Large-model-driven intelligent old-age care internet-of-things system and chronic disease personalized intervention system

Abstract

The invention relates to the technical field of health management, in particular to a large-model-driven intelligent old-age care internet-of-things system and a chronic disease personalized intervention system. Comprising a multi-dimensional physiological data acquisition module, an edge preprocessing and real-time alarm module, a data transmission and time sequence database storage module, an individual feature matching module, a dynamic personalized intervention module, an intervention execution and compliance monitoring module, an intervention effect evaluation and model iteration module and a deployment and compliance guarantee module. Data application is upgraded from fragmented storage to deep value conversion, chronic disease risk prediction accuracy is improved, an intervention scheme is upgraded from universal guidance to dynamic personalized adaptation, health management practicability is improved, a technical system is upgraded from a single link to a full-link closed loop, landing feasibility and scene adaptability are improved, and the risk prediction accuracy of chronic diseases is improved. And the business value is remarkable, and the chronic disease prevention and control and old-age service efficiency improvement is facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of health management, in particular to a large model driven smart elderly care Internet of Things system and a chronic disease personalized intervention system. BACKGROUND

[0002] Chronic diseases have become a major global health burden. Although the 2018-2025 forecast shows that some chronic diseases (such as COPD and diabetes) have a downward trend, the overall prevention and control situation is still severe. At present, more than 190 million elderly people in China suffer from chronic diseases.

[0003] Traditional home-based elderly care models in the field of health management, such as, face two challenges: data fragmentation (sleep, heart rate, blood pressure, blood sugar, etc. Indicators are scattered and recorded in real time), and intervention is extensive (health recommendations lack personalization); On the one hand, in the context of smart elderly care, the physiological health data of the elderly shows a significant fragmentation feature: covering sleep (such as sleep duration, sleep stage, number of night awakenings), heart rate (such as resting heart rate, exercise heart rate, heart rate variability), blood pressure (such as systolic pressure, diastolic pressure, day / night blood pressure fluctuations), blood sugar (such as fasting blood sugar, 2-hour postprandial blood sugar, random blood sugar) and other core health indicators. These indicators are not only scattered and recorded in different monitoring devices (such as smart bracelets, blood pressure meters, blood glucose meters, sleep monitors) or data terminals, lacking unified integrated management, and the data generation has real-time nature - that is, dynamic data is continuously generated as the elderly's daily physiological state changes (such as heart rate updated every second, blood sugar fluctuates in real time with diet / exercise, sleep data is dynamically recorded according to sleep period), resulting in a fragmented form of "multiple sources, scattered, real-time high frequency, lack of correlation".

[0004] On the other hand, there are problems of extensive intervention in the health intervention link of smart elderly care: the output of health recommendations lacks personalized design - it is not matched according to the individual characteristics of the elderly (such as age, gender, history of underlying diseases), long-term physiological monitoring data (such as heart rate variability trend, blood sugar control stability, sleep structure characteristics) and chronic disease management goals (such as blood pressure control threshold for hypertension patients, blood sugar target for diabetes patients), only provides universal health guidance (such as "eat more fruits and vegetables" "avoid staying up late"), which cannot adapt to the risk level of different elderly people and cannot fit their living habits and physical tolerance, resulting in health intervention that cannot accurately reach individual needs, making it difficult to achieve efficient chronic disease prevention and control and health management.

[0005] Therefore, a large model driven smart elderly care Internet of Things system and a chronic disease personalized intervention system are needed to improve the above problems. SUMMARY

[0006] To solve the above technical problems, the present application provides a big model driven smart elderly care Internet of Things system and a chronic disease personalized intervention system, which can upgrade the data application from fragmented storage to deep value conversion, improve the accuracy of chronic disease risk prediction, upgrade the intervention scheme from universal guidance to dynamic personalized adaptation, improve the practicality of health management, upgrade the technical system from a single link to a full-link closed loop, improve the feasibility and scene adaptability of landing, and significantly improve the business value, helping to prevent and control chronic diseases and improve the efficiency of elderly care services.

[0007] The big model driven smart elderly care Internet of Things system and the chronic disease personalized intervention system of the present application comprise: A multi-dimensional physiological data acquisition module: long-term and stable acquisition of core physiological data of the elderly, such as sleep, heart rate, blood pressure and blood glucose, and home behavior data, to provide continuous and reliable original data sources for subsequent analysis; An edge preprocessing and real-time alarm module: preliminary cleaning and anomaly detection of data are completed at the edge, reducing the amount of invalid data uploaded, and real-time alarm for emergency health risks is provided to ensure the safety of the elderly; A data transmission and time series database storage module: the preprocessed data are uploaded to the cloud through a stable transmission protocol, and are stored in a time series database to ensure consistent data order and traceability, and support quick query by user, time and data type; A physiological data fusion analysis module: multi-source time series physiological data are fused and analyzed to predict future short-term and medium-term chronic disease risks, output risk level, warning window and key contributing factors, and solve the problem of data fragmentation and lack of depth application; An individual feature matching module: static information, disease history and living habits of the elderly are integrated to establish an individual archive, which provides individual basis for personalized intervention and avoids disconnection between intervention scheme and individual conditions; A dynamic personalized intervention module: combining the risk prediction results of chronic diseases and individual archives, readable suggestions are generated through LLM, and the clinical safety is verified by a rule engine, and a personalized intervention scheme matched with risks and individuals is output to solve the problem of extensive intervention; An intervention execution and compliance monitoring module: intervention tasks are issued to users or caregivers through multiple terminals, task execution data are collected in real time, compliance scores are automatically calculated, and intervention landing conditions are mastered; An intervention effect evaluation and model iteration module: the influence of intervention on physiological indicators and chronic disease risks is quantified, intervention strategies are optimized through A / B testing and causal inference, and the risk prediction model is iteratively updated to form a closed loop of intervention-evaluation-optimization; A deployment and compliance guarantee module: a hybrid deployment mode of edge+cloud is adopted to balance data processing real-time performance and privacy security, meet the compliance requirements of GDPR and other regulations, and ensure long-term stable operation of the system.

[0008] Preferably, the data processing flow of the multi-dimensional physiological data acquisition module comprises: Step one, hardware selection and data acquisition range determination: The intelligent sleep band monitors sleep duration, sleep staging, number of night awakenings, sleep heart rate and sleep respiration rate. The six-in-one vital sign monitor monitors blood pressure, blood glucose, blood oxygen, uric acid, heart rate and body temperature data. The infrared sensor can monitor the length of time and the number of activities of the user in different spaces in the home care scene, ensuring that the data covers physiological indicators and behavioral indicators. Step two, real-time data acquisition and preliminary labeling: The hardware collects data at a set frequency and automatically labels data quality. 0 indicates normal data, 1 indicates slight noise, and 2 indicates data loss or device abnormality, providing a basis for subsequent cleaning.

[0009] Preferably, the data processing flow of the edge preprocessing and real-time alarm module comprises: Step one, data denoising and repeated data rejection: High-frequency data is processed using a 3-sample moving average filter to remove transient spike noise. Repeated data is determined by combining device_id + timestamp + sensor_type keys. If the data value difference is <5% under the same combination key, the first collected data is retained and the subsequent repeated data is rejected, reducing data transmission volume. Step two, emergency exception detection and alarm triggering: 1) Low blood oxygen detection: Real-time monitoring of blood oxygen data, if the blood oxygen saturation is <88% and the duration is >5 minutes, an emergency alarm is triggered, and the blood oxygen fluctuation curve in this time period is recorded as the basis for the alarm. 2) Blood pressure abnormality detection: If the systolic pressure is >180 mmHg or the diastolic pressure is >110 mmHg, a medium-level alarm is triggered, and the deviation rate is calculated by comparing the user's blood pressure average value in the past 7 days.

[0010] Preferably, the data processing flow of the data transmission and time series database storage module comprises: Step one, data transmission protocol selection and reliability guarantee: The LoRa low-power Internet of Things gateway and MQTT transmission protocol of the prior patent are used to ensure stable data transmission in a weak home network environment. The at-least-once transmission strategy is adopted. If the gateway detects that the data has not been confirmed by the cloud, it will be retransmitted every 5 minutes until it is successfully received. The transmission time and reception time of each data are recorded, the transmission delay is calculated, and if the delay is >30 seconds, it is marked as a transmission delay data and stored separately. Step two, time series database selection and table structure design: ClickHouse is selected as the time series database, and the table structure is designed according to the user-data type-time dimension, and the joint primary key index is established to improve the efficiency of multi-condition query.

[0011] Preferably, the data processing flow of the physiological data fusion analysis module comprises: Step one, data preprocessing: 1) Time series data alignment and resampling: resample data of different frequencies with 1 hour as the time granularity, fill in missing data by linear interpolation, and mark missing segments with missing_segment if the missing time is more than 3 hours; 2) Feature engineering: extract time series features, periodic features and text features; Step two, time series Transformer model training and inference: 1) Model architecture: linear projection and relative time position encoding are performed on the input layer, TransformerEncoder captures long-term dependence of data, cross-attention layer fuses static features and text embedding, and output layer outputs multi-label chronic disease risk probability through MLP; 2) Training strategy: user stratification K-fold cross-validation is used, FocalLoss is used to weight rare events, sample imbalance is balanced, data augmentation is performed through time mask and noise injection to improve model robustness; 3) Inference process: input the user's recent 30-day time series data, and the model outputs the chronic disease risk probability in the next 30 days or 90 days; Step three, explainability analysis: 1) Attention weight visualization: output the contribution weight of each time step and each data type to risk prediction, and locate the high-contribution time period and data type; 2) SHAP value analysis: calculate the SHAP value of static features to quantify the influence of features on risk; 3) Counterfactual generation: simulate the increase of SpO2 by 2% at night, the model recalculates the risk, and outputs the risk reduction amplitude, which provides basis for subsequent intervention.

[0012] Preferably, the data processing flow of the individual feature matching module comprises: Step one, archive data collection and structuring: 1) Standardization of basic information: convert height and weight to BMI, classify according to WHO standard, and convert birth date to age; 2) Structuring of chronic disease history: synchronize ICD-10 coding from EMR system, and associate disease diagnosis time and treatment plan; 3) Habit tagging: Convert the user's reported vegetarianism and intolerance of intense exercise, as well as information from other users, into tags to facilitate subsequent rule matching; 4) Index design and matching logic for the archive: Establish a joint index of chronic disease history and lifestyle habits to support quick queries for intervention plan templates for similar users; 5) Real-time update mechanism: When the user's medication changes or lifestyle habits are adjusted, the archive is updated in real time, with the update time and source marked.

[0013] Preferably, the data processing flow of the dynamic personalized intervention module includes: Step 1: LLM suggestion generation: 1) Prompt construction: integrate key information; 2) LLM output: generate suggestions based on open-source LLaMA-3, including risk warnings, emergency actions, short-term tasks, and long-term suggestions; Step 2: Rule engine verification: 1) Hard rule verification: call the clinical rule library to perform safety verification on the LLM output. If the suggestion does not include the emergency prompt of oxygen saturation <88% for 5 minutes, automatically supplement the emergency prompt. If the suggestion increases exercise intensity, but the individual's profile indicates only low-intensity exercise tolerance, modify it to a 10-minute post-meal walk to avoid jogging; 2) Priority sorting: sort the intervention items according to urgency and executability; Step 3: Intervention plan structuring: Convert the verified suggestions into task card format, clearly defining task description, completion time window, verification method, and adherence scoring weight.

[0014] Preferably, the data processing flow of the intervention execution and adherence monitoring module includes: Step 1: Multi-terminal task distribution: 1) Channel adaptation: APP pushes task cards, smart speakers broadcast tasks through voice, and for users without smart devices, send task reminders through SMS, and synchronize family members to assist in supervision; 2) Time adaptation: adjust the distribution time according to the user's lifestyle habits to avoid disturbing rest; Step 2: Execution data collection and verification: 1) Automatic verification: connect sensor data and automatically mark completed tasks; 2) Manual assistance verification: for diet records and other tasks that cannot be automatically verified, caregivers or community nurses review photos through the background and provide feedback on verification results within 24 hours; Step 3: Adherence score calculation: 1) Single task score: calculated according to task weight x completion degree; 2) Overall compliance score: take the average score of all tasks in the last 7 days, if there are overdue tasks, deduct the total score by overdue days x 0.1, the final score range is 0-100 points.

[0015] Preferably, the data processing flow of the intervention effect evaluation and model iteration module comprises: Step one, intervention effect quantitative evaluation: 1) Physiological index change analysis: compare the average of key physiological indicators before and after intervention, and calculate the change rate; 2) Impact evaluation of chronic disease events: statistics of hospitalization rate and emergency rate of intervention group and control group, using propensity score matching to control confounding factors; 3) Risk prediction accuracy verification: compare the risk value predicted by the model after intervention with the actual event occurrence, and calculate the risk reduction rate; Step two, A / B test optimization intervention strategy: 1) Test design: randomly divide users into A group and B group, A group for voice reminder and photo verification, B group for text reminder and data verification, each group has 100 people, test period is 1 month, core indicators are task compliance score and physiological index improvement rate; 2) Result analysis: if the average compliance score of A group is 75 points and that of B group is 62 points; the physiological index improvement rate of A group is 35%, and that of B group is 22%, then the conclusion is that the intervention strategy of voice reminder and photo verification is more effective, and the strategy should be promoted in the future; Step three, risk prediction model iteration: 1) Data update: add the physiological indicators and event records after intervention to the training data set to expand the sample size; 2) Model retraining: use the original training architecture, adjust the hyperparameters, and retrain the model; use the new test set to verify the performance, if the AUC rises, it means that the model iteration is effective, and the old model should be replaced for subsequent prediction.

[0016] Preferably, the data processing flow of the deployment and compliance guarantee module comprises: Step one, hybrid deployment architecture design: 1) Edge deployment: responsible for real-time data preprocessing and emergency alarm, using lightweight hardware to ensure alarm delay <10 seconds, storing high-frequency data in the last 7 days to reduce cloud storage pressure; 2) Cloud deployment: responsible for risk prediction model training and inference, individual archive management and intervention scheme generation, using Kubernetes containerized deployment to support elastic expansion, using object storage for long-term data, partitioned by user ID+year for easy query; 3) Data flow control: edge only uploads clean effective data, sensitive data uses local encryption combined with cloud decryption transmission to avoid leakage during transmission; Step two, privacy control measures: 1) Data encryption: TLS1.3 encryption is used in the transmission layer, AES-256 encryption is used in the storage layer, and user sensitive information is stored using desensitization; 2) User consent management: record user consent for data use, consent scope can be modified at any time, and data use permissions are updated in real time after modification; 3) Data minimization and deletion: only collect data necessary for risk prediction and intervention, delete all data on the cloud and edge within 24 hours when the user applies for data deletion, and issue a deletion certificate; Step three, system operation and monitoring: 1) Key indicator monitoring: use Prometheus+Grafana to monitor system availability, data transmission delay and model inference time, and trigger operation and maintenance alarm if the indicators exceed the threshold; 2) Model drift monitoring: calculate the deviation of model prediction results and actual events every month, and if the deviation is >10%, determine that the model has drifted and trigger model retraining.

[0017] Compared with the prior art, the beneficial effects of the present application are: 1) Long-term data can be collected for risk prediction, and after prediction, LLM can be used for dynamic adaptation and data interpretation; 2) Data application is upgraded from fragmented storage to deep value conversion, improving the accuracy of chronic disease risk prediction; 3) Intervention programs are upgraded from universal guidance to dynamic personalized adaptation, improving the practicality of health management; 4) The technical system is upgraded from a single link to a full-link closed loop, improving the feasibility and scenario adaptability of landing; 5) The business value is significant, helping to improve the efficiency of chronic disease prevention and control and pension services. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of the present application. DETAILED DESCRIPTION

[0019] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the related drawings. The present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0020] As Figure 1 shown, the large model driven smart pension Internet of Things system and chronic disease personalized intervention system comprises: Multi-dimensional physiological data acquisition module: long-term and stable acquisition of core physiological data of the elderly, such as sleep, heart rate, blood pressure and blood glucose, and home behavior data, providing continuous and reliable raw data sources for subsequent analysis; Edge preprocessing and real-time alarm module: preliminary cleaning and anomaly detection of data are completed at the edge, reducing the amount of invalid data upload, and real-time alarm for urgent health risks is provided to ensure the safety of the elderly; Data transmission and time series database storage module: the preprocessed data is uploaded to the cloud through a stable transmission protocol, and is stored in a time series database to ensure data consistency and traceability, and support fast query by user, time and data type; Physiological data fusion analysis module: multi-source time series physiological data are fused and analyzed to predict future short-term and medium-term chronic disease risks, output risk level, warning window and key contributing factors, and solve the problem of data fragmentation and lack of depth application; Individual feature matching module: static information, chronic disease history and living habits of the elderly are integrated to establish an individual archive, which provides individual basis for personalized intervention and avoids disconnection between intervention scheme and individual conditions; Dynamic personalized intervention module: combining the risk prediction results of chronic diseases and individual archives, the LLM generates readable suggestions, which are then verified for clinical safety by a rule engine to output a personalized intervention scheme that matches the risk and individual, and solves the problem of extensive intervention; Intervention execution and compliance monitoring module: intervention tasks are assigned to users or caregivers through multiple terminals, real-time task execution data is collected, and compliance scores are automatically calculated to monitor the landing of interventions; Intervention effect evaluation and model iteration module: the impact of intervention on physiological indicators and chronic disease risks is quantified, intervention strategies are optimized through A / B testing and causal inference, and risk prediction models are iteratively updated to form a closed loop of intervention-evaluation-optimization; Deployment and compliance guarantee module: a hybrid deployment mode of edge and cloud is adopted to balance data processing real-time and privacy security, meet compliance requirements such as GDPR, and ensure long-term stable operation of the system; The data processing process of the multi-dimensional physiological data acquisition module includes: Input: direct acquisition of physical signals by hardware; Output: raw sensor data (including device identifier, user identifier, timestamp, data value, unit, quality label); Step 1: hardware selection and determination of data acquisition range: The hardware architecture of the prior art patent (CN111045409A) is reused. The smart sleep band monitors sleep duration, sleep staging, number of night awakenings, sleep heart rate, sleep respiratory rate, etc. The six-in-one vital sign monitor monitors blood pressure, blood glucose, blood oxygen, uric acid, heart rate, and body temperature, etc. The infrared sensor can monitor the user's activity duration and frequency in different spaces in the home care scene, ensuring that the data covers both physiological indicators and behavioral indicators. Step two, real-time data collection and preliminary labeling: The hardware collects data at a set frequency, such as heart rate every second, blood glucose every 30 minutes, and sleep data recorded dynamically according to the sleep cycle (about 90 minutes per cycle), while automatically labeling data quality. 0 indicates normal data, 1 indicates slight noise, and 2 indicates data missing or device abnormality, providing a basis for subsequent cleaning. Data format example

[0021] (Note: sample_rate unit is "times / second", blood glucose is collected once every 30 minutes, i.e. 1 / (30*60)=0.00083 times / second) Quality control: Regularly (every 24 hours) self-check the device. If the data quality flag is 2 for 3 consecutive times, trigger device failure alarm and remind maintenance personnel to troubleshoot. The data processing flow of the edge preprocessing and real-time alarm module includes: Input: Raw sensor data output by the collection module (JSON format); Output: Preprocessed data (remove noise / repeat data), real-time alarm signal (including alarm type, emergency level, and trigger basis); Step one, data denoising and repeat removal: Use 3-sample moving average filter to process high-frequency data (such as heart rate) to remove transient spike noise such as heart rate transient jump to 150bpm due to device loosening. Use device_id+timestamp+sensor_type combination key to judge repeated data. If the data value difference is <5% under the same combination key, keep the first collected data and remove the subsequent repeated data to reduce data transmission volume. Step two, emergency exception detection and alarm triggering: 1) Low blood oxygen detection: Real-time monitoring of blood oxygen (SpO2) data. If the blood oxygen saturation (SpO2) is <88% and the duration is >5 minutes, trigger an emergency alarm, and record the blood oxygen fluctuation curve (1 data point per 1 minute) in this time period as the alarm basis. 2) Abnormal blood pressure detection: If the systolic pressure is > 180 mmHg or the diastolic pressure is > 110 mmHg, trigger a medium-level alarm, and compare the user's blood pressure average in the past 7 days to calculate the deviation rate (e.g. if the current blood pressure is 20% higher than the average, prioritize the alarm priority); Data format example (after preprocessing)

[0022] Quality control: After the alarm is triggered, the edge will retain the original data segment (such as the acceleration data 10 seconds before and 20 seconds after the fall), which can be used for subsequent manual review of alarm accuracy and to avoid false alarms; The data processing flow of the data transmission and time series database storage module includes: Input: Cleaned data after edge preprocessing, alarm signal log; Output: Structured data stored in time series database (including data table index), data upload log (records transmission status); Step 1: Data transmission protocol selection and reliability guarantee: The LoRa low-power Internet of Things gateway and MQTT transmission protocol used in the previous patent are used to ensure stable data upload in a weak home network environment. The at-least-once transmission strategy is adopted. If the gateway detects that the data has not been confirmed by the cloud, it will be retransmitted every 5 minutes until it is successfully received. At the same time, the transmission time and reception time of each data are recorded, and the transmission delay is calculated. If the delay is > 30 seconds, it is marked as transmission delay data and stored separately; Step 2: Time series database selection and table structure design: ClickHouse is selected as the time series database, and the table structure is designed according to the user-data type-time dimension, and a joint primary key index is established to improve the efficiency of multi-condition queries (e.g. query blood glucose data of user U123 from October 1 to 7, 2025) 1) Core data table (timeseries) fields: user_id (user identifier), device_id (device identifier), sensor_type (data type, such as HR / blood_glucose), ts (data collection timestamp, TIMESTAMPTZ type), value (data value), unit (unit), sample_rate (sampling frequency), quality_flag (data quality), preprocess_flag (preprocessing flag), trans_delay (transmission delay, unit: seconds); 2) Alert log table (alerts) fields: alert_id, user_id, alert_type (e.g. low_SpO2 / fall), alert_level (emergency level: emergency / intermediate / low), trigger_data (original data segment triggering the alert, JSONB type), alert_time, handle_status (processing status: unprocessed / processed); Data format example (database table record) user_id device_id sensor_type ts value unit sample_rate quality_flag preprocess_flag trans_delay U123 B3 blood_glucose 2025-10-01T08:03:00Z 6.2 mmol / L 0.00083 0 none 8 Quality control: integrity check of the uploaded data of the previous day at 3 am every day, statistics of the missing rate of each user and each data type (missing data amount / should upload data amount), if the missing rate >5%, trigger device data retransmission instruction; The data processing flow of the physiological data fusion analysis module includes: Input: cleaned data in the time series database, user static features (age, gender, chronic disease history), text medical record abstract (such as previous hospitalization record); Output: chronic disease risk prediction result (multi-label probability, such as heart failure risk 0.38), explainability analysis report (key contribution features, attention weight visualization); Step one, data preprocessing: 1) Time series data alignment and resampling: resample data of different frequencies according to 1 hour as the time granularity, such as taking the hourly mean of heart rate (1 per second) and the maximum value of blood glucose (1 per 30 minutes) every hour, linear interpolation is used to supplement the missing data, and missing data> 3 hours is marked as missing_segment, and the model is informed by a mask; 2) Feature engineering: extract time series features (such as hourly mean, daily fluctuation standard deviation, and night (22:00-6:00) and daytime mean difference of heart rate), periodic features (such as sleep day rhythm consistency index-sleep time standard deviation<30 minutes for rhythm stability), and text features (convert medical record abstract to 768-dimensional embedding vector using sentence-transformer); Step two, time series Transformer model training and inference: 1) Model architecture: linear projection and relative time position encoding for time series data in the input layer, TransformerEncoder (6 layers, 8 attention heads) to capture long-term dependence of data (such as the association between sleep abnormalities in a week and blood pressure rise 1 month later), cross-attention layer to fuse static features and text embedding, and output layer to output multi-label chronic disease risk probability (such as hypertension, diabetes, and heart failure) through MLP; 2) Training strategy: User stratification K-fold cross-validation (same user data does not cross fold, avoid data leakage), use FocalLoss weighting for rare events (such as heart failure), balance sample imbalance, data augmentation through time mask (randomly mask 10% of the time steps), noise injection (add ±5% Gaussian noise), improve model robustness; 3) Inference process: Input user's recent 30-day time series data, model outputs the risk probability of chronic disease in the next 30 days or 90 days, if the risk>0.3 is high risk, 0.1-0.3 is medium risk, <0.1 is low risk; Step three, explainability analysis: 1) Attention weight visualization: Output the contribution weight of each time step (such as the last 7 days) and each data type (such as sleep, blood pressure) to the risk prediction, locate the high contribution time period and data type, such as the night oxygen data from October 5th to 7th is the main contributing factor of heart failure risk); 2) SHAP value analysis: Calculate the SHAP value of static features (such as age, BMI), quantify the impact of features on risk (such as BMI>28, SHAP value=0.08, indicating that obesity will increase the risk of diabetes); 3) Counterfactual generation: Simulate if the night SpO2 increases by 2%, the model recalculates the risk, and outputs the risk reduction (such as from 0.38 to 0.31), which provides the basis for subsequent intervention; Data format example (risk prediction result)

[0023] Quality control: Verify the model performance with 100 new user data every month, if AUC<0.75 (target AUC≥0.78), trigger model retraining, update model parameters; The data processing process of the individual feature matching module includes: Input: Basic information filled in by users during registration (age, gender, height, weight), chronic disease history / drug records synchronized from hospital EMR system, and self-reported lifestyle habits (such as vegetarian, exercise contraindications); Output: Structured individual profile (including index, supporting matching by chronic disease history + lifestyle habits); Step one, profile data collection and structuring: 1) Basic information standardization: Convert height (cm) and weight (kg) to BMI (BM = weight / height²), classify according to WHO standards (<18.5: underweight, 18.5-23.9: normal, 24-27.9: overweight, ≥28: obese), and convert date of birth (DOB) to age (accurate to year); 2) Chronic disease history structuring: Sync ICD-10 codes from EMR system (e.g. I10: Essential hypertension, E11: Type 2 diabetes), correlate with disease diagnosis time, treatment plan; 3) Lifestyle habit tagging: Convert user-reported vegetarian and intolerance of intense exercise, and other users' information into tags (e.g. diet:vegetarian, exercise:low_intensity_only) for subsequent rule matching; 4) Index design and matching logic for archive library: Establish a joint index of chronic disease history and lifestyle habits (e.g. I10+vegetarian, E11+low_intensity_only) to support quick query of intervention plan templates for similar users; 5) Real-time update mechanism: When the user's medication changes (e.g. adds a hypoglycemic drug) or lifestyle habits adjust (e.g. from vegetarian to omnivore), the archive library is updated in real time, with the update time and update source marked (e.g. 2025-10-01, hospital EMR synchronization); Data format example (individual archive)

[0024] Quality control: Check the consistency of archive data and hospital EMR system every quarter, if "medication record inconsistency" is found (e.g. archive shows ACEI, EMR shows ARB has been replaced), trigger manual review and correct archive data; The data processing flow of the dynamic personalized intervention module includes: Input: Risk prediction results of physiological data fusion analysis module (including interpretability), user information of individual archive library, clinical rule library (e.g. SpO2<88% need emergency medical treatment); Output: Structured personalized intervention plan (including priority, time window, execution requirements), rule verification log (record whether the safety rule is triggered); Step 1, LLM suggestion generation (Prompt engineering + content generation): 1) Prompt construction: Integrate key information, for example: User U123, 72-year-old male, BMI 26.9 (overweight), medical history: hypertension (diagnosed in 2018, treated with ACEI), heart failure (diagnosed in 2020, treated with diuretics), lifestyle habits: vegetarian, only tolerates low-intensity exercise, heart failure risk in the next 30 days is 0.38 (high risk), key contributing factors: average nighttime blood oxygen is low in the last 7 days, weight increased by 2.8 kg in the last 2 weeks, counterfactual analysis shows that a 2% increase in nighttime blood oxygen can reduce the risk of heart failure to 0.31, please generate personalized intervention recommendations, including urgent priority actions, short-term tasks (within 24 hours), long-term behavior adjustments (within 1 week), language is simple and suitable for the understanding ability of the elderly; 2) LLM output: generate recommendations based on open-source LLaMA-3 (70B parameters), including risk warnings (high risk of heart failure, pay attention to nighttime blood oxygen and weight), urgent actions (contact family doctor to review diuretic dosage), short-term tasks (weigh in the morning and evening and upload, record urine volume and dyspnea score), and long-term recommendations (limit sodium intake <2g / day, elevate bed head 10-15 degrees); Step two, rule engine verification: 1) Hard rule verification: call the clinical rule library to perform safety verification on the LLM output, if the suggestion does not include blood oxygen saturation <88% for 5 minutes, automatically supplement the emergency prompt, if the suggestion increases exercise intensity but the individual profile indicates only low-intensity exercise tolerance, modify it to walk for 10 minutes after meals instead of jogging; 2) Priority sorting: sort the intervention items according to urgency (urgent / short-term / long-term) and executability (whether it requires assistance from others), such as "contact doctor" set as "urgent (within 1 hour)", "weigh and upload" set as "short-term (within 24 hours)", "sodium control" set as "long-term (within 1 week)"; Step three, intervention scheme structuring: Convert the verified suggestions into task card format, clearly describe the task description, completion time window, verification method, and adherence scoring weight (e.g. "weighing completion weight 0.6, photo uploading weight 0.3, time consistency weight 0.1"); Data format example (personalized intervention scheme)

[0025]

[0026] Quality control: for "high-risk" users, the intervention scheme must be reviewed by a doctor (within 24 hours), and only after the review is passed can it be issued, to avoid unsafe suggestions generated by LLM; The data processing flow of the intervention execution and adherence monitoring module includes: Input: Task cards generated by the dynamic personalized intervention module, execution data uploaded by the user through the APP / smart speaker (such as weight photos, step records); Output: Task execution status (completed / unfinished / overdue), user adherence score (single task score + overall score), execution data log; Step 1, multi-terminal task distribution: 1) Channel adaptation: APP pushes task cards (including text and picture instructions), smart speaker broadcasts tasks through voice (such as "at 9 am, take a photo and upload it"), for users without smart devices, send task reminders through SMS, and inform family members to assist in supervision; 2) Time adaptation: adjust the distribution time according to the user's living habits, such as distributing weight tasks to the user's daily wake-up (7:00) and bedtime (21:30) periods to avoid disturbing rest; Step 2, execution data collection and verification: 1) Automatic verification: connect sensor data, automatically mark completed tasks (such as 1500 steps), read electronic scale upload data for weight tasks, compare photo weight value and sensor data, difference <0.5kg, then verify; 2) Artificial assistance verification: for diet records and other tasks that cannot be automatically verified, the caregiver or community nurse verifies the photo through the background, and feedbacks the verification result within 24 hours; Step 3, adherence score calculation: 1) Single task score: calculate according to task weight x completion, such as "weight task" complete weight (0.6) + upload photo (0.3), time consistent (0.1), then score = 1.0; only complete weight, not upload photo, score = 0.6 / 0.9≈0.67 (denominator is the total weight set); 2) Overall adherence score: take the average of all task scores of the user in the last 7 days, if there are overdue tasks, deduct the total score by overdue days x 0.1 (such as overdue for 2 days, deduct 0.2), the final score range is 0-100 points (60 points or more is "good adherence"); Data format example (adherence score result)

[0027] Quality control: if the user's overall adherence score is less than 40 points for 3 consecutive days, trigger "caregiver follow-up", understand the reason for not executing (such as task difficulty is too high, device operation is not skilled), adjust the intervention scheme; The data processing process of the intervention effect evaluation and model iteration module includes: Input: User adherence data, post-intervention physiological indicator changes (e.g., blood pressure, blood glucose), chronic disease event records (e.g., hospitalization, emergency), historical prediction results of risk prediction models; Output: Intervention effect evaluation report (e.g., hospitalization rate reduction), A / B test conclusion (e.g., "push voice reminders have 20% higher compliance than text reminders"), model iteration update log; Step 1: Quantitative evaluation of intervention effect: 1) Physiological indicator change analysis: Compare the mean values of key physiological indicators before and after intervention (e.g., 1 month before intervention vs. 1 month after intervention) and calculate the change rate, such as "after intervention, the average SpO2 at night increased from 86% to 89%, with a change rate of +3.5%; body weight decreased from 78 kg to 76.5 kg, with a change rate of -1.9%"; 2) Chronic disease event impact assessment: Calculate the hospitalization rate and emergency rate of the intervention group (receiving personalized intervention) and the control group (only regular monitoring), and use propensity score matching (matching age, gender, and underlying diseases) to control confounding factors. If the 30-day hospitalization rate of the intervention group is 5% and that of the control group is 12%, then the intervention reduces the hospitalization rate by 7 percentage points; 3) Risk prediction accuracy verification: Compare the risk values predicted by the model after intervention with the actual event occurrence, calculate the risk reduction rate (e.g., heart failure risk decreased from 0.38 to 0.25, a decrease of 0.13), and evaluate the improvement effect of intervention on risk; Step 2: A / B test to optimize intervention strategy: 1) Test design: Randomly divide users into A and B groups, A group uses voice reminders and photo verification, B group uses text reminders and data verification, each group has 100 people, test period is 1 month, core indicators are task compliance score and physiological indicator improvement rate; 2) Result analysis: If the average compliance score of A group is 75 and that of B group is 62; the physiological indicator improvement rate of A group is 35% and that of B group is 22%, then the conclusion is that the intervention strategy of voice reminders and photo verification is more effective, and the strategy should be promoted; Step 3: Risk prediction model iteration: 1) Data update: Add post-intervention physiological indicators and event records to the training data set, expand the sample size (e.g., add 500 heart failure event data); 2) Model retraining: Use the original training architecture, adjust hyperparameters (e.g., increase the number of attention heads to 10), and retrain the model; use the new test set (1000 users) to verify the performance, if AUC increases from 0.78 to 0.81, it means that the model iteration is effective, replace the old model for subsequent prediction; Data format example (intervention effect evaluation report)

[0028] Quality control: the evaluation data need to pass the "data integrity check (no missing key indicators), statistical method correctness audit (such as balance test of covariates for propensity score matching)", to ensure the reliability of the evaluation results; The data processing flow of the deployment and compliance guarantee module includes: Input: deployment requirements of each module (such as real-time alarm for edge, storage of large amount of data for cloud), privacy regulation requirements (such as data encryption, user consent management); Output: deployment architecture scheme, privacy control measures, system operation and monitoring data (such as availability, delay); Step one, hybrid deployment architecture design: 1) Edge deployment: responsible for real-time data preprocessing (such as denoising, duplicate removal), emergency alarm (such as fall, low blood oxygen), using lightweight hardware (such as edge gateway) to ensure alarm delay <10 seconds; store the high-frequency data of the last 7 days (such as heart rate, blood oxygen) to reduce the storage pressure of the cloud; 2) Cloud deployment: responsible for risk prediction model training and inference, individual profile management, intervention scheme generation, using Kubernetes containerized deployment to support elastic scaling (such as automatically increasing computing nodes when the number of users increases from 1000 to 10000); using object storage (such as S3) to store long-term data (such as physiological indicators and intervention records for more than 1 year), partitioned by "user ID + year" for easy query; 3) Data flow control: edge only uploads cleaned valid data, sensitive data (such as medical record summary) is transmitted using "local encryption + cloud decryption" to avoid leakage during transmission; Step two, privacy control measures: 1) Data encryption: TLS1.3 encryption at the transmission layer, AES-256 encryption at the storage layer (both time series database and object storage are enabled); user sensitive information (such as ID number) uses "desensitization storage" (only the last 4 digits are kept, the first 14 digits are replaced with *); 2) User consent management: record user's consent for data use (such as "agree to use physiological data for risk prediction" "do not agree to use for third-party research"), the consent scope can be modified at any time, and the data usage permission is updated in real time after modification; 3) Data minimization and deletion: only collect "necessary" data for risk prediction and intervention (such as not collect user social information); when the user applies for data deletion, delete all data on the cloud and edge within 24 hours, and issue a deletion certificate; Step three, system operation and monitoring: 1) Key indicators monitoring: use Prometheus + Grafana to monitor system availability (target 99.9%), data transmission delay (target <30 seconds), model inference time (target <1 second). If the indicators exceed the threshold, trigger operation and maintenance alarm; 2) Model drift monitoring: calculate the deviation between model prediction results and actual events every month (e.g. the actual event occurrence rate of high-risk users decreases from 30% to 15%). If the deviation is >10%, it is determined that "model drift" has occurred, triggering model retraining; Data format example (deployment and compliance monitoring report)

[0029] Quality control: conduct a "comprehensive compliance audit" every quarter, invite a third-party agency to check the compliance of data encryption, user consent management, etc., issue an audit report, and rectify the problems found.

[0030] Embodiment: Example A: High risk - heart failure (predicted 30-day risk 38%) • System highlights: continuous low SpO2 at night, weight increase of 2.8 kg in 2 weeks, 50% decrease in activity • Suggestions (urgent priority): 1. Immediately - contact family doctor / caregiver today: arrange 24-hour phone follow-up and review medication (whether diuretic dose is appropriate).

[0031] 2. Executable tasks within 24 hours: weigh and upload in the morning and evening; record urine volume and shortness of breath (1-5 score).

[0032] 3. Behavior suggestions: limit sodium intake (<2g per day), elevate bed head 10-15 degrees to relieve shortness of breath at night.

[0033] 4. If: shortness of breath, cyanosis, or SpO2 <90% for >5 minutes, please seek medical attention or call emergency services immediately.

[0034] 5. The system will arrange a remote nurse visit within 48 hours and track weight / SpO2 curve.

[0035] Example B: Medium risk - diabetic complication risk (predicted 90-day risk 12%) • System highlights: large postprandial blood glucose fluctuations, low activity at night, chronic foot sensation decline (questionnaire) • Suggestions: 1. Adjust diet: reduce refined carbohydrates, increase 10-minute pre-meal walks; upload three-day diet photos for nutritionist evaluation.

[0036] 2. Medication Reminders: Confirm insulin / oral medication on time, open medication for taking photos and check-in.

[0037] 3. Schedule: Blood glucose / hemoglobin A1c check-up in two weeks; remind foot care and instruct daily self-check (and upload photos for AI screening of wounds).

[0038] Example C: Low risk - fall risk (predicted 3% in 30 days) • System highlights: occasional night-time getting up, slightly slow pace, a few slippery hazards in the home • Recommendations: 1. Home safety: Install night lights and non-slip mats in the bathroom and beside the bed.

[0039] 2. Exercise prescription: Perform balance training 2 times a week (10-15 minutes), with simple videos with voice prompts in the App.

[0040] 3. Regular review: If the pace slows down by >20% or it is difficult to turn over at night, upgrade to medium risk and automatically trigger a visit.

[0041] The main functions implemented by the present application are: Module 1: Physiological data fusion analysis module (solves "data application limitations"): The self-attention mechanism of time series Transformer is good at capturing long-distance dependencies in sequence data, and can effectively align and fuse fragmented physiological data of different times and different sources (such as time series indicators such as blood pressure, blood glucose, heart rate, etc.). By arranging fragmented data into sequence input according to the time axis, the model can automatically learn the relevance between data (such as the implicit correlation between fluctuations in certain indicators and disease risk), solving the problem of high requirements for data integrity in traditional methods. Through model iterative learning, it can output "hypertension / diabetes risk level (such as low / medium / high risk)" "risk warning time window (such as risk within the next 3 months)", realizing the transformation from "data" to "risk signal"; Module 2: Individual feature matching module (supporting personalized intervention): Add "elderly individual archives", integrate information such as their chronic disease history (such as whether they have diabetes), eating habits (such as whether they are vegetarian), physical contraindications (such as whether they are intolerant to exercise), etc. to provide "individual basis" for intervention plan customization; Module 3: Dynamic personalized intervention module (solve "intervention extensive limitations"): based on the chronic disease risk results output by module 1, through LLM suggestion module & rule engine, risk prediction is converted into personalized suggestions, which can be completed through the design of Prompt, access to LLM, and the addition of clinical rules, and the generation of personalized suggestion modules, and combined with individual archive information, to automatically generate adaptive solutions (such as: for "diabetes high risk + vegetarian" old people, recommend low GI vegetarian menu + 1 hour after meal light activity reminder; for "high blood pressure high risk + poor sleep" old people, push the pre-sleep relaxation audio + morning blood pressure monitoring reminder), and through APP, smart speaker and other terminals Real-time access.

[0042] The above only describes the preferred embodiments of the present application, and it should be noted that for ordinary skilled persons in the art, without departing from the technical principles of the present application, a number of improvements and modifications can be made, and these improvements and modifications should be considered as the protection scope of the present application.

Claims

1. A large-scale model-driven smart elderly care IoT system and a personalized intervention system for chronic diseases, characterized in that: include: Multi-dimensional physiological data acquisition module: Long-term and stable collection of core physiological data such as sleep, heart rate, blood pressure and blood sugar, as well as home behavior data of the elderly, providing a continuous and reliable source of raw data for subsequent analysis; Edge preprocessing and real-time alarm module: Performs preliminary data cleaning and anomaly detection at the edge to reduce the amount of invalid data uploaded, while providing real-time alarms for urgent health risks to ensure the safety of the elderly; Data transmission and time-series database storage module: Uploads pre-processed data to the cloud through a stable transmission protocol, and uses a time-series database for storage to ensure data consistency and traceability, and supports fast querying by user, time, and data type; Physiological data fusion and analysis module: performs fusion analysis on multi-source time-series physiological data, predicts the risk of chronic diseases in the short and medium term, and outputs the risk level, early warning window and key contributing factors, solving the problems of data fragmentation and lack of in-depth application; Individual characteristic matching module: Integrates data such as static information, chronic disease history, and lifestyle habits of the elderly to establish an individual profile database, providing an individual basis for personalized intervention and avoiding the disconnect between intervention plans and individual conditions; Dynamic Personalized Intervention Module: Combining chronic disease risk prediction results with individual profiles, it generates readable suggestions through LLM, verifies clinical safety through a rule engine, and outputs a personalized intervention plan that matches risk and individual profiles, thus solving the problem of inefficient intervention. Intervention Implementation and Compliance Monitoring Module: Intervention tasks are distributed to users or caregivers through multiple terminals, task implementation data is collected in real time, compliance scores are automatically calculated, and the implementation status of the intervention is monitored; Intervention effect evaluation and model iteration module: Quantify the impact of intervention on physiological indicators and chronic disease risk, optimize intervention strategies through A / B testing and causal inference, and iteratively update the risk prediction model to form a closed loop of intervention-evaluation-optimization; Deployment and Compliance Assurance Module: Adopts an edge + cloud hybrid deployment model to balance the real-time performance of data processing with privacy and security, while meeting compliance requirements such as GDPR and ensuring the long-term stable operation of the system.

2. The large-scale model-driven smart elderly care IoT system and personalized intervention system for chronic diseases as described in claim 1, characterized in that, The data processing flow of the multi-dimensional physiological data acquisition module includes: Step 1: Hardware Selection and Data Acquisition Scope Determination The smart sleep belt monitors sleep duration, sleep stages, number of nighttime awakenings, sleep heart rate, and sleep breathing rate. It also monitors six vital signs in one device: blood pressure, blood sugar, blood oxygen, uric acid, heart rate, and body temperature. An infrared sensor can monitor the duration and frequency of user activities in different spaces within a home-based elderly care setting, ensuring that the data covers both physiological and behavioral indicators. Step 2: Real-time data acquisition and preliminary labeling: The hardware collects data at a set frequency and automatically marks the data quality: 0 indicates normal data, 1 indicates slight noise, and 2 indicates missing data or equipment malfunction, providing a basis for subsequent cleaning.

3. The large-scale model-driven smart elderly care IoT system and personalized intervention system for chronic diseases as described in claim 1, characterized in that, The data processing flow of the edge preprocessing and real-time alarm module includes: Step 1: Data Denoising and Duplicate Removal High-frequency data is processed using a 3-sample moving average filter to remove instantaneous spike noise. Duplicate data is identified by the combination key device_id+timestamp+sensor_type. If the difference in data values ​​under the same combination key is less than 5%, the first data collection is retained and subsequent duplicate data is removed to reduce the amount of data transmission. Step 2: Emergency Anomaly Detection and Alarm Triggering: 1) Low blood oxygen detection: Real-time monitoring of blood oxygen data. If blood oxygen saturation is <88% and lasts for >5 minutes, an emergency alarm is triggered. At the same time, the blood oxygen fluctuation curve during this period is recorded as the basis for alarm. 2) Blood pressure abnormality detection: If the systolic blood pressure is >180 mmHg or the diastolic blood pressure is >110 mmHg, a medium-level alarm is triggered. At the same time, the deviation rate is calculated by comparing the user's average blood pressure over the past 7 days.

4. The large-scale model-driven smart elderly care IoT system and personalized intervention system for chronic diseases as described in claim 1, characterized in that, The data processing flow of the data transmission and time-series database storage module includes: Step 1: Selection of Data Transmission Protocol and Reliability Assurance It uses the prior patented LoRa low-power IoT gateway + MQTT transmission protocol to ensure stable data upload in home weak network environment. It adopts at-least-once transmission strategy. If the gateway detects that the data has not been confirmed by the cloud, it will retransmit after 5 minutes until it is successfully received. At the same time, it records the transmission time and reception time of each data and calculates the transmission delay. If the delay is >30 seconds, it is marked as transmission delay data and stored separately. Step Two: Time Series Database Selection and Table Structure Design ClickHouse was chosen as the time-series database. The table structure was designed according to the user-data type-time dimension, and a composite primary key index was established to improve the efficiency of multi-condition queries.

5. The large-scale model-driven smart elderly care IoT system and personalized intervention system for chronic diseases as described in claim 1, characterized in that, The data processing flow of the physiological data fusion and analysis module includes: Step 1: Data Preprocessing 1) Time series data alignment and resampling: Data at different frequencies are resampled with a time granularity of 1 hour. Missing data is supplemented by linear interpolation. Missing data > 3 hours is marked as missing_segment and the model is informed by a mask. 2) Feature engineering: Extracting temporal features, periodic features, and text features; Step 2: Temporal Transformer Model Training and Inference: 1) Model architecture: The input layer performs linear projection and relative time position encoding on the time series data, the TransformerEncoder captures long-term data dependencies, the cross-attention layer fuses static features and text embeddings, and the output layer outputs multi-label chronic disease risk probabilities through MLP. 2) Training strategy: User-stratified K-fold cross-validation is adopted, FocalLoss weighting is used for rare events to balance sample imbalance, and data augmentation is performed through time masking and noise injection to improve model robustness. 3) Inference process: Input the user's time series data for the most recent 30 days, and the model outputs the probability of chronic disease risk for the next 30 or 90 days; Step 3: Interpretability Analysis: 1) Attention weight visualization: Output the contribution weight of each time step and each data type to risk prediction, and locate the time period and data type with high contribution; 2) SHAP value analysis: Calculate the SHAP value of static features to quantify the impact of features on risk; 3) Counterfactual generation: If SpO2 increases by 2% at night, the model recalculates the risk and outputs the magnitude of the risk reduction, providing a basis for subsequent intervention.

6. The large-scale model-driven smart elderly care IoT system and personalized intervention system for chronic diseases as described in claim 1, characterized in that, The data processing flow of the individual feature matching module includes: Step 1: Data Collection and Structure of Archives 1) Standardization of basic information: convert height and weight to BMI, classify according to WHO standards, and convert date of birth to age; 2) Structured chronic disease history: Synchronize ICD-10 codes from the EMR system and link them to the time of disease diagnosis and treatment plan; 3) Lifestyle Habit Tagging: Convert user-reported vegetarianism and intolerance to strenuous exercise, as well as other user information, into tags to facilitate subsequent rule matching; 4) Archive index design and matching logic: Establish a joint index of chronic disease history and lifestyle habits to support quick query of intervention plan templates for similar users; 5) Real-time update mechanism: When a user's medication changes or lifestyle habits are adjusted, the database is updated in real time, and the update time and source are marked.

7. The large-scale model-driven smart elderly care IoT system and personalized intervention system for chronic diseases as described in claim 1, characterized in that, The data processing flow of the dynamic personalized intervention module includes: Step 1: LLM Recommendation Generation 1) Prompt building: Integrating key information; 2) LLM Output: Based on open-source LLaMA-3, recommendations are generated, including risk warnings, urgent actions, short-term tasks, and long-term recommendations; Step 2: Rule Engine Verification 1) Hard rule verification: Call the clinical rule base to perform safety verification on LLM output. If the suggestion does not include the need to seek medical attention if blood oxygen saturation is <88% for 5 minutes, the emergency prompt will be automatically added. If the suggestion is to increase exercise intensity, but the individual profile indicates that the individual can only tolerate low-intensity exercise, the suggestion will be corrected to a 10-minute walk after meals, avoiding brisk walking. 2) Prioritization: Rank intervention items according to urgency and feasibility; Step 3: Structuring the intervention plan: The validated recommendations are converted into task card format, clearly defining the task description, completion time window, verification method, and compliance score weight.

8. The large-scale model-driven smart elderly care IoT system and personalized intervention system for chronic diseases as described in claim 1, characterized in that, The data processing flow of the intervention implementation and compliance monitoring module includes: Step 1: Multi-terminal task distribution: 1) Channel adaptation: The APP pushes task cards, and the smart speaker announces the task via voice. For users without smart devices, task reminders are sent via SMS, and family members are simultaneously notified to assist in supervision. 2) Time adaptation: Adjust the delivery time according to users' lifestyles to avoid disturbing their rest; Step 2: Perform data collection and verification. 1) Automatic verification: By connecting to sensor data, the marking task is automatically marked as completed; 2) Manual verification: For dietary records and other tasks that cannot be automatically verified, caregivers or community nurses review the photos through the backend and provide verification results within 24 hours. Step 3: Compliance score calculation: 1) Single task score: calculated by task weight × completion rate; 2) Overall compliance score: The average score of all tasks completed by the user in the last 7 days is taken. If there are any overdue tasks, the total score will be reduced by 0.1 times the number of overdue days. The final score range is 0-100 points.

9. The large-scale model-driven smart elderly care IoT system and personalized intervention system for chronic diseases as described in claim 1, characterized in that, The data processing flow of the intervention effect evaluation and model iteration module includes: Step 1: Quantitative evaluation of intervention effectiveness: 1) Analysis of changes in physiological indicators: Compare the mean values ​​of key physiological indicators before and after the intervention, and calculate the rate of change; 2) Impact assessment of chronic disease events: Hospitalization rate and emergency room rate were compared between the intervention group and the control group, and propensity score matching was used to control for confounding factors; 3) Verification of the accuracy of risk prediction: Compare the risk values ​​re-predicted by the model after intervention with the actual occurrence of events, and calculate the extent of risk reduction; Step 2: A / B testing to optimize intervention strategies: 1) Test Design: Users will be randomly divided into Group A and Group B. Group A will receive voice prompts and photo verification, while Group B will receive text prompts and data verification. Each group will have 100 users. The test period will be one month. The core indicators will be task compliance score and physiological indicator improvement rate. 2) Results analysis: If the average compliance score of group A is 75 points and that of group B is 62 points; and the improvement rate of physiological indicators is 35% in group A and 22% in group B, then the conclusion is that the intervention strategy of voice reminder and photo verification is more effective, and this strategy should be promoted in the future. Step 3: Iteration of the risk prediction model: 1) Data update: Add post-intervention physiological indicators and event records to the training dataset to expand the sample size; 2) Model retraining: Using the original training architecture, adjust the hyperparameters and retrain the model; use a new test set to verify the performance. If the AUC increases, it indicates that the model iteration is effective, and the old model can be replaced for subsequent predictions.

10. The large-scale model-driven smart elderly care IoT system and personalized intervention system for chronic diseases as described in claim 1, characterized in that, The data processing flow of the deployment and compliance assurance module includes: Step 1: Hybrid Deployment Architecture Design 1) Edge deployment: Responsible for real-time data preprocessing and emergency alarms, using lightweight hardware to ensure alarm latency <10 seconds, storing high-frequency data from the last 7 days to reduce cloud storage pressure; 2) Cloud Deployment: Responsible for risk prediction model training and inference, individual profile management and intervention plan generation. It adopts Kubernetes containerized deployment, supports elastic scaling, uses object storage for long-term data, and partitions by user ID + year for easy querying; 3) Data flow control: Only cleaned and valid data is uploaded at the edge. Sensitive data is encrypted locally and decrypted in the cloud before transmission to prevent leakage during transmission. Step Two: Privacy Control Measures 1) Data encryption: The transport layer uses TLS 1.3 encryption, the storage layer uses AES-256 encryption, and sensitive user information is stored in anonymized form; 2) User consent management: Records user consent to the use of data. The scope of consent can be modified at any time, and the data usage permissions are updated in real time after modification. 3) Data minimization and deletability: Only data necessary for risk prediction and intervention is collected. When a user requests data deletion, all data on the cloud and edge devices is deleted within 24 hours, and a deletion certificate is issued. Step 3: System Operation and Monitoring 1) Monitoring key metrics: Use Prometheus + Grafana to monitor system availability, data transmission latency, and model inference time. If the metrics exceed the threshold, trigger an operation and maintenance alarm. 2) Model drift monitoring: The deviation between the model prediction results and the actual events is calculated monthly. If the deviation is >10%, it is judged as model drift and model retraining is triggered.

Images