Slow obstructive pulmonary disease patient leaving hospital monitoring system

Abstract

The invention relates to the technical field of medical health monitoring and neural networks, and provides a chronic obstructive pulmonary disease patient leaving hospital monitoring system which comprises the steps that multi-source data of a patient is acquired, and the multi-source data at least comprises daily health monitoring data and electronic medical record data; preprocessing and feature extraction are conducted on the multi-source data through a data processing module, health feature data of the patient are obtained, and the health feature data are features related to the chronic obstructive pulmonary disease condition selected from the multi-source data and comprise features based on respiratory functions, features based on lifestyles and behaviors and features based on complications and complications; and features based on drug use; integrating and updating the health feature data through a patient health condition evaluation module, and performing clustering analysis on the integrated and updated real-time health data by adopting a density-based clustering algorithm DBSCAN to obtain a health portrait of the patient; the health portrait is analyzed through a medical large model analysis module, an illness state evaluation result and a health management suggestion are obtained, and a deep learning model designed by combining a convolutional neural network CNN with a multi-modal model architecture of an attention mechanism Transform is used in the medical large model analysis module; remote monitoring, interactive communication and personalized intervention operation are executed through the remote medical intervention module according to the illness state evaluation result.

Description

Technical Field

[0001] The present invention relates to the technical fields of medical health monitoring and neural networks, and particularly to a monitoring system for discharged patients with chronic obstructive pulmonary disease (COPD). Background Art

[0002] Chronic obstructive pulmonary disease (COPD) is a common respiratory disease, and patients often need to receive regular examinations and treatments in the hospital. The health management of COPD patients faces challenges such as fragmented data, inaccurate assessment, and untimely intervention. Traditional methods often rely on a single data source and are difficult to comprehensively grasp the patient's health status. At the same time, due to the complex and changeable condition of COPD, it is difficult to accurately judge the development trend of the disease based on a simple assessment model. In addition, the existing health management models are mostly passive follow-up, unable to achieve real-time monitoring and timely intervention. These problems lead to poor health management effects for COPD patients and increase the risk of disease deterioration and acute exacerbation. How to integrate multi-source heterogeneous data to construct a comprehensive patient health profile? How to use advanced artificial intelligence technology to improve the accuracy of condition assessment? How to establish an active health management mechanism to achieve timely intervention? The solution to these problems is of great significance for improving the quality of life of COPD patients and reducing medical costs. However, in practical applications, how to balance the comprehensiveness of data collection and patient privacy protection, how to ensure the stability and interpretability of the AI model in a complex and changeable clinical environment, and how to design a scientific and user-friendly intervention mechanism are all technical problems that need to be deeply explored. Summary of the Invention

[0003] In order to improve the efficiency and quality of monitoring discharged COPD patients, reduce the number of times patients travel back and forth to the hospital, and improve the efficiency of disease management, the present invention provides a monitoring system for discharged COPD patients, mainly including:

[0004] Obtain multi-source data of the patient, where the multi-source data at least includes daily health monitoring data and electronic medical record data; preprocess and extract features from the multi-source data through a data processing module to obtain the patient's health feature data. The health feature data is the features related to the condition of chronic obstructive pulmonary disease (COPD) selected from the multi-source data, including: features based on respiratory function, features based on lifestyle and behavior, features based on complications and comorbidities, and features based on drug use; integrate and update the health feature data by a patient health status assessment module according to the health feature data, and perform clustering analysis on the integrated and updated real-time health data using the density-based spatial clustering of applications with noise (DBSCAN) algorithm. The DBSCAN algorithm can identify clusters with arbitrary shapes and can effectively handle noise and outliers; analyze the health portrait through a medical large model analysis module to obtain a disease condition evaluation result and health management suggestions. The medical large model analysis module uses a deep learning model designed with a multi-modal model architecture that combines a convolutional neural network (CNN) and an attention mechanism Transformer; perform remote monitoring, interactive communication, and intervention operations according to the disease condition evaluation result through a remote medical intervention module. The remote monitoring includes real-time monitoring and abnormal alarm, the interactive communication includes online consultation and video conferencing, and the personalized intervention includes treatment plan adjustment and health education.

[0005] Further, the preprocessing and feature extraction of the multi-source data by the data processing module includes: obtaining structured data and unstructured data from the multi-source data. The structured data includes the patient's age, gender, smoking history, respiratory rate, heart rate, blood oxygen saturation, weight, and exercise volume. The unstructured data includes the patient's medication record, past medical history, laboratory test results, and imaging examination reports; perform data cleaning operations on the structured data and the unstructured data to obtain first-processed data. The data cleaning operations include data deduplication, filling missing values, and outlier processing; perform attribute numericalization and standardization processing on the first-processed data to obtain second-processed data. The attribute numericalization uses label encoding or one-hot encoding, and the standardization uses mean and standard deviation calculation; select features related to the condition of COPD from the second-processed data through a feature selection algorithm to obtain the health feature data. The feature selection algorithm includes chi-square test. The larger the chi-square statistic value in the chi-square test, the stronger the correlation between the feature and the condition of COPD.

[0006] Further, the features based on respiratory function include the ratio change R of the forced expiratory volume in one second (FEV1) to the forced vital capacity (FVC) and the overall respiratory function (ORF) index. The overall respiratory function ORF index is a comprehensive index of oxygen therapy time, finger pulse oxygen saturation, and respiratory function exercise; the calculation formula for the ratio change R is: where (FEV1 / FVC) follow-up is the ratio of the forced expiratory volume in the first second (FEV1) to the forced vital capacity (FVC) measured during the follow-up period (such as after treatment or after rehabilitation intervention), and (FEV1 / FVC) baseline is the ratio of the forced expiratory volume in the first second (FEV1) to the forced vital capacity (FVC) measured during the baseline period (before treatment or intervention); the calculation formula for the overall respiratory function ORF index is: where, T oxygen is the oxygen therapy time, T max is the maximum oxygen therapy time, SpO2 is the fingertip pulse oxygen saturation, and E exercise indicates whether respiratory function exercise is performed, taking 1 for exercise and 0 for no exercise. ω1, ω2, and ω3 are the weights of oxygen therapy time, fingertip pulse oxygen saturation, and respiratory function exercise respectively; the characteristic based on lifestyle and behavior is the smoking index SI, which is defined as the product of the number of years of smoking and the daily cigarette consumption, and is used to evaluate the impact of smoking on COPD. The higher the value of the smoking index SI, the more serious the cumulative damage of smoking to the lungs; the characteristic based on complications and comorbidities is the comorbidity score CCI, which evaluates the long-term health of patients by assigning different weights to different diseases; the calculation formula for the comorbidity score CCI is: where C i is the i-th comorbidity, ω i is the weight assigned to each comorbidity according to its impact on the patient's health, and n is the total number of comorbidities; the characteristics based on drug use include long-term medication adherence LMA and the frequency of use of inhaled drugs FIU. The characteristic definition of long-term medication adherence LMA evaluates whether the patient takes medicine on time and follows the prescribed dose and frequency, and the characteristic definition of the frequency of use of inhaled drugs FIU is to record the frequency of the patient's use of short-acting or long-acting inhaled drugs.

[0007] Further, the patient health status evaluation module integrates and updates the health feature data, and uses the density-based clustering algorithm DBSCAN to perform clustering analysis on the integrated and updated real-time health data to obtain the patient's health portrait, including: integrating the health feature data through a data integration unit to obtain integrated data; updating the integrated data through a real-time update mechanism to obtain real-time health data, and the real-time update mechanism is used to automatically update the patient's health status to ensure that the information in the system is consistent with the patient's current actual situation when new health feature data is generated. The update process of the real-time update mechanism includes data interface input data, cache and database management, data processing, automatic reminder mechanism, and regular summary and reporting; using the density-based clustering algorithm DBSCAN through a health portrait construction unit to perform clustering analysis on the real-time health data to obtain the patient's health portrait, and generating descriptive labels according to the health portrait, and the descriptive labels are used to characterize the severity of the patient's condition, including mild chronic obstructive pulmonary disease patients, moderate chronic obstructive pulmonary disease patients, and severe chronic obstructive pulmonary disease patients.

[0008] Further, the deep learning model includes an input layer, a data preprocessing layer, an encoding layer, a feature fusion layer, a thinking logic reasoning layer, and an output layer. Analyzing the health portrait through the medical large model analysis module includes: inputting the image data, text data, time series data, and knowledge graph in the health portrait through the input layer of the deep learning model; preprocessing the image data, text data, time series data, and knowledge graph respectively through the data preprocessing layer; encoding the image data through the convolutional neural network CNN encoder in the encoding layer to obtain the spatial features of the image, and encoding the text data and time series data through the attention Transformer encoder to capture the long-range dependencies between features through the self-attention mechanism (that is, the correlation between elements that are far apart in the data, such as cross-sentence references in text or the impact of early events on the future in time series), and encoding the structural information of the knowledge graph through the knowledge graph encoder to extract the relationship features between entities such as symptoms, causes, and drugs; using the self-attention mechanism through the feature fusion layer to splice and fuse the feature data extracted in the encoding layer to obtain fused feature data; performing logical reasoning on the fused feature data based on the medical knowledge and logical reasoning rules in the knowledge graph through the thinking logic reasoning layer to obtain the disease condition features of patient classification, and performing anomaly detection on the fused feature data to predict the risks of different diseases to obtain a prediction result, and finally obtaining the disease condition evaluation result; classifying the disease condition features of patient classification through the output layer of the deep learning model and outputting the corresponding descriptive labels, converting the prediction result into a disease probability and outputting it, and combining the descriptive labels, disease probability, and medical advice rules in the knowledge graph to output health management opinions.

[0009] The technical solution provided by the embodiment of the present invention may include the following beneficial effects:

[0010] The present invention discloses a monitoring system for chronic obstructive pulmonary disease (COPD) patients after discharge. First, by obtaining the patient's daily health monitoring data and the electronic medical record data of the hospital information system, and at the same time selecting the features related to the COPD condition from multi-source data, including: features based on respiratory function, features based on lifestyle and behavior, features based on complications and comorbidities, and features based on drug use, to ensure the comprehensiveness and accuracy of the data. Further integrate and update the health feature data, and use the density-based clustering algorithm DBSCAN to perform clustering analysis on the integrated and updated real-time health data to effectively process noise points and outliers and improve data accuracy, providing reliable data support for subsequent condition evaluation and diagnosis. Secondly, use a deep learning model designed with a multi-modal model architecture that combines the convolutional neural network (CNN) and the attention mechanism Transformer, and combine it with a knowledge graph for condition evaluation and diagnostic suggestions, improving the accuracy and personalization of the analysis, and providing a more accurate health management plan for patients. Finally, through remote medical intervention, real-time monitoring and timely intervention of the patient's condition are realized, reducing the risk of disease deterioration, improving the patient's self-management ability, thereby realizing intelligent and personalized health management of COPD patients after leaving the hospital, improving the efficiency of condition management and helping to prevent disease deterioration. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the working process of a monitoring system for chronic obstructive pulmonary disease (COPD) patients after discharge according to the present invention;

[0012] Figure 2 It is a schematic diagram of the update process of the real-time update mechanism according to the present invention; Figure 3 It is a model architecture diagram according to the present invention; Figure 4 It is a general flowchart of the method according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0013] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this specification.

[0014] The present invention aims to provide a monitoring system for chronic obstructive pulmonary disease (COPD) patients after discharge. By combining the collection of patients' daily health indicators and electronic medical record data for mining and analysis, using a medical large model for disease evaluation, giving preliminary diagnoses, treatment suggestions, and medical guidance suggestions, and adding links such as doctors' remote follow-up interventions, efficient and intelligent management of COPD patients can be achieved.

[0015] In the present invention, the monitoring system for COPD patients after discharge includes: a data processing module, a patient health status assessment module, a medical large model analysis module, and a remote medical intervention module. As Figure 1 shown, the working process of the monitoring system for COPD patients after discharge is as follows:

[0016] S101. Obtain multi-source data of the patient, including: daily health monitoring data and electronic medical record data.

[0017] Specifically, the daily health monitoring data can be from the patient's daily health monitoring devices, such as smart watches, pulse oximeters, etc., and the electronic medical record data can be from the hospital information system.

[0018] S102. Preprocess and extract features from the patient's multi-source data through the data processing module to obtain the patient's health feature data.

[0019] The health feature data are features related to the COPD condition selected from the multi-source data (i.e., daily health monitoring data and electronic medical record data), including: features based on respiratory function, features based on lifestyle and behavior, features based on complications and comorbidities, and features based on drug use.

[0020] The data types of the multi-source data can be classified into structured data and unstructured data. The structured data include the patient's age, gender, smoking history, respiratory rate, heart rate, blood oxygen saturation, weight, exercise volume, etc.; the unstructured data include medication records, past medical history, laboratory test results, imaging examination reports, etc.

[0021] Furthermore, the preprocessing and feature extraction of multi-source data by the data processing module specifically include: obtaining structured data and unstructured data from the multi-source data. The structured data includes the patient's age, gender, smoking history, respiratory rate, heart rate, blood oxygen saturation, weight, and exercise volume. The unstructured data includes the patient's medication records, past medical history, laboratory test results, and imaging examination reports; performing data cleaning operations on the structured data and unstructured data to obtain the first processed data. The data cleaning operations include data deduplication, filling missing values, and outlier handling; performing attribute numericalization and standardization processing on the first processed data to obtain the second processed data. Attribute numericalization uses label encoding or one-hot encoding, and standardization uses mean and standard deviation calculations; selecting features related to the condition of chronic obstructive pulmonary disease from the second processed data through a feature selection algorithm to obtain healthy feature data. The feature selection algorithm includes chi-square test. The larger the chi-square statistic value in the chi-square test, the stronger the correlation between the feature and the condition of chronic obstructive pulmonary disease.

[0022] Specifically, data deduplication: Repeated records may lead to statistical result biases. Therefore, it is necessary to detect and delete these repeated records through unique identifiers (such as ID, composite keys, etc.). Filling missing values: For the missing values of some non-critical attributes, statistical methods can be used for filling. For example, continuous data can be filled using mean, median, mode, etc.; for categorical variables, the mode (the category with the most occurrences) can be used for filling. Outlier handling: Outliers are values in the dataset that are significantly different from most of the data, which may be caused by data entry errors, system failures, or other accidental factors. Common outliers include extreme values (such as illogical extremely high or low values) or unreasonable records (such as negative age). Outlier handling mainly includes deleting outliers, replacing outliers, and transforming outliers to reduce the impact of outliers.

[0023] Attribute numericalization is to numerically process non-numerical attributes such as the patient's gender (male / female), whether taking medicine regularly, etc., using two numericalization methods: label encoding or one-hot encoding. Label Encoding: For ordered categorical data (such as disease stage), different categories can be represented by integers. For example, the disease course stages ["initial stage", "mid-stage", "late stage"] can be encoded as [0, 1, 2]. One-Hot Encoding: For unordered categorical data (such as gender, whether smoking, etc.), one-hot encoding can be used to represent each category as a binary vector, avoiding introducing unnecessary order relationships between categories.

[0024] Standardization: Different features usually have different units and dimensions. Standardize the data with different dimensions. Use Z-score standardization: Standardize the data according to its mean and standard deviation, so that the data of each feature has a distribution with a mean of 0 and a standard deviation of 1. The formula is: Where X is the feature value, μ is the mean, and σ is the standard deviation.

[0025] The calculation formula of the chi-square statistic in the chi-square test is: Where, O i : The actually observed frequency, E i : The theoretically expected frequency, calculated assuming independence between the feature and the target variable, χ 2 : The chi-square statistic. The larger the value, the stronger the association between the feature and the target variable.

[0026] Calculate the chi-square statistic for each feature, and screen out the feature with the strongest correlation with the condition of chronic obstructive pulmonary disease according to the size of the chi-square value. Generally speaking, the larger the chi-square value of a feature, the stronger its association with the target variable.

[0027] S103. Integrate and update the health feature data through the patient health status evaluation module, and use the density-based clustering algorithm DBSCAN to perform clustering analysis on the integrated and updated real-time health data to obtain the patient's health portrait.

[0028] Specifically, the construction of the health portrait is as follows: Use the machine learning clustering algorithm DBSCAN (Density-Based Spatial Clustering of Applications with Noise) to cluster according to the feature density of the patient, forming different groups. These groups can be used for preliminary disease stratification and disease course management, can identify clusters with arbitrary shapes, and can effectively handle noise and outliers. The specific clustering process is as follows:

[0029] 1) Given the data set D = {x1, x2,..., x n}, where x i represents the feature vector of the patient.

[0030] 2) Select the parameters ∈ and MinPts. Where ∈ defines the radius of the neighborhood, representing the neighborhood range of a point, and MinPts represents the minimum number of points required to form a cluster within the ∈ neighborhood.

[0031] 3) Calculate the distance dist(x i , x j) By using the weighted Manhattan distance method, weights are assigned to different features to reflect their importance. In medical data, the impacts of different features on patient classification may vary. Therefore, the classification accuracy can be improved through the weighted Manhattan distance. The formula for the weighted Manhattan distance is as follows: where: w i is the weight of the i-th feature. Usually, w i ≥ 0; a i and b i are the values of the feature vectors A and B in the i-th dimension; |a i - b i | represents the absolute difference between the two samples in the i-th dimension.

[0032] 4) Taking each point as the center, determine the set of points within its ∈-neighborhood. Its expression is: N ∈ (x) = {y ∈ D | dist(x, y) ≤ ∈}. If |N ∈ (x)| ≥ MinPts, then the point x is regarded as a core point.

[0033] 5) Starting from the core points, expand the clustering and label all density-reachable points as the same cluster until no new density-reachable points can be found.

[0034] 6) Finally, all points that are not assigned to any cluster are labeled as noise points.

[0035] Furthermore, label generation: According to the identified features, create descriptive labels for each group, such as "severe COPD patients", "mild COPD patients".

[0036] The patient health status assessment module includes a data integration unit and a health profile construction unit. The data integration unit is used to perform operations such as information integration and real-time update, and the health profile construction unit is used to construct the patient's health profile.

[0037] Specifically, information integration means: integrating all the patient's data into a single view for subsequent analysis. Real-time update is: designing a real-time update mechanism. When new monitoring data is generated, automatically update the patient's health status to ensure that the information in the system is consistent with the patient's current actual situation.

[0038] The update process of the real-time update mechanism specifically includes: data interface, cache and database management, data processing, automatic reminder mechanism, and regular summary and reporting, as Figure 2 shown.

[0039] Among them, the data interface: is used for the system to dock with the patient's monitoring devices and medical record systems, adopting standard protocols such as API interfaces or HL7, and receiving data from wearable devices and hospital systems in real time.

[0040] Cache and Database Management: Real-time data first enters the cache for quick display in a short time; meanwhile, the data is also regularly written into the database for long-term analysis.

[0041] Data Processing: Clean, denoise, and preprocess new monitoring data, and update relevant health parameters according to the set algorithms, such as recalculating the smoking index and updating the medication compliance score.

[0042] Automatic Reminder Mechanism: When certain key indicators (such as blood oxygen saturation, heart rate) exceed the normal range, the system can issue an alarm to remind medical staff or the patient himself / herself to intervene.

[0043] Regular Summarization and Reporting: The system can automatically generate health reports regularly, summarizing the patient's health trends and indicator changes into a brief report for doctors and patients to regularly view the overall condition changes.

[0044] S104. Analyze the health portrait through the medical large model analysis module to obtain the disease condition evaluation result and health management suggestions.

[0045] The medical large model analysis module uses a multi-modal model architecture deep learning model that combines the convolutional neural network CNN and Transformer, and uses a high-quality training data set labeled by professional doctors for model training. Using the trained model, compare the patient's health portrait and output the disease condition evaluation result and health management plan. The labeled data used for the training data set is a high-quality training data set labeled by professional doctors, including diagnosed cases and their corresponding detailed clinical information. Further, data augmentation is performed on the labeled data by simulating different clinical scenarios to increase data diversity and improve the generalization ability of the model. The specific method of data augmentation by simulating different clinical scenarios is as follows: At different disease stages, create different symptom manifestations of patients with mild and severe chronic obstructive pulmonary disease, simulate the frequency and impact of acute exacerbation events, or change clinical parameters, and generate new sample data by adjusting clinical parameters (such as oxygen therapy time, drug dosage, etc.) to reflect the patient's response under different treatment plans.

[0046] Specifically, the deep learning model includes an input layer, a data preprocessing layer, an encoding layer, a feature fusion layer, a thinking logic reasoning layer, and an output layer. The health portrait is analyzed through the medical large model analysis module, including: the image data, text data, time series data, and knowledge graph in the health portrait are input through the input layer of the deep learning model; the image data, text data, time series data, and knowledge graph are preprocessed respectively through the data preprocessing layer; the convolutional neural network (CNN) encoder in the encoding layer encodes the image data to obtain the spatial features of the image, and the attention Transformer encoder encodes the text data and time series data to capture the long-range dependence relationship between features through the self-attention mechanism (i.e., the correlation between elements that are far apart in the data, such as cross-sentence references in text or the impact of early events on the future in time series), and the knowledge graph encoder encodes the structural information of the knowledge graph to extract the relationship features between entities such as symptoms, causes, and drugs; the feature fusion layer uses the self-attention mechanism to splice and fuse the feature data extracted in the encoding layer to obtain the fused feature data; the thinking logic reasoning layer performs logical reasoning on the fused feature data based on the medical knowledge and logical reasoning rules in the knowledge graph to obtain the disease condition features of the patient classification, and performs anomaly detection on the fused feature data to predict the risks of different diseases to obtain the prediction results, and finally obtains the disease condition evaluation results; the output layer of the deep learning model classifies the disease condition features of the patient classification and outputs the corresponding descriptive labels, converts them into disease probabilities according to the prediction results and outputs them, and combines the descriptive labels, disease probabilities, and medical advice rules in the knowledge graph to output the health management opinions.

[0047] S105. Execute remote monitoring, interactive communication, and personalized intervention operations according to the disease condition evaluation results through the remote medical intervention module.

[0048] The remote monitoring includes real-time monitoring and abnormal alarm, the interactive communication includes online consultation and video conferencing, and the personalized intervention includes treatment plan adjustment and health education.

[0049] The following is a detailed description of the health feature data in this application:

[0050] 1. The features based on respiratory function include the ratio change R of the forced expiratory volume in one second (FEV1) to the forced vital capacity (FVC), and the overall respiratory function ORF index. The overall respiratory function ORF index is a comprehensive index of oxygen therapy time, finger pulse oxygen saturation, and respiratory function exercise. The calculation formula for the ratio change R is: where (FEV1 / FVC) follow-up is the ratio of the forced expiratory volume in one second (FEV1) to the forced vital capacity (FVC) measured during the follow-up period (such as after treatment, after rehabilitation intervention), and (FEV1 / FVC) baselineIt is the ratio of the forced expiratory volume in the first second (FEV1) measured during the baseline period (before treatment or intervention) to the forced vital capacity (FVC). The calculation formula for the overall respiratory function ORF index is: where, T oxygen is the oxygen therapy time, T max is the maximum oxygen therapy time, SpO2 is the fingertip pulse oximetry saturation, E exercise indicates whether respiratory function exercise is carried out, taking 1 for exercise and 0 for no exercise, and ω1, ω2, and ω3 are the weights of oxygen therapy time, fingertip pulse oximetry saturation, and respiratory function exercise respectively.

[0051] 2. The characteristic based on lifestyle and behavior is the smoking index SI, which is defined as the product of the number of years of smoking and the daily smoking amount, and is used to evaluate the impact of smoking on chronic obstructive pulmonary disease (COPD). The higher the value of the smoking index SI, the more severe the cumulative damage of smoking to the lungs; the calculation formula for the smoking index SI is: SI = SY × DSA, where, SY (Smoking Years): the number of years of smoking (i.e., the number of years from the start of smoking to the present); DSA (Daliy Smoking Amounts): the daily smoking amount (usually expressed in the number of cigarettes).

[0052] 3. The characteristic based on complications and comorbidities: the comorbidity score CCI.

[0053] Characteristic definition: CCI (Charlson Complication Index), which calculates the comorbidity burden of patients based on the Charlson comorbidity index. This index evaluates the long-term prognosis of patients by assigning different weights to different diseases. The calculation formula for the comorbidity score CCI is: where, C i : the i-th comorbidity (such as heart disease, diabetes, etc.), w i : the weight assigned to each comorbidity according to its impact on the patient's health, n: the total number of comorbidities.

[0054] 4. The characteristics based on drug use include long-term medication adherence LMA and the frequency of using inhaled drugs FIU.

[0055] Specifically, the characteristic definition of long-term medication adherence LMA is: LMA (Long-term Medication Adherence), which evaluates whether the patient takes medicine on time and follows the prescribed dose and frequency. The higher the adherence, usually the higher the patient's persistence during the treatment course, which helps in the control and remission of the disease.

[0056] The calculation formula for long-term medication adherence LMA is: Among them, AD (Actual Dose) represents the actual dose of medicine taken, and SD (Scheduled Dose) represents the scheduled dose of medicine.

[0057] Feature definition of the frequency FIU of using inhaled drugs: FIU (Frequency of Inhaled drug Use) records the frequency of a patient using short-acting or long-acting inhaled drugs, and can evaluate the patient's drug dependence, disease stability, as well as the patient's inhalation treatment effect and compliance.

[0058] The calculation formula for the frequency FIU of using inhaled drugs is: Among them, TIU (Total Inhaler Use) represents the total number of times of using the inhaler, which indicates the number of times a patient uses short-acting or long-acting inhaled drugs within a certain period (such as daily, weekly, etc.); FP (Follow-up Period) represents the follow-up period, usually in days, weeks or months, which indicates the time period for recording the frequency of the patient's inhaler use.

[0059] Next, the deep learning model in the present invention will be described as follows:

[0060] The deep learning model is a multi-modal model architecture that uses a convolutional neural network CNN combined with Transformer to process image and clinical data. Its model architecture design includes: an input layer, a data preprocessing layer, an encoding layer, a feature fusion layer, a thinking logic reasoning layer, and an output layer. Its model architecture diagram is as Figure 3 shown, and the specific functions of each layer are as follows:

[0061] The input layer is used to input data, including: image data (such as CT images), time series data (such as respiratory rate and oxygen saturation), medical record text data, laboratory test results, and knowledge graphs;

[0062] The data preprocessing layer is used for image preprocessing, time series data processing, text preprocessing, and knowledge graph preprocessing. Image preprocessing is to perform scaling, normalization, and noise filtering on CT images, etc. Time series data processing is to complete, interpolate, and smooth irregular time data to generate a standard time series. Text preprocessing: perform word segmentation, stop word removal, and embedding conversion on medical records and diagnostic texts. Knowledge graph preprocessing: convert entities and relationships in the knowledge graph into low-dimensional vectors through graph embedding methods (such as TransE, node2vec) for the model to use.

[0063] The encoding layer includes a CNN encoder, a Transformer encoder, and a knowledge graph encoder. The CNN encoder extracts the spatial features of images through multiple convolutional layers and pooling layers. The Transformer encoder encodes time series data and text data. The data or text at each time step is input into the multi-head self-attention module of the Transformer after embedding, and the long-range dependencies between features are captured through the self-attention mechanism. The knowledge graph encoder encodes the structural information of the knowledge graph and extracts the relationship features between entities such as symptoms, causes, and drugs.

[0064] The feature fusion layer fuses the feature vectors output by each encoding layer. The self-attention fusion module is used to splice the features.

[0065] The thinking logic reasoning layer is used for logical rule reasoning, dynamic reasoning update, and anomaly detection.

[0066] Logical rule reasoning: Use the medical knowledge in the knowledge graph (such as disease symptom relationships, drug efficacy, clinical diagnosis rules, etc.) to deduce the disease characteristics of the patient. Through a series of predefined medical logic rules, the input patient data is combined with medical knowledge for preliminary reasoning.

[0067] Dynamic reasoning update: Adjust the reasoning chain according to the patient's real-time data or dynamically changing information.

[0068] Anomaly detection: Identify abnormal patterns in the patient's input data. When abnormal data patterns are detected, this module will trigger an early warning mechanism to prompt possible disease deterioration.

[0069] The output layer is used for patient classification output, disease probability output, and health management advice output.

[0070] Patient classification output: Use the output features of the deep learning model to classify the patient. Common classification labels can include "mild COPD patient", "moderate COPD patient", "severe COPD patient", etc.

[0071] Disease probability output: Based on the inference of the model, the probability regression method is used to predict and output the risk values of different diseases, such as "the probability of acute exacerbation is 70%", "the risk of developing serious complications is 50%", etc.

[0072] Health management advice: Combine the model's assessment of the patient's current condition and the medical advice rules in the knowledge graph to output a personalized health management plan, such as exercise methods to improve respiratory function, nutritional advice, and lifestyle adjustments.

[0073] Transfer learning: If the data volume is insufficient, the model can be pre-trained on a larger public medical dataset first, and then fine-tuned on a specific task.

[0074] In the present invention, the model training of the deep learning model uses a hyperparameter optimization method, specifically, the Bayesian optimization method is used to find the optimal hyperparameter configuration. The model training method is as follows:

[0075] 1) Define the objective function: Minimize the loss or error on the validation set, use cross-entropy loss for classification tasks, and use mean squared error for regression tasks. Its expression is as follows: Where: λ represents the hyperparameter vector (such as learning rate, regularization strength, network structure, etc.); ytrue represents the true label of the validation set; ypred represents the predicted result of the validation set; Lval represents the loss function on the validation set.

[0076] 2) Construct a surrogate model: Use a Gaussian process (GP) as an approximate surrogate for the objective function f(λ). Its construction process is expressed as: f(λ) ∼ GP(m(λ), k(λ, λ′)), where m(λ) represents the mean function, initialized to zero; k(λ, λ′) represents the kernel function, and the RBF kernel is used to measure the similarity between hyperparameter points.

[0077] 3) Select the acquisition function: Use the Expected Improvement (EI) to search for points that have the potential to exceed the current optimal value. Its expression is:

[0078] 4) Update the surrogate model: Evaluate the best point λ of the acquisition function and update the Gaussian process model with the true objective function value f(λ).

[0079] 5) Stopping condition: Repeat steps (3) and (4) until either of the following conditions is met: reaching the preset maximum number of iterations, or the improvement of the objective function is less than the threshold.

[0080] Verification and testing: Use the K-fold cross-validation method to evaluate the model performance and verify its accuracy on an independent test set. The method is as follows: First, divide the training data set into K non-overlapping subsets (folds), and successively select one of the subsets as the validation set, and the rest as the training set; Second, for each fold k, calculate the validation loss, and then average the validation losses of all folds to obtain the cross-validation performance index; Finally, use the optimal hyperparameter configuration in the cross-validation to evaluate the model performance on the independent test set and output the test error.

[0081] Furthermore, the logical reasoning rules in the thinking logic reasoning layer include: the rule from symptoms to diseases, the rule from examination results to diseases, and the disease combination rule. The rule from symptoms to diseases infers the probability of suffering from a specific disease through the patient's symptoms. The rule from examination results to diseases infers whether the patient has a specific disease by matching the patient's examination results with the diagnostic criteria of the disease. The disease combination rule infers the possible complex conditions caused by the coexistence of certain specific diseases.

[0082] Specifically, the logical rule reasoning in the thinking logic reasoning layer of the present invention is: a series of logical rules are defined based on the knowledge graph to form a thinking link from the patient's input information to the disease condition assessment, ensuring the coherence and logic of the information processing process. The specific reasoning steps are as follows:

[0083] 1) Knowledge graph construction:

[0084] Define entity nodes (such as patients, symptoms, diseases, examination results, treatment plans, etc.) and attribute nodes (numerical or categorical attributes associated with entity nodes, such as the number of days of fever persistence, blood oxygen saturation, disease risk score, etc.).

[0085] Define edges and relationships such as a patient has symptoms (indicating the patient's current main complaint symptoms), symptoms are associated with diseases (the causal relationship from symptoms to possible diseases), and diseases are associated with treatments (indicating recommended treatment plans), etc.

[0086] 2) Logical rule design, including the rule from symptoms to diseases, the rule from examination results to diseases, and the disease combination rule.

[0087] The rule from symptoms to diseases:

[0088] If multiple symptoms S = {s1, s2,..., sk} of a patient are highly correlated with a certain disease D, then the probability of disease D is inferred to be relatively high, which is expressed as:

[0089] Among them, weight(si, D) represents the correlation weight between the symptom and the disease; θ represents the correlation threshold of symptom - disease; P(D∣S) represents the probability of disease D under the given symptom S; γ represents the lower limit of significant probability.

[0090] The rule from examination results to diseases:

[0091] If the examination results C = {c1, c2,..., cm} of a patient meet the diagnostic criteria of a certain disease D, then the existence of the disease is inferred. It is expressed as: Among them: ReferenceRange(D) represents the reference range of disease D; P(D∣C) represents the probability of disease D under the given examination results C.

[0092] Disease combination rule:

[0093] The co - existence of certain diseases D1, D2, …, Dn may trigger a new complex disease D′, expressed as:

[0094]

[0095] Where: CombinationProbability(D′) represents the complex disease group; P(D′∣D): represents the probability of the complex disease D′ occurring under the given combined disease D.

[0096] 3) Reasoning method: Use the adjacency matrix A to represent the structure of the knowledge graph:

[0097] Where: w ij represents the edge weight between nodes v i and v j .

[0098] The disease condition assessment can be expressed as: Risk Score = ∑ j,j A jj ·MatchingDgree(v i , v j ), where: MatchingDegree(v i , v j ) represents the matching degree of the nodes.

[0099] 4) Output the assessment result: According to the reasoning result, output the likelihood ranking of diseases, recommended examinations and treatment plans.

[0100] In the present invention, dynamic reasoning update: According to the real - time feedback and historical data of the patient, dynamically adjust the parameter settings in the thinking chain to improve the personalization of decision - making. The specific adjustment method is as follows:

[0101] 1) The weight ωi of each feature included in the model is adjusted according to historical data and real - time feedback: Where ωi is the weight of feature xi, is the predicted output of the model.

[0102] 2) Use the gradient descent method to update the weight parameters:

[0103] 3) Where η is the learning rate and L is the loss function, is the partial derivative of the loss function with respect to the weight.

[0104] In the present invention, anomaly detection: Perform anomaly detection in the thinking chain, identify abnormal patterns in the patient's input data, and timely warn of possible disease deterioration.

[0105] Anomaly detection uses the statistical method Z-score to identify abnormal data by calculating the Z-score of each feature and comparing it with the predetermined threshold. If the Z-score of a feature exceeds the predetermined threshold, it is considered an anomaly and the system will detect the anomaly and trigger an alarm.

[0106] The disease probability prediction in the output layer of the present invention specifically includes:

[0107] 1) Multimodal fusion: Multiple types of information such as text (such as medical history descriptions), images (such as CT scans), and numerical values (such as blood test results) are fused together as model input.

[0108] 2) Modal feature extraction includes the extraction of text features, image features and numerical features.

[0109] Text feature extraction: Input text data into Transformer and output the embedding vector Htext.

[0110] Image feature extraction: The image data is input into a convolutional neural network (CNN) and the embedded vector Himage is output.

[0111] Numerical feature extraction: Directly use the fully connected layer to process the numerical features and output the embedding vector Hnum.

[0112] 3) Modal fusion: The weighted fusion method is used to fuse the embedding vectors of different modalities to form a unified representation, and the modal weights are dynamically adjusted based on the attention mechanism: ω = Softmax ([ω text ,ω image ,ω num ]), H fusion =∑ m∈{text,image,num} ω m ·H m ,where: ,Hfusion: fused feature representation, which combines the information of all modalities as the final input of the model; ω: modality weight vector (containing the weights of all modalities), indicating the importance of each modality in the final prediction.

[0113] 4) Output explanation: The model outputs a probability distribution of various possible diseases, along with explanatory information (such as which features best support the diagnosis).

[0114] 5) Output disease probability distribution: The fusion features are processed by a multi-layer perceptron (MLP), and the model output is normalized to probability through a Softmax activation function to obtain the final disease category probability distribution: Output example: ŷ = {0.2, 0.1, 0.05, 0.6, 0.05}, indicating that the model believes the probability that the patient has the fourth type of disease is 60%, the probability of having the first type of disease is 20%, and so on.

[0115] Output explanatory information: The Shapley value calculates the global importance of each feature by traversing all feature combinations and computing the contribution of each feature in different situations, and finally outputs the positive or negative contribution of each feature to the prediction result. The calculation formula is as follows: Among them, F: the set of all input features (such as medical history descriptions, image features, blood indicators, etc.); S: a subset of features that does not include feature i; f(S): the model prediction value (disease probability) when using the feature subset S;

[0116] f(S ∪ {i}): the model prediction value after adding feature i to the subset S; φi: the Shapley value of feature i, representing the average marginal contribution of feature i to the model prediction in all possible feature combinations.

[0117] The example is as follows:

[0118] Suppose the model predicts that the probability of a certain patient having chronic obstructive pulmonary disease is 85%, and the baseline probability (the average disease probability of all samples) is 50%. The Shapley values of key features are as follows:

[0119] Decline in respiratory function (FEV1 / FVC ratio < 70%): +20%, long-term smoking history: +15%, history of repeated cough and expectoration (text description): +10%, imaging features showing emphysema (CT scan): +25%, no significant family genetic history: -5%,

[0120] Output explanation:

[0121] The model believes that the probability of the patient having chronic obstructive pulmonary disease is relatively high (85%), mainly because the decline in the FEV1 / FVC ratio, long-term smoking history, and the CT findings of emphysema features strongly support the diagnosis.

[0122] Output of diagnostic suggestions, which includes personalized suggestions and medical referral suggestions, as follows:

[0123] Personalized suggestions: Provide personalized health management suggestions according to the model prediction results, such as lifestyle adjustments, drug dosage adjustments, etc.

[0124] Medical referral suggestions consist of two parts: urgency assessment and medical institution recommendation. Urgency assessment: Evaluate whether immediate medical attention is required based on the severity of the condition. Medical institution recommendation: Recommend the nearest or most suitable medical institution according to factors such as the patient's location and the service capabilities of medical institutions.

[0125] The telemedicine intervention module has functions of remote monitoring, interactive communication, and personalized intervention. Specifically as follows:

[0126] Remote monitoring includes two parts: real-time monitoring and abnormal alarm. Real-time monitoring: Doctors can view the patient's health status in real time through the system, including physiological indicators, symptom changes, etc. Abnormal alarm: When the system detects an abnormal situation, it automatically sends an alarm message to the doctor.

[0127] Interactive communication includes two parts: online consultation and video conferencing. Online consultation: Doctors can communicate with patients online, understand the current situation of patients, and give guidance. Video conferencing: Supports video call function, and doctors can communicate face-to-face with patients to observe the patient's status.

[0128] Personalized intervention includes two parts: treatment plan adjustment and health education. Treatment plan adjustment: According to the patient's latest condition, doctors can adjust the treatment plan, including drug prescriptions, rehabilitation suggestions, etc. Health education: Provide health education materials to help patients better understand disease management and self-care.

[0129] In the COPD patient discharge monitoring system of the present invention, the system implementation mainly includes the following parts:

[0130] The front-end interface includes user interface design and interaction feedback mechanism.

[0131] User interface design: Design an intuitive and easy-to-use operation interface to facilitate patients or their families to quickly view the health status.

[0132] Interaction feedback mechanism: Provide instant feedback to display the change trend of health monitoring data.

[0133] 2) Background processing includes: high-performance computing and distributed storage.

[0134] High-performance computing: Utilize cloud computing resources to accelerate data processing and analysis. Distributed storage: Use a distributed database to store massive medical data to ensure system stability and scalability.

[0135] 3) Decision support includes: visualization tools and diagnostic assistance.

[0136] Visualization tools: Provide forms such as charts to display the recommendation results and help doctors understand the recommendation basis. Auxiliary diagnosis: Based on the opinions of the medical large model, the system can assist in generating preliminary diagnosis suggestions.

[0137] 4) The feedback mechanism includes: closed-loop feedback mechanism and continuous improvement mechanism.

[0138] Closed-loop feedback mechanism: Collect actual monitoring effect data for optimizing the recommendation algorithm.

[0139] Continuous improvement mechanism: Regularly evaluate the system performance and adjust the model parameters according to the feedback.

[0140] Based on the monitoring system for chronic obstructive pulmonary disease (COPD) patients after discharge, the present application also provides a corresponding method for monitoring COPD patients after discharge. The overall flow chart of the method is as Figure 4 shown as follows:

[0141] (1) Data input: Patients or their families input the daily health monitoring data of the patients through the application on the mobile device.

[0142] (2) Analysis and processing: The system automatically reads the electronic medical record data of the patient, combines the input information, and constructs a patient health profile.

[0143] (3) Disease evaluation: The system compares the profile with the medical large model and outputs the disease evaluation result.

[0144] (4) Treatment suggestions: The system provides preliminary diagnosis, treatment suggestions, and medical guidance suggestions according to the evaluation result.

[0145] (5) Doctor's remote follow-up: The doctor monitors the patient's condition in real time through the system and provides remote intervention measures.

[0146] (6) Monitoring feedback: Patients or their families perform corresponding treatments according to the suggestions and feedback the treatment results to the system for subsequent optimization.

[0147] As mentioned above, the above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A monitoring system for patients with chronic obstructive pulmonary disease (COPD) after discharge, characterized in that, Including: Obtain multi-source data of the patient, where the multi-source data at least includes daily health monitoring data and electronic medical record data; Preprocess and extract features from the multi-source data through a data processing module to obtain the patient's health feature data. The health feature data is the features related to the condition of chronic obstructive pulmonary disease (COPD) selected from the multi-source data, including: features based on respiratory function, features based on lifestyle and behavior, features based on complications and comorbidities, and features based on drug use; Integrate and update the health feature data through a patient health status assessment module, and perform clustering analysis on the integrated and updated real-time health data using the density-based clustering algorithm DBSCAN to obtain the patient's health portrait. The density-based clustering algorithm DBSCAN can identify clusters with arbitrary shapes and can effectively handle noise and outliers; Analyze the health portrait through a medical large model analysis module to obtain a disease condition evaluation result and health management suggestions. The medical large model analysis module uses a deep learning model designed with a multi-modal model architecture that combines a convolutional neural network (CNN) and an attention mechanism (Transformer); Execute remote monitoring, interactive communication, and personalized intervention operations according to the disease condition evaluation result through a remote medical intervention module. The remote monitoring includes real-time monitoring and abnormal alarm, the interactive communication includes online consultation and video conferencing, and the personalized intervention includes treatment plan adjustment and health education.

2. The COPD patient discharge monitoring system according to claim 1, wherein The preprocessing and feature extraction of the multi-source data by the data processing module includes: Obtain structured data and unstructured data from the multi-source data. The structured data includes the patient's age, gender, smoking history, respiratory rate, heart rate, blood oxygen saturation, weight, and exercise volume. The unstructured data includes the patient's medication record, past medical history, laboratory test results, and imaging examination reports; Perform data cleaning operations on the structured data and the unstructured data to obtain first-processed data. The data cleaning operations include data deduplication, filling missing values, and outlier processing; Perform attribute numericalization and standardization processing on the first-processed data to obtain second-processed data. The attribute numericalization uses label encoding or one-hot encoding, and the standardization uses mean and standard deviation calculation; Select features related to the condition of COPD from the second-processed data through a feature selection algorithm to obtain the health feature data. The feature selection algorithm includes chi-square test. The larger the chi-square statistic value in the chi-square test, the stronger the correlation of the feature with the condition of COPD.

3. The COPD patient discharge monitoring system according to claim 1, characterized in that The features based on respiratory function include the ratio change R of forced expiratory volume in one second (FEV1) to forced vital capacity (FVC), and the overall respiratory function ORF index. The overall respiratory function ORF index is a comprehensive index of oxygen therapy time, finger pulse oxygen saturation, and respiratory function exercise; The calculation formula for the ratio change R is as follows: where (FEV1 / FVC) follow-up is the ratio of the forced expiratory volume in the first second FEV1 to the forced vital capacity FVC measured during the follow-up period (such as after treatment, after rehabilitation intervention), and (FEV1 / FVC) baseline is the ratio of the forced expiratory volume in the first second FEV1 to the forced vital capacity FVC measured during the baseline period (before treatment or intervention); The calculation formula for the overall respiratory function ORF index is as follows: Where T oxygen is the oxygen therapy time, T max is the maximum oxygen therapy time, SpO2 is the finger pulse oxygen saturation, and E exercise indicates whether respiratory function exercise is carried out. Take 1 for exercise and 0 for no exercise. ω1, ω2, and ω3 are the weights of oxygen therapy time, finger pulse oxygen saturation, and respiratory function exercise respectively; The feature based on lifestyle and behavior is the Smoking Index SI, which is defined as the product of the number of years of smoking and the daily cigarette consumption, and is used to evaluate the impact of smoking on chronic obstructive pulmonary disease (COPD). The higher the value of the Smoking Index SI, the more severe the cumulative damage to the lungs caused by smoking; The feature based on complications and comorbidities is the Charlson Comorbidity Index CCI, which evaluates the long-term health of patients by assigning different weights to different diseases; The calculation formula of the comorbidity score CCI is as follows: Where C i is the i-th comorbidity, ω i is the weight assigned to each comorbidity according to its impact on the patient's health, and n is the total number of comorbidities; The features based on medication use include Long-term Medication Adherence LMA and Frequency of Inhaled Medication Use FIU. The feature definition of Long-term Medication Adherence LMA evaluates whether patients take medications on time and follow the prescribed dosage and frequency. The feature definition of Frequency of Inhaled Medication Use FIU is to record the frequency of patients' use of short-acting or long-acting inhaled medications.

4. The COPD patient discharge monitoring system according to claim 1, wherein, The patient health status evaluation module integrates and updates the health feature data, and uses the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm to perform clustering analysis on the integrated and updated real-time health data to obtain a patient health portrait, including: The data integration unit performs information integration on the health feature data to obtain integrated data; The real-time update mechanism updates the integrated data to obtain real-time health data. The real-time update mechanism is used to automatically update the patient's health status when new health feature data is generated to ensure that the information in the system is consistent with the patient's current actual situation. The update process of the real-time update mechanism includes data interface input data, cache and database management, data processing, automatic reminder mechanism, and regular summary and reporting; The health portrait construction unit uses the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm to perform clustering analysis on the real-time health data to obtain a patient health portrait, and generates descriptive labels based on the health portrait. The descriptive labels are used to characterize the severity of the patient's condition, including mild COPD patients, moderate COPD patients, and severe COPD patients.

5. The COPD patient discharge monitoring system according to claim 1, wherein The deep learning model includes an input layer, a data preprocessing layer, an encoding layer, a feature fusion layer, a thinking logic reasoning layer, and an output layer. The medical large model analysis module analyzes the health portrait, including: The image data, text data, time series data, and knowledge graph in the health portrait are input through the input layer of the deep learning model; The data preprocessing layer preprocesses the image data, text data, time series data, and knowledge graph respectively; The Convolutional Neural Network (CNN) encoder in the encoding layer encodes the image data to obtain the spatial features of the image, and the Attention Transformer encoder encodes the text data and time series data to capture the long-range dependence relationship between features through the self-attention mechanism. The long-range dependence relationship refers to the correlation between elements that are far apart in the data. The knowledge graph encoder encodes the structural information of the knowledge graph to extract the relationship features between entities such as symptoms, causes, and medications; The feature fusion layer uses the self-attention mechanism to splice and fuse the feature data extracted in the encoding layer to obtain fused feature data; Through the above-mentioned thinking logic reasoning layer, logical reasoning is performed on the fused feature data based on medical knowledge and logical reasoning rules in the knowledge graph to obtain the disease characteristics of patient classification, and anomaly detection is performed on the fused feature data to predict the risks of different diseases to obtain a prediction result, and finally the disease evaluation result is obtained; Through the output layer of the deep learning model, the disease characteristics of patient classification are classified and corresponding descriptive labels are output, converted into disease probabilities according to the prediction result and output, and health management opinions are output in combination with the descriptive labels, disease probabilities, and medical advice rules in the knowledge graph.

6. The chronic obstructive pulmonary disease patient discharge monitoring system according to claim 5, characterized in that, In the thinking logic reasoning layer, the entity nodes of the knowledge graph are defined as patients, symptoms, diseases, examination results, and treatment plans, and its attribute nodes are defined as numerical or categorical attributes associated with the entity nodes, including the number of days of persistent fever, blood oxygen saturation, and disease risk scores; In the thinking logic reasoning layer, the edges and relationships of the knowledge graph are defined as patients having symptoms, symptoms being associated with diseases, and diseases being associated with treatments. The statement that a patient has symptoms represents the current main complaint symptoms of the patient. The statement that symptoms are associated with diseases represents the causal relationship from symptoms to possible diseases. The statement that diseases are associated with treatments represents recommended treatment plans.

7. The COPD patient discharge monitoring system according to claim 5 or 6, characterized in that, The logical reasoning rules in the thinking logic reasoning layer include: rules from symptoms to diseases, rules from examination results to diseases, and disease combination rules. The rules from symptoms to diseases infer the likelihood of a patient suffering from a specific disease through the patient's symptoms. The rules from examination results to diseases infer whether a patient has a specific disease by matching the patient's examination results with the diagnostic criteria of the disease. The disease combination rules infer the possible complex diseases caused by the coexistence of certain specific diseases.

8. The COPD patient discharge monitoring system according to claim 5, characterized in that, The hyperparameter optimization method of the deep learning model is the Bayesian optimization method, and its model training method includes: Defining the objective function f(λ): minimizing the loss or error on the validation set, using cross-entropy loss for classification tasks and mean squared error for regression tasks; Constructing a surrogate model: using the Gaussian process GP as an approximate surrogate for the objective function f(λ); Selecting an acquisition function: using the expected improvement EI to search for points that have the potential to exceed the current optimal value; Updating the surrogate model: evaluating the best point λ of the acquisition function and updating the Gaussian process model with the true objective function value f(λ); Stopping condition: repeating the operations of the above acquisition function and updating the surrogate model until the preset maximum number of iterations is reached or the improvement of the objective function is less than the threshold to stop training.

Images