Chronic asthma off-hospital monitoring system based on large medical model
Through the chronic asthma exit monitoring system based on medical big models, combined with multimodal data for real-time monitoring and precise management, the problem of difficulty in timely detection of changes in chronic asthma patients after discharge is solved, and personalized health management and abnormal detection are achieved.
Patent Information
- Application Number
- CN202510320437.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-18
- Publication Date
- 2025-07-08
AI Technical Summary
Patients with chronic asthma lack continuous monitoring after discharge, which makes it difficult to detect changes in the disease in a timely manner, and the accuracy of self-report is low, which affects the timing of treatment.
The chronic asthma exit monitoring system based on medical big model is adopted, and through the input layer, data preprocessing and coding layer, feature fusion layer, dynamic adjustment layer and thinking chain reasoning layer, combined with text data, time series data and medical knowledge graph data, the doctor's diagnostic thinking process is simulated and personalized health management suggestions are generated.
Real-time monitoring and precise management of chronic asthma patients has been achieved, self-management ability and medical efficiency have been improved, the number of patients traveling to and from the hospital has been reduced, and the changes in the condition are detected in a timely manner and alarms have been sent.
Smart Images
Figure CN120280138A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical fields of medical monitoring and artificial intelligence, and particularly to a chronic asthma discharge monitoring system based on a medical large model. Background Art
[0002] With the booming development and remarkable progress in the fields of artificial intelligence technologies such as natural language processing, machine learning, and deep learning, artificial intelligence technologies have gradually been applied to the field of medical health monitoring technologies.
[0003] After being discharged from the hospital, the health status of chronic asthma patients still needs to be monitored. Currently, the daily monitoring of chronic asthma patients mainly relies on the patients' self-reports and regular hospital reexaminations. Due to the patients' objectively lacking medical knowledge and being affected by their subjective feelings, memory biases, or expression abilities, the accuracy of self-reports is low. At the same time, the lack of continuous monitoring after discharge makes it difficult to detect changes in the condition in a timely manner, which may lead to delays in the best treatment opportunity.
[0004] It is not difficult to see that for patients, during the period of leaving the hospital, how to achieve real-time monitoring and precise management of the patients' health status, improve the management efficiency of chronic asthma, ensure that patients receive timely and effective treatment, and thus help prevent the further deterioration of the condition is of great significance. Summary of the Invention
[0005] Aiming at the problems in the prior art, the present invention provides a chronic asthma discharge monitoring method and system based on a medical large model, which solves the problems such as low data accuracy, lack of continuous monitoring, and difficulty in timely detecting changes in the condition existing in the management of traditional chronic asthma patients.
[0006] The present invention aims to provide a more intelligent, precise, and real-time new solution to effectively improve the self-management ability and medical efficiency of chronic asthma patients, while reducing the number of times patients travel to and from the hospital and alleviating the medical burden.
[0007] The present invention provides a chronic asthma discharge monitoring system based on a medical large model, mainly including:
[0008] The input layer is used to input text data, time series data, medical knowledge graph data, and prompt words; the data preprocessing and encoding layer is used to preprocess the input data and encode the preprocessed data. The encoding includes text encoding, time series encoding, and knowledge graph encoding; the feature fusion layer is used to fuse the encoded features of the text data, time series data, and medical knowledge graph data using an attention mechanism; the dynamic adjustment layer is used to extract specific features from the fused multimodal data and calculate dynamic weights based on historical data and current data to weight the extracted specific features; the thought chain reasoning layer uses a Transformer model for reasoning, simulates the doctor's diagnostic thought process, and gradually conducts reasoning based on the disease condition score, risk level, and prompt words as guidance to obtain the disease condition score and risk level; the output layer is used to generate personalized health management suggestions according to the disease condition score and risk level. The suggestions include medication adjustment suggestions, lifestyle change suggestions, and follow-up plan suggestions.
[0009] Further, the input layer includes: text data, which includes the patient's symptom descriptions, medical records, etc.; time series data, which includes the patient's medication usage, physiological indicators, etc.; medical knowledge graph data, which includes entity and relationship data related to chronic asthma; and prompt words, which are designed in advance according to medical diagnosis logic and are used to guide the model to conduct reasoning.
[0010] Further, the data preprocessing and encoding layer includes: a data preprocessing module, which is used to clean the text data, standardize the time series data, and update the medical knowledge graph data; a text encoder, which uses a pre-trained BERT model to encode the text data; a time series encoder, which uses an LSTM model to encode the time series data; and a knowledge graph encoder, which uses a graph neural network GCN model to encode the entity and relationship data in the medical knowledge graph data.
[0011] Further, the feature fusion layer uses a dot product-based attention mechanism to perform weighted fusion on the text data, time series data, and medical knowledge graph data by calculating attention weights to obtain a fused feature representation.
[0012] Furthermore, the dynamic adjustment layer includes: a feature extraction module for extracting specific features from the fused multi-modal data, where the specific features include patient symptom features corresponding to historical data, medication compliance features, etc.; a dynamic weight calculation module for calculating dynamic weights based on historical data and current data using a preset formula; a weighted feature extraction module for performing weighted calculation on the extracted specific features according to the dynamic weights; and an anomaly detection module for detecting whether the features of the current data are abnormal, and if an anomaly is detected, dynamically increasing the weights of the corresponding features using an anomaly factor.
[0013] Furthermore, the thought chain reasoning layer includes a score calculation module and a logical reasoning module; the score calculation module is used to calculate the final disease score based on data such as the patient's symptoms, medication status, physiological indicators, etc.; the logical reasoning module uses a Transformer model for multi-step reasoning, and based on the disease score, risk level, and prompt words as inputs, gradually performs reasoning through operations such as multi-head attention mechanism and residual connection, and outputs the reasoning result.
[0014] Furthermore, the scores calculated by the score calculation module include: asthma control level score, ACT score conversion score, medication and follow-up situation score, main symptom and allergy history / complication score, and the disease score is calculated based on the above scores and corresponding dynamic weights.
[0015] Furthermore, the output layer generates personalized health management suggestions according to the disease score and risk level, in combination with preset thresholds and rules. The suggestions include: judging whether it is necessary to adjust medications and change lifestyles according to the disease score and risk level, and giving specific medication adjustment suggestions and lifestyle change suggestions; giving personalized follow-up plan suggestions according to the disease score and risk level, in combination with follow-up rules.
[0016] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:
[0017] 1. Real-time monitoring and precise management: Through the medical large model and thought chain reasoning technology, real-time monitoring and precise management of chronic asthma patients are achieved, improving the patient's self-management ability and medical efficiency.
[0018] 2. Personalized health management suggestions: Through dynamic adjustment and multi-modal data fusion, personalized health management suggestions are generated, including medication adjustment, lifestyle change, and follow-up plan, which helps patients better manage their conditions.
[0019] 3. Anomaly detection and timely alert: Through the anomaly detection module, changes in the patient's condition are detected in a timely manner, and an alert is sent through the notification module to ensure that patients and doctors can take measures in a timely manner to avoid the deterioration of the condition. Brief Description of the Drawings
[0020] Figure 1 This is the model framework diagram of the chronic asthma out-of-hospital monitoring system based on the medical large model in the present invention. Detailed Description of the Invention
[0021] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments, but the protection scope of the present invention is not limited to the content described.
[0022] The purpose of the present invention is to provide a chronic asthma out-of-hospital monitoring system based on a medical large model to solve the above problems existing in the prior art.
[0023] Refer to Figure 1 , the chronic asthma out-of-hospital monitoring system based on the medical large model in the embodiment of the present invention includes:
[0024] An input layer, a data preprocessing and encoding layer, a feature fusion layer, a dynamic adjustment layer, a thought chain reasoning layer, and an output layer;
[0025] The input layer is used to input text data, time series data, medical knowledge graph data, and prompt words;
[0026] The data preprocessing and encoding layer is used to preprocess the input data and encode the preprocessed data, including text encoding, time series encoding, and knowledge graph encoding;
[0027] The feature fusion layer uses the attention mechanism to fuse the encoded features of text data, time series data, and medical knowledge graph data;
[0028] The dynamic adjustment layer extracts specific features from the fused multimodal data, and calculates dynamic weights based on historical data and current data to weight the extracted specific features;
[0029] The thought chain reasoning layer uses the Transformer model for reasoning, simulates the doctor's diagnosis thinking process, and gradually conducts reasoning based on the calculated disease condition score, risk level, and prompt words as guidance to obtain the disease condition score and risk level;
[0030] The output layer generates personalized health management suggestions according to the disease condition score and risk level, including medication adjustment suggestions, lifestyle change suggestions, and follow-up plan suggestions.
[0031] Each layer in the chronic asthma out-of-hospital monitoring system of the present invention will be further described one by one as follows:
[0032] 1) Input layer:
[0033] The text data includes text data such as the patient's symptom descriptions and medical records; the time series data includes the patient's medication usage, physiological indicators, etc.; the medical knowledge graph data includes entity and relationship data, where entities refer to things with clear identities, such as diseases, drugs, symptoms, patients, medical facilities, etc. Here, the entities related to chronic asthma may include: asthma, bronchodilators, cough, wheezing, etc. The relationship data between entities refers to the associations between entities, such as the relationship between diseases and symptoms, the relationship between drugs and diseases, etc.; the prompt words are designed in advance according to the medical diagnosis logic and are used to guide the model to make inferences.
[0034] Specifically, the patient inputs information such as symptoms and medication usage through an application on the mobile device, and sensors on the mobile device or wearable device collect the patient's physiological index data, such as heart rate, blood oxygen saturation, respiratory rate, etc.
[0035] 2) Data preprocessing and encoding layer:
[0036] Data preprocessing includes the cleaning of text data, the standardization of time series data, and the update of medical knowledge graph data. The preprocessed data is input into the corresponding encoders for encoding to obtain the corresponding encoded features.
[0037] Specifically, the data encoding layer includes a text encoder, a time series encoder, and a knowledge graph encoder. The text encoder uses the pre-trained BERT model to encode the text data, the time series encoder uses the LSTM model to encode the time series data, and the knowledge graph encoder uses the graph neural network GCN model to encode the entity and relationship data in the medical knowledge graph data.
[0038] The text data is encoded as text data by the text encoder, The time series data is encoded as time series data by the time series encoder, The medical knowledge graph data is encoded as medical knowledge graph data by the knowledge graph encoder,
[0039] 3) Feature fusion layer:
[0040] Feature fusion specifically performs the following operations:
[0041] 1. Attention weight calculation: Calculate the attention weight α using dot product-based attention ij , and its calculation formula is:
[0042] Where: Q i is the query vector, K j is the key vector, and i, j represent the indices of different modality data.
[0043] 2. Feature Fusion: Apply the attention weight to the value vector to obtain the fused feature representation H fused : The expression is as follows: where: V j represents the value vector, and n, m, and k respectively represent the number of rows of the text data matrix, the time series feature matrix, and the medical knowledge graph data matrix.
[0044] 4) Dynamic Adjustment Layer:
[0045] In the present invention, the dynamic adjustment layer is used to perform operations such as feature extraction, dynamic weight adjustment, weighted feature extraction, and weighted abnormal data, as follows:
[0046] Feature Extraction: Extract specific features from the fused multi-modal data. Specifically, it includes the patient symptom features of the current input corresponding to the historical data, such as: ACT score conversion features, medication and follow-up situation features, main symptom and allergy history / complication features, etc.; medication compliance features.
[0047] Dynamic Weight Adjustment: Based on historical data and current data, calculate the dynamic weight using a clearly defined formula (see 4.1 - 4.3 below).
[0048] Weighted Feature Extraction: For multi-modal features, the dynamic weight is used for weighting:
[0049] where: H i represents the specific features extracted above, w i represents the weight of the feature, and n represents the number of specific features.
[0050] The role of weighted feature extraction is to re-weight the specific features extracted from the fused multi-modal data according to the calculated weights, so as to more accurately reflect the importance of different features in the overall data evaluation, making subsequent analysis and decision-making more scientific and reasonable.
[0051] Weighted Abnormal Data: If an abnormal pattern is detected, use the enhancement factor β:
[0052] where: w i represents the weight of the specific feature, and β is a constant.
[0053] When it is detected that the features of the current data are significantly different from the features of the historical data or the features of common data of the same type, the weight parameters of this part of the features need to be dynamically increased.
[0054] In order to better combine the user's current data and historical records, this patent also designs a set of scores suitable for dynamic adjustment, as follows:
[0055] 4.1 Adjustment of Scoring Weights Based on Historical Data:
[0056] For each scoring item, we define a dynamic weight w dynamic , where:
[0057] w dynamic = f(historical data)
[0058] The specific calculation formula is as follows:
[0059] If a certain symptom appears in more than half of the historical data of a patient, the system will add a dynamic weight to this item.
[0060] For example, if in the past three follow - ups, the patient reported the symptom of waking up breathless at night twice, the dynamic weight can be defined as:
[0061]
[0062] If the record of waking up breathless at night is 2 times in the past three follow - ups, then:
[0063]
[0064] When calculating the final score, this dynamic weight is used for the scoring of this item:
[0065] Score for waking up breathless at night = Score × w dynamic，夜间憋醒
[0066] 4.2 Personalized Scoring of Medication Adherence:
[0067] For medication adherence, a compliance weight w adherence :
[0068] If the number of times of irregular medication use by the patient in history exceeds half, it can be defined as:
[0069] w adherence = 1 + 0.2 × number of historical records of irregular medication use
[0070] The score for the medication adherence item is then:
[0071] Score for medication adherence = Score × w adherence
[0072] 4.3 Weighted Adjustment of Abnormal Data:
[0073] Add a weight to the abnormal data, and define the abnormal weight w abnormal . If the system detects an abnormal situation (such as an increase in the SABA usage frequency), then multiply the score of this item by a weighting coefficient:
[0074] wabnormal = 1 + 0.3 × Abnormal frequency
[0075] Suppose that in the past follow - up, the abnormal frequency of SABA use occurred 2 times, then:
[0076] w abnormal = 1 + 0.3 × 2
[0077] Then, the score of SABA use frequency is:
[0078] SASB score = Score × w abnormal
[0079] 4.4 Dynamic assessment threshold adjustment:
[0080] By dynamically adjusting the alarm threshold T based on historical data alert for dynamic adjustment
[0081]
[0082] The default alarm threshold T base = 15, and if the patient is recorded as high - risk in 3 out of the past 5 follow - ups, then:
[0083]
[0084] When the total score exceeds 13.8, the system will trigger an alarm instead of the default 15 points.
[0085] The calculation formula for the total score:
[0086] TotalScore = C + A + M + S
[0087] Based on the total score, the disease level can be classified:
[0088] 0 - 4 points: Good condition.
[0089] 5 - 10 points: Need attention.
[0090] 11 - 15 points: Medium risk.
[0091] 16 points and above: Uncontrolled condition or high risk, urgent intervention required.
[0092] The reasoning process in the thought chain will compare the total score with the grading criteria to determine the patient's disease status and provide personalized management suggestions or alarms according to different disease grades.
[0093] 5) Thought reasoning layer:
[0094] The thinking and reasoning layer includes two parts: scoring calculation and logical reasoning. The scoring calculation is specifically as follows: Calculate the asthma control level score, ACT score conversion score, medication and follow-up situation score, main symptom and allergy history / complication score based on data such as the patient's symptoms, medication use, and physiological indicators, and calculate the final disease condition score based on these scores.
[0095] The scoring calculation formula is specifically as follows: C final = w c C + w A A + w M M + w S S, where: C represents the asthma control level score; A represents the ACT score conversion score; M represents the medication and follow-up situation score; S represents the main symptom and allergy history / complication score, and w c 、w A 、w M 、w S are the respective dynamic weights, and the dynamic weights here are the weights calculated by the dynamic adjustment layer.
[0096] The specific calculation methods of the above various evaluation scores are as follows:
[0097] First, the asthma control level score C:
[0098] C = daytime asthma symptom score + nighttime waking up due to breathlessness score + SABA use score + activity limitation score
[0099] Among them:
[0100] Daytime asthma symptom score: If there are asthma symptoms more than 2 times a week in the past 4 weeks, count 1 point, otherwise count 0 point.
[0101] Nighttime waking up due to breathlessness score: If waking up due to asthma at night in the past 4 weeks, count 1 point, otherwise count 0 point.
[0102] SABA use score: If using SABA more than 2 times a week in the past 4 weeks, count 1 point, otherwise count 0 point.
[0103] Activity limitation score: If there is activity limitation due to asthma in the past 4 weeks, count 1 point, otherwise count 0 point.
[0104] Second, the ACT score conversion score A:
[0105]
[0106] Third, the medication and follow-up situation score M:
[0107] M = medication situation score + medication awareness score + outpatient follow-up score + regular review score
[0108] Among them:
[0109] Medication situation score: If the medication is irregular, 2 points are counted; otherwise, 0 points are counted.
[0110] Drug awareness score: If the usage method and precautions of the drug are not known, 1 point is counted; otherwise, 0 points are counted.
[0111] Outpatient follow-up score: If regular outpatient follow-up is not carried out, 1 point is counted; otherwise, 0 points are counted.
[0112] Regular review score: If there is no regular review or the recent results are abnormal, 1 - 2 points are counted (depending on the specific situation).
[0113] Fourth, the scores of the main symptoms and allergy history / complications S:
[0114] S = symptom score + complication score
[0115] Among them:
[0116] Symptom score: 1 point is counted for each symptom (such as cough, wheezing, shortness of breath, chest tightness, etc.).
[0117] Complication score: If there are serious complications (such as respiratory failure, chronic pulmonary heart disease, etc.), 2 points are counted; otherwise, 0 points are counted.
[0118] Furthermore, the corresponding risk level RiskLevel is divided according to the calculated disease score, and the specific division method is as follows:
[0119]
[0120] The process of logical reasoning is as follows:
[0121] 5.1 Input data processing: During the reasoning process, the disease score, risk level, and prompt word information input into the Transformer model are processed into an embedding vector X input , and its expression is:
[0122] X input = [C final , RiskLevel, PromptWords], where: C final is the disease score, RiskLevel is the risk level determined by the score, and PromptWords are the prompt words that the model needs to refer to for disease reasoning.
[0123] 5.2 Use the Transformer multi - head attention mechanism for step - by - step reasoning, including:
[0124] Calculation of query, key, and value: For each input feature (such as disease condition score, risk level, prompt word), query Q, key K, and value V are obtained through linear transformation respectively. The calculation formula is: Q = X input W q , K =
[0125] X input W k , V = X input W v , where: W q , W k , W v are the parameter matrices learned by the model during training.
[0126] Calculation of attention weights: Calculate the correlation between the query and the key to obtain the attention weight α ij . Note that the attention weight here is different from the attention weight in the feature fusion layer. Note the attention weight in the feature fusion layer: It aims to fuse multi-modal data into a unified feature representation, providing a basis for the subsequent model to process comprehensive information and enhancing the model's ability to understand and process complex information. The attention weight of the Transformer multi-head attention mechanism: Its main goal is to perform feature screening and weighted aggregation within the input features, highlighting the influence of key features on the inference result, helping the model better capture important information in the input features, and improving the accuracy and effectiveness of inference:
[0127] where: d is the dimension of the vector. Here, d is the dimension of the K (or Q) vector. In a high-dimensional vector space, the result of the dot product may become very large, which will cause the gradient of the softmax function to become very small, making it difficult for the model to train. And introducing the scaling factor can scale the dot product result to an appropriate range, avoid the problem of gradient disappearance, and enable the model to learn the correlation between features more stably. The essence of this calculation is to measure the similarity between each query and all keys, and convert the similarity into a probability distribution through the softmax function, that is, the attention weight. The attention weight reflects the importance of each value under the current query.
[0128] Calculation of weighted value: According to the calculated attention weight α ij , perform weighted summation on the value V j to obtain the output Z of the current layer: This step is to apply the attention weights obtained from the previous calculation to the values. Through weighted summation, the values related to the current query are highlighted, thereby realizing the screening and aggregation of the input features and obtaining a more representative output. The calculation of the query, key, and value is the basis. Through them, the attention weights are calculated, which in turn guide the calculation of the weighted values. The whole process is closely linked. The purpose is to enable the model to focus on the key information in the input features and provide a more effective feature representation for subsequent reasoning.
[0129] 5.3 Multi-layer stacking and residual connections
[0130] In the present invention, the deep reasoning of the Transformer model is carried out by means of multi-layer stacking. Each layer includes a self-attention mechanism and a feed-forward neural network (Feed-Forward Network). The output of each layer is passed to the next layer through residual connections and layer normalization, and its expression is:
[0131] X l+1 = LayerNorm(Z l + X l ), where: X l is the input of the l-th layer, and Z l is the output of the l-th layer.
[0132] 5.4 Inference output
[0133] After multiple layers of reasoning, the final output is the inference result of the model, which is used to calculate the disease score and recommend a treatment plan. This result undergoes a linear transformation and an activation function to generate the inference result R (final disease inference), which represents the final disease score:
[0134] R = sigmoid(W o Z L + b o ),
[0135] where Z L is the output of the last layer, and W o and b o are the weights and biases of the output layer.
[0136] The inference result R here is the disease score of the final patient, which is a numerical score. The risk level RiskLevel is obtained according to the following formula.
[0137]
[0138] 6) Output layer:
[0139] The output layer will generate personalized health management suggestions based on the inferred disease score and risk level inference module. These suggestions include:
[0140] Medication adjustment:
[0141] MedicationAdvice = f med (R, RiskLevel)
[0142] For example, based on the inferred disease score and risk level, advice such as whether to adjust medications, increase medication dosage, etc. is given.
[0143] Lifestyle changes:
[0144] LifestyleAdvice = f life (R, RiskLevel)
[0145] Based on the severity of the disease, advice on whether to change lifestyle (such as diet, exercise, etc.) is given.
[0146] Follow-up plan:
[0147] Follow-upAdvice = f follow (R, RiskLevel)
[0148] A follow-up plan for the patient, such as the frequency of reexamination, etc., is given.
[0149] Finally, the content output by the inference module includes:
[0150] Disease score R;
[0151] Personalized health management advice (medication, lifestyle, follow-up, etc.);
[0152] Risk level (determined by the disease score).
[0153] FinalOutput = {R, MedicationAdvice, LifestyleAdvice, Follow-upAdvice, RiskLevel}.
[0154] The present invention discloses a method for the condition assessment and management of chronic asthma patients based on multi-modal data. This method obtains the text, time series, and medical knowledge graph data of patients, and after preprocessing, uses the BERT, LSTM, and GCN models respectively for feature extraction. The attention mechanism is used to fuse multi-modal features, and the feature weights are dynamically adjusted by combining the patient's historical and current data. The Transformer model is innovatively introduced to simulate the doctor's diagnostic thinking for condition scoring and risk level assessment. Finally, according to the assessment results, personalized suggestions for medication adjustment, lifestyle changes, and follow-up plans are generated. The present invention realizes the accurate assessment and personalized management of chronic asthma patients, effectively improves the accuracy and efficiency of chronic disease management, and provides more scientific and reasonable health guidance for patients.
[0155] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and details without departing from the scope defined by the claims of the present invention.
Claims
1. A chronic asthma discharge monitoring system based on a medical large model, characterized in that, It includes: An input layer for inputting text data, time series data, medical knowledge graph data, and prompt words; A data preprocessing and encoding layer for preprocessing the input data and encoding the preprocessed data. The encoding includes text encoding, time series encoding, and knowledge graph encoding; A feature fusion layer for using an attention mechanism to fuse the encoded features of the text data, time series data, and medical knowledge graph data; A dynamic adjustment layer for extracting specific features from the fused multi-modal data and calculating dynamic weights based on historical data and current data to weight the extracted specific features; A thought chain reasoning layer that uses a Transformer model for reasoning, simulates the doctor's diagnostic thinking process, and gradually conducts reasoning based on the disease condition score, risk level, and prompt words as guidance to obtain the disease condition score and risk level; An output layer for generating personalized health management suggestions according to the disease condition score and risk level. The suggestions include medication adjustment suggestions, lifestyle change suggestions, and follow-up plan suggestions.
2. The system according to claim 1, wherein, The input layer includes: Text data, which includes the patient's symptom description and medical records; Time series data, which includes the patient's medication situation and physiological indicators; Medical knowledge graph data, which includes entity and relationship data related to chronic asthma; And prompt words, which are designed in advance according to medical diagnosis logic and are used to guide the model to conduct reasoning.
3. The system according to claim 1, wherein, The data preprocessing and encoding layer includes: A data preprocessing module for cleaning the text data, standardizing the time series data, and updating the medical knowledge graph data; A text encoder that uses a pre-trained BERT model to encode the text data; A time series encoder that uses an LSTM model to encode the time series data; A knowledge graph encoder that uses a graph neural network GCN model to encode the entity and relationship data in the medical knowledge graph data.
4. The system according to claim 1, wherein The feature fusion layer uses a dot product-based attention mechanism to weight and fuse the text data, time series data, and medical knowledge graph data by calculating attention weights to obtain a fused feature representation.
5. The system according to claim 1, wherein The dynamic adjustment layer includes: A feature extraction module for extracting specific features from the fused multi-modal data. The specific features include patient symptom features and medication compliance features corresponding to historical data; A dynamic weight calculation module for calculating dynamic weights based on historical data and current data using a preset formula; A weighted feature extraction module for performing weighted calculation on the extracted specific features according to the dynamic weights; An anomaly detection module for detecting whether the features of the current data are abnormal. If an anomaly is detected, the weight of the corresponding feature is dynamically increased using an anomaly factor.
6. The system according to claim 1, wherein The thought chain reasoning layer includes a score calculation module and a logical reasoning module; The score calculation module is used to calculate the final disease condition score according to the patient's symptoms, medication situation, and physiological indicators; The described logical reasoning module uses a Transformer model for multi-step reasoning. Based on the disease condition score, risk level, and prompt words as inputs, it gradually conducts reasoning through multi-head attention mechanisms and residual connection operations, and outputs the reasoning results.
7. The system according to claim 6, wherein The scores calculated by the described score calculation module include: Asthma control level score, ACT score conversion score, medication and follow-up situation score, main symptom and allergy history / complication score.
8. The system according to claim 1, characterized in that, The output layer generates personalized health management suggestions according to the disease condition score and risk level, in combination with preset thresholds and rules. The suggestions include: Based on the disease condition score and risk level, determine whether drug adjustment or lifestyle change is needed, and give specific drug adjustment suggestions and lifestyle change suggestions; Based on the disease condition score and risk level, in combination with the follow-up rules, give personalized follow-up plan suggestions.