Auxiliary diagnosis and treatment system based on artificial intelligence

Abstract

The invention belongs to the technical field of medical artificial intelligence, and discloses an artificial intelligence-based auxiliary diagnosis and treatment system, which comprises a multi-modal data acquisition module, a dynamic learning module, a diagnosis reasoning module, a privacy protection module, an interactive decision module and an early warning monitoring module, the output end of the multi-modal data acquisition module is connected with the input end of the privacy protection module, the output end of the privacy protection module is connected with the input end of the dynamic learning module, the output end of the dynamic learning module is connected with the input end of the diagnostic reasoning module, and the output end of the diagnostic reasoning module is connected with the input end of the interactive decision module. And the early warning monitoring module monitors abnormal data in real time and performs bidirectional interaction with the diagnosis reasoning module. According to the method, multi-source medical data are integrated, and high-precision real-time auxiliary diagnosis is realized by adopting a dynamic incremental learning and privacy encryption technology; the medical worker cooperation efficiency is improved through an interactive interface, the safety is guaranteed in combination with real-time monitoring and early warning, and the system can remarkably improve the diagnosis and treatment efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the field of medical artificial intelligence technology, and specifically to an auxiliary diagnosis and treatment system based on multimodal data fusion, dynamic incremental learning and privacy protection technology, which is suitable for disease diagnosis and treatment decision support in hospitals, clinics and telemedicine scenarios. Background Art

[0002] The current medical field faces multiple challenges: 1) Data silos: Patients' medical images, electronic medical records, and genetic data are scattered across disparate systems, lacking effective integration, leading to incomplete diagnostic information. 2) Model staticity: Traditional AI diagnostic models rely on historical data for training and are unable to adapt to new cases or updated medical knowledge in real time, resulting in diagnostic lags. 3) Privacy risks: Medical data contains sensitive information, and existing systems often use centralized storage, making it vulnerable to data leaks and unauthorized access. 4) Insufficient explainability: The diagnostic conclusions output by black-box AI models lack transparency, making it difficult for doctors to trust and collaborate on decision-making. For example, patent publication number CN117649935A, titled "A Deep Learning-Based CT Image-Assisted Diagnosis System," uses convolutional neural networks to analyze medical images but fails to address multimodal data fusion and dynamic model updates. Patent publication number CN106027248B proposes a medical data encryption method, but it is not deeply integrated into the AI ​​diagnostic process, resulting in inefficient data processing. Therefore, there is an urgent need for an intelligent diagnosis and treatment system that integrates data fusion, dynamic learning, privacy protection, and explainable decision-making. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention proposes an artificial intelligence-based auxiliary diagnosis and treatment system, which solves the above problems through the following innovative solutions: 1) Dynamic fusion of multimodal data: integrating multi-source data such as medical images, electronic medical records, and physiological signals, and adopting attention mechanism weighted fusion to improve the comprehensiveness of feature expression. 2) Collaboration of incremental learning and federated learning: supporting real-time model updates and ensuring data privacy, and preventing historical knowledge from being forgotten through elastic weight solidification technology. 3) Hierarchical privacy protection architecture: combining differential privacy, homomorphic encryption and distributed storage to achieve data security management throughout the entire life cycle. 4) Interactive explainable decision-making: providing diagnostic basis heat maps and reasoning chain visualization to enhance doctors' trust in AI conclusions and the efficiency of correction. 5) Real-time monitoring and early warning: triggering hierarchical alarms based on patients' physiological indicators and abnormal diagnosis and treatment processes to reduce medical risks.

[0004] The present invention provides the following specific technical solutions: An artificial intelligence-based auxiliary diagnosis and treatment system, whose core modules include a multimodal data acquisition module, a privacy protection module, a dynamic learning module, a diagnostic reasoning module, an interactive decision-making module, and an early warning monitoring module; The multimodal data acquisition module is used to collect the patient's medical imaging data, electronic medical record text data, gene sequencing data and physiological signal data in real time, and standardize the data into a unified format; The privacy protection module is used to desensitize, encrypt, and distribute multimodal data, supporting data processing based on homomorphic encryption and model training under the federated learning framework; The dynamic learning module is used to update the diagnostic model parameters in real time through an incremental learning algorithm, and optimize the model's generalization ability by combining the medical knowledge graph and the historical case database; The diagnostic reasoning module is used to generate preliminary diagnostic conclusions based on multimodal data fusion and optimize diagnostic confidence through Bayesian networks and deep reinforcement learning algorithms; The interactive decision-making module is used to provide a visual diagnosis and treatment suggestion interface, support doctors to correct and provide feedback on AI diagnosis results, and input feedback data into the dynamic learning module to complete model iteration; The early warning monitoring module is used to analyze the patient's physiological indicators and abnormal signals in the diagnosis and treatment process in real time, trigger graded early warnings and push them to the doctor's terminal: Among them, the output end of the multimodal data acquisition module is connected to the input end of the privacy protection module, the output end of the privacy protection module is connected to the input end of the dynamic learning module, the output end of the dynamic learning module is connected to the input end of the diagnostic reasoning module, the output end of the diagnostic reasoning module is connected to the input end of the interactive decision-making module, and the early warning monitoring module and the diagnostic reasoning module have two-way data interaction.

[0005] Furthermore, the multimodal data acquisition module includes an image parsing unit, a text analysis unit, and a signal processing unit. The image parsing unit uses an improved 3DU-Net network to segment lesions from CT and MRI images, adapting to imaging differences between different devices through transfer learning, achieving a segmentation accuracy of 98.2% (Dice coefficient). The text analysis unit uses the BioBERT model to extract key entities (such as symptoms and medications) from electronic medical records, linking ICD-10 codes with the SNOMEDCT terminology library, achieving an F1 score of 92.5% for entity recognition. The signal processing unit performs wavelet transform denoising on electrocardiogram signals, extracting time-frequency features such as RR interval and ST segment slope, and combining this with an LSTM network to predict arrhythmia risk.

[0006] Furthermore, the privacy protection module includes data desensitization, homomorphic encryption calculation and federated learning framework. Among them, the privacy protection module, data desensitization: performs k-anonymization (k≥5) on PII information such as patient names and ID numbers to ensure that a single record cannot uniquely identify an individual; homomorphic encryption calculation: performs feature extraction and model inference in an encrypted state to meet , where E is the encryption function and f is the computational operation; Federated learning framework: Each hospital node trains the model locally and only uploads the encrypted gradient parameters to the central server for aggregation. The data does not leave the domain and the model accuracy loss is <1.5%.

[0007] Furthermore, the dynamic learning module includes an incremental learning algorithm and knowledge graph verification. The incremental learning algorithm uses an online random forest to dynamically expand the decision tree, generating a new tree for every 100 new cases, with a weight decay coefficient of 0.9 for the historical tree. The knowledge graph verification compares the incremental learning results with the medical knowledge graph (e.g., disease-symptom-treatment relationships), triggering a manual review process if any logical conflicts are found.

[0008] Furthermore, the diagnostic reasoning module includes a multimodal fusion model and confidence optimization. The multimodal fusion model defines the feature vectors of images, texts, and physiological signals. , through the attention weight Dynamic weighting is achieved, and after fusion, the data is input into a multi-layer perceptron (MLP) to output the disease probability distribution; confidence optimization: a Bayesian network is used to calculate the uncertainty of the diagnosis. If the confidence is <85%, the doctor is prompted to review, and a reinforcement learning strategy is initiated to adjust the model parameters.

[0009] Furthermore, the interactive decision-making module includes a visual interface design and feedback mechanism. The visual interface features: a heat map that highlights key areas in the image that influence diagnosis (such as the malignancy score of lung nodules); a reasoning chain that uses a tree diagram to display the logical path of "cough → lung shadow → bacterial pneumonia → antibiotic A"; and a feedback mechanism: After the doctor corrects the diagnosis, the system automatically records the correction and generates incremental learning samples to promote model iteration.

[0010] Furthermore, the early warning monitoring module includes an anomaly detection algorithm and process compliance verification. The anomaly detection algorithm uses an isolation forest to identify physiological indicator anomalies. For example, a sudden drop in blood oxygen saturation exceeding 10% / min triggers a level 1 warning. The process compliance verification compares diagnostic and treatment procedures with clinical guidelines. If an overdose or incompatible medication is detected, a pop-up warning window is immediately displayed and prescription submission permissions are locked.

[0011] The artificial intelligence-based auxiliary diagnosis and treatment system of the present invention has the following significant advantages over the existing technology: 1) It can directly significantly improve the hospital's diagnostic accuracy. The multimodal fusion model has high accuracy in lung cancer diagnosis tasks and has obvious advantages over single imaging models. Incremental learning significantly improves the model's performance in diagnosing newly discovered diseases.

[0012] 2) Privacy compliance is significantly enhanced. Under the federated learning framework, the risk of data leakage at each node is greatly reduced, and the performance of the model remains stable after aggregation. Homomorphic encryption processing meets real-time diagnosis needs.

[0013] 3) Clinical efficiency has been significantly optimized. The efficiency of doctors in correcting AI conclusions through the interactive interface has been significantly improved. The early warning module has significantly shortened the rescue response time for critically ill patients and effectively reduced the incidence of postoperative complications.

[0014] 4) The system's adaptability is significantly enhanced. Incremental learning supports the model to quickly adapt to new cases, and knowledge graph verification effectively avoids logical error diagnoses.

[0015] 5) The system's interpretability has been significantly improved. Heat maps and reasoning chains have significantly increased doctors' trust in AI conclusions, and counterfactual report generation has quickly assisted doctors in verifying the rationality of their diagnoses. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which: Figure 1 This is a schematic diagram of the overall system architecture of an artificial intelligence-based auxiliary diagnosis and treatment system of the present invention; Figure 2 This is a diagram of the improved 3DRes-U-Net network structure of the artificial intelligence-based auxiliary diagnosis and treatment system of the present invention; Figure 3 This is a flow chart of the multimodal data fusion attention mechanism of an artificial intelligence-based auxiliary diagnosis and treatment system of the present invention; Figure 4 This is a flowchart of the federated learning training of an artificial intelligence-based auxiliary diagnosis and treatment system of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] See also Figure 1-4As shown, an artificial intelligence-based auxiliary diagnosis and treatment system includes: a multimodal data acquisition module, a dynamic learning module, a diagnostic reasoning module, a privacy protection module, an interactive decision-making module and an early warning monitoring module; The multimodal data acquisition module is used to collect the patient's medical imaging data, electronic medical record text data, gene sequencing data and physiological signal data in real time, and standardize the data into a unified format; The privacy protection module is used to desensitize, encrypt, and distribute multimodal data, supporting data processing based on homomorphic encryption and model training under the federated learning framework; The dynamic learning module is used to update the diagnostic model parameters in real time through an incremental learning algorithm, and optimize the model's generalization ability by combining the medical knowledge graph and the historical case database; The diagnostic reasoning module is used to generate preliminary diagnostic conclusions based on multimodal data fusion and optimize diagnostic confidence through Bayesian networks and deep reinforcement learning algorithms; The interactive decision-making module is used to provide a visual diagnosis and treatment suggestion interface, support doctors to correct and provide feedback on AI diagnosis results, and input feedback data into the dynamic learning module to complete model iteration; The early warning monitoring module is used to analyze the patient's physiological indicators and abnormal signals in the diagnosis and treatment process in real time, trigger graded early warnings and push them to the doctor's terminal: Among them, the output end of the multimodal data acquisition module is connected to the input end of the privacy protection module, the output end of the privacy protection module is connected to the input end of the dynamic learning module, the output end of the dynamic learning module is connected to the input end of the diagnostic reasoning module, the output end of the diagnostic reasoning module is connected to the input end of the interactive decision-making module, and the early warning monitoring module and the diagnostic reasoning module have two-way data interaction.

[0019] Example 1: Implementation details of the multimodal data acquisition module The multimodal data acquisition module includes an image analysis unit, a text analysis unit and a signal processing unit; The image analysis unit is responsible for 3D reconstruction and lesion feature extraction of medical images (such as CT, MRI, and ultrasound). The specific implementation steps are as follows: 1) Data preprocessing: Standardize the original DICOM format images, including adjusting the window width and window position (e.g., lung CT is set to a window width of 1500HU and a window position of -600HU), spatial resampling (resolution is uniformly set to 1mm³), and grayscale normalization (range [-1, 1]). An improved U-Net-based network (3DRes-U-Net) is used for lesion segmentation. The network structure is as follows: Figure 2As shown: Encoder: Contains 5 residual blocks, each block contains 2 3×3×3 convolution layers, the number of channels is 32, 64, 128, 256, 512, and the downsampling operation is a step of 2. Decoder: Upsampling through transposed convolution, each layer is skipped with the corresponding layer of the encoder, and the number of output channels is reduced to 256, 128, 64, and 32. Output layer: 1×1×1 convolution and Sigmoid activation function are used to generate the lesion probability map, and the threshold is set to 0.5 to determine the positive area. Training details: Dataset: From the public LUNA16 lung nodule dataset (888 CT scans) and 1,000 liver MRI data from an internal hospital, the training set, validation set, and test set are divided into 8:1:1. Loss function: Dice loss + Focal Loss, the formula is: ,in, is the predicted probability, The ground-truth label is represented by the training set, α = 0.25, and γ = 2. The optimizer uses Adam, with an initial learning rate of 0.001 and a 50% decay every 20 epochs. Performance metrics: The Dice coefficient for lung nodule segmentation reaches 98.2%, and the Dice coefficient for liver tumor segmentation reaches 96.8%, outperforming the traditional U-Net (94.1%).

[0020] 2) Feature extraction: Extract morphological features (such as volume, surface area, sphericity), texture features (based on contrast and energy of gray-level co-occurrence matrix), and dynamic enhancement features (such as the slope of the time-density curve of CT perfusion imaging) of the lesion from the segmentation results. Feature vector dimension: a total of 128 dimensions, expressed as .

[0021] The text analysis unit is used to parse the unstructured text data in the electronic medical record. The key steps are as follows: 1) Entity Recognition and Relation Extraction: A pre-trained BioBERT model (fine-tuned on PubMed abstracts and MIMIC-III medical records) was used for named entity recognition (NER). It identified entities such as symptoms (e.g., "cough"), diseases (e.g., "pneumonia"), and medications (e.g., "amoxicillin"). Entity types were mapped to ICD-10 codes (e.g., "J18.9" corresponds to "pneumonia, pathogen unspecified"). Dependency parsing was used to extract relationships between entities, such as "cough → persists for 3 days → worsens." Performance indicators: Entity recognition F1 score reached 92.5%, and relation extraction accuracy reached 87.3%.

[0022] 2) Text vectorization: Use Sentence-BERT to encode medical record text into a 768-dimensional vector , and focuses on key descriptions through attention pooling.

[0023] The signal processing unit performs real-time analysis on physiological signals (such as electrocardiogram, electroencephalogram, and blood oxygen): 1) ECG signal processing and noise reduction: Wavelet transform (Daubechies6 wavelet) is used to remove baseline drift and power frequency interference, improving the signal-to-noise ratio to 35dB. Feature extraction: RR interval variation (RMSSD): Calculates the standard deviation of the interval between adjacent R waves, reflecting autonomic nervous system function. ST segment slope: Identifies myocardial ischemia by linearly fitting the voltage change of the ST segment (80ms after the J point). An LSTM network is used to predict arrhythmia type (such as atrial fibrillation and premature ventricular contractions). The input is a 10-second ECG segment (sampling rate 250Hz), and the output is a five-category probability distribution. Performance: Atrial fibrillation detection sensitivity is 98.7% and specificity is 99.2%.

[0024] 2) Blood oxygen signal processing: Calculate blood oxygen saturation based on photoplethysmography (PPG) ( ), a moving average filter (window length 5 seconds) is used to eliminate motion artifacts. The feature vector Vbio∈R64 includes Mean, minimum value, rate of decline, etc.

[0025] Example 2: Implementation details of the privacy protection module The privacy protection module includes data desensitization and encryption and a federated learning framework.

[0026] Among them, data desensitization and encryption include differential privacy processing part, homomorphic encryption part and computational time testing part.

[0027] 1) Differential privacy processing: Perform k-anonymization (k=5) on patient identity information (name, ID number), ensuring that each record is indistinguishable from at least four other records in terms of quasi-identifiers (such as age, gender, and zip code). Laplace noise is added during the feature calculation phase, and the amount of noise is determined by the privacy budget. Control (Default =1)0), satisfying: ,in =10-5.

[0028] 2) Homomorphic encryption: The Paillier algorithm is used to encrypt feature vectors, supporting addition and scalar multiplication operations in the ciphertext state. Example of feature calculation after encryption: Computational time test: Single encryption takes 15ms, and decryption takes 8ms (Intel Xeon Gold 6248R processor).

[0029] Among them, the federated learning framework includes system architecture part, training process part and performance comparison part.

[0030] System Architecture: Central Server: Responsible for aggregating model parameters of each node and generating a global model. Hospital Node: Locally stores data, trains the model and uploads encrypted gradients. Training Process: Step 1: Central Server Initializes the Global Model , distributed to each node. Step 2: Each node uses local data Di to calculate the gradient , and encrypted as Step 3: Central server aggregates gradients , update the global model after decryption: Performance comparison : The AUC of the lung cancer diagnosis model under federated learning is 0.973, and that under centralized training is 0.982, with a difference of only 0.009.

[0031] Example 3: Implementation details of the dynamic learning module The dynamic learning module includes incremental learning algorithm and knowledge graph verification.

[0032] Among them, the incremental learning algorithm includes the online random forest part, the node splitting standard part, the elastic weight consolidation (Elastic Weight Consolidation, EWC) part and the effect verification part.

[0033] Online random forest: A new decision tree is generated for every 100 new cases received, and the weight of the historical tree decays exponentially (decay coefficient λ=0.9). Node splitting criteria: Based on Impurity is dynamically adjusted, the formula is: ,in For category Elastic weight solidification: Calculate historical data parameters Fisher information matrix , constraining the new parameters Update: . Effect verification: EWC uses the model to verify the new variant data of COVID-19. The value increased from 68.4% to 91.1%.

[0034] Knowledge graph verification involves constructing a medical knowledge graph containing 120,000 entities (diseases, symptoms, and medications) and 2.3 million relationships (e.g., "pneumonia - causes - fever"). If the incremental learning results conflict with the knowledge graph (e.g., "diagnosed with pneumonia despite no cough symptoms"), a manual review process is triggered, with an 82% approval rate.

[0035] Example 4: Implementation details of the diagnostic reasoning module The diagnostic reasoning module includes multimodal data fusion and confidence optimization.

[0036] Among them, multimodal data fusion includes the attention mechanism weighted part, the weight coefficient calculation part, the fusion feature vector part, the structure part and the training data part.

[0037] Attention Mechanism Weighting: Defining the Trainable Parameter Matrix 、 、 . Calculate the weight coefficient: ,in =64 is the scaling factor. Fusion feature vector: Multi-layer Perceptron (MLP): Structure: Input layer (256 dimensions) → Fully connected layer (128 dimensions, ReLU) → Dropout layer (rate = 0.3) → Output layer (number of disease categories, Softmax). Training data: The fusion feature achieved an accuracy of 96.3% in the lung cancer diagnosis task (test set N = 2000).

[0038] Among them, confidence optimization includes the Bayesian network calculation diagnosis uncertainty part.

[0039] Bayesian networks compute diagnostic uncertainty: ,in, is the prior probability (from the historical case database), Output from the multimodal fusion model. If the confidence level is less than 85%, the system prompts the doctor to review and initiates reinforcement learning to adjust the model parameters.

[0040] Example 5: Implementation details of the interactive decision module The interactive decision-making module includes a visual interface design and a feedback mechanism.

[0041] Among them, the visual interface design includes the diagnostic basis heat map part and the treatment recommendation reasoning chain part.

[0042] 1) Diagnosis based on heatmap: Gradient-weighted class activation map (Grad-CAM) is calculated for each pixel of the CT image. The formula is: ,in For the feature maps, For category Feature Map Example: In lung cancer diagnosis, the heat icon Shows the spiculation area at the edge of the lung nodule (contribution accounted for 62%).

[0043] 2) Reasoning chain of treatment recommendations: Use a tree diagram to show the reasoning path, such as: fever (38.5℃) → elevated white blood cell count (15×10 9 / L) → Ground-glass opacity in the lungs → Bacterial pneumonia (87% probability) → Ceftriaxone (2g every 12 hours) is recommended. Click a node to view the source of evidence (e.g., "elevated white blood cell count" linked to an electronic medical record and laboratory report).

[0044] The feedback mechanism stores doctor revision records as a triplet: "original diagnosis, revised diagnosis, reason for revision." For example, "AI diagnosis: gastric ulcer, doctor revision: gastric cancer, reason: biopsy pathology suggests atypical cells." This revised data is used to generate incremental learning samples, with a 24-hour model iteration cycle.

[0045] Example 6: Implementation details of the early warning monitoring module The early warning monitoring module includes a hierarchical early warning strategy and an anomaly detection algorithm.

[0046] The graded warning strategy includes level one warning, level two warning and level three warning.

[0047] 1) Level 1 Warning (Red Alert): Triggering conditions: Blood oxygen saturation < 90% for 5 minutes or heart rate > 140 bpm for 10 minutes. Response: Push alert to the attending physician's mobile app and automatically call the nursing station.

[0048] 2) Level 2 Warning (Yellow Alert): Triggering conditions: Exceeding the drug dosage limit (e.g., ceftriaxone >4g / day) or incompatible drug combination (e.g., warfarin combined with aspirin). Response: Locking the prescription submission function and displaying a pop-up window suggesting an alternative (e.g., switching to rivaroxaban).

[0049] 3) Level 3 Warning (Blue Alert): Triggering condition: AI diagnosis confidence <85% and no doctor review within 30 minutes. Response: Automatically initiate a multidisciplinary consultation request and mark the case as high-risk.

[0050] The anomaly detection algorithm uses an isolation forest model to detect abnormal physiological indicators. This algorithm constructs a random tree, where shorter paths indicate a higher probability of an anomaly. The area under the receiver operating characteristic (ROC) curve (AUC) for detecting sudden drops in blood oxygen levels reached 0.974.

[0051] Example 7: Implementation details of the model interpretation module The model explanation module includes LIME local explanations and counterfactual reasoning reports.

[0052] Among them, LIME local explanation: for the image classification results, randomly perturb the input pixels, and fit the linear model to explain the importance of local features: ,in is the perturbation sample, is the feature weight. Example: In pneumonia diagnosis, the weight of the right upper lobe infiltrate is 45%.

[0053] The counterfactual reasoning report includes the generation of hypothetical scenarios and the optimization target. Generate hypothetical scenarios: "If the patient has no cough symptoms, how will the diagnosis probability change?" Calculation method: Find the counterfactual point ' closest to the original data x in the feature space ,satisfy ≠ Optimization goal: , output report Example: "When there is no cough, the probability of bacterial pneumonia drops from 87% to 52%, and it is recommended to check for tuberculosis."

[0054] Example 8: System deployment and testing The system deployment and testing includes a hardware configuration part and a performance testing part.

[0055] The hardware configuration includes the server, GPU, memory, storage, tablet, and display resolution. Server: CPU: Intel Xeon Gold 6248R (40 cores), GPU: NVIDIA A100 (80GB of video memory); Memory: 512GB of DDR4; Storage: 10TB of NVMe SSD (encrypted partition); Doctor Terminal: Tablet: iPad Pro (M1 chip, 5G module); Display resolution: 2732×2048, with stylus annotation support.

[0056] Among them, the performance test includes real-time part, stability part and clinical verification part.

[0057] 1) Real-time performance: The average latency from data acquisition to generating diagnostic recommendations is 2.8 seconds (1.5 seconds for CT image analysis, 0.7 seconds for text processing, and 0.6 seconds for fusion reasoning).

[0058] 2) Stability: Continuous operation for 72 hours without any problems, with CPU / GPU load peaks of 78% / 92% respectively.

[0059] 3) Clinical validation: A double-blind trial was conducted on 300 lung cancer patients in a tertiary hospital. The diagnostic consistency rate between AI and the expert group was 94.6%, and the Kappa coefficient was 0.91.

[0060] Example 9: Comparative Experiment and Effect Verification Comparative experiments and effect verification include the comparison between multimodal fusion and single modality, the comparison between federated learning and centralized learning, and the incremental learning effect.

[0061] The specific results of the comparison between multimodal fusion and single modality are shown in Table 1 below: Table 1: Comparison between multimodal fusion and single modality The specific results of the comparison between federated learning and centralized learning are shown in Table 2 below: Table 2: Comparison of Federated Learning and Centralized Learning The specific results of incremental learning effects are shown in Table 3 below: Table 3: Incremental learning effect The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art will be able to modify the technical solutions described in the foregoing embodiments or to substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0062] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0063] In the description of the present invention, it should be understood that the terms "first", "second", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0064] In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0065] In the description of the present invention, “several” means one or more, and “a large number” means two or more.

[0066] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0067] The formulas in this manual are all dimensionless and calculated using numerical values. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field based on actual conditions.

[0068] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. An artificial intelligence-based auxiliary diagnosis and treatment system, characterized in that: The system includes: a multimodal data acquisition module, a dynamic learning module, a diagnostic reasoning module, a privacy protection module, an interactive decision-making module and an early warning monitoring module; The multimodal data acquisition module is used to collect the patient's medical imaging data, electronic medical record text data, gene sequencing data and physiological signal data in real time, and standardize the data into a unified format; The privacy protection module is used to desensitize, encrypt, and distribute multimodal data, supporting data processing based on homomorphic encryption and model training under the federated learning framework; The dynamic learning module is used to update the diagnostic model parameters in real time through an incremental learning algorithm, and optimize the model's generalization ability by combining the medical knowledge graph and the historical case database; The diagnostic reasoning module is used to generate preliminary diagnostic conclusions based on multimodal data fusion and optimize diagnostic confidence through Bayesian networks and deep reinforcement learning algorithms; The interactive decision-making module is used to provide a visual diagnosis and treatment suggestion interface, support doctors to correct and provide feedback on AI diagnosis results, and input feedback data into the dynamic learning module to complete model iteration; The early warning monitoring module is used to analyze the patient's physiological indicators and abnormal signals in the diagnosis and treatment process in real time, trigger graded early warnings and push them to the doctor's terminal: Among them, the output end of the multimodal data acquisition module is connected to the input end of the privacy protection module, the output end of the privacy protection module is connected to the input end of the dynamic learning module, the output end of the dynamic learning module is connected to the input end of the diagnostic reasoning module, the output end of the diagnostic reasoning module is connected to the input end of the interactive decision-making module, and the early warning monitoring module and the diagnostic reasoning module have two-way data interaction.

2. The artificial intelligence-based auxiliary diagnosis and treatment system according to claim 1, characterized in that: The multimodal data acquisition module includes an image analysis unit, a text analysis unit and a signal processing unit; The image analysis unit is used to perform three-dimensional reconstruction and lesion feature extraction on CT, MRI, and ultrasound images, and uses an improved U-Net network to segment key anatomical structures; The text analysis unit is used to extract symptom descriptions, medication records, and medical history keywords in electronic medical records through the BERT model and associate them with ICD-10 disease codes; The signal processing unit is used to perform noise reduction and time-frequency domain feature extraction on the electrocardiogram, electroencephalogram, and blood oxygen saturation signals, and eliminate baseline drift in combination with wavelet transformation.

3. The artificial intelligence-based auxiliary diagnosis and treatment system according to claim 1, characterized in that: The privacy protection module uses a differential privacy algorithm to desensitize patient identity information and uses homomorphic encryption technology to realize feature calculation of data in an encrypted state, meeting GDPR and HIPAA compliance requirements; the privacy protection module supports model training under the federated learning framework. The local data of each medical node does not leave the original storage location, and only the encrypted model gradient parameters are uploaded.

4. The artificial intelligence-based auxiliary diagnosis and treatment system according to claim 1, characterized in that: The incremental learning algorithm of the dynamic learning module includes the following steps: Step 1: Based on the online random forest algorithm, the decision tree branch weights are dynamically adjusted according to the newly input case data; Step 2: Use elastic weight solidification technology to retain the key feature weights of historical data to prevent catastrophic forgetting; Step 3: Combine the disease-symptom-treatment relationship in the medical knowledge graph to logically verify the incremental learning results.

5. The artificial intelligence-based auxiliary diagnosis and treatment system according to claim 1, characterized in that: The multimodal data fusion method of the diagnostic reasoning module is: setting the medical image feature vector to , the electronic medical record text feature vector is , the physiological signal feature vector is ; Calculate the weight coefficient of each modality through the attention mechanism ,in is a trainable parameter; Fusion Eigenvector , input into the multi-layer perceptron to generate the diagnosis probability distribution.

6. The artificial intelligence-based auxiliary diagnosis and treatment system according to claim 1, characterized in that: The visualization interface of the interactive decision-making module includes: a diagnostic basis heat map, which shows the contribution of the lesion area in the image to the diagnostic conclusion; a treatment recommendation reasoning chain, which presents the key reasoning path from symptoms to treatment plans in the form of a tree diagram; and a doctor correction record unit, which records the doctor's modifications to the AI ​​suggestions and marks the reasons for the corrections.

7. The artificial intelligence-based auxiliary diagnosis and treatment system according to claim 1, characterized in that: The hierarchical early warning strategy of the early warning monitoring module is: Level 1 warning: If the patient's physiological indicators exceed the preset safety threshold (such as blood oxygen saturation <90%), a red alert will be triggered and sent to the attending physician's mobile terminal; Level 2 warning: If there is a deviation from the operating procedures in the diagnosis and treatment process (such as drug incompatibility), a yellow alarm will be triggered and a pop-up window will be displayed on the system interface; Level 3 warning: The AI ​​diagnosis confidence level is lower than the set threshold (e.g., <85%), triggering a blue alert and recommending the initiation of a multidisciplinary consultation process.

8. The artificial intelligence-based auxiliary diagnosis and treatment system according to claim 1, characterized in that: The system also includes a model interpretation module for generating an explainable report of the diagnostic conclusion, specifically including: performing local feature importance analysis on image classification results based on the LIME algorithm; generating a comparative report on "changes in diagnostic probability if a certain symptom does not exist" through counterfactual reasoning; and performing compliance verification of AI-recommended treatment plans based on clinical guidelines.

Images