM network. The medical record coding branch is used to extract temporal features from electronic medical record data using a BERT model. The gene analysis branch is used to extract genetic features from gene sequencing data using a GAT network. The cross-modal alignment module is used to align and fuse image features, periodic features, temporal features, and genetic features using contrastive learning loss to obtain a joint representation.

[0123] The feature interaction layer obtains interaction features by performing feature interaction on the joint representation through a cross-attention mechanism, and obtains fused features by fusing the interaction features through a dynamic graph neural network. The feature distillation module compresses redundant features in the fused features.

[0124] Feature interaction layers are an indispensable part of modern deep learning models. They enhance the model's overall understanding of data and predictive performance by promoting information exchange and fusion between different features. Feature interaction layers play an important role in fields such as multimodal learning, multi-task learning, recommendation systems, and natural language processing.

[0125] The formula for the decision loss function is:

[0126] L2=β1*L diag +β2*L path +β3*L time ;

[0127] Where L2 represents the loss value of the decision loss function; L diag L represents the diagnostic cross-entropy loss, used to measure the difference between the recommended medical strategy and the true label; path The regularization loss of the clinical pathway is represented by L, and the KL divergence is used to measure the deviation between the model's decision path and the standard clinical pathway; time β1, β2, and β3 represent the temporal consistency loss, used to ensure the rationality between successive decision steps; β1, β2, and β3 all represent weighting coefficients.

[0128] The formula for the decision reward function is:

[0129] R2=γ1*R acc2 +γ2*R path +γ3*R risk +γ4*R eff2 ;

[0130] Where R² represents the reward value of the decision reward function; R acc2 Rewards for diagnostic accuracy are expressed using weighted F1 scores; R path The reward for path rationality is represented by cosine similarity; R risk Indicates a risk-averse reward; R eff2 This represents the reward for treatment efficiency; γ1, γ2, γ3, and γ4 all represent weighting coefficients.

[0131] By fusing contrast loss and focus loss through the labeling loss function, class imbalance and modality consistency are effectively addressed; by fusing KL divergence and temporal consistency loss through the decision loss function, the rationality of clinical pathways and decision coherence are effectively constrained; by introducing indicators such as user approval and treatment efficiency through the reward function, end-to-end optimization is achieved in combination with reinforcement learning.

[0132] Step S5 specifically involves:

[0133] The dataset is divided into a second training set, a second validation set, and a second test set in a ratio of 8:1:1 using stratified sampling. The medical decision-making model is trained using the second training set. In each training round, the decision loss function and the decision reward function are used alternately 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 the preset second early cessation condition is met.

[0134] The decision accuracy of the medical decision-making model is calculated using the second validation set to validate the model. If the validation fails, the second training set is expanded for continued training; if the validation succeeds, then:

[0135] 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.

[0136] By combining weight pruning, weight quantization, and parameter sharing during the model training phase, the model size is effectively compressed, computational resource consumption is reduced, and it is adapted to edge device deployment.

[0137] Step S6 specifically involves:

[0138] The trained multi-modal data annotation model and medical decision-making model are deployed locally using containerization technology. A large amount of real medical data is collected, and each piece of real medical data is de-identified, preprocessed, and labeled. Based on the labeled real medical data, the deployed multi-modal data annotation model and medical decision-making model are subjected to model drift training. The performance of the multi-modal data annotation model and medical decision-making model after model drift training is verified by accuracy, recall, F1 score, and ROC curve.

[0139] By dividing the training set, validation set, and test set into an 8:1:1 ratio, overfitting is avoided; the training set is automatically expanded when validation fails, effectively improving generalization ability; and the model performance is verified from multiple perspectives, including accuracy, recall, F1 score, and ROC curve, to ensure clinical reliability.

[0140] Step S8 specifically involves:

[0141] Through a visual interface, input medical questions carrying medical imaging data, detection waveform data, electronic medical record data, or gene sequencing data are obtained. After preprocessing the medical questions through a streaming computing engine, they are input into a deployed medical decision-making model. The medical decision-making model, combined with GPU acceleration technology and memory caching technology, infers medical strategy suggestions from the medical questions and provides feedback on the medical strategy suggestions through the visual interface to conduct medical Q&A. The Q&A log is recorded in real time, including at least the medical question, medical strategy suggestions, Q&A time, questioning account, and misdiagnosis feedback.

[0142] Step S9 specifically involves:

[0143] The question-and-answer logs are automatically labeled by the deployed multi-modal data labeling model. When the number of labeled question-and-answer logs reaches a preset threshold, the medical decision-making model is automatically incrementally trained using the labeled question-and-answer logs within a preset time period.

[0144] Real-time medical data and question-and-answer logs are automatically labeled and fed back to the model, forming a closed-loop optimization of "data-model". Long-term performance stability is ensured by using model drift training combined with ROC curves and other indicators for verification.

[0145] The question-and-answer log is encrypted to obtain an encrypted log, the encrypted log is distributed and stored, the data fingerprint of the encrypted log is calculated using the HMAC algorithm, and the data fingerprint is uploaded to the blockchain;

[0146] By employing a multi-modal data annotation model to automatically annotate question-and-answer logs carrying misdiagnosis feedback, annotation efficiency is effectively improved; by generating data fingerprints using the HMAC algorithm and uploading them to the blockchain, the traceability and tamper-proof nature of misdiagnosis cases are ensured; by setting a quantity threshold through an incremental training trigger mechanism, the model is automatically triggered to update when the number of misdiagnosis cases accumulates to a clinically significant level, effectively solving the problem of missing feedback, and thus constructing a closed loop of "misdiagnosis cases - annotation feedback - model iteration".

[0147] The specific steps for encrypting the question-and-answer log to obtain the encrypted log are as follows:

[0148] The medical image data in the question-and-answer log is subjected to DCT transformation to extract frequency domain coefficients. Based on the frequency domain coefficients, Zigzag scanning is performed to convert the medical image data into a one-dimensional sequence. The text data, including electronic medical record data, medical strategy suggestions, and misdiagnosis feedback, is encoded using UTF-8 encoding and then compressed using Huffman coding to obtain compressed data. Wavelet packet decomposition is performed on the detection waveform data to extract feature coefficients. The gene sequencing data is converted to FASTQ format and encoded using base sequence encoding to obtain encoded data.

[0149] The one-dimensional sequence, compressed data, feature coefficients, and encoded data are each subjected to SM3 hash calculation to obtain corresponding sub-check codes. Then, each sub-check code is subjected to SM3 hash calculation to obtain the master check code.

[0150] The one-dimensional sequence is encrypted into first encrypted data using the AES-GCM algorithm, the compressed data is encrypted into second encrypted data using the SM9 algorithm, the feature coefficients are encrypted into third encrypted data using the ECIES algorithm, and the encoded data is encrypted into fourth encrypted data using the Paillier algorithm.

[0151] Using the SM4 and Serpent algorithms, the first encrypted data, the second encrypted data, the third encrypted data, the fourth encrypted data, and the master checksum are sandwiched together to obtain the encrypted log.

[0152] By combining two SM3 hash calculations during the encrypted log generation process, and employing AES-GCM, SM9, ECIES, and Paillier algorithms for different data types, and finally using SM4 and Serpent algorithms for sandwich encryption, and uploading the data fingerprint of the encrypted log to the blockchain, multi-level encryption, integrity verification, and tamper-proofing of question-and-answer logs are achieved, greatly improving the security of question-and-answer log storage.

[0153] A preferred embodiment of the medical strategy generation system combining data flywheel reinforcement learning according to the present invention includes the following modules:

[0154] The multimodal data annotation model creation module is used to create a multimodal 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 to set the annotation loss function and annotation reward function of the multimodal data annotation model.

[0155] By adopting a modular structure for each layer, such as feature extraction, fusion, and enhancement, it is easy to add new data modalities (such as voice recordings and sensor data) or task types (such as drug recommendations) without reconstructing the overall framework; by supporting the dynamic integration of the latest medical knowledge and user feedback (misdiagnosis feedback), it adapts to the rapid iteration characteristics of the medical field.

[0156] 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; and the gene feature extraction module is used to extract gene features from gene sequencing data.

[0157] The modality fusion layer is constructed based on a spatiotemporal alignment module and a multi-scale feature fusion module. The spatiotemporal alignment module is used to perform spatiotemporal alignment operations on image features, waveform features, case features, and gene features. The multi-scale feature fusion module is used to fuse the spatiotemporally aligned image features, waveform features, case features, and gene features to obtain fused features.

[0158] 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.

[0159] The self-supervised contrastive learning module works somewhat like a "spot the difference" game. It identifies two sets of samples from the data: one set is "positive sample pairs," which are essentially similar (e.g., data from two patients with the same disease); the other set is "negative sample pairs," which are essentially different (e.g., data from a sick patient and a healthy patient). By comparing these two sets of samples, it learns which features are important, thereby enhancing the data's features and obtaining "first-level enhanced features." The knowledge graph enhancement module can be understood as a "medical knowledge base," containing various medical knowledge (e.g., relationships between diseases, associations between symptoms and diseases). This module utilizes a 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, thus obtaining "second-level enhanced features."

[0160] In practice, the dynamic medical knowledge graph can be automatically updated based on the real-time update mechanism of data flow, automatically updated through the automatic extraction technology of natural language processing (NLP), or automatically updated through the incremental update mechanism of machine learning.

[0161] 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 medical record data and gene sequencing data respectively based on the secondary enhancement features. The uncertainty estimation module is used to predict the uncertainty of each annotation result, filter the annotation results based on the uncertainty, and output the annotation results carrying the uncertainty.

[0162] The multi-task annotation prediction module is built on a multi-branch classification / regression network with shared underlying features. It is used to predict multiple annotation types (such as lesion location, disease type, etc.) at the same time, and adopts a modular design to support flexible expansion of new annotation tasks.

[0163] The uncertainty estimation module is built based on Bayesian neural networks or deep ensemble methods and is used to quantify and predict uncertainty to assist clinical decision-making.

[0164] By predicting the uncertainty of the annotation results, high-confidence data is automatically filtered to reduce the negative impact of noisy annotations on the model; by using multi-expert consensus for historical data annotation, the bias of a single annotator is reduced and the reliability of training data is improved.

[0165] The feedback optimization layer is used to optimize the model parameters of the multi-modal data annotation model;

[0166] The feedback optimization layer is constructed based on an active learning module and a continuous learning module. The active learning module is based on uncertainty sampling and domain adaptation strategies, and combines medical expert knowledge to design sampling strategies for intelligently selecting the most valuable data for annotation. The continuous learning module is based on elastic weight solidification and knowledge distillation, and combines a robust processing mechanism for changes in the distribution of medical data to support incremental learning of the model without forgetting existing knowledge.

[0167] The core objective of the active learning module is to enable the model to automatically select the most valuable data for human experts to annotate, avoiding wasting effort on unimportant data. Uncertainty sampling (prioritizing learning from areas of model uncertainty): When the model predicts, it assigns a "confidence score" to each sample. The active learning module prioritizes these "low-confidence" samples for doctors or experts to annotate, as these data may represent the model's "knowledge blind spots." Diversity sampling (avoiding repetitive learning of similar data): If the model has already learned some common cases (such as pneumonia), it will proactively seek out uncommon cases (such as rare diseases) for expert annotation, avoiding continuously learning repetitive information. Domain adaptation (optimization for the specificities of medicine): Medical data may contain noise (such as blurry images or incorrect medical records). The active learning module identifies these "difficult" samples and prioritizes expert confirmation, preventing the model from being misled.

[0168] The core objective of the continuous learning module is to enable the model to continuously learn new data without forgetting previously learned knowledge. Elastic Weight Consolidation (EWC) (protecting old knowledge): The model adds a "protective shield" to previously learned "important knowledge points," preventing them from being easily overwritten during training with new data; for example, if the model has already learned the "features of pneumonia," it won't completely forget the diagnostic methods for pneumonia when a new batch of "heart disease" data arrives. Knowledge Distillation (old model teaching new model): When training a new model, the old model acts as the "teacher," supervising the new model's predictions and ensuring that the new model doesn't deviate from the correct direction; for example, after the new model learns new features of "diabetes," it still needs to be checked by the old model to ensure it can still correctly diagnose "hypertension." Incremental Learning (gradual updates): Instead of retraining the model with all new data at once, it updates in batches and gradually to avoid "learning collapse"; for example, let the model learn 100 new cases first to test the effect before learning the next batch, which is more stable.

[0169] The active learning module allows the model to select challenging data for expert labeling, thereby optimizing model parameters, improving efficiency, and reducing repetitive work. The continuous learning module enables the model to continuously learn new things without forgetting old knowledge, ensuring long-term stability and reliability.

[0170] When only one modality of data is input in a practical application, this multimodal medical data annotation model can still perform automatic annotation because each modality has its own dedicated feature extraction module.

[0171] The dataset construction module is used to acquire a large amount of historical medical data, including medical imaging data, detection waveform data, electronic medical record data, and gene sequencing data. The dataset is constructed after the historical medical data is desensitized, preprocessed, and labeled.

[0172] The multi-modal data annotation model training module is used to train the multi-modal data annotation model using the dataset, annotation loss function, and annotation reward function, and to compress the multi-modal data annotation model during the training process;

[0173] The medical decision model creation module is used to create a medical decision model based on a multimodal data fusion layer, a feature interaction layer, and a diagnostic decision layer, and to set the decision loss function and decision reward function of the medical decision model.

[0174] The multimodal data fusion layer is used to extract joint representations from medical imaging 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 layer is used to output medical strategy recommendations carrying disease classification, prognostic assessment, and treatment suggestions based on the fused features.

[0175] The medical decision-making model training module is used to train the medical decision-making model using the dataset, decision loss function, and decision reward function, and to compress the medical decision-making model during the training process.

[0176] The model deployment module is used to deploy the trained multi-modal data annotation model and medical decision-making model, and to perform model drift training on the deployed multi-modal data annotation model and medical decision-making model.

[0177] The incremental knowledge training module 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 incrementally train the medical decision-making model using the annotated real-time medical data and the latest medical knowledge.

[0178] The medical question-and-answer module is used to conduct medical question-and-answer through the deployed medical decision-making model, and to record question-and-answer logs in real time, including at least the medical question, medical strategy suggestions, question-and-answer time, question account, and misdiagnosis feedback.

[0179] The question-and-answer log management module is used to annotate the question-and-answer logs using the deployed multi-modal data annotation model, incrementally train the medical decision-making model using the annotated question-and-answer logs, and encrypt and distribute the question-and-answer logs.

[0180] 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 contextual information into the initial image features through a Transformer to obtain contextual features, and filters the contextual features through an attention gating mechanism to obtain image features.

[0181] In essence, 3D CNN acts as the "basic worker," responsible for extracting raw features; Transformer acts as the "integration expert," responsible for piecing together features into a more meaningful whole; and attention gating mechanism acts as the "quality inspector," responsible for filtering out the most important features. Through this collaborative work, the medical image data feature extraction module can extract key information from images more efficiently and accurately, providing strong support for subsequent diagnosis or analysis.

[0182] The waveform feature extraction module extracts local waveform features from the detected waveform data through a 1D temporal convolutional network, performs sequential regularization on each of the local features through causal convolution, and captures the temporal dependencies from each of the sequentially regularized local features through a recurrent neural network to obtain waveform features.

[0183] In essence, 1D CNN acts as the "starter," responsible for extracting local features from waveform data; causal convolution is the "relay runner," ensuring that the feature extraction process doesn't "cheat" (i.e., it doesn't use future information); and RNN is the "sprinter," responsible for integrating time-series information and capturing the dynamic changes in the waveform. Through this collaborative work, the waveform data feature extraction module can extract key features from the waveform more efficiently and accurately, providing strong support for subsequent analysis or detection.

[0184] The case feature extraction module extracts case features from electronic medical record data using a Transformer model that incorporates pre-trained medical term vectors.

[0185] In essence, the Transformer model acts as the "basic translator," responsible for converting electronic medical record data into feature vectors that computers can understand; the medical terminology embedding pre-training is the "professional translator," helping the model better understand medical terminology. Through this collaborative work, the electronic medical record data feature extraction module can extract key features from medical records more efficiently and accurately, providing strong support for subsequent diagnosis, treatment, or research.

[0186] The gene feature extraction module extracts key gene features from gene sequencing data through a convolutional autoencoder, learns the interaction between the key gene features through a graph neural network to obtain gene enhancement features, and integrates semantic information for each gene enhancement feature through a gene ontology knowledge graph to obtain gene features.

[0187] In essence, the convolutional autoencoder acts as the "basic detective," responsible for extracting preliminary features from gene sequencing data; the graph neural network is the "relationship expert," further refining gene feature representations by analyzing interactions between genes; and the gene ontology knowledge graph is the "knowledge base expert," providing rich semantic information for genes to help us better understand their biological significance. 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 development.

[0188] By integrating medical imaging data, waveform data, electronic medical record data, and gene sequencing data, this approach comprehensively captures patient information, avoiding the limitations of a single modality. Image features are designed using a combination of 3D CNN and Transformer to balance local details and global context. Waveform features are designed using 1D temporal and causal convolutions to handle temporal dependencies. Medical record features are integrated with pre-trained medical terminology vectors to enhance semantic understanding. Gene features are combined with graph neural networks and knowledge graphs to analyze gene interactions and biological significance. Finally, cross-modal comparative learning and dynamic weighted fusion are used to strengthen multimodal synergy and effectively improve feature extraction capabilities.

[0189] The spatiotemporal 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 warp alignment network to complete the spatiotemporal alignment operation.

[0190] Deformable convolutions act as "spatial adjusters," responsible for aligning different modal features in space; temporal warp alignment networks act as "temporal adjusters," responsible for aligning different modal features in time. Through this collaborative work, the spatiotemporal alignment module can more efficiently and accurately align multimodal features, providing more consistent feature representations for subsequent tasks (such as multimodal fusion, video understanding, or behavior recognition).

[0191] The multi-scale feature fusion module extracts relevant features from spatiotemporally aligned image features, waveform features, case features, and gene features through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of the image features, waveform features, case features, and gene features based on the relevant features to obtain fused features. The complementary feature expression of the fused features is enhanced through intermodal contrastive learning.

[0192] In essence, the cross-attention mechanism acts as the "communicator," enabling different features to understand each other; the gated fusion unit acts as the "regulator," adjusting the degree of feature fusion based on importance; and the intermodal contrastive learning acts as the "optimizer," ensuring that the fused features are more consistent and effective. Through this collaborative work, the multi-scale feature fusion module can more efficiently and accurately fuse features from different sources, providing more powerful feature representations for subsequent tasks (such as classification, detection, or generation).

[0193] By combining spatiotemporal alignment and multi-scale fusion, specifically by using deformable convolution and temporal warping alignment to address spatiotemporal inconsistency, and by dynamically adjusting weights through a cross-attention mechanism, the quality of fused features is effectively improved, thereby enhancing the quality of annotation.

[0194] The formula for the labeling loss function is:

[0195] L1=λ1*L vis +λ2*L text +λ3*L cons +λ4*L fusionLtotal ;

[0196] Where L1 represents the loss value of the labeled loss function; L vis This represents the contrast loss in medical imaging data, used to constrain the correlation between key image features and labels; L text This represents the focus loss in electronic medical record data, used to mitigate the class imbalance problem; L cons L represents the cross-modal consistency loss, used to constrain the consistency of prediction distributions across different modes using KL divergence constraints; fusionLtotal λ1, λ2, λ3, and λ4 represent the fusion enhancement loss, used to enhance multimodal synergistic effects using dynamic weighted cross-entropy; λ1, λ2, λ3, and λ4 all represent weight coefficients.

[0197] The formula for the labeled reward function is:

[0198] R1=α1*R acc1 +α2*R comp +α3*R eff1 +α4*R userRtotal ;

[0199] Where R1 represents the reward value of the labeled reward function; R acc1Indicates the accuracy of the annotation; R comp Represents modal complementarity; R eff1 Indicates annotation efficiency; R userRtotal α1, α2, α3, and α4 represent user approval; α1, α2, α3, and α4 all represent weighting coefficients.

[0200] The dataset construction module is specifically used for:

[0201] Acquire a large amount of historical medical data, including medical imaging data, test waveform data, electronic medical record data, and gene sequencing data;

[0202] The names in the metadata carried by each medical image data are deleted, and the facial regions in each medical image data are blurred to complete the desensitization of each medical image data. The desensitized medical image data is then preprocessed by noise reduction, grayscale transformation, image enhancement, image normalization and image segmentation. Based on the preset image annotation specifications and the multi-annotator fusion method, the preprocessed medical image data is annotated with patient regions, disease classifications, prognostic assessments and treatment suggestions to generate corresponding labels.

[0203] The patient identifiers carried in each of the detection waveform data are deleted to complete the desensitization. The desensitized detection waveform data are then preprocessed by noise reduction, baseline correction, mean removal, linear trend removal, waveform peaking, and data segmentation. Based on the preset waveform annotation specifications and the multi-annotator fusion method, the preprocessed detection waveform data are annotated with patient waveforms, disease classifications, prognostic assessments, and treatment suggestions to generate corresponding labels.

[0204] The patient identifiers carried in each of the electronic medical records are deleted to complete the desensitization. The desensitized electronic medical records are preprocessed by deleting irrelevant content, filling missing values, correcting erroneous data, unifying format, and tagging parts of speech. Based on the preset medical record tagging specifications and the multi-annotator fusion method, the preprocessed electronic medical records are annotated with disease classification, prognosis assessment and treatment suggestions to generate corresponding tags.

[0205] The patient identifiers carried in each of the gene sequencing data are deleted to complete the desensitization. The desensitized gene sequencing data are preprocessed by removing low-quality sequences, repetitive sequences, adapter sequences and performing error correction. Based on the preset gene annotation specifications and the multi-annotator fusion method, the preprocessed gene sequencing data are annotated with abnormal sequences, disease classification, prognostic assessment and treatment suggestions to generate corresponding tags.

[0206] A dataset is constructed based on the labeled historical medical data.

[0207] The multi-modal data annotation model training module is specifically used for:

[0208] The dataset is divided into a first training set, a first validation set, and a first test set in a ratio of 8:1:1 using stratified sampling. The multi-modal data labeling model is trained using the first training set. In each training round, the labeling loss function and the labeling reward function are used alternately to optimize the multi-modal data labeling model. The training process combines weight pruning, weight quantization, and parameter sharing to compress the multi-modal data labeling model until the preset first early stopping condition is met.

[0209] The annotation accuracy of the multi-modal data annotation model is calculated using the first validation set to validate the model. If the validation fails, the first training set is expanded and training continues; if the validation passes, then:

[0210] The label confidence of the multi-modal data labeling model is calculated using the first test set to test the multi-modal data labeling model. If the test fails, the first training set is expanded and training continues; if the test passes, training ends.

[0211] In the medical decision-making model creation module, the multimodal data fusion layer is constructed based on a medical imaging branch, a waveform processing branch, a medical record coding branch, a gene analysis branch, and a cross-modal alignment module. The medical imaging branch is used to extract image features from medical image data using a 3D ResNet-50 network; the waveform processing branch is used to extract periodic features from detected waveform data using a 1D CNN-BiLSTM network; the medical record coding branch is used to extract temporal features from electronic medical record data using a BERT model; the gene analysis branch is used to extract genetic features from gene sequencing data using a GAT network; and the cross-modal alignment module is used to align and fuse image features, periodic features, temporal features, and genetic features using contrastive learning loss to obtain a joint representation.

[0212] The feature interaction layer obtains interaction features by performing feature interaction on the joint representation through a cross-attention mechanism, and obtains fused features by fusing the interaction features through a dynamic graph neural network. The feature distillation module compresses redundant features in the fused features.

[0213] Feature interaction layers are an indispensable part of modern deep learning models. They enhance the model's overall understanding of data and predictive performance by promoting information exchange and fusion between different features. Feature interaction layers play an important role in fields such as multimodal learning, multi-task learning, recommendation systems, and natural language processing.

[0214] The formula for the decision loss function is:

[0215] L2=β1*L diag +β2*L path +β3*L time ;

[0216] Where L2 represents the loss value of the decision loss function; L diag L represents the diagnostic cross-entropy loss, used to measure the difference between the recommended medical strategy and the true label; path The regularization loss of the clinical pathway is represented by L, and the KL divergence is used to measure the deviation between the model's decision path and the standard clinical pathway; time β1, β2, and β3 represent the temporal consistency loss, used to ensure the rationality between successive decision steps; β1, β2, and β3 all represent weighting coefficients.

[0217] The formula for the decision reward function is:

[0218] R2=γ1*R acc2 +γ2*R path +γ3*R risk +γ4*R eff2 ;

[0219] Where R² represents the reward value of the decision reward function; R acc2 Rewards for diagnostic accuracy are expressed using weighted F1 scores; R path The reward for path rationality is represented by cosine similarity; R risk Indicates a risk-averse reward; R eff2 This represents the reward for treatment efficiency; γ1, γ2, γ3, and γ4 all represent weighting coefficients.

[0220] By fusing contrast loss and focus loss through the labeling loss function, class imbalance and modality consistency are effectively addressed; by fusing KL divergence and temporal consistency loss through the decision loss function, the rationality of clinical pathways and decision coherence are effectively constrained; by introducing indicators such as user approval and treatment efficiency through the reward function, end-to-end optimization is achieved in combination with reinforcement learning.

[0221] The medical decision-making model training module is specifically used for:

[0222] The dataset is divided into a second training set, a second validation set, and a second test set in a ratio of 8:1:1 using stratified sampling. The medical decision-making model is trained using the second training set. In each training round, the decision loss function and the decision reward function are used alternately 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 the preset second early cessation condition is met.

[0223] The decision accuracy of the medical decision-making model is calculated using the second validation set to validate the model. If the validation fails, the second training set is expanded for continued training; if the validation succeeds, then:

[0224] 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.

[0225] By combining weight pruning, weight quantization, and parameter sharing during the model training phase, the model size is effectively compressed, computational resource consumption is reduced, and it is adapted to edge device deployment.

[0226] The model deployment module is specifically used for:

[0227] The trained multi-modal data annotation model and medical decision-making model are deployed locally using containerization technology. A large amount of real medical data is collected, and each piece of real medical data is de-identified, preprocessed, and labeled. Based on the labeled real medical data, the deployed multi-modal data annotation model and medical decision-making model are subjected to model drift training. The performance of the multi-modal data annotation model and medical decision-making model after model drift training is verified by accuracy, recall, F1 score, and ROC curve.

[0228] By dividing the training set, validation set, and test set into an 8:1:1 ratio, overfitting is avoided; the training set is automatically expanded when validation fails, effectively improving generalization ability; and the model performance is verified from multiple perspectives, including accuracy, recall, F1 score, and ROC curve, to ensure clinical reliability.

[0229] The medical Q&A module is specifically used for:

[0230] Through a visual interface, input medical questions carrying medical imaging data, detection waveform data, electronic medical record data, or gene sequencing data are obtained. After preprocessing the medical questions through a streaming computing engine, they are input into a deployed medical decision-making model. The medical decision-making model, combined with GPU acceleration technology and memory caching technology, infers medical strategy suggestions from the medical questions and provides feedback on the medical strategy suggestions through the visual interface to conduct medical Q&A. The Q&A log is recorded in real time, including at least the medical question, medical strategy suggestions, Q&A time, questioning account, and misdiagnosis feedback.

[0231] The question-and-answer log management module is specifically used for:

[0232] The question-and-answer logs are automatically labeled by the deployed multi-modal data labeling model. When the number of labeled question-and-answer logs reaches a preset threshold, the medical decision-making model is automatically incrementally trained using the labeled question-and-answer logs within a preset time period.

[0233] Real-time medical data and question-and-answer logs are automatically labeled and fed back to the model, forming a closed-loop optimization of "data-model". Long-term performance stability is ensured by using model drift training combined with ROC curves and other indicators for verification.

[0234] The question-and-answer log is encrypted to obtain an encrypted log, the encrypted log is distributed and stored, the data fingerprint of the encrypted log is calculated using the HMAC algorithm, and the data fingerprint is uploaded to the blockchain;

[0235] By employing a multi-modal data annotation model to automatically annotate question-and-answer logs carrying misdiagnosis feedback, annotation efficiency is effectively improved; by generating data fingerprints using the HMAC algorithm and uploading them to the blockchain, the traceability and tamper-proof nature of misdiagnosis cases are ensured; by setting a quantity threshold through an incremental training trigger mechanism, the model is automatically triggered to update when the number of misdiagnosis cases accumulates to a clinically significant level, effectively solving the problem of missing feedback, and thus constructing a closed loop of "misdiagnosis cases - annotation feedback - model iteration".

[0236] The specific steps for encrypting the question-and-answer log to obtain the encrypted log are as follows:

[0237] The medical image data in the question-and-answer log is subjected to DCT transformation to extract frequency domain coefficients. Based on the frequency domain coefficients, Zigzag scanning is performed to convert the medical image data into a one-dimensional sequence. The text data, including electronic medical record data, medical strategy suggestions, and misdiagnosis feedback, is encoded using UTF-8 encoding and then compressed using Huffman coding to obtain compressed data. Wavelet packet decomposition is performed on the detection waveform data to extract feature coefficients. The gene sequencing data is converted to FASTQ format and encoded using base sequence encoding to obtain encoded data.

[0238] The one-dimensional sequence, compressed data, feature coefficients, and encoded data are each subjected to SM3 hash calculation to obtain corresponding sub-check codes. Then, each sub-check code is subjected to SM3 hash calculation to obtain the master check code.

[0239] The one-dimensional sequence is encrypted into first encrypted data using the AES-GCM algorithm, the compressed data is encrypted into second encrypted data using the SM9 algorithm, the feature coefficients are encrypted into third encrypted data using the ECIES algorithm, and the encoded data is encrypted into fourth encrypted data using the Paillier algorithm.

[0240] Using the SM4 and Serpent algorithms, the first encrypted data, the second encrypted data, the third encrypted data, the fourth encrypted data, and the master checksum are sandwiched together to obtain the encrypted log.

[0241] By combining two SM3 hash calculations during the encrypted log generation process, and employing AES-GCM, SM9, ECIES, and Paillier algorithms for different data types, and finally using SM4 and Serpent algorithms for sandwich encryption, and uploading the data fingerprint of the encrypted log to the blockchain, multi-level encryption, integrity verification, and tamper-proofing of question-and-answer logs are achieved, greatly improving the security of question-and-answer log storage.

[0242] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended 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 of protection of the claims of the present invention.

Claims

1. A method for generating medical strategies by combining data flywheel reinforcement learning, characterized in that: Includes the following steps: Step S1: Create a multi-modal data annotation model based on the feature extraction layer, modality fusion layer, feature enhancement layer, annotation prediction layer, and feedback optimization layer; and set the annotation loss function and annotation reward function of the multi-modal data annotation model. The formula for the labeling loss function is: L1=λ1*L vis +λ2*L text +λ3*L cons +λ4*L fusionLtotal ; Where L1 represents the loss value of the labeled loss function; L vis This represents the contrast loss in medical imaging data, used to constrain the correlation between key image features and labels; L text This represents the focus loss in electronic medical record data, used to mitigate the class imbalance problem; L cons L represents the cross-modal consistency loss, used to constrain the consistency of prediction distributions across different modes using KL divergence constraints; fusionLtotal λ1, λ2, λ3, and λ4 represent the fusion enhancement loss, used to enhance multimodal synergistic effects using dynamic weighted cross-entropy; λ1, λ2, λ3, and λ4 all represent weight coefficients. The formula for the labeled reward function is: R1=α1*R acc1 +α2*R comp +α3*R eff1 +α4*R userRtotal ; Where R1 represents the reward value of the labeled reward function; R acc1 Indicates the accuracy of the annotation; R comp Represents modal complementarity; R eff1 Indicates annotation efficiency; R userRtotal Indicates user approval; α1, α2, α3, and α4 all represent weighting coefficients. Step S2: Acquire a large amount of historical medical data, including medical imaging data, detection waveform data, electronic medical record data, and gene sequencing data; desensitize, preprocess, and label each of the historical medical data to construct a dataset. Step S3: Train the multi-modal data labeling model using the dataset, labeling loss function, and labeling reward function, and compress the multi-modal data labeling model during the training process; Step S4: Create a medical decision-making model based on the multimodal data fusion layer, feature interaction layer, and diagnostic decision layer, and set the decision loss function and decision reward function of the medical decision-making model; The multimodal data fusion layer is used to extract joint representations from medical imaging 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 layer is used to output medical strategy recommendations carrying disease classification, prognostic assessment, and treatment suggestions based on the fused features. 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 L represents the diagnostic cross-entropy loss, used to measure the difference between the recommended medical strategy and the true label; path The regularization loss of the clinical pathway is represented by L, and the KL divergence is used to measure the deviation between the model's decision path and the standard clinical pathway; time β1, β2, and β3 represent the temporal consistency loss, used to ensure the rationality between successive decision steps; β1, β2, and β3 all represent weighting coefficients. The formula for the decision reward function is: 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 Rewards for diagnostic accuracy are expressed using weighted F1 scores; R path The reward for path rationality is represented by cosine similarity; R risk Indicates a risk-averse reward; R eff2 This represents the reward for treatment efficiency; γ1, γ2, γ3, and γ4 all represent weighting coefficients. 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 labeling model and medical decision-making model locally using containerization technology, collect a large amount of real medical data, de-identify, preprocess and label each piece of real medical data, perform model drift training on the deployed multi-modal data labeling model and medical decision-making model based on the labeled real medical data, and verify the performance of the multi-modal data labeling model and medical decision-making model after model drift training by accuracy, recall, F1 score and ROC curve. Step S7: Collect real-time medical data and the latest medical knowledge, automatically label the real-time medical data and the latest medical knowledge using the deployed multi-modal data labeling model, and incrementally train the medical decision-making model using the labeled real-time medical data and the latest medical knowledge. Step S8: Perform medical question and answer through the deployed medical decision model, and record question and answer logs in real time, including at least the medical question, medical strategy suggestions, question and answer time, question account, and misdiagnosis feedback; Step S9: Automatically annotate the question-and-answer logs using the deployed multi-modal data annotation model. When the number of annotated question-and-answer logs reaches a preset threshold, the medical decision-making model is automatically incrementally trained using the annotated question-and-answer logs within a preset time period. The question-and-answer log is encrypted to obtain an encrypted log, which is then distributed and stored. The data fingerprint of the encrypted log is calculated using the HMAC algorithm, and the data fingerprint is uploaded to the blockchain.

2. The medical strategy generation method combining data flywheel reinforcement learning as described in claim 1, characterized in that: 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 medical record data; and the gene feature extraction module is used to extract gene features from gene sequencing data. The modality fusion layer is constructed based on a spatiotemporal alignment module and a multi-scale feature fusion module. The spatiotemporal alignment module is used to perform spatiotemporal alignment operations on image features, waveform features, case features, and gene features. The multi-scale feature fusion module is used to fuse the spatiotemporally aligned image features, waveform features, case features, and gene features 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 primary enhancement features through a dynamic medical knowledge graph to obtain secondary enhancement 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 medical record data and gene sequencing data respectively based on the secondary enhancement features. The uncertainty estimation module is used to predict the uncertainty of each annotation result, filter 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 contextual information into the initial image features through a Transformer to obtain contextual features, and filters the contextual features through an attention gating mechanism to obtain image features. The waveform feature extraction module extracts local waveform features from the detected waveform data through a 1D temporal convolutional network, performs sequential regularization on each of the local features through causal convolution, and captures the temporal dependencies from each of the sequentially regularized local features through a recurrent neural network to obtain waveform features. The case feature extraction module extracts case features from electronic medical record data using a Transformer model that incorporates pre-trained medical term vectors. The gene feature extraction module extracts key gene features from gene sequencing data through a convolutional autoencoder, learns the interaction between the key gene features through a graph neural network to obtain gene enhancement features, and integrates semantic information for each gene enhancement feature through a gene ontology knowledge graph to obtain gene features. The spatiotemporal 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 warp alignment network to complete the spatiotemporal alignment operation. The multi-scale feature fusion module extracts relevant features from spatiotemporally aligned image features, waveform features, case features, and gene features through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of the image features, waveform features, case features, and gene features based on the relevant features to obtain fused features. The complementary feature expression of the fused features is enhanced through intermodal contrastive learning.

3. The medical strategy generation method combining data flywheel reinforcement learning as described in claim 1, characterized in that: Step S2 specifically involves: Acquire a large amount of historical medical data, including medical imaging data, test waveform data, electronic medical record data, and gene sequencing data; The names in the metadata carried by each medical image data are deleted, and the facial regions in each medical image data are blurred to complete the desensitization of each medical image data. The desensitized medical image data is then preprocessed by noise reduction, grayscale transformation, image enhancement, image normalization and image segmentation. Based on the preset image annotation specifications and the multi-annotator fusion method, the preprocessed medical image data is annotated with patient regions, disease classifications, prognostic assessments and treatment suggestions to generate corresponding labels. The patient identifiers carried in each of the detection waveform data are deleted to complete the desensitization. The desensitized detection waveform data are then preprocessed by noise reduction, baseline correction, mean removal, linear trend removal, waveform peaking, and data segmentation. Based on the preset waveform annotation specifications and the multi-annotator fusion method, the preprocessed detection waveform data are annotated with patient waveforms, disease classifications, prognostic assessments, and treatment suggestions to generate corresponding labels. The patient identifiers carried in each of the electronic medical records are deleted to complete the desensitization. The desensitized electronic medical records are preprocessed by deleting irrelevant content, filling missing values, correcting erroneous data, unifying format, and tagging parts of speech. Based on the preset medical record tagging specifications and the multi-annotator fusion method, the preprocessed electronic medical records are annotated with disease classification, prognosis assessment and treatment suggestions to generate corresponding tags. The patient identifiers carried in each of the gene sequencing data are deleted to complete the desensitization. The desensitized gene sequencing data are preprocessed by removing low-quality sequences, repetitive sequences, adapter sequences and performing error correction. Based on the preset gene annotation specifications and the multi-annotator fusion method, the preprocessed gene sequencing data are annotated with abnormal sequences, disease classification, prognostic assessment and treatment suggestions to generate corresponding tags. A dataset is constructed based on the labeled historical medical data. Step S3 specifically involves: The dataset is divided into a first training set, a first validation set, and a first test set in a ratio of 8:1:1 using stratified sampling. The multi-modal data labeling model is trained using the first training set. In each training round, the labeling loss function and the labeling reward function are used alternately to optimize the multi-modal data labeling model. The training process combines weight pruning, weight quantization, and parameter sharing to compress the multi-modal data labeling model until the 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 validate the model. If the validation fails, the first training set is expanded and training continues; if the validation passes, then: The label confidence of the multi-modal data labeling model is calculated using the first test set to test the multi-modal data labeling model. If the test fails, the first training set is expanded and training continues; if the test passes, training ends.

4. The medical strategy generation method combining data flywheel reinforcement learning as described in claim 1, characterized in that: In step S4, the multimodal data fusion layer is constructed based on a medical imaging branch, a waveform processing branch, a medical record coding branch, a gene analysis branch, and a cross-modal alignment module. The medical imaging branch is used to extract image features from medical image data using a 3D ResNet-50 network. The waveform processing branch is used to extract periodic features from detected waveform data using a 1D CNN-BiLSTM network. The medical record coding branch is used to extract temporal features from electronic medical record data using a BERT model. The gene analysis branch is used to extract genetic features from gene sequencing data using a GAT network. The cross-modal alignment module is used to align and fuse image features, periodic features, temporal features, and genetic features using contrastive learning loss to obtain a joint representation. The feature interaction layer obtains interaction features by performing feature interaction on the joint representation through a cross-attention mechanism, and obtains fused features by fusing the interaction features through a dynamic graph neural network. The feature distillation module compresses redundant features in the fused features. Step S5 specifically involves: The dataset is divided into a second training set, a second validation set, and a second test set in a ratio of 8:1:1 using stratified sampling. The medical decision-making model is trained using the second training set. In each training round, the decision loss function and the decision reward function are used alternately 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 the preset second early cessation condition is met. The decision accuracy of the medical decision-making model is calculated using the second validation set to validate the model. If the validation fails, the second training set is expanded for continued training; if the validation succeeds, 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.

5. The medical strategy generation method combining data flywheel reinforcement learning as described in claim 1, characterized in that: Step S8 specifically involves: Through a visual interface, input medical questions carrying medical imaging data, detection waveform data, electronic medical record data, or gene sequencing data are obtained. After preprocessing the medical questions through a streaming computing engine, they are input into a deployed medical decision-making model. The medical decision-making model, combined with GPU acceleration technology and memory caching technology, infers medical strategy suggestions from the medical questions and provides feedback on the medical strategy suggestions through the visual interface to conduct medical Q&A. The Q&A log is recorded in real time, including at least the medical question, medical strategy suggestions, Q&A time, questioning account, and misdiagnosis feedback. In step S9, encrypting the question-and-answer log to obtain the encrypted log specifically involves: The medical image data in the question-and-answer log is subjected to DCT transformation to extract frequency domain coefficients, and Zigzag scanning is performed 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, compressed data is obtained by Huffman coding; Wavelet packet decomposition is performed on the detected waveform data to extract feature coefficients; Gene sequencing data is converted to FASTQ format and encoded using base sequences to obtain encoded data. The one-dimensional sequence, compressed data, feature coefficients, and encoded data are each subjected to SM3 hash calculation to obtain corresponding sub-check codes. Then, each sub-check code is subjected to SM3 hash calculation to obtain the master check code. The one-dimensional sequence is encrypted into first encrypted data using the AES-GCM algorithm, the compressed data is encrypted into second encrypted data using the SM9 algorithm, the feature coefficients are encrypted into third encrypted data using the ECIES algorithm, and the encoded data is encrypted into fourth encrypted data using the Paillier algorithm. Using the SM4 and Serpent algorithms, the first encrypted data, the second encrypted data, the third encrypted data, the fourth encrypted data, and the master checksum are sandwiched together to obtain the encrypted log.

6. A medical strategy generation system combining data flywheel reinforcement learning, characterized in that: Includes the following modules: The multimodal data annotation model creation module is used to create a multimodal 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 to set the annotation loss function and annotation reward function of the multimodal data annotation model. The formula for the labeling loss function is: L1=λ1*L vis +λ2*L text +λ3*L cons +λ4*L fusionLtotal ; Where L1 represents the loss value of the labeled loss function; L vis This represents the contrast loss in medical imaging data, used to constrain the correlation between key image features and labels; L text This represents the focus loss in electronic medical record data, used to mitigate the class imbalance problem; L cons L represents the cross-modal consistency loss, used to constrain the consistency of prediction distributions across different modes using KL divergence constraints; fusionLtotal λ1, λ2, λ3, and λ4 represent the fusion enhancement loss, used to enhance multimodal synergistic effects using dynamic weighted cross-entropy; λ1, λ2, λ3, and λ4 all represent weight coefficients. The formula for the labeled reward function is: R1=α1*R acc1 +α2*R comp +α3*R eff1 +α4*R userRtotal ; Where R1 represents the reward value of the labeled reward function; R acc1 Indicates the accuracy of the annotation; R comp Represents modal complementarity; R eff1 Indicates annotation efficiency; R userRtotal Indicates user approval; α1, α2, α3, and α4 all represent weighting coefficients. The dataset construction module is used to acquire a large amount of historical medical data, including medical imaging data, detection waveform data, electronic medical record data, and gene sequencing data. The dataset is constructed after the historical medical data is desensitized, preprocessed, and labeled. The multi-modal data annotation model training module is used to train the multi-modal data annotation model using the dataset, annotation loss function, and annotation reward function, and to compress the multi-modal data annotation model during the training process; The medical decision model creation module is used to create a medical decision model based on a multimodal data fusion layer, a feature interaction layer, and a diagnostic decision layer, and to set the decision loss function and decision reward function of the medical decision model. The multimodal data fusion layer is used to extract joint representations from medical imaging 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 layer is used to output medical strategy recommendations carrying disease classification, prognostic assessment, and treatment suggestions based on the fused features. 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 L represents the diagnostic cross-entropy loss, used to measure the difference between the recommended medical strategy and the true label; path The regularization loss of the clinical pathway is represented by L, and the KL divergence is used to measure the deviation between the model's decision path and the standard clinical pathway; time β1, β2, and β3 represent the temporal consistency loss, used to ensure the rationality between successive decision steps; β1, β2, and β3 all represent weighting coefficients. The formula for the decision reward function is: 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 Rewards for diagnostic accuracy are expressed using weighted F1 scores; R path The reward for path rationality is represented by cosine similarity; R risk Indicates a risk-averse reward; R eff2 This represents the reward for treatment efficiency; γ1, γ2, γ3, and γ4 all represent weighting coefficients. The medical decision-making model training module is used to train the medical decision-making model using the dataset, decision loss function, and decision reward function, and to compress the medical decision-making model during the training process. The model deployment module is used to locally deploy the trained multi-modal data labeling model and medical decision-making model using containerization technology. It collects a large amount of real medical data, desensitizes, preprocesses, and labels each piece of real medical data, performs model drift training on the deployed multi-modal data labeling model and medical decision-making model based on the labeled real medical data, and verifies the performance of the multi-modal data labeling model and medical decision-making model after model drift training by accuracy, recall, F1 score, and ROC curve. The incremental knowledge training module 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 incrementally train the medical decision-making model using the annotated real-time medical data and the latest medical knowledge. The medical question-and-answer module is used to conduct medical question-and-answer through the deployed medical decision-making model, and to record question-and-answer logs in real time, including at least the medical question, medical strategy suggestions, question-and-answer time, question account, and misdiagnosis feedback. The question-and-answer log management module is used to automatically annotate the question-and-answer logs through the deployed multi-modal data annotation model. When the number of annotated question-and-answer logs reaches a preset threshold, the medical decision-making model is automatically incrementally trained using the annotated question-and-answer logs within a preset time period. The question-and-answer log is encrypted to obtain an encrypted log, which is then distributed and stored. The data fingerprint of the encrypted log is calculated using the HMAC algorithm, and the data fingerprint is uploaded to the blockchain.

7. A medical strategy generation system combining data flywheel reinforcement learning as described in 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 medical record data; and the gene feature extraction module is used to extract gene features from gene sequencing data. The modality fusion layer is constructed based on a spatiotemporal alignment module and a multi-scale feature fusion module. The spatiotemporal alignment module is used to perform spatiotemporal alignment operations on image features, waveform features, case features, and gene features. The multi-scale feature fusion module is used to fuse the spatiotemporally aligned image features, waveform features, case features, and gene features 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 primary enhancement features through a dynamic medical knowledge graph to obtain secondary enhancement 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 medical record data and gene sequencing data respectively based on the secondary enhancement features. The uncertainty estimation module is used to predict the uncertainty of each annotation result, filter 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 contextual information into the initial image features through a Transformer to obtain contextual features, and filters the contextual features through an attention gating mechanism to obtain image features. The waveform feature extraction module extracts local waveform features from the detected waveform data through a 1D temporal convolutional network, performs sequential regularization on each of the local features through causal convolution, and captures the temporal dependencies from each of the sequentially regularized local features through a recurrent neural network to obtain waveform features. The case feature extraction module extracts case features from electronic medical record data using a Transformer model that incorporates pre-trained medical term vectors. The gene feature extraction module extracts key gene features from gene sequencing data through a convolutional autoencoder, learns the interaction between the key gene features through a graph neural network to obtain gene enhancement features, and integrates semantic information for each gene enhancement feature through a gene ontology knowledge graph to obtain gene features. The spatiotemporal 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 warp alignment network to complete the spatiotemporal alignment operation. The multi-scale feature fusion module extracts relevant features from spatiotemporally aligned image features, waveform features, case features, and gene features through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of the image features, waveform features, case features, and gene features based on the relevant features to obtain fused features. The complementary feature expression of the fused features is enhanced through intermodal contrastive learning.

8. A medical strategy generation system combining data flywheel reinforcement learning as described in claim 6, characterized in that: The dataset construction module is specifically used for: Acquire a large amount of historical medical data, including medical imaging data, test waveform data, electronic medical record data, and gene sequencing data; The names in the metadata carried by each medical image data are deleted, and the facial regions in each medical image data are blurred to complete the desensitization of each medical image data. The desensitized medical image data is then preprocessed by noise reduction, grayscale transformation, image enhancement, image normalization and image segmentation. Based on the preset image annotation specifications and the multi-annotator fusion method, the preprocessed medical image data is annotated with patient regions, disease classifications, prognostic assessments and treatment suggestions to generate corresponding labels. The patient identifiers carried in each of the detection waveform data are deleted to complete the desensitization. The desensitized detection waveform data are then preprocessed by noise reduction, baseline correction, mean removal, linear trend removal, waveform peaking, and data segmentation. Based on the preset waveform annotation specifications and the multi-annotator fusion method, the preprocessed detection waveform data are annotated with patient waveforms, disease classifications, prognostic assessments, and treatment suggestions to generate corresponding labels. The patient identifiers carried in each of the electronic medical records are deleted to complete the desensitization. The desensitized electronic medical records are preprocessed by deleting irrelevant content, filling missing values, correcting erroneous data, unifying format, and tagging parts of speech. Based on the preset medical record tagging specifications and the multi-annotator fusion method, the preprocessed electronic medical records are annotated with disease classification, prognosis assessment and treatment suggestions to generate corresponding tags. The patient identifiers carried in each of the gene sequencing data are deleted to complete the desensitization. The desensitized gene sequencing data are preprocessed by removing low-quality sequences, repetitive sequences, adapter sequences and performing error correction. Based on the preset gene annotation specifications and the multi-annotator fusion method, the preprocessed gene sequencing data are annotated with abnormal sequences, disease classification, prognostic assessment and treatment suggestions to generate corresponding tags. A dataset is constructed based on the labeled 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 in a ratio of 8:1:1 using stratified sampling. The multi-modal data labeling model is trained using the first training set. In each training round, the labeling loss function and the labeling reward function are used alternately to optimize the multi-modal data labeling model. The training process combines weight pruning, weight quantization, and parameter sharing to compress the multi-modal data labeling model until the 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 validate the model. If the validation fails, the first training set is expanded and training continues; if the validation passes, then: The label confidence of the multi-modal data labeling model is calculated using the first test set to test the multi-modal data labeling model. If the test fails, the first training set is expanded and training continues; if the test passes, training ends.

9. A medical strategy generation system combining data flywheel reinforcement learning as described in claim 6, characterized in that: In the medical decision-making model creation module, the multimodal data fusion layer is constructed based on a medical imaging branch, a waveform processing branch, a medical record coding branch, a gene analysis branch, and a cross-modal alignment module. The medical imaging branch is used to extract image features from medical image data using a 3DResNet-50 network; the waveform processing branch is used to extract periodic features from detected waveform data using a 1D CNN-BiLSTM network; the medical record coding branch is used to extract temporal features from electronic medical record data using a BERT model; the gene analysis branch is used to extract genetic features from gene sequencing data using a GAT network; and the cross-modal alignment module is used to align and fuse image features, periodic features, temporal features, and genetic features using contrastive learning loss to obtain a joint representation. The feature interaction layer obtains interaction features by performing feature interaction on the joint representation through a cross-attention mechanism, and obtains fused features by fusing the interaction features through a dynamic graph neural network. The feature distillation module compresses redundant features in the fused features. 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 in a ratio of 8:1:1 using stratified sampling. The medical decision-making model is trained using the second training set. In each training round, the decision loss function and the decision reward function are used alternately 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 the preset second early cessation condition is met. The decision accuracy of the medical decision-making model is calculated using the second validation set to validate the model. If the validation fails, the second training set is expanded for continued training; if the validation succeeds, 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.

10. A medical strategy generation system combining data flywheel reinforcement learning as described in claim 6, characterized in that: The medical Q&A module is specifically used for: Through a visual interface, input medical questions carrying medical imaging data, detection waveform data, electronic medical record data, or gene sequencing data are obtained. After preprocessing the medical questions through a streaming computing engine, they are input into a deployed medical decision-making model. The medical decision-making model, combined with GPU acceleration technology and memory caching technology, infers medical strategy suggestions from the medical questions and provides feedback on the medical strategy suggestions through the visual interface to conduct medical Q&A. The Q&A log is recorded in real time, including at least the medical question, medical strategy suggestions, Q&A time, questioning account, and misdiagnosis feedback. In the question-and-answer log management module, encrypting the question-and-answer log to obtain the encrypted log specifically involves: The medical image data in the question-and-answer log is subjected to DCT transformation to extract frequency domain coefficients, and Zigzag scanning is performed 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, compressed data is obtained by Huffman coding; Wavelet packet decomposition is performed on the detected waveform data to extract feature coefficients; Gene sequencing data is converted to FASTQ format and encoded using base sequences to obtain encoded data. The one-dimensional sequence, compressed data, feature coefficients, and encoded data are each subjected to SM3 hash calculation to obtain corresponding sub-check codes. Then, each sub-check code is subjected to SM3 hash calculation to obtain the master check code. The one-dimensional sequence is encrypted into first encrypted data using the AES-GCM algorithm, the compressed data is encrypted into second encrypted data using the SM9 algorithm, the feature coefficients are encrypted into third encrypted data using the ECIES algorithm, and the encoded data is encrypted into fourth encrypted data using the Paillier algorithm. Using the SM4 and Serpent algorithms, the first encrypted data, the second encrypted data, the third encrypted data, the fourth encrypted data, and the master checksum are sandwiched together to obtain the encrypted log.

Images