AI agent driven method and system for home health management of copd patients

Abstract

The application discloses an AI agent driven COPD patient home health management method and system, belongs to the technical field of intelligent medical health management, and relates to the technical field of intelligent medical health management.The method comprises the following steps: a multi-modal health data sensing and fusion step, an individualized baseline model construction and dynamic updating step, a multi-step chain reasoning acute exacerbation risk assessment step, an autonomous hierarchical intervention decision and execution step, and a tool calling type contextual service and feedback optimization step.The system deploys an AI agent as a virtual health butler of a patient, realizes all-weather intelligent management of the home health of the COPD patient through a five-stage closed loop of sensing-reasoning-decision-execution-feedback, and reduces the 30-day re-hospitalization rate by about 28%.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical and health management technology, and in particular to an AI agent-driven home health management method and system for patients with chronic obstructive pulmonary disease. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a chronic respiratory disease characterized by persistent airflow limitation, affecting over 380 million people worldwide, with a prevalence rate as high as 13.7% in people aged 40 and above in my country. COPD patients discharged after hospitalization for acute exacerbations face a lack of professional and continuous health management guidance. Acute exacerbations in COPD are often accompanied by progressive deterioration of physiological indicators and symptom changes. Early identification and intervention can significantly reduce the risk of readmission due to disease control. However, traditional post-discharge follow-up models typically rely on regular outpatient visits or telephone follow-ups. These methods have long follow-up intervals, collect incomplete information, and lack real-time updates, making it difficult to capture the dynamic changes in the patient's health status. This results in a persistently high readmission rate within 30 days of discharge for COPD patients. It is worth noting that the triggers for acute exacerbations of COPD are often multidimensional. In addition to internal factors such as respiratory infections, environmental factors such as sudden drops in temperature and air pollution are also important exogenous triggers. However, traditional follow-up management models lack the ability to incorporate external environmental information into the comprehensive assessment of patients' health risks.

[0003] In recent years, with the rapid development of wearable health monitoring devices and artificial intelligence technology, chronic disease management solutions based on remote monitoring have gradually become a research hotspot. Chinese patent CN118471481A discloses a medical multimodal assisted diagnosis and treatment system based on a large model and an artificial intelligence agent. This system includes a data input and preprocessing module, a comprehensive intelligent analysis module, a diagnostic report generation module, and a report export and data packaging module. Through an artificial intelligence agent, it understands and classifies medically relevant clinical problems, breaks down tasks into multiple sub-tasks, and calls upon an expert tool library to execute them, ultimately generating a diagnostic report. However, the above technical solution is mainly geared towards image-assisted diagnosis scenarios within hospitals. Its core function as an artificial intelligence agent focuses on the passive analysis and report generation of existing medical data, and does not address the continuous and proactive perception of patients' health status, individualized baseline modeling, and autonomous intervention decisions based on multi-step reasoning in home settings.

[0004] Furthermore, most existing COPD home monitoring studies limit artificial intelligence technology to single-dimensional data analysis. For example, they may only use deep learning models to predict acute exacerbations from physiological signals collected by wearable devices, or simply use rule-based expert systems to classify patient-reported symptoms. In addition, current protocols generally do not incorporate external environmental factors (such as temperature changes and air pollution levels) into the comprehensive assessment of patient health risks, even though environmental factors play a significant role in triggering acute exacerbations of COPD. These solutions suffer from the following shortcomings: First, they lack the ability to deeply integrate physiological parameter monitoring with patients' subjective symptom reports, failing to achieve a comprehensive understanding of the health status of COPD patients. Second, they use fixed thresholds at the group level rather than dynamic baselines based on individual patient differences, making it difficult to simultaneously consider the sensitivity and specificity of risk warnings. Third, the risk assessment process uses a single-step judgment logic, lacking the chain-like analysis capabilities of clinicians' step-by-step screening and reasoning, making it difficult to distinguish between different types of acute exacerbation precursors. Fourth, intervention measures after risk identification rely on manual intervention, lacking the ability for AI-driven autonomous decision-making and graded response. Fifth, a closed-loop feedback pathway from intervention effects to risk assessment models has not been established, preventing the system from continuously self-optimizing through accumulated management experience. Therefore, there is an urgent need for an AI-driven home health management solution for COPD patients that integrates perception, reasoning, decision-making, execution, and feedback to address the technical problem of insufficient intelligence in existing home management technologies. Summary of the Invention

[0005] The purpose of this invention is to provide an AI agent-driven home health management method and system for COPD patients, in order to solve the technical problem in the prior art that the lack of autonomous intelligent agents for continuous perception, chain reasoning and graded autonomous intervention in the home management of COPD patients after discharge leads to a high rate of readmission due to acute exacerbations.

[0006] To achieve the above objectives, this invention provides an AI agent-driven home health management method for COPD patients, comprising the following steps:

[0007] Step S1, Multimodal Health Data Perception and Fusion Step: Deploy an AI agent to continuously connect to wearable devices worn by COPD patients, collect data on patients' blood oxygen saturation, respiratory rate, activity level, and sleep quality as physiological parameter data, and proactively ask patients about their cough and sputum production, degree of dyspnea, and medication administration status through a voice interaction channel to obtain voice inquiry data. Perform time alignment and feature-level fusion on the collected physiological parameter data and voice inquiry data to generate a multimodal health status feature vector.

[0008] Step S2, Individualized Baseline Model Construction and Dynamic Update: Based on the patient's clinical assessment data at discharge and the multimodal health status feature vectors accumulated in Step S1, an individualized baseline model is constructed for the patient, and the parameters are continuously updated using an online learning mechanism.

[0009] Step S3, multi-step chain reasoning for acute exacerbation risk assessment: The AI ​​agent's reasoning engine receives multimodal health status feature vectors, calls the individualized baseline model to calculate the deviation, performs multi-step chain reasoning matching with the COPD acute exacerbation early warning rule base, and outputs acute exacerbation risk levels including normal, mild abnormal, moderate abnormal, and severe abnormal. The acute exacerbation risk level is jointly determined by the deviation of the current health status and the matching confidence of each rule chain obtained by the multi-step chain reasoning matching.

[0010] Step S4, Autonomous Tiered Intervention Decision and Execution Steps: The AI ​​agent autonomously makes decisions and executes tiered intervention responses based on the risk level. For mild abnormalities, it pushes guidance and reminders; for moderate abnormalities, it contacts family members and suggests follow-up; and for severe abnormalities, it alerts doctors and assists with remote consultations.

[0011] Step S5, Tool-invoking contextualized service and feedback optimization steps: The AI ​​agent autonomously calls external tool interfaces to obtain environmental data and drug knowledge, provides contextualized health services, and feeds back the intervention effect evaluation to step S3 to calibrate the risk assessment threshold. At the same time, after the patient's risk level falls back to normal, its recovery period data is fed back to step S2 to update the individualized baseline model, forming a closed loop.

[0012] This invention also provides an AI agent-driven home health management system for COPD patients, comprising: a multimodal health data perception and fusion module, a personalized baseline model construction and dynamic update module, a multi-step chain-based inference risk assessment module, a self-regulating tiered intervention decision-making and execution module, and a tool-invoking contextualized service and feedback optimization module. Each module corresponds one-to-one with each step in the aforementioned method. Specifically, the multimodal health data perception and fusion module simultaneously outputs feature vectors to both the personalized baseline model construction and dynamic update module and the multi-step chain-based inference risk assessment module. The multi-step chain-based inference risk assessment module outputs risk levels to the self-regulating tiered intervention decision-making and execution module. The tool-invoking contextualized service and feedback optimization module sends feedback data back to both the personalized baseline model construction and dynamic update module and the multi-step chain-based inference risk assessment module, forming a complete closed-loop data flow.

[0013] The beneficial effects of this invention are as follows: First, by deploying an AI agent with a closed-loop perception-decision-execution capability as a virtual health manager for COPD patients, it achieves 24 / 7 proactive perception and multimodal fusion analysis of patients' health status, overcoming the shortcomings of traditional solutions in terms of one-sided information collection and lack of real-time performance. Second, through the construction of an individualized baseline model and an online learning dynamic update mechanism, it overcomes the deficiency of traditional fixed threshold solutions in simultaneously achieving both sensitivity and specificity, enabling the system to provide accurate and tailored early warning services for patients with different degrees of severity. Third, through a multi-step chain-based inference engine and hierarchical matching verification of multiple rule chains, it achieves... It possesses risk assessment capabilities similar to those of clinicians, involving step-by-step screening and analysis, and can distinguish between different types of acute exacerbation prodromes, such as infectious, bronchospasm, and mixed types. Fourth, through an autonomous tiered intervention decision-making mechanism, it eliminates the human delay between risk detection and intervention implementation, reducing the intervention response time from tens of hours in traditional methods to minutes. Fifth, through tool-based access capabilities, it incorporates external environmental data into the risk assessment system, enabling collaborative analysis of environmental and physiological factors. Sixth, through a dual-pathway feedback closed-loop mechanism, the system can continuously self-optimize as management experience accumulates, and the overall solution reduces the 30-day readmission rate of COPD patients by approximately 28%. Attached Figure Description

[0014] Figure 1 This is a flowchart of the AI ​​agent-driven home health management method for COPD patients according to the present invention.

[0015] Figure 2 This is an architecture diagram of the AI ​​agent-driven home health management system for COPD patients according to the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0017] Reference Figure 1 This invention provides an AI agent-driven home health management method for COPD patients. The core of this method lies in deploying an AI agent with a five-stage closed-loop capability encompassing perception, reasoning, decision-making, execution, and feedback, acting as a virtual health manager for the patient. In one embodiment, the AI ​​agent runs on a hybrid architecture combining a cloud server and a patient-side edge gateway. The edge gateway handles low-latency preprocessing of real-time data, while the cloud server undertakes the complex computational tasks of the inference engine and knowledge base. The steps of the method are described in detail below.

[0018] Step S1: Multimodal health data perception and fusion step. In this step, the AI ​​agent comprehensively collects the health status of COPD patients through two parallel data perception channels. The first channel is the wearable device data acquisition channel. The AI ​​agent continuously connects to the wearable device worn on the patient's wrist or fingertip via Bluetooth Low Energy protocol. This wearable device has a built-in pulse oximeter, accelerometer, and respiratory motion sensor. Preferably, the blood oxygen saturation is collected every 10 minutes to balance monitoring continuity and device power consumption; the respiratory rate is collected every 3 minutes, as changes in respiratory rate have high early warning value for acute exacerbations of COPD; activity level is characterized by the integral value of the triaxial signal from the accelerometer over a 15-minute time window. This integral value reflects the patient's exercise tolerance changes in daily life. It is calculated by taking the absolute values ​​of the triaxial acceleration signals and then integrating them numerically within the time window. The Euclidean norm of the triaxial integral value is used as the final activity level indicator, with units of g·s; sleep quality is comprehensively assessed based on the number of body movements and the fluctuation range of blood oxygen saturation during the period from 22:00 to 06:00 the next day. In one embodiment of the present invention, a decrease in sleep quality is defined as a nighttime body movement frequency exceeding 30 times / hour or a blood oxygen fluctuation range exceeding 4%. The wearable device data acquisition channel also performs signal quality assessment. When the signal quality index of the pulse oximeter is lower than a preset threshold (e.g., perfusion index is lower than 0.3%), the data for that period is marked as low confidence. In the subsequent feature fusion and risk assessment stages, the low confidence data is downweighted to avoid misjudgment caused by loose sensor wear or motion artifacts.

[0019] The second channel is a voice-interactive proactive inquiry channel. The AI ​​agent initiates proactive voice inquiries twice daily, from 8:00 AM to 9:00 AM and from 4:00 PM to 5:00 PM. The inquiry content is structured and includes three dimensions: First, the degree of dyspnea, using a modified version of the UK Medical Research Council Dyspnea Scale (mMRC), guiding patients to self-assess on a scale of 0-4; second, the cough and sputum dimension, recording the frequency of coughs and changes in the color and characteristics of sputum within 24 hours, with color categorized into three levels: white mucoid sputum, yellow purulent sputum, and yellow-green purulent sputum; and third, medication adherence, confirming the actual number of times the patient used inhaled corticosteroids and long-acting bronchodilators that day. The AI ​​agent's built-in natural language understanding module performs semantic analysis on the patient's voice responses, extracting structured symptom feature values.

[0020] In one embodiment of the present invention, time alignment and feature-level fusion operations are performed on the data acquired from the two channels mentioned above. Specifically, based on the acquisition timestamp, the continuously acquired data from the wearable device is segmented and aggregated according to the time window of the voice inquiry, and the mean and rate of change of each segment are calculated. The rate of change of blood oxygen saturation, the rate of change of respiratory rate, and the rate of change of symptom score are all defined as the ratio of the difference between the mean of the corresponding feature in the current time window and the mean of the corresponding feature in the previous time window to the time interval between the two windows. Subsequently, the physiological parameter features (mean blood oxygen, rate of change of blood oxygen, mean respiratory rate, rate of change of respiratory rate, activity integral value, and sleep quality score, a total of 6 dimensions) are concatenated with the voice inquiry features (mMRC score, cough frequency, sputum level, and medication adherence rate, a total of 4 dimensions) to form a 10-dimensional multimodal health status feature vector. subscript This indicates the number of the data collection time window.

[0021] It should be further explained that, during the feature-level fusion process, to address the heterogeneity in the collection frequency and data type of physiological parameters and symptom inquiry data, this invention introduces a time window adaptive alignment strategy. After a patient completes a voice inquiry, the AI ​​agent automatically backtracks the wearable device data sequence from 2 hours prior to the inquiry time to the inquiry time itself and calculates the statistical features within that time window. Preferably, the statistical features extracted for the blood oxygen saturation dimension include four sub-features: mean, minimum value, percentage of cumulative duration below 90%, and slope of change. The percentage of cumulative duration below 90% is obtained by judging and accumulating each sample point; this sub-feature has high detection sensitivity for intermittent hypoxia events. Two sub-features, mean and peak value, are extracted for the respiratory rate dimension. These sub-features are compressed into 6-dimensional physiological parameter features after dimensionality reduction using principal component analysis, and then concatenated with the 4-dimensional voice inquiry features to form the final 10-dimensional multimodal health status feature vector. This fusion strategy ensures the consistency of different modalities of data across time scales while retaining the most clinically significant statistical features in each modality.

[0022] Step S2: Individualized Baseline Model Construction and Dynamic Update Step. This step aims to establish a healthy baseline model that reflects the individual characteristics of each COPD patient, distinguishing it from the uniform threshold used in traditional approaches. In one embodiment of the present invention, the construction of the individualized baseline model is divided into an initialization phase and a dynamic update phase.

[0023] During the initialization phase, the AI ​​agent acquires clinical assessment data recorded by the pulmonologist at the time of patient discharge as the initial baseline anchor. This clinical assessment data includes, but is not limited to: the measured FEV1 value and FEV1 / FVC ratio at discharge, the 6-minute walk distance (6MWD), PaO2 and PaCO2 values ​​from arterial blood gas analysis, and the COPD assessment test (CAT) score. This clinical assessment data is obtained from the electronic medical record system through a standardized data interface. The AI ​​agent performs format verification and reasonableness checks on the acquired data before storing it in the patient's personal health record. Preferably, continuous monitoring data for the first 14 days after discharge is used as the baseline calibration window. During this window, the patient is usually in a stable recovery period, and their various indicators can reliably reflect the patient's baseline status after discharge. Within the baseline calibration window, the multimodal health status feature vector sequence output in step S1 is processed. Calculate the sliding window mean for each dimension separately. with standard deviation Thus for the first Parameters establish individualized normal ranges ,in For the sensitivity coefficient, in one embodiment of the present invention, it is taken as... For severe patients with GOLD grade III-IV, the tightness can be reduced to... To improve early warning sensitivity. Furthermore, within the baseline calibration window, the trend slope of each parameter is fitted using linear regression. It is used to identify whether a patient is in a stable phase or a slow deterioration phase.

[0024] During the dynamic update phase, the AI ​​agent continuously updates the baseline parameters using an Exponentially Weighted Moving Average (EWMA) online learning mechanism. The specific update formula is as follows: ,in: For the updated number Baseline mean of the parameter; For the current moment, the first Measured values ​​of the parameters; For the previous version Baseline mean of the parameter; The learning rate ranges from 0.01 to 0.10, and is preferably set to a value of 0.01 in this invention. This value allows the baseline model to adapt to slow changes in patient condition over approximately 20 acquisition cycles without baseline drift due to occasional outliers. In one embodiment of the invention, the aforementioned EWMA update operation for the baseline mean and standard deviation is triggered only when the risk level output in step S3 is normal, thereby avoiding the inclusion of abnormal data in baseline calculations during acute exacerbations. Update using a similar EWMA mechanism.

[0025] It is worth noting that the individualized baseline model also includes a seasonal correction submodule. Because the respiratory function of COPD patients is significantly affected by seasonal changes, especially during the autumn-winter transition, the drop in temperature and dry air can cause patients' baseline oxygen saturation to generally decrease by 0.5% to 1.5% and their respiratory rate to increase by 1 to 3 breaths / min. Therefore, this invention embeds a seasonal compensation coefficient into the baseline model. superscript This indicates the current seasonal interval (spring, summer, autumn, or winter), and the compensation coefficient is obtained by fitting the patient's historical data from the same period. When a patient first uses the system and there is no historical data from the same period, a population statistic based on region and COPD severity is used as the initial value for the seasonal compensation coefficient, and it gradually transitions to an individualized value as subsequent data accumulates. In addition, the baseline model maintains an event-marked log to record confirmed acute exacerbations, medication adjustment events, and external interference events (such as concurrent colds). Data within the marked period will be excluded from the baseline update calculation to maintain the robustness of the baseline model.

[0026] Step S3: Multi-step chain-reasoning for acute exacerbation risk assessment. This step is the core of the intelligent risk assessment in this invention. Unlike the single-step threshold judgment or simple classifier methods used in the prior art, the inference engine built into the AI ​​agent of this invention adopts a multi-step chain-reasoning strategy to simulate the step-by-step screening and diagnosis process of a respiratory clinician for acute exacerbations of COPD.

[0027] First, the inference engine calculates the multidimensional normalized deviation of the current multimodal health state feature vector relative to the individualized baseline. For the... The normalized deviation of the term parameter is defined as follows: ,in: For the first The normalized deviation of the parameter is dimensionless. For the current moment, the first Measured values ​​of the parameters; This is the individualized baseline mean of this parameter; This is the individualized baseline standard deviation of the parameter. This normalization operation maps parameters of different dimensions to a comparable scale space.

[0028] Subsequently, the inference engine employs an adaptive weighted fusion strategy to combine the deviations of each dimension into a comprehensive deviation index. The calculation formula is: ,in: This is a dimensionless index representing the overall deviation. For the first The adaptive weighting coefficients of the term parameters satisfy the following: The weighting coefficients are dynamically adjusted based on the patient's COPD severity grading and historical exacerbation patterns. For example, for patients whose primary exacerbation pattern is infection-related, the weighting coefficients for changes in sputum color and respiratory rate are adjusted accordingly. In one embodiment of this invention, the default weighting allocation is as follows: blood oxygen saturation related parameters ( , The total accounted for 0.30, and respiratory rate-related parameters ( , The total percentage was 0.25, and the activity level and sleep parameters ( , The total percentage was 0.15, and the symptom score parameters ( , , The total percentage was 0.25%, and the medication adherence rate was ( ) accounts for 0.05.

[0029] Next, we move on to the core step of multi-step chain-based reasoning and matching. The AI ​​agent's inference engine has a built-in COPD acute exacerbation warning rule base containing three typical inference rule chains, each corresponding to one of the three main types of acute exacerbations:

[0030] The first rule chain is an infection-related aggravation rule chain. Its reasoning logic is as follows: (Chain node C1-1) Determine whether the sputum color changes from white to yellow or yellowish-green, and whether the sputum volume increases by more than 50% compared to the baseline; if so, proceed to (Chain node C1-2) to determine whether the respiratory rate increases by more than 20% compared to the baseline; if so, proceed to (Chain node C1-3) to determine whether the blood oxygen saturation drops below the individualized baseline lower limit; output the matching confidence score of this rule chain based on the matching degree of each node. The value ranges from 0 to 1.

[0031] The second rule chain is the airway spasm exacerbation rule chain, and its reasoning logic is as follows: (Chain node C2-1) Determine whether the mMRC score has increased by 2 grades or more from the baseline, and is manifested as increased wheezing after activity; if so, proceed to (Chain node C2-2) Determine whether the number of nighttime sleep interruptions has increased by more than 100% from the baseline; if so, proceed to (Chain node C2-3) Determine whether the as-needed use of inhaled bronchodilators has increased by more than 2 times / day from the baseline; output the matching confidence score. .

[0032] The third rule chain is a hybrid weighted rule chain, which combines the matching results of some nodes from the first two rule chains for cross-validation and outputs the matching confidence score. Specifically, the reasoning logic of the hybrid aggravation rule chain is as follows: if both chain node C1-1 of the infection rule chain and chain node C2-1 of the airway spasm rule chain are satisfied, and the matching confidence of both is not lower than 0.4, then the cross-validation subprocess is initiated. After comprehensively evaluating the co-occurrence patterns of abnormal indicators across various dimensions, the output is... This design stems from observations in clinical practice that approximately 30% of acute exacerbations of COPD are driven by a combination of infection and bronchospasm.

[0033] It is important to note that the matching confidence of each node in each rule chain is calculated using a fuzzy membership function. Taking node C1-1 of the infection-type rule chain as an example, the change in sputum color from white to yellow-green is not determined by binary judgment, but rather mapped to a continuous interval of 0 to 1 using a fuzzy membership function. For example, white sputum corresponds to a membership degree of 0, light yellow to 0.3, yellow to 0.6, and yellow-green to 0.9. The increase in sputum volume is also mapped proportionally between 10% and 50%. The matching confidence of a node is the geometric mean of the fuzzy membership degrees of all sub-conditions contained in that node. The technical significance of using the geometric mean instead of the arithmetic mean is that when the membership degree of any sub-condition is close to zero, the confidence of the entire node will be significantly reduced, thus avoiding the situation where a single strong indicator masks the absence of other important weak indicators.

[0034] Furthermore, in the multi-step chain-based reasoning process, the inference engine also introduces a temporal consistency verification mechanism. This mechanism requires that the abnormal features corresponding to each chain node exhibit a reasonable sequential evolution relationship in the time dimension. For example, in the infection-type exacerbation rule chain, changes in sputum color usually precede a significant increase in respiratory rate, and an increase in respiratory rate usually precedes a significant decrease in blood oxygen saturation. In one embodiment of the present invention, when the inference engine detects that the temporal relationship of abnormal indicators is inconsistent with the expected evolution temporal sequence of the rule chain (e.g., blood oxygen decreases first but sputum shows no significant change), it will reduce the final matching confidence of the rule chain and mark it as a suspected atypical exacerbation pattern requiring manual review by a clinician. This temporal consistency verification mechanism effectively reduces false alarms caused by temporary sensor errors or occasional patient activities.

[0035] Ultimately, the inference engine determines the overall deviation. The matching confidence with the three rule chains is used to output the acute exacerbation risk level using the following decision logic: when And all When the risk level is normal; when Or there exists any And all At that time, the risk level was mildly abnormal; when Or there exists any And all At that time, the risk level was moderately abnormal; when Or there exists any Or, if the absolute value of blood oxygen saturation is below 88%, the risk level is severe abnormality. In one embodiment of the present invention, , , The aforementioned threshold can be dynamically calibrated through the feedback path in step S5. Wherein: The matching confidence of the j-th rule chain in the COPD acute exacerbation early warning rule base is a dimensionless quantity ranging from 0 to 1, which is calculated by the chain node matching based on the aforementioned fuzzy membership function. j=1, 2, and 3 correspond to the infection-type exacerbation rule chain, the airway spasm-type exacerbation rule chain, and the mixed exacerbation rule chain, respectively. , , These are the first risk threshold, the second risk threshold, and the third risk threshold, all of which are dimensionless and satisfy the following conditions: < < This is used to classify the overall deviation D into different risk ranges. The Chinese meaning of the above decision logic is: when the overall deviation D is less than the first risk threshold... When the confidence scores of all three rule chains are below 0.3, the patient's health status is considered stable, and the risk level is normal. When the overall deviation D is between the first and second risk thresholds, or when the confidence score of any rule chain is 0.3 or higher but not 0.6, the patient is considered to have early abnormal signs, and the risk level is mildly abnormal. When the overall deviation D is between the second and third risk thresholds, or when the confidence score of any rule chain is 0.6 or higher but not 0.85, the patient's risk of acute exacerbation is considered significantly increased, and the risk level is moderately abnormal. When the overall deviation D reaches the third risk threshold or higher, or when the confidence score of any rule chain is 0.85 or higher, or when the absolute value of blood oxygen saturation is below 88%, the patient is considered to be at risk of acute exacerbation, and the risk level is severely abnormal.

[0036] Step S4: Autonomous Tiered Intervention Decision-Making and Execution Steps. This step demonstrates the core autonomous action capability that distinguishes the AI ​​agent from traditional passive monitoring systems. Based on the acute exacerbation risk level output in Step S3, the AI ​​agent autonomously executes tiered intervention measures that match the risk level without human intervention.

[0037] When the risk level is mildly abnormal, the AI ​​agent autonomously executes the first-level response. Specifically, the AI ​​agent pushes a personalized breathing training guidance plan through the patient's terminal application. This plan includes textual and audio guidance on pursed-lip breathing and diaphragmatic breathing exercises, with a recommended training frequency of 3 to 4 times per day, each session lasting 5 to 10 minutes. Simultaneously, the AI ​​agent checks the patient's current medication records for any missed medications. If a missed medication is detected, it immediately pushes a medication reminder along with a demonstration of the correct usage of the medication. Preferably, during the mildly abnormal period, the AI ​​agent will also increase the monitoring frequency for the following 24 hours from the regular mode to an encrypted mode, for example, increasing the blood oxygen saturation monitoring frequency from once every 10 minutes to once every 5 minutes, in order to more closely track changes in the patient's condition.

[0038] When the risk level is moderately abnormal, the AI ​​agent autonomously executes a Level 2 response. Building upon all measures of the Level 1 response, the AI ​​agent proactively sends a notification of changes in the patient's condition to the patient's pre-registered emergency contact (usually a family member) via SMS or instant messaging. The notification includes the current risk level, a summary of key abnormal indicators, and recommended measures. Simultaneously, the AI ​​agent queries the outpatient appointment information of the contracted community health service center via an interface, automatically generating outpatient follow-up suggestions and providing optional appointment slots for the patient. In one embodiment of this invention, if the moderately abnormal state persists for more than 48 hours without showing any improvement, the AI ​​agent automatically upgrades the risk level to severe abnormality.

[0039] When the risk level is deemed severely abnormal, the AI ​​agent autonomously executes a Level 3 response. The AI ​​agent immediately sends a structured alert report via system messaging to the patient's contracted family physician or respiratory specialist, containing a trend chart of the patient's physiological parameters over the past 72 hours, records of symptom changes, and the analysis conclusions of the inference engine. Simultaneously, the AI ​​agent assists the patient in establishing a remote video consultation channel with the physician, continuously monitoring the patient's condition while waiting for the physician to connect, and providing emergency response guidance, such as instructing the patient to adopt a comfortable posture and administer an emergency inhalation of a short-acting bronchodilator.

[0040] Furthermore, in one embodiment of the present invention, the autonomous graded intervention decision-making also incorporates an intervention intensity progression mechanism. Specifically, after an intervention measure of a certain risk level is implemented, the AI ​​agent continuously tracks the immediate effect of the intervention. If the patient's overall deviation is within 4 hours after the intervention is implemented, the AI ​​agent will further monitor the effect. If the deviation does not decrease or even continues to increase, the AI ​​agent will automatically trigger an escalation of the intervention intensity. For example, if a breathing training guide for mild abnormalities is sent out, and an assessment 4 hours later reveals that the deviation is still increasing, the AI ​​agent will automatically send out more detailed postural drainage guidance and schedule the next voice inquiry in advance. Similarly, if a moderate abnormality is detected and the family is notified, and an assessment 12 hours later reveals that the patient has still not sought medical attention and the deviation continues to worsen, the AI ​​agent will proactively initiate a second contact and simultaneously send an alert to the contracted physician. This progressive mechanism ensures that the AI ​​agent's intervention is not a one-off trigger, but a complete intervention process with continuous tracking and autonomous adjustment capabilities.

[0041] Preferably, the AI ​​agent automatically generates and stores intervention event logs when executing interventions at all levels. The logs include the intervention trigger time, the risk level and overall deviation value at the time of triggering, a list of specific intervention measures, patient confirmation feedback (if any), the deviation trend at 4 hours and 24 hours after the intervention, and the final outcome label (improvement, stabilization, or deterioration). This log data constitutes the key input source for feedback optimization in step S5.

[0042] Step S5: Tool-invoking contextualized service and feedback optimization step. This step empowers the AI ​​agent to autonomously invoke external tools, enabling it to proactively acquire contextual information related to the patient's environment and medication during the health management process, thereby providing comprehensive health services that go beyond a single physiological monitoring dimension.

[0043] Regarding environmental context services, the AI ​​agent periodically calls meteorological and air quality monitoring services through a standardized RESTful API interface. Preferably, the AI ​​agent calls the meteorological service every 6 hours to obtain the temperature forecast data for the patient's city for the next 24 to 72 hours, and calls the air quality monitoring service every 3 hours to obtain the current air quality index (AQI) and PM2.5 concentration. When the AI ​​agent detects a sudden drop in temperature exceeding 10°C or an AQI exceeding 150 within the next 24 hours, it automatically generates an environmental warning reminder, which includes suggestions to reduce outdoor activities, pay attention to keeping warm, and, if necessary, increase the frequency of inhaled medication use. In one embodiment of the present invention, when environmental factors trigger a warning, the AI ​​agent also simultaneously adjusts the risk assessment sensitivity parameter of the inference engine in step S3, and adjusts the comprehensive deviation threshold. and The risk of acute exacerbation caused by environmental factors will be reduced by 15% to detect these factors in advance.

[0044] In terms of pharmaceutical knowledge services, the AI ​​agent accesses a structured pharmaceutical knowledge base, which covers information such as indications, dosage, adverse reactions, drug interactions, and storage precautions for commonly used COPD medications. When a patient raises medication questions via voice, the AI ​​agent performs intent recognition and entity extraction on the patient's question, retrieves matching information from the pharmaceutical knowledge base, and answers the patient in natural language. For example, when a patient asks what to do about throat discomfort after using budesonide inhaler, the AI ​​agent can identify the medication entity as budesonide inhaler, the symptom entity as oropharyngeal discomfort, and match the medication guidance from the knowledge base that the patient should rinse their mouth thoroughly after use.

[0045] Regarding feedback optimization, step S5 constructs a dual-pathway feedback loop from intervention effects to model parameters. The first feedback path points to the individualized baseline model in step S2: the AI ​​agent continuously tracks the patient's response for 24 to 72 hours after each tiered intervention, recording the changes in multimodal health status feature vectors before and after the intervention as feedback data, which is used to adjust the trend slope parameter in the baseline model. It should be noted that the first feedback pathway adjusts the trend slope parameter, which reflects the direction of disease progression. The baseline mean is not updated directly with abnormal data during acute exacerbations. with standard deviation When a patient's risk level returns from abnormal to normal after intervention, their multimodal health status feature vector during the recovery period is used as normal state data and is reincorporated into the baseline mean according to the EWMA mechanism described in step S2. with standard deviation The update allows post-intervention recovery data to be incorporated into baseline optimization. Therefore, the constraint in step S2 that the baseline mean update is triggered only when the risk level is normal does not contradict the feedback mechanism in this step that returns the intervention results to step S2. Both mechanisms act on different parameters of the baseline model and at different time periods, jointly ensuring the establishment of a dual-pathway feedback loop. The second feedback path points to the inference engine in step S3: the AI ​​agent statistically analyzes the actual outcomes (improvement, stabilization, or deterioration) after intervention at each risk level. When the statistics show that the false positive or false negative rate for a certain risk level exceeds a preset threshold, it automatically adjusts the corresponding deviation threshold. , or And the matching confidence weight of the rule chain. In one embodiment of the present invention, a feedback evaluation cycle is set at 30 days. If the false alarm rate for minor anomalies exceeds 40%, then... Increase by 10%; if the false negative rate for severe anomalies exceeds 5%, then... The reduction is 15%. Through the aforementioned dual-path feedback mechanism, the AI ​​agent can continuously improve the accuracy of risk assessment and the appropriateness of intervention decisions as management experience accumulates, thereby truly realizing the system's self-evolution capability.

[0046] More specifically, the feedback optimization module also maintains an intervention efficacy evaluation matrix. This matrix has three risk levels (mild, moderate, and severe) in its rows and three outcomes (improvement, stability, and deterioration) in its columns. Matrix elements include event counts for each period. By analyzing this matrix, the AI ​​agent can identify systematic evaluation bias patterns. For example, when the improvement column count for a mild abnormal row is significantly greater than the sum of the stability and deterioration column counts, it suggests that the trigger threshold for that level may be too sensitive and needs to be adjusted upwards. Conversely, when the deterioration column count for a moderate abnormal row exceeds 25%, it suggests that the intervention measures for that level may be insufficient or the threshold setting may be too lenient. In one embodiment of this invention, the AI ​​agent automatically generates a monthly management report based on the intervention efficacy evaluation matrix and pushes the report to the contracted physician for clinical evaluation reference. Physicians can review and confirm the threshold adjustment suggestions from the AI ​​agent or manually correct them, thereby establishing a collaborative mechanism between automatic optimization and manual supervision.

[0047] Furthermore, the tool-calling capability in step S5 extends to health education push functionality. Based on the patient's current health status and historical interaction records, the AI ​​agent proactively pushes educational content from a pre-built COPD health education content library that best matches the patient's current needs. For example, when the AI ​​agent detects a continuous decline in the patient's activity level for three consecutive days, it proactively pushes links to home-based pulmonary rehabilitation exercise guidance videos suitable for the current lung function level; when the patient's medication adherence rate is below 80%, it pushes information on the efficacy of the medication and the benefits of consistent treatment. These proactively pushed health education contents complement the aforementioned passive medication Q&A, together forming a comprehensive, contextualized health service system provided by the AI ​​agent.

[0048] In summary, the AI ​​agent-driven home health management method for COPD patients of this invention forms a complete closed-loop system of perception-reasoning-decision-execution-feedback. Multimodal perception in step S1 provides continuously updated data input for baseline modeling in step S2 and risk assessment in step S3; the individualized baseline model in step S2 provides a dynamic reference benchmark for deviation calculation in step S3; the risk assessment results in step S3 directly drive the tiered intervention decision in step S4; the intervention execution process and results in step S4 are fed back to steps S2 and S3 respectively via the dual-pathway feedback mechanism in step S5, achieving adaptive updating of baseline parameters and dynamic calibration of risk assessment thresholds. This deeply coupled closed-loop architecture allows improvements in any part of the system to propagate to the entire system through the feedback pathway, resulting in a synergistic effect where overall performance is greater than the simple sum of the performance of individual modules.

[0049] Preferably, in practical deployment, the method of the present invention can be flexibly configured at the parameter level according to the characteristics of different patient groups. For example, for mild to moderate COPD patients with GOLD classification I to II, the monitoring frequency can be appropriately extended and the deviation threshold relaxed to reduce the intervention frequency and avoid excessive intervention affecting the patient's quality of life; for severe and very severe COPD patients with GOLD classification III to IV, as well as patients with a history of frequent exacerbations, the monitoring frequency and deviation threshold are tightened, and the automatic escalation judgment time for moderate abnormalities is shortened from 48 hours to 24 hours. In addition, for COPD patients with concurrent cardiovascular disease, a heart rate variability monitoring dimension can be added to the data perception stage in step S1, and a reasoning rule chain related to cardiopulmonary interaction can be added to the rule base in step S3, thereby achieving more precise health management for patients with complex comorbidities.

[0050] Reference Figure 2 This invention also provides an AI agent-driven home health management system for COPD patients. This system includes a multimodal health data perception and fusion module, a personalized baseline model construction and dynamic update module, a multi-step chain-based inference risk assessment module, a self-regulating hierarchical intervention decision-making and execution module, and a tool-invoking contextualized service and feedback optimization module. Each module corresponds one-to-one with steps S1 to S5 in the aforementioned method embodiments. The specific composition of each module is described below.

[0051] The multimodal health data perception and fusion module performs the functions described in step S1 above. In one embodiment of the present invention, this module includes a wearable device data access submodule, a voice interaction inquiry submodule, and a multimodal data fusion submodule. The wearable device data access submodule connects to the pulse oximeter and motion sensor worn by the patient via Bluetooth Low Energy protocol, receives raw data on blood oxygen saturation, respiratory rate, activity level, and sleep quality at a preset acquisition frequency, and performs preprocessing operations such as noise reduction and outlier removal on the raw data. The voice interaction inquiry submodule integrates a natural language understanding engine, actively initiates voice interaction according to a set inquiry schedule, and collects the patient's mMRC score, cough and sputum status, and medication administration status. The multimodal data fusion submodule executes the aforementioned time window adaptive alignment strategy, performs feature-level fusion of physiological parameter features and voice inquiry features, and outputs a 10-dimensional multimodal health status feature vector.

[0052] The individualized baseline model construction and dynamic update module performs the functions described in step S2 above. This module includes an initial baseline construction submodule, an online learning update submodule, and a seasonal correction submodule. The initial baseline construction submodule imports the patient's clinical assessment data at discharge and calculates the individualized normal ranges for each parameter within the baseline calibration window. The online learning update submodule continuously updates the baseline mean and standard deviation under normal conditions using the aforementioned exponentially weighted moving average online learning mechanism. The seasonal correction submodule applies compensation coefficients to the baseline parameters based on the seasonal interval of the current calendar month. Preferably, the individualized baseline model construction and dynamic update module also includes an event labeling management submodule, used to maintain a labeling log for acute exacerbation events, medication adjustment events, and external interference events, ensuring that data from labeled periods are excluded from baseline update calculations.

[0053] The multi-step chain-based inference risk assessment module performs the functions described in step S3 above. This module includes a deviation calculation submodule, a rule base management submodule, and a chain-based inference engine submodule. The deviation calculation submodule receives multimodal health status feature vectors and calls the current individualized baseline model parameters. It calculates the deviation of each parameter dimension-by-dimensional according to the aforementioned normalized deviation calculation formula and outputs the comprehensive deviation through an adaptive weighted fusion strategy. The rule base management submodule stores and manages three inference rule chains—infectious, airway spasm, and mixed—and their corresponding fuzzy membership function parameters, supporting versioned management and online updates of the rule chain content. The chain-based inference engine submodule sequentially traverses the chain nodes of each rule chain according to the multi-step inference process, calculates the matching confidence of each node, and, combined with a temporal consistency verification mechanism, finally outputs the acute exacerbation risk level.

[0054] The autonomous tiered intervention decision-making and execution module is used to perform the functions of the aforementioned step S4. In one embodiment of the present invention, this module includes a decision engine submodule and a multi-channel execution submodule. The decision engine submodule maps the received risk level to the corresponding intervention response plan. This mapping relationship is stored in the form of a configurable rule table, supporting contracted physicians to make personalized adjustments to the intervention plan according to changes in the patient's condition. The multi-channel execution submodule integrates a patient terminal push interface, a family notification interface, a physician warning interface, and a remote consultation bridging interface. It can simultaneously or sequentially call multiple intervention channels to complete the execution of the tiered response. Each channel has a built-in sending confirmation and timeout retry mechanism to ensure reliable delivery of intervention information. This module also includes an intervention tracking submodule, which is used to continuously monitor changes in the patient's condition after the intervention is executed and determine whether the intervention intensity needs to be escalated according to the aforementioned progressive mechanism.

[0055] The tool-invoking contextualized service and feedback optimization module performs the functions described in step S5 above. This module includes an external tool invoking submodule, a drug knowledge base query submodule, a health education push submodule, and a feedback optimization submodule. The external tool invoking submodule connects to meteorological and air quality monitoring services via a RESTful API interface, supporting parallel queries and result consistency verification from multiple data sources. The drug knowledge base query submodule accesses a structured COPD drug knowledge base, which employs a knowledge graph architecture based on medical ontology and covers complete attribute information for over 200 commonly used COPD drugs. The health education push submodule matches appropriate educational content based on the patient's current state. The feedback optimization submodule constructs the aforementioned dual-pathway feedback loop, feeding intervention effect data back to the individualized baseline model construction and dynamic update module and the multi-step chain-based inference risk assessment module, respectively, to achieve continuous self-optimization of the system.

[0056] The data flow between the above modules constitutes the following: Figure 2 The closed-loop architecture shown is as follows: the multimodal health data perception and fusion module simultaneously outputs feature vectors to the individualized baseline model construction and dynamic update module and the multi-step chain-inference risk assessment module; the risk level output of the multi-step chain-inference risk assessment module is sent to the autonomous hierarchical intervention decision and execution module; the intervention log of the autonomous hierarchical intervention decision and execution module is output to the tool-invoking contextualized service and feedback optimization module; the tool-invoking contextualized service and feedback optimization module sends feedback data back to the individualized baseline model construction and dynamic update module and the multi-step chain-inference risk assessment module respectively, completing the entire closed loop.

[0057] In one embodiment of the present invention, the deployment architecture of the above system adopts a three-layer collaborative model of cloud-edge-device. The terminal layer consists of wearable devices worn by the patient and smart terminals (mobile phones or tablets) with AI agent client applications installed, responsible for raw data collection and user interface presentation. The edge layer consists of edge gateway devices deployed in the patient's home network. These devices carry the data preprocessing and feature extraction calculation tasks in the multimodal health data perception and fusion module to ensure that basic data processing capabilities can be maintained even under network fluctuations. The cloud layer carries the core computing tasks of the individualized baseline model construction and dynamic update module, the multi-step chain-based inference risk assessment module, and the tool-calling contextualized service and feedback optimization module, providing sufficient computing resources to support the complex calculations of the inference engine and network calls to external tool interfaces. The autonomous hierarchical intervention decision and execution module runs across the edge and cloud layers. Its decision engine submodule is deployed in the cloud to obtain full data support, while the multi-channel execution submodule retains local emergency response capabilities for mild anomalies at the edge layer, ensuring that basic intervention measures can still be executed under extreme network interruption conditions. This three-layer collaborative architecture ensures the system's intelligence level while also taking into account the actual needs of network connectivity reliability and real-time data processing in home scenarios.

[0058] In summary, the AI ​​agent-driven home health management method for COPD patients proposed in this invention achieves end-to-end intelligent closed-loop management from patient data collection to risk assessment, and from intervention decision-making to effect feedback through deep coupling and synergy of five stages: multimodal perception, individualized baseline modeling, multi-step chain reasoning, autonomous hierarchical intervention, and closed-loop feedback optimization. The tight coupling between each stage produces a significant synergistic effect. Compared with existing solutions that implement each functional module in isolation, the technical effect of this invention far exceeds the simple superposition of individual modules, fully demonstrating the system-level synergistic advantage of one plus one being greater than two.

[0059] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. An AI agent-driven home-based health management method for COPD patients, characterized in that, Includes the following steps: Step S1, Multimodal Health Data Perception and Fusion Step: Deploy an AI agent to continuously access wearable devices to collect patients' blood oxygen saturation, respiratory rate, activity level and sleep quality data as physiological parameter data, and actively inquire about patients' cough and sputum, degree of breathing difficulty and medication execution status through voice interaction channel to obtain voice inquiry data. Perform time alignment and feature-level fusion of physiological parameter data and voice inquiry data to generate multimodal health status feature vector; Step S2, Individualized Baseline Model Construction and Dynamic Update: Based on the patient's clinical assessment data at discharge and the multimodal health status feature vectors accumulated in Step S1, an individualized baseline model is constructed, which includes individualized normal ranges of various physiological parameters and symptom change trend parameters. The model parameters are continuously updated using an online learning mechanism. Step S3, multi-step chain reasoning acute exacerbation risk assessment step: The AI ​​agent's reasoning engine receives the multimodal health status feature vector output in step S1, calls the individualized baseline model in step S2 to calculate the current health status deviation, performs multi-step chain reasoning matching with the COPD acute exacerbation early warning rule base, and outputs acute exacerbation risk levels including normal, mild abnormal, moderate abnormal and severe abnormal. The acute exacerbation risk level is jointly determined by the current health status deviation and the matching confidence of each rule chain obtained by the multi-step chain reasoning matching. Step S4, Autonomous Graded Intervention Decision and Execution Steps: The AI ​​agent autonomously executes graded interventions based on the risk level output in Step S3. When there is a mild abnormality, it pushes breathing training guidance and medication reminders. When there is a moderate abnormality, it proactively contacts the family and suggests outpatient follow-up. When there is a severe abnormality, it sends an alert to the contracted physician and assists in establishing a remote consultation channel. Step S5, Tool-based contextualized service and feedback optimization steps: The AI ​​agent autonomously calls external tool interfaces to query weather and air quality data and reminds patients to reduce going out when pollution warnings are issued. It calls the drug knowledge base to answer medication questions and simultaneously sends the intervention execution results and patient feedback back to step S2 to update the symptom change trend parameters in the individualized baseline model. After the patient's risk level drops from abnormal to normal, the recovery period data is included in the update of the individualized baseline model. The intervention effect evaluation is fed back to step S3 to calibrate the risk assessment threshold, forming a closed-loop collaboration.

2. The AI ​​agent-driven home health management method for COPD patients according to claim 1, characterized in that, In step S1, the blood oxygen saturation is collected every 5 to 15 minutes, the respiratory rate is collected every 1 to 5 minutes, the activity level is characterized by the triaxial signal integral value of the accelerometer and is collected every 15 to 30 minutes cumulatively, and the sleep quality is comprehensively evaluated based on the number of nighttime body movements and the amplitude of blood oxygen fluctuations.

3. The AI ​​agent-driven home health management method for COPD patients according to claim 1, characterized in that, In step S1, the voice interaction channel actively initiates inquiries to the patient at a frequency of no less than twice a day. The inquiries include the modified UK Medical Research Council Dyspnea Scale score, the frequency and color of cough within 24 hours, and the number of times inhaled medications are used that day.

4. The AI ​​agent-driven home health management method for COPD patients according to claim 1, characterized in that, In step S2, the construction of the individualized baseline model includes: using the patient's pulmonary function test value, 6-minute walk test result and most recent blood gas analysis value at discharge as the initial baseline anchor point, using the continuous monitoring data from the first 7 to 14 days after discharge as the baseline calibration window, and using the sliding window mean and standard deviation to calculate the individualized normal range of each parameter.

5. The AI ​​agent-driven home health management method for COPD patients according to claim 1, characterized in that, In step S3, the multi-step chain-based reasoning matching includes: First, based on the multimodal health status feature vector relative to the individualized baseline model, extracting the rate of change of blood oxygen saturation, the rate of change of respiratory rate, and the change of symptom score as reasoning input evidence at the current moment; Second, matching the reasoning input evidence with the infection-type exacerbation rule chain, the airway spasm-type exacerbation rule chain, and the mixed-type exacerbation rule chain in the COPD acute exacerbation early warning rule base; Third, combining the matching confidence of each rule chain, and using a weighted fusion strategy to output the acute exacerbation risk level.

6. The AI ​​agent-driven home health management method for COPD patients according to claim 5, characterized in that, The reasoning logic of the infection-type exacerbation rule chain in the COPD acute exacerbation early warning rule base is as follows: if the sputum color changes from white to yellow-green, and the sputum volume increases by more than 50% compared to the baseline, and the body temperature is higher than the baseline temperature, it is inferred to be an infectious exacerbation precursor. Based on this, it is determined whether the blood oxygen saturation is lower than the lower limit of the individualized normal range of blood oxygen saturation in the individualized baseline model, and the matching confidence of the infection-type exacerbation rule chain is determined according to the degree to which the blood oxygen saturation is lower than the lower limit.

7. The AI ​​agent-driven home health management method for COPD patients according to claim 1, characterized in that, In step S3, the deviation is calculated using a multidimensional normalized deviation fusion algorithm. After mapping the deviation of various physiological parameters and the deviation of symptom scores to the same scale, the deviation is weighted and summed using an adaptive weighting coefficient. The weighting coefficient is dynamically adjusted according to the patient's COPD severity level and historical aggravation pattern.

8. The AI ​​agent-driven home health management method for COPD patients according to claim 1, characterized in that, In step S4, the triggering condition for mild abnormality is that the fusion deviation corresponding to the acute exacerbation risk level is between the first threshold and the second threshold; the triggering condition for moderate abnormality is that the fusion deviation is between the second threshold and the third threshold; and the triggering condition for severe abnormality is that the fusion deviation exceeds the third threshold or the absolute value of blood oxygen saturation is lower than the preset safety lower limit.

9. The AI ​​agent-driven home health management method for COPD patients according to claim 1, characterized in that, In step S5, the AI ​​agent calls the meteorological service through a standardized application programming interface to obtain the forecast values ​​of the temperature drop and air quality index for the next 24 to 72 hours. When the temperature drop exceeds the preset temperature threshold or the air quality index exceeds the preset pollution threshold, an environmental warning is generated and the risk assessment sensitivity parameters in step S3 are adjusted simultaneously.

10. An AI agent-driven home health management system for COPD patients, used to implement the AI ​​agent-driven home health management method for COPD patients as described in any one of claims 1-9, characterized in that, include: The multimodal health data perception and fusion module is used to continuously access wearable devices through AI agents to collect patients' blood oxygen saturation, respiratory rate, activity level and sleep quality data as physiological parameter data. It also actively asks patients about their cough and sputum, the degree of breathing difficulty and medication execution status through a voice interaction channel to obtain voice inquiry data. The physiological parameter data and voice inquiry data are time-aligned and feature-level fused to generate a multimodal health status feature vector. The individualized baseline model construction and dynamic update module is used to build and continuously update an individualized baseline model for patients based on their clinical assessment data at discharge and the multimodal health status feature vectors accumulated by the multimodal health data perception and fusion module. The multi-step chain-based reasoning risk assessment module is used to receive multimodal health status feature vectors and call an individualized baseline model to calculate the deviation of the current health status. The deviation is then matched with the COPD acute exacerbation early warning rule base through multi-step chain-based reasoning, and the acute exacerbation risk level is output, including normal, mild abnormal, moderate abnormal and severe abnormal. The acute exacerbation risk level is jointly determined by the current health status deviation and the matching confidence of each rule chain obtained by the multi-step chain-based reasoning. The autonomous tiered intervention decision-making and execution module is used to autonomously make decisions and execute tiered intervention responses based on the risk level of acute exacerbation. The tool-invoking contextualized service and feedback optimization module is used by the AI ​​agent to autonomously call external tool interfaces to obtain environmental data and call the drug knowledge base to provide contextualized health services. It also sends the intervention execution results and patient feedback data back to the personalized baseline model construction and dynamic update module to update the symptom change trend parameters in the personalized baseline model. After the patient's risk level falls back to normal, the recovery period data is included in the update of the personalized baseline model. Finally, the intervention effect evaluation results are fed back to the multi-step chain inference risk assessment module to calibrate the risk assessment threshold.

Images