Rare child disease management method and system based on artificial intelligence and electronic equipment

Abstract

The invention discloses a rare child disease management method and system based on artificial intelligence and electronic equipment, and relates to the technical field of artificial intelligence and health cross, and the method comprises the steps: constructing and training a first preset deep learning model, and obtaining an intelligent question and answer model; and acquiring rare disease associated data of the rare sick child patient in a question and answer mode by using the intelligent question and answer model, and analyzing the rare disease associated data to generate a corresponding personalized health management scheme. According to the method, the intelligent question and answer model is obtained by constructing and training the first preset deep learning model, the problem that in the prior art, a targeted efficient tool is lacked to obtain the associated data of the rare sick children is solved, and scattered data can be effectively integrated. The intelligent question and answer model is used for obtaining data in a question and answer mode and analyzing the data, then the personalized health management scheme is generated, the defects that a traditional method depends on artificial experience and is difficult to meet the requirement for rapidly and accurately formulating the scheme are overcome, and the scientificity and accuracy of the health management scheme are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and health, in particular to a rare disease management method, system and electronic device based on artificial intelligence. BACKGROUND

[0002] At present, the health management of rare disease children patients faces many challenges. On the one hand, there are many types of rare diseases, but the number of patients with each disease is limited, which leads to scattered and difficult to integrate relevant data, and traditional methods are difficult to effectively collect and analyze relevant data. On the other hand, it is necessary to comprehensively consider a large amount of disease information, individual characteristics and other factors to develop individualized health management plans, which is a huge workload and requires high professional knowledge. In the prior art, although there are conventional medical data analysis methods, there is a lack of targeted and efficient tools to specifically collect and analyze data of rare disease children. At the same time, in the aspect of developing health management plans, it is mainly relied on artificial experience, and lacks intelligent and automatic model support, which is difficult to meet the demand of quickly and accurately developing individualized plans, and cannot fully consider the complex individual differences and disease characteristics of rare disease children. SUMMARY

[0003] The technical problem to be solved by the present application is to overcome the deficiencies of the prior art, and specifically to provide a rare disease management method, system and electronic device based on artificial intelligence, as follows: 1) In a first aspect, the present application provides a rare disease management method, system and electronic device based on artificial intelligence, and the specific technical solutions are as follows: A first preset deep learning model is constructed and trained to obtain an intelligent question and answer model. The intelligent question and answer model is used to obtain rare disease related data of rare disease children patients in a question and answer manner and analyze the data to generate a corresponding individualized health management plan.

[0004] The rare disease management method based on artificial intelligence provided by the present application has the following beneficial effects: By constructing and training the first preset deep learning model to obtain the intelligent question and answer model, the problem of lacking targeted and efficient tools to collect rare disease children related data in the prior art is solved, and the scattered data can be effectively integrated. By using the intelligent question and answer model to collect data in a question and answer manner and analyze the data, an individualized health management plan is generated, which overcomes the defects of traditional methods relying on artificial experience and being difficult to meet the demand of quickly and accurately developing a plan. The individual differences and disease characteristics of rare disease children can be fully considered, and the scientificity and accuracy of the health management plan are improved.

[0005] On the basis of the above-mentioned scheme, the rare disease management method based on artificial intelligence provided by the present application can be further improved as follows.

[0006] Further, the method further comprises: constructing and training a second preset deep learning model to obtain an emotion recognition model; using the emotion recognition model to obtain an emotion recognition result of the rare disease child patient, and generating a corresponding emotion management scheme according to the emotion recognition result.

[0007] The beneficial effect of the above further scheme is that the emotion recognition model is obtained by constructing and training the second preset deep learning model, which solves the problem of lack of intelligent tools for emotion management of rare disease children in the prior art, and realizes accurate recognition of the emotions of rare disease children. The emotion recognition result is obtained by using the emotion recognition model, and a corresponding emotion management scheme is generated according to the result, which makes up for the defect of lack of attention to emotion management in traditional health management schemes, can respond to the emotional changes of rare disease children in time, provides more comprehensive and personalized emotional support for overall health management, and improves the quality of life.

[0008] Further, the method further comprises: constructing and training a third preset deep learning model in combination with a child language corpus to obtain a psychological support model; when the user intervenes in the rare disease child patient by using the generated emotion management scheme, the psychological support model is used to convert the expression content of the user into expression content suitable for the understanding of children.

[0009] The beneficial effect of the above further scheme is that the psychological support model is obtained by constructing and training the third preset deep learning model in combination with the child language corpus, which solves the problem of lack of effective language conversion tools for child psychological support in the prior art, and enables the model to more accurately understand the language characteristics of children. When the user intervenes in the rare disease child patient by using the generated emotion management scheme, the psychological support model is used to convert the expression content of the user into expression content suitable for the understanding of children, which overcomes the defect that the expression content in the traditional intervention method is not close to the cognitive level of children, improves the applicability and effect of the emotion management scheme, and is more conducive to the acceptance and understanding of children, thereby better promoting their mental health and emotional improvement.

[0010] Further, the method further comprises: constructing and training a fourth preset deep learning model to obtain a gene variation semantic understanding model; using the gene variation semantic understanding model to analyze the genetic data of the rare disease child patient to generate a corresponding genetic risk assessment result.

[0011] The beneficial effect of adopting the above further scheme is that by constructing and training the fourth preset deep learning model, the genetic variation semantic understanding model is successfully obtained, solving the limitations of the prior art in rare disease pediatric genetic data analysis, enabling the model to more accurately understand the complex semantics in genetic data. The genetic variation semantic understanding model is used to analyze the genetic data of rare disease pediatric patients, generating corresponding genetic risk assessment results. Not only does it improve the accuracy and scientificity of genetic risk assessment, but it also provides a deeper genetic basis for health management programs, overcoming the limitations of traditional methods in fully exploiting the value of genetic data.

[0012] 2) In a second aspect, the present application also provides an artificial intelligence-based rare disease management system for children, and the specific technical solutions are as follows: The system comprises a model construction and training module and a model application module. The model construction and training module is configured to construct and train the first preset deep learning model to obtain an intelligent question and answer model. The model application module is configured to use the intelligent question and answer model to obtain and analyze rare disease-related data of rare disease pediatric patients in a question and answer manner, and generate a corresponding individualized health management program.

[0013] On the basis of the above scheme, the artificial intelligence-based rare disease management system for children of the present application can be further improved as follows.

[0014] Further, the model construction and training module is further configured to construct and train the second preset deep learning model to obtain an emotion recognition model. The model application module is further configured to use the emotion recognition model to obtain an emotion recognition result of the rare disease pediatric patient, and generate a corresponding emotion management program according to the emotion recognition result.

[0015] Further, the model construction and training module is further configured to construct and train the third preset deep learning model in combination with a child language corpus to obtain a psychological support model. The model application module is further configured to use the psychological support model to convert the user's expression content into expression content suitable for children to understand when the user uses the generated emotion management program to intervene in the rare disease pediatric patient.

[0016] Further, the model construction and training module is further configured to construct and train the fourth preset deep learning model to obtain a genetic variation semantic understanding model. The model application module is further configured to use the genetic variation semantic understanding model to analyze the genetic data of rare disease pediatric patients, generating corresponding genetic risk assessment results.

[0017] 3) In a third aspect, the present application also provides an electronic device, which comprises a processor and a memory coupled to the processor, and the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to enable the electronic device to implement any of the above artificial intelligence-based rare pediatric disease management methods.

[0018] 4) In a fourth aspect, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement any of the above artificial intelligence-based rare pediatric disease management methods.

[0019] It should be noted that the technical solutions of the second to fourth aspects of the present application and the corresponding possible implementation manners have the beneficial effects described above for the first aspect and its corresponding possible implementation manners, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the description of the embodiments of the present application will be briefly introduced as follows: Figure 1 A flowchart of a rare pediatric disease management method based on artificial intelligence according to an embodiment of the present application; Figure 2 A structure diagram of a rare pediatric disease management system based on artificial intelligence according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] The principles and features of the present application are described below, and the examples are only used to explain the present application and not to limit the scope of the present application.

[0022] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the drawings.

[0023] As shown in the figure, a rare pediatric disease management method based on artificial intelligence according to an embodiment of the present application comprises the following steps: Figure 1 S10, constructing and training the first preset deep learning model to obtain an intelligent question and answer model. S10, constructing and training the first preset deep learning model to obtain an intelligent question and answer model.

[0024] The first preset deep learning model can be a BERT (Bidirectional Encoder Representations from Transformers) model based on Transformer, or other deep learning models can be selected according to actual conditions, and the training process is as follows: A large amount of medical field text data is collected, including rare disease related literature, encyclopedic knowledge, authoritative medical website content, etc. The data is preprocessed, segmented, labeled and other preprocessing operations to form a format suitable for model training. Then, the preprocessed data is input into the BERT model for pre-training, through mask language model (MLM) and next sentence prediction (NSP) tasks, etc. to make the model learn the semantic information and language rules of medical text.

[0025] After pre-training, in order to make the model more suitable for the rare disease child health management question and answer scene, further fine-tuning (Fine-tuning) is carried out. Prepare targeted question and answer data sets, questions cover rare disease children's diet, exercise, diagnosis and treatment, etc. The answers are provided by professional doctors or researchers. The question and answer pairs are input into the model, the parameters of the model are adjusted, and the loss function is optimized, such as using cross-entropy loss function, so that the model learns to generate accurate answers related to rare disease child health management according to the question. During training, optimization algorithms such as gradient descent method are used, and appropriate learning rate, batch size and other hyperparameters are set. After multiple rounds of iterative training, when the performance indicators (such as accuracy, recall rate, etc.) of the model on the validation set reach the ideal level, the training is stopped, and the intelligent question and answer model is obtained.

[0026] S11, using the intelligent question and answer model, obtaining rare disease related data of rare disease child patients in a question and answer manner and analyzing, generating corresponding individualized health management scheme.

[0027] In practical application, users (such as doctors, parents, etc.) interact with the intelligent question and answer model through the interactive interface. The user raises questions about rare disease children, which can cover many aspects of rare disease children's diet, exercise, diagnosis and treatment, such as "what food is suitable for children with a certain rare disease?" "How should a child with a rare disease exercise properly?" "What are the common treatment methods for this rare disease?" etc. The intelligent question and answer model receives the question, uses its semantic understanding and question and answer generation ability of medical text to retrieve relevant information from the learned knowledge, and generates accurate and detailed answers. These question and answer data are recorded to form a rare disease related data set.

[0028] The obtained rare disease association data is analyzed. The analysis methods include text clustering, topic extraction, etc. Through text clustering, similar questions and answers are classified into a category, and the common needs and concerns of rare disease children patients in different aspects are found; through topic extraction, the main topics and key information in the data are identified, and the potential problems of rare disease children patients in diet, exercise, diagnosis and treatment, etc. are understood. For example, it is found that children with a certain rare disease generally have special needs for certain foods, or a certain exercise method has a significant effect on the rehabilitation of children with a specific rare disease, etc.

[0029] Based on the potential problems of rare disease children patients in diet, exercise, diagnosis and treatment, etc., combined with the individual characteristics of rare disease children patients (such as age, gender, disease severity, physical condition, etc.), a corresponding personalized health management plan is generated. The personalized health management plan includes diet matching plan, exercise plan and diagnosis and treatment plan.

[0030] The rare disease association data includes the types of rare diseases, disease severity, disease development stage, age, gender, height, weight, body fat rate, dietary preference data, muscle strength, muscle atrophy degree, joint mobility, basic physical indicators (balance ability, coordination ability, flexibility, endurance, etc.), historical exercise data, disease diagnosis data and treatment response data, etc.

[0031] Optionally, in the above technical solution, it further comprises: S20, constructing and training the second preset deep learning model to obtain an emotion recognition model, specifically comprising the following steps: S200, collect multi-modal data of rare disease children patients in different emotional states, including facial expression pictures, voice tone clips and text descriptions, etc. Clean these data to remove samples with poor quality or not meeting the requirements. Standardize the facial expression pictures, unify the image size, grayscale, etc.; extract features from voice tone clips, such as pitch, timbre, and speech rate; perform preprocessing operations such as word segmentation and stop word removal on text descriptions, and label emotion categories such as happy, sad, angry, and anxious, to form a structured training data set.

[0032] S201, a convolutional neural network (CNN) is selected to construct a second preset deep learning model. CNN performs well in processing image and sequence data, and can automatically extract features in the data. The model structure includes an input layer, multiple convolutional layers, pooling layers, and fully connected layers. Taking a facial expression picture as an example, the input layer receives preprocessed image data; the convolutional layer uses convolution kernels of different sizes and numbers to perform convolution operations on the image, extracting features such as edges and textures in the image; the pooling layer reduces the dimension of the feature map after convolution, reducing the amount of calculation while preserving important features; the fully connected layer integrates the extracted features and outputs the probability distribution of the emotion category. It should be noted that other deep learning models can also be selected as the second preset deep learning model according to actual conditions.

[0033] S202, input the labeled multi-modal data into the constructed CNN model for training. In the training process, a cross-entropy loss function is used to measure the difference between the model's predicted emotion category and the true emotion category. The stochastic gradient descent (SGD) or Adam optimization algorithm is used to optimize the model parameters, and by adjusting the learning rate, batch size, and other hyperparameters, the model can quickly converge and improve the accuracy. After each training cycle, the performance of the model is evaluated on the validation set, and the accuracy, recall rate, and F1 value are used as indicators to judge the performance of the model. After multiple rounds of iterative training, when the performance of the model on the validation set reaches an ideal level and tends to be stable, the training is stopped, and the emotion recognition model is obtained.

[0034] S21, using the emotion recognition model, the emotion recognition result of the rare disease child patient is recognized, and the corresponding emotion management scheme is generated according to the emotion recognition result, which includes the following steps: S210, in practical application, the real-time data of the rare disease child patient (such as newly taken facial expression pictures, newly recorded voice tone clips, or newly input text descriptions) is input into the trained emotion recognition model. The model will perform the same preprocessing and feature extraction operations on the input data, and then output the current emotion recognition result of the child patient through the learned feature pattern and emotion category mapping relationship. For example, the probability of the child being in an "anxious" emotional state is 80%, and the probability of being in a "calm" emotional state is 20%, so the current main emotion of the child is determined to be anxiety.

[0035] S211, according to the emotion recognition result, combined with the individual characteristics of the rare disease child patient (such as age, disease condition, etc.), the corresponding emotion management scheme is generated. This step can be completed by artificial means, for example, if the emotion recognition result is anxiety, an emotion management scheme guiding the child to perform deep breathing relaxation training can be generated.

[0036] In another implementation manner, S20 to S21 include the following steps: S2000, collect and label a large-scale general child emotion dataset, including synchronously collected facial video streams, audio streams, and basic physiological signals; construct a multi-modal fusion deep learning model based on a Transformer architecture as the second deep learning model, a video branch in the multi-modal fusion deep learning model uses a 3D-CNN to extract spatiotemporal features, an audio branch of the multi-modal fusion deep learning model uses a time-domain convolutional network to extract acoustic features, and a physiological signal branch of the multi-modal fusion deep learning model uses a time series encoder to extract physiological pattern features; cross-modal feature interaction is achieved through a cross-attention mechanism, model pre-training is completed on a general dataset, an initial model with basic emotion recognition capability is formed, and a Softmax function is used in an output layer to generate an emotion probability distribution.

[0037] Among them, through professional data acquisition equipment, facial video streams, audio streams and basic physiological signals (such as heart rate, skin conductance, etc.) of children in various emotional states are synchronously collected. Ensure the diversity of data, covering children of different ages, genders, and races, and rich emotional scenes (such as happiness, sadness, anger, surprise, fear, etc.). Adopting a manual annotation combined with machine assistance method, the collected data is labeled with emotional labels. The annotators are trained professionally, and according to the behavior, expression, language and other characteristics of children, each data sample is assigned to the corresponding emotional category.

[0038] Among them, the facial video stream data is taken as the input of the video branch, and the 3D-CNN is used to extract the spatiotemporal features. The 3D-CNN can capture the information in the spatial and temporal dimensions of the video at the same time, and gradually extract the key features of the facial expression, such as the movement changes of the eyes, eyebrows, and mouth, etc., to form the feature representation of the video modality.

[0039] Among them, the audio branch uses a time-domain convolutional network to extract acoustic features in the audio stream, including pitch, timbre, loudness, speech rate, etc. These features can reflect the changes in children's speech under emotional states, forming the feature representation of the audio modality.

[0040] Among them, the physiological signal branch uses a time series encoder to extract physiological pattern features in the basic physiological signals. The time series encoder can process the time series characteristics of the physiological signals, analyze the change trend and pattern of the physiological signals under different emotions, such as the acceleration or deceleration of heart rate, the rise or fall of skin conductance, etc., to form the feature representation of the physiological signal modality.

[0041] On the basis of extracting features of each modality, a cross-attention mechanism is introduced to realize cross-modal feature interaction. By calculating the attention weights between different modalities, the model can focus on the key information related to emotion recognition in each modality and fuse these information to form more discriminative feature representation. For example, the changes in facial expressions in the video are related to the fluctuations in tone in the audio, and the changes in heart rate in the physiological signal are echoed by the emotional state. The cross-attention mechanism can capture these cross-modal associated information.

[0042] The model is pre-trained on a general dataset to form an initial model with basic emotion recognition capability. During the pre-training process, a large amount of multi-modal emotion data is used, and a Softmax function is used as the output layer to generate an emotion probability distribution. By optimizing the loss function (such as cross-entropy loss), the parameters of the model are adjusted so that the model can learn the feature patterns under different emotion categories, thereby possessing the ability to recognize basic emotions. After pre-training, the initial model obtained can provide a basis for further model optimization and application.

[0043] S2001, based on the initial model, personalized data of target rare disease children is collected through a compliant mobile terminal; the bottom feature extraction layer of the initial model is frozen, and an adapter layer is inserted to realize efficient parameter migration; a disease-aware attention mechanism is introduced for the top classifier to dynamically weight the contribution of different modalities to the emotional expression of the target patient; an elastic weight consolidation algorithm is used to constrain the fine-tuning process to generate an emotion recognition model.

[0044] Among them, based on the initial model, the bottom feature extraction layer is frozen, including the 3D-CNN of the video branch, the time domain convolution network of the audio branch, and the time series encoder of the physiological signal branch. These frozen layers have learned the general emotion feature extraction pattern in the pre-training process, and after being frozen, they can be used as basic components for feature extraction, maintaining their stability and universality. An adapter layer is inserted between the frozen feature extraction layer and the top classifier. The adapter layer consists of two fully connected layers with an activation function in between. Its role is to map the general features to a space specific to the rare disease children emotion recognition task, realizing efficient parameter migration. The introduction of the adapter layer can quickly adapt to new tasks without changing the original model structure, while reducing the computational load and storage requirements during model fine-tuning.

[0045] Among them, for the top-level classifier, a disease-aware attention mechanism is introduced. This mechanism dynamically calculates the contribution weight of each modality to the target patient's emotional expression by modeling the correlation degree of different modality features and rare disease-related information. During the training process, different modality features are weighted and fused according to these weights, so that the model can pay more attention to the modality information closely related to the emotional expression of rare disease children, thereby improving the accuracy and relevance of emotion recognition. For example, for some rare disease children whose main emotional expression feature is physiological signal change, the weight of the physiological signal modality will be appropriately increased, and vice versa, for children whose emotional expression is mainly through facial expressions and voice, the weights of the video and audio modalities will be increased accordingly.

[0046] During the model fine-tuning process, an elastic weight consolidation algorithm is used for constraint. This algorithm limits the update of model parameters by adding a regularization term to the loss function, preventing the model from overfitting to the personalized data of rare disease children, while consolidating the general features learned in the initial model. The elastic weight consolidation algorithm can gradually adapt the model to the emotion recognition task of rare disease children while preserving the general features of the initial model, ensuring the generalization ability and stability of the model. By reasonably setting the regularization coefficient, the fitting degree of the model to new and old data is balanced, ensuring that the fine-tuned model has good performance in the emotion recognition task of rare disease children.

[0047] S2002, deploy the emotion recognition model to an online active learning framework; real-time calculation of the prediction divergence of unlabeled data, when high uncertainty samples are detected, trigger the data completion protocol to synchronize the capture of multi-modal data stream and label by clinical experts; the new samples are injected into the fine-tuning dataset after data augmentation, and the incremental retraining cycle is started to update the emotion recognition model.

[0048] Among them, the well-adapted emotion recognition model is deployed to an online active learning framework, usually on a cloud server or a local server with sufficient computing power. Ensure that the model can receive multi-modal data stream (face video stream, audio stream and basic physiological signal) from mobile terminal in real time and output emotion recognition result. For this purpose, an efficient data transmission interface needs to be designed to ensure low-latency data transmission and enable the model to respond quickly.

[0049] During the model inference process, the prediction divergence of unlabeled data is calculated in real time. Specifically, the prediction divergence is calculated using the multi-round prediction output of the model or the prediction results of different model components (such as each modality branch). For example, the variance and entropy value of the prediction results can be used as statistical indicators to measure the divergence. When the prediction divergence exceeds the set threshold, it is determined to be a high uncertainty sample. The determination of the threshold can be obtained through experimental analysis on the validation set, after balancing the recall rate and false positive rate.

[0050] Wherein, after triggering the data completion protocol, multi-modal data stream needs to be captured synchronously. This requires the corresponding data acquisition module on the mobile terminal to respond quickly to collect facial video, audio and physiological signals at a high sampling rate and in a lossless format, and upload them to the server in a timely manner. At the same time, clinical experts are notified to intervene in the labeling process, and a professional labeling interface is provided to ensure that experts can efficiently and accurately assign emotion labels to samples. During the labeling process, strict labeling specifications and quality control mechanisms need to be established to avoid negative impacts of labeling errors on model training.

[0051] Data augmentation is performed on the new samples, and different augmentation strategies are designed for different modalities. For facial video data, image enhancement methods such as random cropping, rotation, and flipping can be used; for audio data, noise can be added, speech rate and pitch can be changed; and for physiological signal data, interpolation and noise injection can be used to expand the sample size. The augmented samples and the original samples are jointly injected into the fine-tuning dataset to expand the dataset size and enhance its representativeness and diversity.

[0052] When starting the incremental retraining cycle, the model training parameters need to be reconfigured, and the bottom feature extraction layer (3D-CNN for video branch, time domain convolution network for audio branch, and time series encoder for physiological signal branch) is frozen, only the adapter layer, top classifier, and disease perception attention mechanism related parameters are updated. The elastic weight consolidation algorithm is used to constrain the model to retain the learned general features and knowledge during the update process. During training, the learning rate is dynamically adjusted, a larger step size is used in the early stage to quickly adjust the parameters, and the learning rate is gradually reduced in the later stage to fine-tune the model, finally realizing the continuous evolution and performance improvement of the emotion recognition model.

[0053] S2003, generating coarse-grained emotion probability based on the emotion recognition model; modeling the time evolution law of emotion expression through the gated recurrent unit, calculating the correlation score of real-time physiological signals and basic emotion prototypes, and dynamically calibrating the confidence of video / audio modalities; in the joint output layer, the time series features and cross-modal weights are fused to generate structured recognition results containing emotion categories, intensity scalar and physiological coordination index.

[0054] Using the existing multi-modal emotion recognition model, the input facial video stream, audio stream and basic physiological signals are preliminarily recognized to obtain coarse-grained emotion probability. The Softmax output layer obtains the probability distribution corresponding to the emotion category, providing basic information for subsequent time series modeling and feature fusion.

[0055] wherein the sequence of features of the video, audio and physiological signals is input into a gated recurrent unit (GRU). The GRU controls the flow of information through internal reset and update gates, and updates the hidden state frame by frame, thereby modeling the temporal evolution of emotional expression. This enables the model to capture the trend of emotional changes in the time dimension, such as the gradual transition from calm to excited.

[0056] wherein the process of calculating the correlation scores of the physiological signals and the basic emotion prototypes is as follows: first, define the physiological feature vectors of the basic emotion prototypes. These prototypes can be based on existing physiological research, including the patterns of typical physiological signals such as heart rate, skin conductance, body temperature under different emotions. Then, similarity calculation is performed between the real-time physiological signal features and the basic emotion prototype vectors, for example, using cosine similarity or correlation coefficient method, to obtain the correlation scores. These scores reflect the degree of association between the physiological signals and different basic emotions, providing a basis for subsequent cross-modal calibration.

[0057] Based on the correlation scores of the physiological signals and the basic emotion prototypes, the confidence of the video and audio modalities is dynamically calibrated. Specifically, for a certain moment, if the correlation score of the physiological signals and a certain emotion prototype is high, it means that the physiological performance of this emotion is more obvious, and then the confidence weight of the video and audio features related to this emotion expression is appropriately increased. For example, if the heart rate increases and the skin conductance rises with a high correlation score of the excited emotion prototype, and the facial expression in the video and the tone in the audio also show excited features, then the weights of these video and audio features in emotion recognition are increased, so that the model pays more attention to the modal information consistent with the physiological signals, and the accuracy of emotion recognition is improved.

[0058] wherein in the joint output layer, the time series feature vector obtained by GRU modeling is fused with the feature weight vector of the video and audio modalities after dynamic calibration. Methods such as feature concatenation or weighted summation are used to integrate the information of different modalities into a comprehensive feature representation, which fully reflects the multi-dimensional features of emotion. Based on the fused comprehensive features, structured recognition results including emotion categories, intensity scalars and physiological coordination indexes are generated through classification and regression operations. The emotion category is determined by the classification layer according to the comprehensive features; the intensity scalar is obtained by quantifying the degree of emotional activation, for example, using the comprehensive index of emotion probability and physiological signal intensity; the physiological coordination index is determined by evaluating the consistency between the physiological signals and the emotional expression, for example, by calculating the matching degree between the trend of physiological signal changes and the expected physiological pattern of the emotion category, thereby providing more comprehensive and detailed quantitative information for the emotion recognition of rare disease children.

[0059] S2004, deploy the emotion recognition model to an edge computing terminal to build a closed-loop verification system; output the structured recognition result to a clinical monitoring dashboard in real time to trigger a preset response protocol; record the original multi-modal data and environmental context of misjudgment samples through a medical feedback interface; periodically analyze the model bias mode based on misjudgment data, update the adversarial training sample library, and start adversarial retraining to optimize the emotion recognition model.

[0060] The emotion recognition model is deployed to an edge computing terminal, such as a mobile device or an Internet of Things device with certain computing capabilities. This requires quantization and compression of the model to adapt to the hardware limitations of the edge device. At the same time, a closed-loop verification system is built to ensure stable operation of the model in the edge environment. The closed-loop verification system needs to include data input module, model inference module, result output module and feedback collection module, which seamlessly connect each other to form a complete verification closed loop.

[0061] On the edge computing terminal, the model processes multi-modal data streams (face video stream, audio stream and basic physiological signals) in real time to generate structured recognition results containing emotion categories, intensity scalar and physiological coordination index. Through the optimized data transmission protocol, the results are transmitted to the clinical monitoring dashboard in real time. The dashboard uses an intuitive visual interface to display the patient's emotional state in real time. At the same time, the preset response protocol is associated with the emotion recognition result, and when the recognition result meets the specific emotion threshold, the corresponding response protocol is automatically triggered, such as when extreme anxiety or depression is detected, medical personnel are immediately notified to intervene.

[0062] A medical feedback interface is designed for medical personnel to conveniently record misjudgment sample information. When medical personnel find that the model output does not match the actual emotional state of the patient, they can record the original multi-modal data (face video clips, audio clips, physiological signal data) and environmental context information (such as the light and sound of the environment where the patient is at the time) through the interface. These information will be saved to a special misjudgment sample database to provide data support for subsequent model optimization.

[0063] The misjudgment sample database is analyzed regularly to find the bias patterns of the model. Through clustering analysis, feature importance evaluation and other methods, the rules of the model's misjudgment in specific emotion categories, specific patient groups or specific environmental conditions are found. Based on these bias patterns, the adversarial training sample library is updated, representative misjudgment samples are added to the original data set, and adversarial enhancement is performed on them, such as geometric transformation of video data and addition of noise to audio data, to generate more challenging adversarial samples. When adversarial retraining is started, the model's bottom feature extraction layer is frozen, and adversarial training loss functions are introduced in the adapter layer and the top classifier, so that the model can gradually overcome the bias in the learning process and improve the accuracy and robustness of emotion recognition for rare pediatric patients.

[0064] Optionally, in the above technical solution, further comprising: S30, constructing and combining a child language corpus to train the third preset deep learning model to obtain a psychological support model, specifically including the following steps: S300, collecting a large-scale child language corpus, including various text data such as children's daily conversations, fairy tales, children's literature works, and language records in the process of children's psychological treatment. These data can cover the language expression characteristics and understanding level of children in different scenarios. The collected data is cleaned to remove adult complex language, professional terms, and content that does not conform to the habits of children's language. Then, preprocessing operations such as word segmentation, part-of-speech tagging, and syntax analysis are performed to label key features in children's language, such as simple vocabulary, repeated sentence patterns, and common expression methods, forming a structured training data set.

[0065] S301, selecting a recurrent neural network (RNN) and its variants (such as LSTM or GRU) to construct the third preset deep learning model. RNN is good at processing sequence data and can capture the context information and time sequence relationship in language, suitable for modeling the generation and understanding of children's language. LSTM (Long Short-Term Memory Network) or GRU (Gated Recurrent Unit) can effectively solve the gradient vanishing problem of RNN when processing long sequence data, and better capture the long-distance dependency relationship in children's language. The model structure includes an input layer, multiple recurrent neural network layers, and an output layer. The input layer receives the preprocessed child language text data and converts it into word vector representation; the recurrent neural network layer processes the word vector sequence through the loop structure to learn the semantic and grammatical features of the language; the output layer outputs the corresponding results according to the task requirements, such as semantic understanding of children's language, generating language suitable for children's understanding, etc. It should be noted that other deep learning models can also be selected as the third preset deep learning model according to actual conditions.

[0066] S302, input the preprocessed child language corpus data into the constructed RNN (LSTM or GRU) model for training. In the training process, appropriate loss functions are used, such as cross-entropy loss function for language generation task, mean square error loss function for semantic understanding task, etc. Random gradient descent (SGD) or Adam optimization algorithm is used to optimize the model parameters, and by adjusting the learning rate, batch size, number of hidden layer neurons and other hyperparameters, the model can quickly converge and improve the understanding and generation ability of child language. In the training process, the model is evaluated by using the validation set, and the perplexity, BLEU (Bilingual Evaluation Understudy) value and other indicators are used to measure the performance of the model on the language generation task, or the accuracy and other indicators are used to evaluate the performance of the semantic understanding task. After multiple rounds of iterative training, when the performance of the model on the validation set reaches the ideal level and tends to be stable, the training is stopped, and the psychological support model is obtained.

[0067] S31, when the user intervenes in the rare disease child patient by using the generated emotional management scheme, the psychological support model is used to convert the expression content of the user into expression content suitable for the understanding of children.

[0068] When a user (such as a doctor, parent, or psychologist, etc.) intervenes in a rare disease child patient using an emotional management program, the user inputs the content he or she wants to express in the form of text into the psychological support model. These expressions may contain complex professional terms, adult expressions, or more abstract concepts, such as "you need to do relaxation training according to this program, which helps to relieve your anxiety and let your brain rest." After receiving the user's input expressions, the psychological support model first preprocesses the text, including word segmentation, part-of-speech tagging, and other operations. Then, the model uses the knowledge trained on the child language corpus to understand and analyze the user's expressions. It will identify parts that do not meet the level of understanding of child language, such as complex words "relaxation training" and "anxiety." Next, the model replaces and reconstructs the expressions according to the characteristics of child language, such as using simpler, more specific, and more vivid words and expressions. For example, "relaxation training" is converted to "let's do some small games that make your body soft together," and "anxiety" is converted to "when you feel uncomfortable in your heart." After conversion, the psychological support model generates expressions suitable for children's understanding and outputs them. The output content will be presented in a way that is closer to children's thinking and language habits, such as "when you feel uncomfortable in your heart, we can do some small games that make your body soft together, so you will feel much better in your heart." Such expressions are easier for rare disease children to accept and understand, thereby improving the effectiveness of the emotional management program intervention, helping children better cooperate with the intervention measures, and improving their emotional state.

[0069] In another implementation, S30 to S31 specifically include the following steps: S3000, collect multi-source language data of rare disease children in psychological intervention scenarios, including doctor-patient dialogue recordings, emotional expression picture books, game interaction texts, and parent feedback records; perform semantic unit segmentation, emotion label annotation, and cognitive complexity classification on the collected multi-source language data, and establish a three-dimensional corpus matrix containing language structure, emotional dimension, and cognitive level; filter high-frequency child cognitive paradigm expressions from the three-dimensional corpus matrix through TF-IDF vectorization and attention mechanism, construct a child language corpus, and store it in a distributed graph database, with node association of clinical metadata such as age, disease stage, and language development level.

[0070] Among them, using professional recording equipment, in the medical scene of psychological intervention, the dialogue between the doctor and the rare disease children in the process of inquiry, communication, treatment guidance, etc. is recorded, and the doctor-patient dialogue recording is obtained. Collect the emotional expression picture books used by rare disease children in psychological intervention. These picture books contain picture creation, text description and other information of children in different emotional states. The picture works completed by children in the psychological treatment process, the picture book templates filled in with emotional description, etc. can be obtained, and the oral explanation or elaboration of the children on the picture book content is recorded. In the process of game interaction between children and psychological therapists (such as role-playing games, puzzle games, building blocks, etc.), the dialogue content between the therapist and the child, the child's self-talk in the game process, and the guiding speech of the therapist to the child's behavior are recorded. Game interaction text. Through the design of corresponding questionnaire, interview outline and other ways, the parents' feedback records on the children's performance, emotional changes, language communication and other aspects in the process of psychological intervention are collected. Let the parents fill in the feedback form regularly, or conduct regular telephone and face-to-face interviews, and convert the interview content into text data, that is, the parents' feedback records.

[0071] Among them, using professional voice recognition technology, such as automatic speech recognition (ASR) system based on deep learning, the collected doctor-patient dialogue recording is converted into text. In the conversion process, the voice signal is preprocessed, such as removing background noise, voice enhancement, etc., to improve the accuracy of voice recognition. The collected text data such as emotional expression picture books, game interaction texts, and parent feedback records are formatted uniformly to ensure consistency in data encoding format, font, etc. At the same time, all data are cleaned to remove redundant spaces, special symbols, irrelevant advertisements or incorrect words, etc. to ensure the neatness and accuracy of the data.

[0072] Among them, the word segmentation technology and semantic analysis method in natural language processing (NLP) are used to divide the preprocessed text data into semantic units. For example, the sentence "I saw a beautiful butterfly in the hospital today" can be divided into "I", "today", "in the hospital", "saw", "beautiful", "butterfly" and other semantic units, ensuring that each unit has independent and clear meaning.

[0073] Then, a sentiment dictionary is constructed, including positive, negative and neutral sentiment words and their corresponding weights. Through text matching algorithm, the words in the text are compared with the sentiment dictionary, and the corresponding sentiment label is marked for each semantic unit. For example, the semantic unit "happy play" is marked as positive emotion, and "fear of pain" is marked as negative emotion.

[0074] Moreover, according to the language development law of children and cognitive theory, the cognitive complexity grading standard is formulated, and multiple levels are set from simple to complex. By analyzing the lexical richness of semantic units, the complexity of sentence structure, the abstract degree of the concepts involved and other factors, the cognitive complexity level of each semantic unit is determined, such as "the sun is hot" is classified as a lower cognitive complexity level, and "Einstein's theory of relativity has changed our cognition of space-time" is classified as a higher cognitive complexity level.

[0075] The acquisition process of the three-dimensional corpus matrix is as follows: The language features of the semantic units are analyzed, including lexical composition, grammatical structure, and part-of-speech tagging, which are taken as the content of the language structure dimension. For example, the distribution of nouns, verbs, adjectives, and other parts of speech in the semantic units is counted, and the subject-predicate-object structure of the sentence and the use of subordinate clauses are analyzed, so as to assign a corresponding language structure feature vector to each semantic unit. The results of sentiment label tagging are integrated to calculate the score of each semantic unit in different sentiment categories such as positive sentiment and negative sentiment, forming a vector representation of the sentiment dimension. For example, a semantic unit scores 0.8 in positive sentiment and 0.2 in negative sentiment, so its sentiment dimension vector can be represented as [0.8, 0.2]. According to the results of cognitive complexity grading, each semantic unit is mapped to the corresponding cognitive level to form the label of the cognitive level dimension. It can be converted into a vector form by using one-hot encoding and other methods, and fused with other dimension vectors to finally construct a three-dimensional corpus matrix containing language structure, sentiment dimension, and cognitive level.

[0076] Among them, the text content in the three-dimensional corpus matrix is subjected to TF-IDF (Term Frequency-Inverse Document Frequency) vectorization processing to calculate the TF-IDF value of the words in each semantic unit. According to the set threshold, the semantic units corresponding to the words with higher TF-IDF values are selected. These semantic units are often representative expressions of children's cognitive paradigm and can reflect the typical language features and cognitive patterns of children with rare diseases in the psychological intervention scene. On the basis of TF-IDF screening, an attention mechanism model is introduced. By training the attention model, the model learns and automatically focuses on those semantic units that are more critical in semantic understanding, emotional expression, and cognitive development, further filtering and determining high-frequency expressions of children's cognitive paradigm, and improving the accuracy and effectiveness of the screening. The selected high-frequency expressions of children's cognitive paradigm are collected to build a corpus of children's language. A distributed graph database is used for storage, with each semantic unit as a node, and the attributes of the node including language structure features, sentiment dimension vectors, and cognitive level labels. At the same time, the association between the node and the clinical metadata such as age, disease stage, and language development level is established. Through the indexing and querying functions of the graph database, the data in the corpus can be conveniently searched, analyzed and utilized to support the research and practice of psychological intervention for children with rare diseases.

[0077] S3001. Using the Transformer model as the third preset deep learning model, train the Transformer model based on the standardized corpus and multi-source language data to obtain a psychological support model.

[0078] A corpus was constructed by extracting previously collected and processed multi-source language data of children with rare diseases from a distributed graph database. This corpus includes text content such as doctor-patient dialogues, emotional picture books, interactive game texts, and parental feedback, along with corresponding clinical metadata such as age, disease stage, and language development level. This data was formatted to ensure it meets the input requirements of the Transformer model; for example, text sequences were truncated or padded according to a uniform length standard to ensure consistent length.

[0079] The Transformer model architecture is primarily composed of encoders and decoders. The encoder contains multiple identical layers, each consisting of two sub-layers: a multi-head self-attention layer to capture the associations between words at different positions in the text, and a position-based feedforward neural network. The decoder also has multiple identical layers, but in addition to the encoder, it inserts a multi-head attention layer between the two sub-layers. This layer focuses on the encoder's output to better understand the semantics of the input sequence, thereby generating a more context-appropriate output.

[0080] The model training process is as follows: ① Utilize data loading tools within deep learning frameworks (such as TensorFlow or PyTorch) to load the prepared corpus data into the model in batches. During each loading, the text data is encoded, converting words into corresponding index numbers to form a numerical input sequence. Simultaneously, a corresponding output sequence is generated based on the training objective. For example, in training a psychological support model, if the input is a child's expression in a certain context, the output might be a suitable example of a psychological intervention response.

[0081] ②: The input sequence is fed into the encoder of the Transformer model. After multiple layers of encoding, a semantic representation is obtained, which is then passed to the decoder. The decoder generates predictions based on the target output sequence. The cross-entropy loss function is used to measure the difference between the model's predicted output and the actual target output. The loss value is calculated, and this loss value can intuitively reflect the model's performance under the current parameters. The smaller the loss value, the closer the model's prediction result is to the true value.

[0082] ③According to the calculated loss value, the gradient of each parameter of the model is calculated by the back propagation algorithm, that is, the direction of parameter adjustment is determined, and then the model parameters are updated according to a certain learning rate by using the Adam optimizer, so as to gradually reduce the loss value and make the model continuously iterate and optimize in the direction of better performance. During the whole training process, it is very important to reasonably set the learning rate and other hyperparameters. If the learning rate is too large, the model may diverge during optimization and cannot converge to the ideal state. If the learning rate is too small, the training speed will be too slow and a lot of time cost will be consumed.

[0083] During the training process, a part of data is reserved as a validation set, and the validation set is used to evaluate the model regularly to monitor the performance indicators (such as accuracy, recall rate, F1 value, etc.) of the model on the validation set, so as to timely find out whether the model has overfitting or underfitting phenomenon. If overfitting occurs, regularization method can be used to alleviate it; if underfitting occurs, the model structure can be adjusted (such as increasing the number of layers of the encoder or decoder, expanding the number of neurons in each layer, etc.), the hyperparameters can be optimized (adjusting the learning rate, changing the batch size, etc.), or the training data can be expanded to improve the performance of the model.

[0084] S3002, when the user inputs the text of the emotional management scheme for rare disease children, the psychological support model is called for three-level semantic conversion. Specifically, first, the adult abstract concepts are stripped, second, the medical terms are mapped to the equivalent expressions of the picture book level in the standardized corpus, and finally, the simple syntax structure is reconstructed based on the cognitive complexity classification results in the clinical metadata; the output content meets the Collins child readability index ≤L3 standard, and the core emotional regulation intention of the original intervention scheme is retained.

[0085] Among them, the pre-trained psychological support model is called, and the multi-head self-attention mechanism in the encoder is used to understand and analyze the semantics of the input text. Through attention weight calculation, the adult abstract concept words in the text that deviate greatly from the cognitive level of children are identified. These words often have unique vector features in the semantic representation of the model, and there is a significant difference from the language mode commonly used by children. At the same time, with the help of the semantic association knowledge trained in the model, the specific position of these abstract concepts in the text and the context information associated with them are determined, which prepares for subsequent processing.

[0086] Among them, a vocabulary table specially used for medical terminology and picture book level language mapping is constructed. The table is generated based on the emotional expression picture book part in the previously established standardized corpus and the related expressions in multi-source language data through manual sorting and machine learning algorithm assisted mining. When the model identifies the medical terminology in the text, it uses a search algorithm to find the corresponding picture book level equivalent expression in the vocabulary table. For example, "anxiety symptoms" is mapped to "there is a little rabbit jumping in my heart", which is a vivid and child-friendly picture book description. In this process, it is ensured that the mapped expression can be maximally equivalent in semantics to the emotional management meaning expressed by the original medical terminology, and it conforms to the language habits and understanding ability of children.

[0087] According to the cognitive complexity classification result in the clinical metadata, the syntax complexity standard corresponding to the cognitive level of the target child is determined. The decoder part of the model is used to analyze the syntax structure of the original text in combination with the attention mechanism, to identify the subject, predicate, object and other components in the sentence and their dependency relationships. Then, the sentence is reorganized according to the requirements of simple syntax structure. For example, the complex compound sentence "if emotional regulation can be done in time, the anxiety of children with rare diseases will be effectively alleviated" is reconstructed into a simple subject-predicate-object structure sentence "regulate emotions, anxiety will decrease". In the reconstruction process, the language development characteristics of children are fully considered to ensure that the sentence is concise and easy to understand, while the core emotional regulation intention of the original intervention plan is not changed.

[0088] The output content after the above three-level semantic conversion processing meets the Collins Child Readability Index ≤L3 standard, and can convey the core points of the emotional management plan in a way that children can easily accept and understand, better meeting the psychological support needs of children with rare diseases in emotional management.

[0089] In practice, to ensure the quality of the converted text, a combination of manual checking and machine automatic evaluation can be used. Manual checking is carried out by a team composed of professional child psychologists, linguists and experts in the field of rare diseases, who review a certain number of conversion samples to check whether they meet the cognitive level of children and accurately convey the emotional regulation intention, etc. Machine automatic evaluation can use existing text readability evaluation tools and algorithms to quantitatively analyze the converted text, such as calculating the average length of the sentence, the vocabulary difficulty index, etc., and comparing it with the Collins Child Readability Index standard, so as to continuously monitor and optimize the conversion effect of the model, and continuously improve the appropriateness and effectiveness of the output text, to ensure that it can provide precise and effective psychological support text content for children with rare diseases, to assist them in emotional management and psychological health maintenance work.

[0090] S3003, collecting feedback data of the child on the reconstructed simple syntactic structure expression in real time through a voice recognition device, including language response duration, emotional fluctuation physiological indicators, and behavior observation records; establishing a feedback quality evaluation model, and when the child's understanding score is lower than the preset threshold, automatically triggering the cognitive complexity classification subset of the standardized corpus for retraining, updating the parameter weight of the psychological support model; at the same time, the optimized expression paradigm is incrementally stored in the standardized corpus to form a knowledge enhancement cycle.

[0091] In the scenario of rare disease children receiving an emotional management program, professional voice recognition devices such as intelligent voice collection terminals with high-sensitivity microphone arrays and advanced noise reduction functions are deployed to collect feedback data of children on the reconstructed simple syntactic structure expression in real time. The voice recognition device converts the child's voice response into text in real time, records the language response duration, i.e. the time interval from the child's response to the end of the response. At the same time, wearable physiological monitoring devices such as smart bracelets or specially designed medical-grade physiological monitoring vests are used to monitor the child's emotional fluctuation physiological indicators in real time during the response process, including heart rate variability, skin conductance response, and respiratory rate, which can reflect the child's emotional arousal state and psychological changes. In addition, professional observers use standardized behavior observation recording tools to record the child's behavior in detail, such as recording the child's body movements, facial expressions, eye movements, and other behavioral characteristics during the response, and observers classify and code the observed behaviors according to the established behavior coding system to form behavior observation records.

[0092] The feedback quality evaluation model is built based on a deep learning framework, which takes the collected language response duration, emotional fluctuation physiological indicators, behavior observation records, and child clinical metadata (age, disease stage, language development level, etc.) as input features. First, pre-process and feature engineer these multi-source heterogeneous data, normalize the language response duration, extract time and frequency domain features from the physiological indicators, and perform one-hot encoding on the behavior observation records, etc. to make them effective inputs to the model. Then, use historical feedback data and corresponding child understanding scores (obtained through professional psychologist evaluation or standardized testing tools) to train the model, use appropriate loss functions (such as mean square error loss function) and optimization algorithms (such as Adam optimizer) to adjust model parameters, so that the model can accurately predict the child's understanding score based on the input feedback features. During model training, reasonably divide the training set, validation set and test set, and evaluate the model's performance through cross-validation and other methods to ensure the model has good generalization ability.

[0093] The preset threshold of the child understanding degree score is set, and when the child understanding degree score predicted by the feedback quality assessment model is lower than the threshold, a series of automatic processes are triggered. First, a subset related to the current emotional management scheme and with a lower cognitive complexity level is extracted from the standardized corpus. This subset contains simpler and more basic language expression patterns and emotional regulation cases. The psychological support model is retrained using these data, and the model's parameter weights are adjusted to optimize its language generation and emotional regulation guidance capabilities in low cognitive complexity scenarios. During the retraining process, the model's performance indicators, such as loss value and accuracy on the validation set, are monitored to ensure that the model effectively learns new knowledge and expression patterns. At the same time, the validated and optimized expression patterns are incrementally stored in the standardized corpus, forming a knowledge enhancement cycle. When storing, the newly added expression patterns are annotated and archived according to the organization structure and metadata annotation specifications of the corpus, including recording information such as the corresponding emotional management theme, cognitive complexity level, and applicable child age range, to facilitate subsequent model training and corpus query usage.

[0094] Optionally, in the above technical solution, further comprising: S40, constructing and training the fourth preset deep learning model to obtain a gene variation semantic understanding model, specifically comprising the following steps: S400, a large amount of rare disease patient gene data and corresponding clinical manifestations, disease-related literature and other information are collected. These gene data include gene sequences, gene variation types (such as single nucleotide polymorphism, insertion / deletion, etc.), gene expression levels, etc. The gene data is cleaned to remove low-quality or incomplete data and standardized to unify the data format. At the same time, the gene variation is annotated to label its semantic information such as association with rare diseases, variation site, and variation impact, forming a labeled training data set.

[0095] S401, a graph neural network (GNN) is selected to construct the fourth preset deep learning model. Gene data can be represented as a graph structure, where genes are nodes and interactions and regulatory relationships between genes are edges. GNN can effectively process graph structure data and capture complex relationships and semantic information between genes. The model structure includes an input layer, multiple graph neural network layers, and an output layer. The input layer receives the preprocessed gene data graph structure; the graph neural network layer propagates and aggregates information between nodes through a message passing mechanism, learns the embedding representation of gene nodes, and captures the semantic features of gene variations; the output layer outputs the semantic understanding result of the association between gene variations and rare diseases, such as the impact of the variation on the disease and the correlation between the variation and a specific disease. It should be noted that other deep learning models can also be selected as the fourth preset deep learning model according to actual conditions.

[0096] S402, input the labeled gene data into the constructed GNN model for training. During the training process, appropriate loss functions are used, such as mean square error loss function for mutation impact prediction task, cross-entropy loss function for disease correlation classification task, etc. Random gradient descent (SGD) or Adam optimization algorithm is used to optimize model parameters, by adjusting learning rate, batch size, number of graph neural network layers and other hyperparameters, so that the model can quickly converge and improve the understanding of gene mutation semantics. During the training process, the model is evaluated using the validation set, and the accuracy, recall, F1 value and other indicators are used to measure the performance of the model in the gene mutation semantic understanding task. After multiple rounds of iterative training, when the performance of the model on the validation set reaches an ideal level and tends to be stable, the training is stopped, and the gene mutation semantic understanding model is obtained.

[0097] S41, analyze the gene data of rare disease children patients using the gene mutation semantic understanding model, and generate corresponding genetic risk assessment results, specifically: After the gene data of rare disease children patients (including gene sequence, known variation, etc.) is preprocessed in the same way as the training data, it is input into the trained gene mutation semantic understanding model. The model first represents the input gene data in a graph structure, constructing gene nodes and their relationship edges. The gene mutation semantic understanding model uses the gene mutation semantic features and gene interaction relationships learned during training to analyze the input gene data. It identifies the mutation sites in the genes and combines the semantic information of these mutation sites (such as mutation type, impact of mutation on protein function, etc.), as well as the interaction relationships between genes, to assess the degree of association between these mutations and rare diseases. For example, the model can analyze whether a certain gene mutation will cause changes in protein structure and function, thereby affecting the normal physiological function of the organism and increasing the risk of rare diseases. Based on the analysis results of the model, a corresponding genetic risk assessment report is generated. The report content includes the gene mutation situation carried by the rare disease children patients, the correlation between each mutation and a specific rare disease, the potential risk assessment of the mutation on the occurrence of the disease (such as high risk, medium risk, low risk, etc.), the possible genetic mode of the disease (such as autosomal dominant inheritance, autosomal recessive inheritance, etc.), and personalized health management recommendations based on genetic risk (such as whether further genetic testing is needed, consultation with a specialist, regular monitoring, etc.). These genetic risk assessment results can provide scientific basis for doctors, parents and others, helping them better understand the genetic risk of rare disease children patients, develop appropriate prevention and intervention measures, and improve the health management and quality of life of the children.

[0098] In the above embodiments, although the steps are numbered, it is only a specific embodiment given by the present application, and those skilled in the art can adjust the execution order of the steps according to the actual situation, which is within the protection scope of the present application. It can be understood that in some embodiments, part or all of the above embodiments can be included.

[0099] As shown in Figure 2 The embodiment of the present application is a rare disease management system for children based on artificial intelligence, which comprises a model construction and training module and a model application module. The model construction and training module is used to construct and train a first preset deep learning model to obtain an intelligent question and answer model. The model application module is used to use the intelligent question and answer model to obtain rare disease related data of rare disease children patients in a question and answer mode and analyze the data to generate a corresponding individual health management plan. Optionally, in the above technical solution, the model construction and training module is further used to construct and train a second preset deep learning model to obtain an emotion recognition model. The model application module is further used to use the emotion recognition model to obtain an emotion recognition result of the rare disease children patients, and generate a corresponding emotion management plan according to the emotion recognition result.

[0100] Optionally, in the above technical solution, the model construction and training module is further used to construct and train a third preset deep learning model in combination with a child language corpus to obtain a psychological support model. The model application module is further used to use the psychological support model to convert the expression content of the user into expression content suitable for children to understand when the user uses the generated emotion management plan to intervene in the rare disease children patients.

[0101] Optionally, in the above technical solution, the model construction and training module is further used to construct and train a fourth preset deep learning model to obtain a gene variation semantic understanding model. The model application module is further used to use the gene variation semantic understanding model to analyze the genetic data of the rare disease children patients to generate a corresponding genetic risk assessment result.

[0102] In another embodiment, the artificial intelligence-based rare pediatric disease management system of the present embodiment is based on pediatric medical needs, integrates AI intelligent question and answer, emotion portrait modeling, peer support, intelligent genetic counseling, and nationwide expert consultation linkage functions, and constructs a multi-dimensional collaborative intelligent guardian new ecology for rare pediatric diseases, specifically including: an intelligent question and answer system for pediatric rare disease emergency and first aid, a Children4Children (C4C) module, a psychological health precise support system, an intelligent navigation system for genetic counseling, and an expert cloud wisdom library linkage module. In the intelligent question and answer system for pediatric rare disease emergency and first aid, a question and answer knowledge base is constructed based on global authoritative guidelines and real-time updated evidence-based medical data; high-risk symptom recognition and automatic emergency treatment suggestion generation under emergency situations are realized; natural language input is supported to assist primary doctors in making quick judgments within the key treatment window. The C4C module uses the language expression system of adolescents to rewrite rare disease knowledge content and create a psychological support platform for "children to children"; it realizes knowledge and emotional interaction between sick children and adolescent volunteers under a safety mechanism, alleviates the sense of loneliness and cognitive anxiety of sick children, and uses the "child heart guarding child heart" method to improve social cognition and popularize rare disease education. The psychological health precise support system identifies the psychological health status of sick children based on emotion recognition and personalized psychological portrait technology; it provides personalized psychological intervention suggestions for sick children in combination with a "family-school-medical" trinity model; it realizes the transition from "passive symptom intervention" to "active psychological resilience construction". The intelligent navigation system for genetic counseling integrates a rich database of pediatric clinical phenotypes and genomic data; it uses a large model to understand and interpret the semantic of variant items in gene reports; it provides doctors with automatically generated diagnosis directions, examination suggestions, and genetic risk prompts. The expert cloud wisdom library linkage module is used to initiate rare disease consultation applications and automatically match relevant experts; it integrates expert diagnosis opinions for unified display, improves decision-making efficiency and authority, and provides equalized high-quality medical resources for primary doctors and patients in remote areas. The beneficial effects are: ① Whole-cycle guardianship: from emergency to rehabilitation, from body to mind, from medicine to family, to build a complete service loop; ② Multi-modal AI support: combining text understanding, semantic question and answer, graph reasoning, and psychological modeling; ③ Child language modeling optimization: introduce child language corpus in model training to ensure that content understanding and output are close to children's cognition; ④ Rare disease knowledge popularization: complex knowledge is expressed in child language through the C4C mechanism to realize mass communication and cognitive popularization.

[0103] It should be noted that the beneficial effects of the above-mentioned embodiment of the system for managing rare diseases of children based on artificial intelligence are the same as those of the above-mentioned method for managing rare diseases of children based on artificial intelligence, and will not be repeated here. In addition, the system provided in the above-mentioned embodiment is only exemplified by the division of the above-mentioned functional modules when realizing its functions, and in actual application, the above-mentioned functions can be completed by different functional modules according to the needs, that is, the system is divided into different functional modules according to the actual situation to complete all or part of the above-described functions. In addition, the system and method embodiments provided in the above-mentioned embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0104] An electronic device according to an embodiment of the present application includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements any of the above-mentioned methods for managing rare diseases of children based on artificial intelligence when executing the computer program.

[0105] A computer readable storage medium according to an embodiment of the present application has a computer program stored thereon, and the computer program is executed by a processor to implement any of the above-mentioned methods for managing rare diseases of children based on artificial intelligence.

[0106] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and represent the limitation of a specific order or sequence. The order of use of similar objects can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described.

[0107] Although the embodiments of the present application have been shown and described above, it should be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

Claims

1. A method for managing rare childhood diseases based on artificial intelligence, characterized in that, include: Construct and train the first preset deep learning model to obtain the intelligent question answering model; Using the intelligent question-and-answer model, rare disease-related data of children with rare diseases are obtained and analyzed in a question-and-answer manner to generate corresponding personalized health management plans.

2. The method for managing rare childhood diseases based on artificial intelligence according to claim 1, characterized in that, Also includes: Construct and train a second preset deep learning model to obtain an emotion recognition model; Using the emotion recognition model, the emotion recognition results of children with rare diseases are obtained, and corresponding emotion management plans are generated based on the emotion recognition results.

3. A method for managing rare childhood diseases based on artificial intelligence according to claim 1 or 2, characterized in that, Also includes: A psychological support model is obtained by constructing and training a third-preset deep learning model using a children's language corpus. When users use the generated emotion management plan to intervene in the rare disease children, a psychological support model is used to transform the user's expression into expression that is suitable for children to understand.

4. A method for managing rare childhood diseases based on artificial intelligence according to claim 1 or 2, characterized in that, Also includes: Construct and train the fourth preset deep learning model to obtain the gene variation semantic understanding model; We use a gene variation semantic understanding model to analyze the genetic data of children with rare diseases and generate corresponding genetic risk assessment results.

5. An artificial intelligence-based management system for rare childhood diseases, characterized in that, It includes a model building and training module and a model application module; The model building and training module is used to: build and train a first preset deep learning model to obtain an intelligent question answering model; The model application module is used to: utilize the intelligent question-and-answer model to obtain and analyze rare disease-related data of children with rare diseases in a question-and-answer manner, and generate corresponding personalized health management plans.

6. The artificial intelligence-based rare disease management system for children according to claim 5, characterized in that, The model building and training module is also used to: build and train a second preset deep learning model to obtain an emotion recognition model; The model application module is also used to: use the emotion recognition model to identify the emotion recognition results of children with rare diseases, and generate corresponding emotion management plans based on the emotion recognition results.

7. A rare childhood disease management system based on artificial intelligence according to claim 5 or 6, characterized in that, The model building and training module is also used to: build and train a third preset deep learning model in conjunction with a children's language corpus to obtain a psychological support model; The model application module is also used to: when a user uses the generated emotion management plan to intervene in the rare disease child patient, use the psychological support model to transform the user's expression into expression that is suitable for children to understand.

8. A rare childhood disease management system based on artificial intelligence according to claim 1 or 2, characterized in that, The model building and training module is also used to: build and train the fourth preset deep learning model to obtain a gene variation semantic understanding model; The model application module is also used to: analyze the genetic data of children with rare diseases using a gene variation semantic understanding model, and generate corresponding genetic risk assessment results.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the artificial intelligence-based method for managing rare childhood diseases as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the artificial intelligence-based method for managing rare childhood diseases as described in any one of claims 1 to 4.

Images