Chronic disease evaluation system and method based on weak AI medical large model platform

Abstract

The invention discloses a chronic disease evaluation system and method based on a weak AI medical large model platform, and particularly relates to the technical field of medical artificial intelligence. The system constructs a multi-modal input matrix and generates fusion features by collecting multi-modal data such as electronic health records, wearable device time sequence signals and medical images; calculating a risk index by using cooperative work of a cloud large model and an edge end lightweight model and combining a current state and a historical trend; an early warning threshold value is dynamically adjusted according to the disease type and historical risk data, and risk early warning is achieved; key features are extracted, similar cases are matched, and an interpretability report is generated; and updating edge end model parameters through federal learning to protect data privacy. The problems of data splitting, static assessment, resource dependence and the like in a traditional chronic disease assessment system are solved, comprehensive, dynamic and accurate assessment of chronic diseases is achieved, and the method has high clinical application value.

Description

Technical Field

[0001] The present invention relates to the field of medical artificial intelligence technology, and more specifically, to a system and method for evaluating chronic diseases based on a weak AI medical large model platform. Background Art

[0002] Chronic diseases are non-communicable diseases with insidious onset, a long course, and persistent symptoms, such as hypertension, diabetes, coronary heart disease, and chronic obstructive pulmonary disease. With the aging population and changing lifestyles, chronic diseases have become a major global health burden. According to the World Health Organization, chronic diseases account for over 70% of all deaths worldwide, severely impacting patients' quality of life and imposing a significant medical and economic burden. The management of chronic diseases requires long-term, continuous health monitoring and assessment to achieve early intervention and slow disease progression.

[0003] The current chronic disease assessment system faces many challenges in practical application, which are mainly reflected in the following aspects: Electronic medical records, data collected by wearable devices, and imaging reports are different types of heterogeneous data. Traditional systems find it difficult to effectively integrate these data, resulting in incomplete evaluation results that cannot fully reflect the patient's overall condition.

[0004] Traditional models mainly rely on single diagnostic results, lack the ability to track patients' conditions in the long term, find it difficult to capture subtle changes in the development of chronic diseases, and are unable to provide patients with timely and accurate assessments and intervention recommendations.

[0005] High-precision evaluation models require massive amounts of labeled data and powerful computing support, which makes it difficult to promote and apply traditional systems in primary medical institutions, limiting the popularization and development of chronic disease management.

[0006] Although existing large models have multimodal understanding capabilities, when directly applied to the medical field, they have problems such as high computational overhead, poor interpretability, and insufficient domain knowledge, and cannot meet the medical industry's strict requirements for accuracy, interpretability, and real-time performance.

[0007] To address the above issues, a chronic disease assessment system and method based on a weak AI medical big model platform is proposed. Summary of the Invention

[0008] In order to overcome the above-mentioned defects of the prior art, the present invention provides a system and method for evaluating chronic diseases based on a weak AI medical large model platform to solve the problems raised in the above-mentioned background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solutions: a chronic disease assessment system based on a weak AI medical large model platform, comprising: (a) Multimodal data acquisition module: Obtaining electronic health record (EHR) text feature vectors , wearable device timing signals , medical imaging features ; (b) Cloud-edge-device collaborative architecture: A large vertical model engine deployed in the cloud outputs the prediction results of the teacher model ; A lightweight evaluation model deployed on the edge to output the student model prediction results ; (c) Dynamic risk assessment module, generating risk index : in, ; in, , Expressed as The multimodal input matrix at time t; is the current state credibility, For the importance of historical trends, is the risk drift, that is, the data distribution at the current moment Relative to the patient's baseline health status The degree of abnormal deviation, is the state memory decay rate; For spatiotemporal memory networks The state matrix at time t, The patient's baseline status; Expressed as the Kullback-Leibler divergence operator; Represented as a long short-term memory neural network; (d) Early warning trigger module: When When an early warning signal is generated, is the warning threshold; (e) Explainability report module: output decision-making feature set ,in, For the characteristic variables, The risk index is The gradient of a feature, is the gradient importance threshold.

[0010] Preferably, the multimodal input matrix Generating fusion features through cross-modal alignment : in, Indicates modality The feature extraction function of Including EHR electronic health records, TS time series signals, and IMG medical images; is a modal set, that is ; For the The weight of each mode; Indicates the The original eigenvectors of the modes; represents a learnable cross-modal alignment matrix used to map different modality features into a unified space; represents the dynamic weight query vector; For the The original eigenvectors of the modes.

[0011] Preferably, the lightweight evaluation model is trained by knowledge distillation, and the loss function of knowledge distillation is: in, is the knowledge distillation loss function, which is used to train the student model to imitate the teacher model; and Represents the logits vectors of the teacher and student models respectively; is the balance coefficient, which is used to adjust the weight between the two terms in the loss function; The temperature coefficient is used in knowledge distillation to “soften” the output of the teacher model by increasing the temperature coefficient, so that the student model can learn richer relationships between categories. Represents the softmax function, for the input vector , No. The softmax value of each element is: , The output vector elements, is an exponential function used to calculate probability weights; represents the cross entropy, which is used to measure the difference between two probability distributions and is defined as: in, is the probability distribution after softening of the teacher model, is the probability distribution after the student model is softened, For the teacher model The softening probability of the class, For students to model Softened probability of the class.

[0012] Preferably, the lightweight evaluation model deployed at the edge updates parameters through federated learning : in, represents the model parameters for the next iteration, represents the current model parameters, is the learning rate, For the A collection of participating medical institutions, is Gaussian noise, For the The gradient of the loss function of each institution, is the noise intensity.

[0013] Preferably, the warning threshold The dynamic calculation formula is: in, is the basic threshold value related to the disease type, is the risk volatility sensitive factor, is the history window length, For historical moments The risk index, It is the average of historical risk indices, used to reflect the long-term risk level.

[0014] Preferably, the feature extraction function for: in, , is the wavelet basis function; is the scale parameter; It is a feature extraction function for time series signals, implemented based on wavelet transform; Timing signals collected for wearable devices; is a wavelet transform function used to extract the time-frequency characteristics of time series signals; is a time variable, used to describe the shape of the wavelet basis function in the time domain; It is a translation parameter that controls the position of the wavelet on the time axis and is used to locate the local features of the signal.

[0015] Preferably, the similar case matching of the explainability reporting module satisfies: in, represents the case similarity score, represents the latent vector of the current patient case, Represents the latent vector of the case in the historical case database, Represents the laboratory index vector of the current patient, A vector of laboratory indicators representing historical cases.

[0016] Preferably, the spatiotemporal memory network update satisfies , For the forgetting factor.

[0017] Preferably, the medical imaging feature comprises a retinal OCT image, and ,in, ResNet50 convolutional neural network. Represents retinal OCT images.

[0018] The evaluation method for the chronic disease assessment system based on the weak AI medical large model platform includes the following steps: S1: Collect patient electronic health records, wearable device timing signals, and medical images to generate feature vectors respectively; S2: Construct a multimodal input matrix and generate fusion features through cross-modal alignment; S3: A large vertical model with a parameter size of ≥70B is deployed on the cloud as the teacher model, and a lightweight model with a parameter size of ≤1B is deployed on the edge as the student model, and the prediction results are output separately. S4: Calculate the risk index of the output prediction result by combining the current credibility and the importance of historical trends; S5: Set a basic threshold based on the disease type, and dynamically adjust the warning threshold based on the fluctuation range of historical risk data. When the risk index exceeds the warning threshold, generate a warning signal; S6: Extract features whose gradient importance is greater than or equal to the threshold to form a decision basis set, and calculate the matching degree of historical similar cases by multiplying the cosine similarity of the case latent vector and the laboratory indicator distance, and generate an explanation report; S7: The edge model updates parameters through the federated learning algorithm, adds Gaussian noise to protect privacy data, and achieves distributed optimization of the model.

[0019] Technical effects and advantages of the present invention: 1. The multimodal data acquisition module integrates heterogeneous data such as electronic health records, wearable device time series signals, and medical images, and generates fusion features through cross-modal alignment, effectively solving the problem of data fragmentation in traditional systems and making the evaluation results more comprehensive and accurate in reflecting the patient's condition.

[0020] 2. The dynamic risk assessment module combines the current state credibility and the importance of historical trends to calculate the risk index. At the same time, the spatiotemporal memory network is used to update the state matrix, which can capture the dynamic changes of the disease, overcome the shortcomings of the static assessment of traditional models, and provide strong support for the long-term management of chronic diseases.

[0021] 3. By deploying a high-precision, large-scale model for a specific vertical domain in the cloud as a teacher model and a lightweight student model on the edge, knowledge distillation is used to enable the edge model to acquire knowledge from the large cloud model. Federated learning is also used to achieve distributed optimization of the edge model. This architecture reduces the reliance on massive amounts of labeled data and powerful computing power, enabling the system to be widely adopted in primary healthcare institutions.

[0022] 4. By setting the basic threshold according to the disease type through the early warning threshold and dynamically adjusting it based on the fluctuation range of historical risk data, it can adapt to the characteristics of different patients and diseases, reduce false alarms and missed alarms, and improve the accuracy and practicality of early warnings.

[0023] 5. The interpretable reporting module extracts features that have a significant impact on decision-making and matches historical similar cases, providing doctors with a clear basis for decision-making, improving the interpretability and credibility of the model, and making it easier for doctors to understand and apply the evaluation results.

[0024] 6. The edge model updates parameters through federated learning, and Gaussian noise is added to protect privacy data. The model is optimized without sharing the original data, meeting the requirements of medical data security. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0026] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0027] The chronic disease assessment system based on the weak AI medical large model platform of the present invention includes: (a) Multimodal data acquisition module: Obtaining electronic health record (EHR) text feature vectors , wearable device timing signals , medical imaging features ; (b) Cloud-edge-device collaborative architecture: A large vertical model engine deployed in the cloud outputs the prediction results of the teacher model ; A lightweight evaluation model deployed on the edge to output the student model prediction results ; (c) Dynamic risk assessment module, generating risk index : in, ; in, , Expressed as The multimodal input matrix at time t; is the current state credibility, For the importance of historical trends, is the risk drift, that is, the data distribution at the current moment Relative to the patient's baseline health status The degree of abnormal deviation, is the state memory decay rate; For spatiotemporal memory networks The state matrix at time t, The patient's baseline status; Expressed as the Kullback-Leibler divergence operator; Represented as a long short-term memory neural network; (d) Early warning trigger module: When When an early warning signal is generated, is the warning threshold; (e) Explainability report module: output decision-making feature set ,in, For the characteristic variables, The risk index is The gradient of a feature, is the gradient importance threshold.

[0028] In the specific implementation, a multimodal data acquisition module is deployed to obtain electronic health records (EHR) from the hospital information system and generate feature vectors by parsing the text. , collect timing signals from wearable devices such as smart watches and bracelets, and generate , process medical images such as CT, MRI, etc., and generate ; Build a cloud-edge-device collaborative architecture, deploy large model engines for vertical fields on the cloud, such as a large medical model with a parameter scale of ≥70B as a teacher model, and deploy a lightweight evaluation model with a parameter scale of ≤1B on the edge as a student model; Construct a dynamic risk assessment module, according to the formula Calculate the risk index, where , The KL divergence is used to calculate the abnormal deviation of the current data distribution relative to the patient's baseline state. Through LSTM and state memory decay rate renew; Set the warning trigger module, when Generate early warning signals when Develop an interpretable reporting module to output gradient importance The feature set .

[0029] Specifically, the multimodal data acquisition module uses natural language processing, signal processing and computer vision technologies to extract features from different data sources respectively; the cloud-edge collaborative architecture provides high-precision predictions through large cloud models, and lightweight edge models achieve localized fast inference, reducing latency and computing power requirements; the dynamic risk assessment module combines current prediction results and historical data trends, quantifies the degree of data deviation through KL divergence, and captures time series features through LSTM to achieve dynamic assessment of chronic disease risks; the early warning trigger module implements risk warnings based on set thresholds; the explainable report module extracts features that have an important impact on decision-making through gradient analysis, thereby improving the interpretability of the model; thereby realizing the fusion processing of multi-source heterogeneous data and improving the comprehensiveness of the assessment. The cloud-edge collaborative architecture enables the system to be deployed in different computing power environments, improving the applicability and scalability of the system, and dynamic risk assessment can timely capture changes in the disease, improving the accuracy and timeliness of the assessment. The explainable report provides doctors with a basis for decision-making and improves the credibility and clinical practicality of the model.

[0030] The multimodal input matrix Generating fusion features through cross-modal alignment : in, Indicates modality The feature extraction function of Including EHR electronic health records, TS time series signals, and IMG medical images; is a modal set, that is ; For the The weight of each mode; Indicates the The original eigenvectors of the modes; represents a learnable cross-modal alignment matrix used to map different modality features into a unified space; represents the dynamic weight query vector; For the The original eigenvectors of the modes.

[0031] In specific implementation, construct a multimodal input matrix , define the feature extraction function , use BERT and other models to extract features for EHR text, use wavelet transform and other methods to extract features for TS time series signals, use ResNet and other convolutional neural networks to extract features for IMG medical images, and define cross-modal alignment matrix and dynamic weight query vector , calculate the weight of each modal feature , and generate fusion features by weighted summation .

[0032] Specifically, cross-modal alignment is achieved through a learnable matrix Mapping features of different modalities to the same feature space, dynamic weight query vector The weights of each modality are dynamically adjusted according to the input data to achieve effective feature fusion. The weighted summation method can integrate the information of each modality and improve the expressiveness of the features, thereby achieving effective fusion of multimodal features and improving the model's ability to understand complex diseases. Dynamic weight adjustment can automatically assign the importance of each modality according to different input data, thereby improving the adaptability of the model.

[0033] The lightweight evaluation model is trained through knowledge distillation, and the loss function of knowledge distillation is: in, is the knowledge distillation loss function, which is used to train the student model to imitate the teacher model; and Represents the logits vectors of the teacher and student models respectively; is the balance coefficient, which is used to adjust the weight between the two terms in the loss function; The temperature coefficient is used in knowledge distillation to “soften” the output of the teacher model by increasing the temperature coefficient, so that the student model can learn richer relationships between categories. Represents the softmax function, for the input vector , No. The softmax value of each element is: , The output vector elements, is an exponential function used to calculate probability weights; represents the cross entropy, which is used to measure the difference between two probability distributions and is defined as: in, is the probability distribution after softening of the teacher model, is the probability distribution after the student model is softened, For the teacher model The softening probability of the class, For students to model Softened probability of the class.

[0034] In specific implementation, define the logits vectors of the teacher model and the student model and , set the balance coefficient and temperature coefficient , calculate the loss function of knowledge distillation , use this loss function to train the student model so that the student model can learn the knowledge of the teacher model.

[0035] Specifically, knowledge distillation transfers complex large model knowledge to a lightweight student model by allowing the student model to learn the output distribution of the teacher model. The mean square error term ensures that the output of the student model is close to that of the teacher model. The cross entropy term softens the output of the teacher model, allowing the student model to learn richer relationships between categories. The temperature coefficient The adjustment can control the degree of softening and plays an important role in the training process, so that the lightweight model on the edge can obtain the knowledge of the large model on the cloud, improve the prediction accuracy of the model, reduce the parameter scale of the model, and increase the inference speed of the model, so that it can be deployed on edge devices.

[0036] The lightweight evaluation model deployed on the edge updates parameters through federated learning : in, represents the model parameters for the next iteration, represents the current model parameters, is the learning rate, For the A collection of participating medical institutions, is Gaussian noise, For the The gradient of the loss function of each institution, is the noise intensity.

[0037] During specific implementation, determine the set of medical institutions participating in federated learning , each medical institution uses local data to calculate the gradient of the loss function , the edge model is based on the formula Update the parameters, adding Gaussian noise Protect private data.

[0038] Specifically, federated learning achieves collaborative optimization of the model through distributed training without sharing the original data. The addition of Gaussian noise implements differential privacy protection and ensures the privacy security of the data. The gradient averaging method can integrate the training results of various medical institutions and improve the generalization ability of the model, thereby realizing distributed updates of the model, protecting the privacy data of medical institutions, enabling the model to be trained on data from different medical institutions, and improving the adaptability and generalization ability of the model.

[0039] The warning threshold The dynamic calculation formula is: in, is the basic threshold value related to the disease type, is the risk volatility sensitive factor, is the history window length, For historical moments The risk index, It is the average of historical risk indices, which is used to reflect the long-term risk level.

[0040] In specific implementation, the basic threshold is set according to the disease type , collect history window length Risk Index , calculate the historical average risk index and risk fluctuation range, according to the formula Dynamically adjust warning thresholds.

[0041] Specifically, the basic threshold According to the characteristics of the disease and clinical needs, it ensures that the risk assessment of different diseases is targeted; the dynamic adjustment part reflects the degree of risk fluctuation and the risk fluctuation sensitivity factor by calculating the standard deviation of historical risk data. The sensitivity to fluctuations can be adjusted according to actual needs; this dynamic adjustment method can adapt to changes in the conditions of different patients and the risk characteristics of different diseases, so that the warning threshold can be dynamically adjusted according to the patient's historical condition and disease characteristics, improving the accuracy and adaptability of the warning, reducing the occurrence of false alarms and missed alarms, and improving the practicality of the system.

[0042] The feature extraction function for: in, , is the wavelet basis function; is the scale parameter; It is a feature extraction function for time series signals, implemented based on wavelet transform; Timing signals collected for wearable devices; is a wavelet transform function used to extract the time-frequency characteristics of time series signals; is a time variable, used to describe the shape of the wavelet basis function in the time domain; It is a translation parameter that controls the position of the wavelet on the time axis and is used to locate the local features of the signal.

[0043] In specific implementation, the timing signals collected by the wearable device Perform wavelet transform and select appropriate wavelet basis function and scale parameters , according to the formula Extract features of time series signals.

[0044] Specifically, wavelet transform can decompose time series signals into components of different frequencies and capture the local characteristics and time-frequency characteristics of the signal; by selecting appropriate wavelet basis functions and scale parameters, useful information in time series signals can be better extracted, thereby effectively extracting features in time series signals, such as heart rate variability, respiratory rate, etc., providing an important basis for chronic disease assessment, improving the processing ability of time series signals, and enhancing the model's ability to capture changes in disease conditions.

[0045] The similar case matching of the explainability reporting module satisfies: in, represents the case similarity score, represents the latent vector of the current patient case, Represents the latent vector of the case in the historical case database, Represents the laboratory index vector of the current patient, A vector of laboratory indicators representing historical cases.

[0046] In specific implementation, the latent vector of the current patient case is extracted and laboratory indicator vector , extract the latent vector of the case from the historical case database and laboratory indicator vector , according to the formula Calculate case similarity scores and match similar cases.

[0047] Specifically, the cosine similarity of latent vectors measures the similarity of cases in the overall feature space, reflecting the degree of overall feature matching of cases. The L1 distance of laboratory indicators measures the difference in specific indicators. The exponential decay function converts the distance into the weight of similarity. The product of the two comprehensively considers the similarity of overall characteristics and specific indicators, improves the accuracy of case matching, and can accurately match historical similar cases, provide reference for doctors, assist in diagnosis and treatment decision-making, improve the practicality of interpretable reports, and enable doctors to better understand the decision-making basis of the model.

[0048] The spatiotemporal memory network updates satisfy , For the forgetting factor.

[0049] When implementing, set the forgetting factor , according to the formula Calculate the state memory decay rate and use this decay rate to update the state matrix when the spatiotemporal memory network is updated .

[0050] Specifically, over time The increase, Gradually increasing means that the memory of historical states gradually weakens, and the weight of current input gradually increases. This dynamically adjusted attenuation rate can adapt to the changes in the importance of information at different time points and better capture the dynamic changes of the disease. As a result, the spatiotemporal memory network can better balance historical information and current information, improve the ability to capture the development trend of the disease, enhance the dynamic adaptability of the model, and better cope with the long-term changes of chronic diseases.

[0051] The medical imaging features include retinal OCT images, and ,in, ResNet50 convolutional neural network. Represents retinal OCT images.

[0052] In specific implementation, retinal OCT images are collected , use ResNet50 convolutional neural network to extract features from OCT images and generate feature vectors .

[0053] Specifically, ResNet50 is a deep convolutional neural network that can automatically extract deep features of images and has strong feature expression capabilities. By processing retinal OCT images, it can extract features related to eye diseases, such as retinal thickness, lesion area, etc., thereby effectively extracting features in medical images, providing an important basis for the evaluation of chronic eye diseases, improving the processing capabilities of medical images, and enhancing the diagnostic accuracy of the model.

[0054] like Figure 1 As shown in FIG, the evaluation method for the chronic disease evaluation system based on the weak AI medical large model platform includes the following steps: S1: Collect patient electronic health records, wearable device timing signals, and medical images to generate feature vectors respectively; S2: Construct a multimodal input matrix and generate fusion features through cross-modal alignment; S3: A large vertical model with a parameter size of ≥70B is deployed on the cloud as the teacher model, and a lightweight model with a parameter size of ≤1B is deployed on the edge as the student model, and the prediction results are output separately. S4: Calculate the risk index of the output prediction result by combining the current credibility and the importance of historical trends; S5: Set a basic threshold based on the disease type, and dynamically adjust the warning threshold based on the fluctuation range of historical risk data. When the risk index exceeds the warning threshold, generate a warning signal; S6: Extract features whose gradient importance is greater than or equal to the threshold to form a decision basis set, and calculate the matching degree of historical similar cases by multiplying the cosine similarity of the case latent vector and the laboratory indicator distance, and generate an explanation report; S7: The edge model updates parameters through the federated learning algorithm, adds Gaussian noise to protect privacy data, and achieves distributed optimization of the model.

[0055] In the specific implementation, the electronic health records, time series signals and medical images of patients are collected through hospital information systems, wearable devices and medical imaging equipment, and the corresponding feature extraction methods are used to generate feature vectors, build a multimodal input matrix, generate fusion features through cross-modal alignment, deploy a large vertical field model with a parameter scale of ≥70B on the cloud as a teacher model, and deploy a lightweight model with a parameter scale of ≤1B on the edge as a student model. The fusion features are predicted respectively and the prediction results are output. According to the current credibility and historical trend importance , combined with the prediction results of the student model and risk drift , calculate the risk index , set basic thresholds according to disease type , dynamically adjust the warning threshold based on the fluctuation range of historical risk data ,when Generate early warning signals and extract gradient importance The features of the decision-making basis set The matching degree of historical similar cases is calculated by multiplying the cosine similarity of the case latent vector and the laboratory indicator distance, and an explanation report is generated. The edge model updates parameters through the federated learning algorithm, adds Gaussian noise to protect privacy data, and realizes distributed optimization of the model.

[0056] This assessment method integrates technologies such as multimodal data processing, cloud-edge collaborative computing, dynamic risk assessment, explainable analysis and federated learning to form a complete chronic disease assessment system. Through the fusion and feature extraction of multimodal data, it improves the comprehensive understanding of the patient's condition; the cloud-edge collaborative architecture realizes efficient computing and deployment; dynamic risk assessment combines current and historical information to improve the accuracy of the assessment; explainable analysis provides decision support for doctors; federated learning ensures data privacy and continuous optimization of the model, realizes a comprehensive, dynamic and accurate assessment of chronic diseases, improves the ability to detect and intervene in the disease early, provides an explainable decision-making basis, enhances doctors' trust in and willingness to use the model, protects patients' privacy data, meets the requirements of medical data security, enables the system to be deployed and applied in different medical environments, and improves the utilization efficiency of medical resources.

[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A chronic disease assessment system based on a weak AI medical large model platform, characterized by: include: (a) Multimodal data acquisition module: Obtaining electronic health record (EHR) text feature vectors , wearable device timing signals , medical imaging features ; (b) Cloud-edge-device collaborative architecture: A large vertical model engine deployed in the cloud outputs the prediction results of the teacher model ; A lightweight evaluation model deployed at the edge to output student model prediction results ; (c) Dynamic risk assessment module, generating risk index : in, ; in, , Expressed as The multimodal input matrix at time t; is the current state credibility, For the importance of historical trends, is the risk drift, that is, the data distribution at the current moment Relative to the patient's baseline health status The degree of abnormal deviation, is the state memory decay rate; For spatiotemporal memory networks The state matrix at time t, The patient's baseline status; Expressed as the Kullback-Leibler divergence operator; Represented as a long short-term memory neural network; (d) Early warning trigger module: When When an early warning signal is generated, is the warning threshold; (e) Explainability report module: output decision-making feature set ,in, For the characteristic variables, The risk index is The gradient of a feature, is the gradient importance threshold.

2. The chronic disease assessment system based on the weak AI medical large model platform according to claim 1 is characterized in that: The multimodal input matrix Generating fusion features through cross-modal alignment : in, Indicates modality The feature extraction function of Including EHR electronic health records, TS time series signals, and IMG medical images; is a modal set, that is ; For the The weight of each mode; Indicates the The original eigenvectors of the modes; represents a learnable cross-modal alignment matrix used to map different modality features into a unified space; represents the dynamic weight query vector; For the The original eigenvectors of the modes.

3. The chronic disease assessment system based on the weak AI medical large model platform according to claim 1 is characterized in that: The lightweight evaluation model is trained through knowledge distillation, and the loss function of knowledge distillation is: in, is the knowledge distillation loss function, which is used to train the student model to imitate the teacher model; and Represents the logits vectors of the teacher and student models respectively; is the balance coefficient, which is used to adjust the weight between the two terms in the loss function; The temperature coefficient is used in knowledge distillation to "soften" the output of the teacher model by increasing the temperature coefficient, so that the student model can learn richer relationships between categories. Represents the softmax function, for the input vector , No. The softmax value of each element is: , The output vector elements, is an exponential function used to calculate probability weights; represents the cross entropy, which is used to measure the difference between two probability distributions and is defined as: in, is the probability distribution after softening of the teacher model, is the probability distribution after the student model is softened, For the teacher model The softening probability of the class, For students to model Softened probability of the class.

4. The chronic disease assessment system based on the weak AI medical large model platform according to claim 1 is characterized in that: The lightweight evaluation model deployed on the edge updates parameters through federated learning : in, represents the model parameters for the next iteration, represents the current model parameters, is the learning rate, For the A collection of participating medical institutions, is Gaussian noise, For the The gradient of the loss function of each institution, is the noise intensity.

5. The chronic disease assessment system based on the weak AI medical large model platform according to claim 1 is characterized in that: The warning threshold The dynamic calculation formula is: in, is the basic threshold value related to the disease type, is the risk volatility sensitive factor, is the history window length, For historical moments The risk index, It is the average of historical risk indices, which is used to reflect the long-term risk level.

6. The chronic disease assessment system based on the weak AI medical large model platform according to claim 2 is characterized in that: The feature extraction function for: in, , is the wavelet basis function; is the scale parameter; It is a feature extraction function for time series signals, implemented based on wavelet transform; Timing signals collected for wearable devices; is a wavelet transform function used to extract the time-frequency characteristics of time series signals; is a time variable, used to describe the shape of the wavelet basis function in the time domain; It is a translation parameter that controls the position of the wavelet on the time axis and is used to locate the local features of the signal.

7. The chronic disease assessment system based on the weak AI medical large model platform according to claim 1 is characterized in that: The similar case matching of the explainability reporting module satisfies: in, represents the case similarity score, represents the latent vector of the current patient case, Represents the latent vector of the case in the historical case database, Represents the laboratory index vector of the current patient, A vector of laboratory indicators representing historical cases.

8. The chronic disease assessment system based on the weak AI medical large model platform according to claim 1 is characterized in that: The spatiotemporal memory network updates satisfy , For the forgetting factor.

9. The chronic disease assessment system based on the weak AI medical large model platform according to claim 1 is characterized in that: The medical imaging features include retinal OCT images, and ,in, ResNet50 convolutional neural network. Represents retinal OCT images.

10. The method for evaluating a chronic disease assessment system using the weak AI medical large model platform according to any one of claims 1 to 9 is characterized in that: The following steps are involved: S1: Collect patient electronic health records, wearable device timing signals, and medical images to generate feature vectors respectively; S2: Construct a multimodal input matrix and generate fusion features through cross-modal alignment; S3: A large vertical model with a parameter size of ≥70B is deployed on the cloud as the teacher model, and a lightweight model with a parameter size of ≤1B is deployed on the edge as the student model, and the prediction results are output separately. S4: Calculate the risk index of the output prediction result by combining the current credibility and the importance of historical trends; S5: Set a basic threshold based on the disease type, and dynamically adjust the warning threshold based on the fluctuation range of historical risk data. When the risk index exceeds the warning threshold, generate a warning signal; S6: Extract features whose gradient importance is greater than or equal to the threshold to form a decision basis set, and calculate the matching degree of historical similar cases by multiplying the cosine similarity of the case latent vector and the laboratory indicator distance, and generate an explanation report; S7: The edge model updates parameters through the federated learning algorithm, adds Gaussian noise to protect privacy data, and achieves distributed optimization of the model.

Images