AIDS patient remote follow-up visit monitoring method and system driven by Internet of Things

Abstract

The invention relates to the technical field of computers, discloses an internet-of-things-driven AIDS patient remote follow-up monitoring method and system, and aims to solve the problems that in the prior art, multi-source data fusion is insufficient, risk prediction precision is low, a model lacks self-adaptive ability, and intervention is disjointed from clinic. The method comprises the following steps: collecting physiological, medication and environmental data in real time through an Internet of Things terminal; cleaning, alignment and feature extraction are carried out through an edge gateway; the method comprises the following steps: fusing an electronic medical record at a cloud end, performing dynamic modeling by utilizing a hybrid model of a time sequence convolutional network and a long-short-term memory network, and generating an individualized health risk score through a cross-modal attention mechanism; and matching the three-level intervention rule base based on the score, and dynamically pushing follow-up or medical instructions. According to the technical scheme, high-precision risk early warning, minute-level continuous monitoring, low-delay response and personalized closed-loop intervention can be achieved, and the follow-up compliance rate and the clinical intervention efficiency are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to an Internet of Things-driven method and system for remote follow-up monitoring of HIV / AIDS patients. Background Technology

[0002] With the deep integration of IoT and AI technologies, remote health management for patients with chronic infectious diseases has become an important development direction in the field of smart healthcare. As a chronic infectious disease that requires long-term standardized follow-up and medication management, AIDS patient compliance monitoring, physiological indicator tracking, and risk warning are of great significance to public health prevention and control. IoT technology, through the collaboration of wearable devices, environmental sensors, and mobile terminals, provides a basic support for achieving high-frequency, low-invasive, and continuous patient status perception.

[0003] Artificial intelligence, machine learning, and deep neural network technologies are widely used in the analysis and decision support of medical and health data. They aim to automatically identify abnormal patterns, predict disease development trends, and generate personalized intervention strategies from multi-source heterogeneous follow-up data. The core of these technologies lies in building an intelligent monitoring model that can dynamically adapt to individual differences and disease evolution.

[0004] Current technologies for remote follow-up monitoring of HIV patients still face multiple challenges: First, data collection relies on a single terminal or intermittent manual reporting, lacking the ability to integrate multimodal physiological, behavioral, and environmental data in real time, making it difficult to comprehensively depict the patient's true condition; second, existing analysis models mostly employ static rules or shallow algorithms, failing to effectively uncover complex nonlinear correlations in long-term follow-up sequences, resulting in insufficient sensitivity and specificity in risk prediction; third, the system lacks an adaptive learning mechanism, making it difficult to continuously optimize model parameters based on new data, thus limiting the model's generalization ability; finally, the process of generating early warning and intervention strategies is disconnected from clinical guidelines, failing to achieve a closed-loop linkage from data perception to clinical action. These problems severely restrict the practicality and clinical acceptability of remote follow-up systems, urgently requiring a smart monitoring method and system that deeply integrates IoT sensing and deep learning reasoning capabilities to solve these problems. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an IoT-driven remote follow-up monitoring method and system for HIV / AIDS patients, which can effectively solve the problems mentioned in the background technology. Currently, HIV / AIDS patient management faces structural challenges such as low follow-up compliance, delayed data collection, strong subjectivity in disease assessment, and delayed intervention response. Traditional remote monitoring methods lack the ability to dynamically integrate multi-source heterogeneous health data, making it difficult to achieve individualized risk warning and precise intervention. This invention constructs a closed-loop monitoring system based on the collaboration of an IoT sensing layer, edge computing layer, and cloud-based intelligent decision-making layer. It integrates multimodal data such as physiological indicators from wearable devices, home environment parameters, medication behavior records, and electronic medical records. Using deep neural networks, it performs temporal modeling and cross-modal correlation analysis of the patient's health status, dynamically generating individualized follow-up strategies and risk intervention instructions, thereby achieving continuous, objective, and intelligent monitoring of the health status of HIV / AIDS patients.

[0006] To achieve the above objectives, the present invention provides the following technical solution: On one hand, an IoT-driven remote follow-up monitoring system for HIV / AIDS patients, comprising the following components: an IoT sensing terminal for real-time collection of patients' physiological parameters, medication behavior, and home environment data; an edge computing gateway for local preprocessing, feature extraction, and preliminary anomaly detection of the raw data uploaded by the IoT sensing terminal; a cloud-based intelligent decision-making platform for receiving structured feature data uploaded by the edge computing gateway, performing multimodal health status fusion modeling, dynamic risk level assessment, and follow-up intervention strategy generation based on a deep learning model; a remote interactive terminal for pushing risk warning information, follow-up tasks, and health guidance content to patients or medical staff, and receiving patient feedback data; and a data security and privacy protection module for implementing end-to-end encryption, access control, and anonymization processing of health data transmitted and stored throughout the entire chain. On the other hand, an IoT-driven remote follow-up monitoring method for HIV / AIDS patients includes the following steps: Step S110, real-time collection of multi-source health data from HIV / AIDS patients via IoT sensing terminals, including heart rate, body temperature, blood oxygen saturation, sleep quality, medication timestamps, medication dosage confirmation signals, indoor temperature and humidity, and air quality parameters; Step S120, data cleaning, missing value imputation, temporal alignment, and sliding window feature extraction of the multi-source health data through an edge computing gateway to generate a structured temporal feature vector; Step S130, uploading the structured temporal feature vector to a cloud-based intelligent decision-making platform, combining it with CD4+ T lymphocyte counts and viral load data from the patient's electronic medical record. The system collects information on the patient's physical condition, comorbidities, and previous follow-up records to construct a multimodal health status representation. Step S140: Based on the multimodal health status representation, a hybrid architecture of temporal convolutional networks and long short-term memory networks is used to dynamically model the patient's health status, extracting the temporal dependency between physiological indicators and medication behavior. Step S150: A cross-modal attention mechanism is introduced to adaptively fuse physiological temporal features, medication behavior features, and clinical indicator features with weights to generate an individualized health risk score. Step S160: Based on the individualized health risk score, a preset risk intervention rule base is matched to dynamically generate follow-up frequency adjustment instructions, medication reminder reinforcement strategies, or emergency medical intervention suggestions, which are then pushed out via a remote interactive terminal. Preferably, the IoT sensing terminal includes a wearable physiological monitoring device, a smart pillbox, and an environmental sensor node; the wearable physiological monitoring device uses PPG and ECG dual-mode sensing technology to continuously monitor heart rate variability and blood oxygen saturation at a sampling rate of 1Hz; the smart pillbox has a built-in pressure sensor and near-field communication module, which automatically records the medication timestamp when the patient opens the pillbox and takes out the medication, and transmits the confirmation signal to the edge computing gateway via Bluetooth 5.0 protocol; the environmental sensor node is deployed in key areas of the patient's residence, collecting temperature, humidity, and PM2.5 concentration data every 5 minutes; Furthermore, the edge computing gateway employs an ARM Cortex-A72 processor and a neural network acceleration unit to run a lightweight feature extraction model; the data cleaning process includes outlier removal based on Hampel filters and missing value imputation based on linear interpolation; the sliding window feature extraction adopts a time-series segmentation strategy with a window length of 60 minutes and a step size of 10 minutes, calculating five statistical features—mean, standard deviation, maximum value, minimum value, and rate of change—for the physiological parameters within each window, forming a structured time-series feature vector with a dimension of 30; Furthermore, the temporal convolutional network deployed on the cloud-based intelligent decision-making platform comprises four causal dilated convolutional layers with dilation coefficients of 1, 2, 4, and 8, a kernel size of 3, and the activation function ReLU, used to capture the dynamic fluctuation patterns of physiological indicators at different time scales; the long short-term memory network contains 128 hidden units, used to model the long-term dependency between medication behavior and viral load; the output feature vectors of the temporal convolutional network and the long short-term memory network are initially fused through a concatenation operation; Preferably, the cross-modal attention mechanism includes a physiological-medication attention submodule and a clinical-temporal attention submodule; the physiological-medication attention submodule calculates the correlation weight between heart rate variability features and medication timestamp features, expressed by the following formula:

[0007] in, This is a query vector for physiological features. Let the key vector be the drug use behavior. The clinical-temporal attention submodule uses CD4+ T lymphocyte count as the query vector to perform weighted aggregation of temporal fusion features, highlighting physiological fluctuation patterns that are highly correlated with immune status. Furthermore, the individualized health risk score is ultimately predicted using a multilayer perceptron, with the input being a fusion feature vector weighted by cross-modal attention, and the output being a continuous risk value from 0 to 100. The risk intervention rule base is pre-set with a three-level response mechanism: when the risk score is below 30, regular monthly follow-up is maintained; when the risk score is between 30 and 70, weekly video follow-up and medication adherence reinforcement reminders are triggered; when the risk score is above 70, the responsible physician is automatically notified and a home assessment process is initiated within 48 hours. In addition, the data security and privacy protection module uses the national cryptographic SM4 algorithm to encrypt transmitted data, and the stored data adopts an attribute-based access control policy, which only authorizes medical staff to decrypt and access the original data in specific diagnosis and treatment scenarios; all data used for model training is k-anonymized to ensure that any individual has at least k=50 indistinguishable neighbors in the dataset. Compared with the prior art, the present invention has the following beneficial effects: Through IoT sensing terminals, continuous and unobtrusive collection of multi-dimensional data on the physiology, behavior, and environment of AIDS patients is achieved, increasing the data collection frequency to the minute level. This is 43,200 times higher than the data density of traditional monthly outpatient follow-up, significantly enhancing the ability to dynamically monitor the disease. We build an edge-cloud collaborative computing architecture that, while ensuring data privacy, offloads more than 90% of the raw data processing tasks to the edge, reducing cloud bandwidth load by more than 60%, and achieves local preliminary anomaly detection through a lightweight model, with response latency controlled within 500 milliseconds. The proposed hybrid architecture of temporal convolutional network and long short-term memory network, combined with cross-modal attention mechanism, effectively models the complex nonlinear relationship between physiological indicators, medication behavior and clinical indicators. The AUC of the individualized health risk score reaches 0.92, which is 15 percentage points higher than the single-modal model. A dynamic, risk-driven, three-tiered intervention response mechanism was established to achieve precise allocation of follow-up resources. Clinical trials showed that the patient's 6-month follow-up compliance rate increased from 68% to 92%, the early warning accuracy rate for high-risk events (such as viral rebound and opportunistic infections) reached 85%, and the average intervention response time was shortened to 2.3 days. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the hybrid architecture of temporal convolutional network and long short-term memory network and cross-modal attention mechanism in this invention. Detailed Implementation

[0009] Example 1 Please refer to Figure 1 and Figure 2 This embodiment aims to comprehensively illustrate an IoT-driven remote follow-up monitoring method and system for HIV patients. This system integrates multi-source heterogeneous data acquisition, edge-side intelligent preprocessing, cloud-based deep learning decision-making, and remote interaction and security mechanisms to achieve continuous, objective, and intelligent monitoring of patients' health status.

[0010] This system first uses an IoT sensing terminal to collect real-time, continuous, and unobtrusive data on the physiological parameters, medication behavior, and home environment of HIV patients. This terminal mainly consists of wearable physiological monitoring devices, a smart pillbox, and environmental sensor nodes.

[0011] The wearable physiological monitoring device employs dual-mode sensing technology of photoplethysmography (PPG) and electrocardiogram (ECG) to continuously monitor the patient's heart rate variability (HRV) and blood oxygen saturation (SpO2) at a sampling rate of 1 Hz. The PPG sensor measures changes in skin blood volume by emitting and receiving light of specific wavelengths, thereby extracting heart rate; the ECG sensor captures cardiac bioelectrical signals through electrodes, providing more accurate heart rate and HRV data. Heart rate variability is an important indicator of autonomic nervous system activity, and its decrease may indicate patient stress or disease progression. Blood oxygen saturation directly reflects pulmonary gas exchange function. The device has a built-in Bluetooth Low Energy 5.0 module that periodically transmits the collected raw pulse waves, ECG waveforms, and SpO2 values ​​in encrypted form to an edge computing gateway. To ensure data accuracy, the device performs a self-check during wear. If there is poor contact or misalignment, it will remind the patient to adjust via indicator lights or vibration and pause data transmission until the correct wearing position is achieved. The device is typically made of medical-grade silicone, has an IP68 waterproof and dustproof rating, and a battery life of up to 7 days. It can be recharged wirelessly.

[0012] The smart pillbox incorporates a pressure sensor and a Near Field Communication (NFC) module. The interior is divided into multiple pill compartments, each with a pressure sensor at the bottom to detect the presence and quantity of pills in real time. When a patient opens the pillbox and removes medication within a preset dosage period, the pressure sensor detects the decrease in dosage, triggering a "medication action" event. This event automatically records the dosage timestamp (accurate to the second) and a confirmation signal (e.g., "pill removed"). The NFC module is used for patient identification and pillbox configuration synchronization. For example, upon first use, the patient pairs with a mobile app via an NFC tag and enters the medication name, dosage, and dosage plan. The pillbox transmits the dosage timestamp, dosage confirmation signal, and pillbox open / close status data to an edge computing gateway via Bluetooth 5.0. If the patient fails to remove medication within the specified time period, the pillbox will alert the patient with an audible and visual alarm and report a "missed dose" event after a certain delay. The pillbox also integrates a small amount of storage space to cache offline data and upload it after network recovery.

[0013] The environmental sensor nodes are deployed in key areas of the patient's residence, such as bedrooms and living rooms, collecting indoor temperature, humidity, and PM2.5 concentration data every 5 minutes. The temperature and humidity sensors utilize high-precision digital sensors (such as the SHT30), with an accuracy of ±0.3 degrees Celsius and ±2% relative humidity. The PM2.5 sensor employs the principle of laser scattering to accurately measure the concentration of fine particulate matter in the air. These sensor nodes typically use Wi-Fi or Zigbee protocols, transmitting data to an edge computing gateway via a home router. The collected environmental data provides objective evidence for assessing the impact of the patient's living environment on physiological indicators; for example, high temperature and humidity environments may affect patient comfort, thereby impacting sleep quality or heart rate.

[0014] Next, the edge computing gateway receives the raw data uploaded by the IoT sensing terminal and performs local preprocessing, feature extraction, and preliminary anomaly detection. This gateway uses a high-performance, low-power ARM Cortex-A72 processor and integrates a neural network acceleration unit (NPU) to enhance local computing capabilities.

[0015] The data processing flow includes: Data cleaning: Outlier removal and missing value imputation are performed on the raw sensor data. Outlier removal employs a statistical method based on the Hampel filter. The Hampel filter identifies and replaces outliers by calculating the median and median absolute deviation (MAD) within a sliding window. Its advantage lies in its insensitivity to outliers and its ability to effectively handle sudden spikes or drifting data. Specifically, for each data point, a window is formed by taking k data points before and after it. The median M and MAD of the data points within this window are calculated. If the difference between the current data point and M exceeds c*MAD (where c is a preset threshold, e.g., 3), it is identified as an outlier and replaced with M. Missing value imputation uses linear interpolation, which is suitable for sporadic missing sensor data caused by signal interruption or transmission problems within a short period of time. It estimates the missing data by using the linear trend of the valid data before and after it to maintain the continuity of the data sequence.

[0016] Timing Alignment: Since different sensors may have inconsistent sampling frequencies or timestamps, the gateway performs timing alignment. For low-frequency data (such as environmental data), forward fill is used to align it with high-frequency data; for high-frequency physiological data, if there is a slight time deviation, it is resampled to a uniform 1Hz or 5-minute time granularity and smoothed using interpolation methods.

[0017] Sliding window feature extraction: A sliding window strategy is used to extract features from the cleaned and aligned time-series data. The window length is set to 60 minutes, with a step size of 10 minutes. For each window, the following five statistical features are calculated for physiological parameters (heart rate, heart rate variability, blood oxygen saturation) and environmental parameters (temperature, humidity, PM2.5): mean, standard deviation, maximum value, minimum value, and rate of change. The rate of change is defined as the difference between the last data point and the first data point within the window divided by the window length, used to capture the trend of parameter changes. For example, for heart rate data, five features will be extracted: mean heart rate, standard deviation heart rate, maximum heart rate, minimum heart rate, and rate of change heart rate. For the three physiological parameters and two environmental parameters, a total of 5*5=25 features are obtained. Combining the medication timestamps and dosage confirmation signals recorded by the smart pillbox, five discrete or continuous features are formed through encoding (e.g., number of medications taken in the past 60 minutes, time since the last medication, deviation between expected and actual medication time, etc.). Ultimately, a structured temporal feature vector with a dimension of 30 is formed.

[0018] Preliminary anomaly detection: The edge gateway runs a lightweight feature extraction model (e.g., based on support vector machines or simplified decision trees) to perform preliminary anomaly detection on the generated structured time-series feature vectors. For example, if a patient's heart rate exceeds the normal physiological range by 3 standard deviations within a short period of time, or if blood oxygen saturation is below 90% for 10 consecutive minutes, it is immediately marked as a potential abnormal event, triggering a local alarm or notifying medical staff for initial attention, effectively reducing the latency and computing resource consumption of cloud processing.

[0019] Subsequently, the cloud-based intelligent decision-making platform receives the structured time-series feature vector uploaded by the edge computing gateway and performs in-depth analysis and decision-making. This platform is deployed based on a cloud-native architecture and possesses elastic scaling and high-concurrency processing capabilities.

[0020] Multimodal health status representation construction: The cloud platform deeply integrates the received structured time-series feature vectors with clinical data from the patient's electronic medical record (EMR). EMR data includes CD4+ T lymphocyte count (an indicator of immune system function), viral load (a disease progression indicator), comorbidity information (such as diabetes, hepatitis, etc.), and previous follow-up records (treatment plans, medication history, adverse reactions). These clinical indicators are obtained from the Hospital Information System (HIS) or Electronic Medical Record System (EMR) through structured queries, standardized, and aligned semantically and temporally with the time-series feature vectors to jointly construct a multimodal health status representation. For example, CD4 counts are typically measured periodically, thus serving as a slowly varying feature or contextual feature input in time-series modeling.

[0021] Dynamic modeling: Based on the multimodal health status representation, the platform uses a hybrid architecture of temporal convolutional network (TCN) and long short-term memory network (LSTM) to dynamically model the patient's health status in order to extract the temporal dependency between physiological indicators and medication behavior.

[0022] Temporal Convolutional Network (TCN): Contains four causal dilated convolutional layers. The dilation rates of each layer are 1, 2, 4, and 8, respectively, and the kernel size is 3. All convolutional layers use the ReLU activation function. Causal convolution ensures that the model uses only past and present information when predicting the current moment, without revealing future information. Dilated convolution increases the receptive field by skipping some values ​​in the input, allowing the model to capture dynamic fluctuation patterns at different time scales, such as short-term heart rate fluctuations and long-term body temperature trends.

[0023] Long Short-Term Memory (LSTM) Network: Contains 128 hidden units. LSTM effectively solves the gradient vanishing or exploding problems present in traditional Recurrent Neural Networks (RNNs) through its unique gating mechanism (forget gate, input gate, output gate), enabling it to better capture the long-term dependencies between medication behavior (e.g., long-term medication adherence, frequency and duration of missed doses) and clinical indicators such as viral load and CD4+ T lymphocyte count. For example, it can analyze whether missing doses for several consecutive days leads to an increased viral load.

[0024] Output concatenation: The output feature vectors of TCN and LSTM are initially fused through concatenation to form a more comprehensive joint feature representation that includes short-term dynamics and long-term dependencies.

[0025] Cross-modal attention mechanism: A cross-modal attention mechanism is introduced to adaptively fuse physiological temporal features, medication behavior features and clinical indicator features with weights to generate a personalized health risk score.

[0026] The physiological-medication attention submodule calculates the correlation weights between heart rate variability features and medication time stamp features. The formula is as follows:

[0027] in, This is a query vector for physiological characteristics (e.g., the heart rate variability vector at the current moment). This is a key vector for medication behavior (e.g., a feature vector of medication time deviation over the past 24 hours). For feature dimensions (ensure the query vector and key vector have the same dimensions). Indicates the first The physiological characteristics and the first Attention weights among individual medication behavior characteristics Dot product (inner product): measures the query vector and key vector The higher the dot product, the more correlated the two features are. Scaling factor: To prevent the dot product from becoming too large and causing the exponential function to explode (numerically unstable), it is usually divided by . Scaling / Scaling Exponential function: Converts the dot product result into a positive number and enhances the influence of highly similar terms. This module can dynamically discover the potential correlation between physiological states (such as reduced heart rate variability due to anxiety) and medication adherence (such as whether medication is taken on time) and assign greater weight to features with high correlation.

[0028] The clinical-temporal attention submodule uses CD4+ T lymphocyte count as the query vector to perform weighted aggregation of temporal fusion features, highlighting physiological fluctuation patterns highly correlated with immune status. For example, when a patient's CD4 count is low, the system pays more attention to physiological abnormalities related to immunosuppression (such as slight increases in body temperature or changes in respiratory rate) and assigns higher weights to these features to detect the risk of opportunistic infections earlier.

[0029] Through these two sub-modules, the system achieves intelligent weight allocation of multimodal features, making the final health status representation more interpretable and predictive.

[0030] Personalized health risk scoring: Finally, a multilayer perceptron (MLP) is used to predict the attention-weighted fused feature vector, outputting a continuous risk value from 0 to 100. The MLP consists of an input layer, two hidden layers (each containing 128 neurons using the ReLU activation function), and an output layer (using the Sigmoid activation function to map the output to 0-1, then multiplying by 100), used to non-linearly combine all features to derive the final risk assessment.

[0031] Risk Intervention Rule Base: Based on the individualized health risk score, the system matches a preset risk intervention rule base to dynamically generate follow-up frequency adjustment instructions, medication reminder reinforcement strategies, or emergency medical intervention suggestions. This rule base has a preset three-level response mechanism: Level 1 Response (Risk Score below 30): The patient's health status is stable and compliance is good. Maintain routine monthly video or telephone follow-ups and provide basic health education.

[0032] Level 2 Response (Risk Score between 30 and 70): The patient's health status fluctuates slightly or medication adherence declines slightly. This triggers weekly video follow-ups and daily reminders to reinforce medication adherence (e.g., via app push notifications or smart speaker voice reminders). Simultaneously, targeted health management articles or videos are pushed to assist the patient in adjusting their lifestyle.

[0033] Level 3 Response (Risk Score Above 70): The patient's health status is at significant risk or compliance issues are severe. The system automatically notifies the attending physician and nursing team and initiates a 48-hour home assessment process. Simultaneously, the system pushes emergency medical advice to the patient's remote interactive terminal, informing them of immediate actions to take, such as measuring body temperature and blood oxygen levels and uploading the data.

[0034] All intervention instructions are executed and pushed through a remote interactive terminal.

[0035] Finally, the remote interactive terminal is used to push risk warning information, follow-up tasks, and health guidance content to patients or medical staff, and to receive patient feedback data. Patient terminals can be smartphone apps or dedicated tablets, providing an intuitive visual interface. Medical staff terminals can be PC workstations or mobile apps.

[0036] For patients: Personalized medication reminders, health education videos / articles, follow-up questionnaires, and symptom self-assessment scales will be sent to them. Patients can submit symptom descriptions, psychological feedback, and questions about the health guidance content through the terminal.

[0037] For healthcare professionals: The system displays an overview of the patient's health risks, detailed physiological data curves, medication adherence reports, and historical follow-up records. Healthcare professionals can remotely adjust follow-up plans, send personalized guidance messages, initiate video calls, and view feedback data submitted by patients.

[0038] Meanwhile, the data security and privacy protection module runs through the entire system, implementing end-to-end encryption, access control, and anonymization processing for health data transmitted and stored throughout the entire chain.

[0039] Transmission Encryption: The national standard SM4 algorithm is used for end-to-end encryption of all data transmitted between IoT sensing terminals, edge computing gateways, and the cloud-based intelligent decision-making platform. SM4 is a 128-bit block cipher algorithm that ensures the confidentiality and integrity of data during transmission. Key management employs the X.509 digital certificate system based on Hardware Security Modules (HSM) for negotiation and distribution.

[0040] Storage Access Control: Data stored on the cloud platform employs an attribute-based access control (ABAC) policy. The system defines user attributes (e.g., doctor, nurse, patient, researcher), data attributes (e.g., physiological data, medication behavior data, clinical records, anonymized data), and environmental attributes (e.g., access time, IP address). Decryption and access to raw data are only permitted when the user's attributes, data attributes, and environmental attributes meet the predefined access policy. For example, only authorized physicians can access patients' original medical record data within a clinical setting.

[0041] Anonymization: All data used for model training and scientific research is k-anonymized. This means that for any given patient in the dataset, there are at least k=50 other records that are indistinguishable on the quasi-identifier attribute. Quasi-identifiers typically include age, gender, postal code, etc. This process aims to ensure that even if an attacker possesses external information, it is difficult to re-identify the anonymized data as a specific individual, thus striking a balance between data analysis and privacy protection.

[0042] Example 2 This embodiment addresses the core pain point of low medication adherence in the management of HIV patients by further optimizing data collection, edge processing, and cloud-based decision-making mechanisms to achieve more accurate medication adherence prediction and personalized intervention. It aims to solve the problem that traditional monitoring only focuses on medication reminders and fails to deeply analyze the behavioral patterns and physiological connections behind adherence.

[0043] The system architecture of this embodiment is basically the same as that of Embodiment 1, but specific enhancements and differentiated designs have been made in the following technical details: First, at the data collection level of IoT sensing terminals, the smart pillbox has been functionally enhanced to obtain more granular medication behavior data.

[0044] In addition to pressure sensors and an NFC module, the smart pillbox integrates a miniature optical sensor array to accurately identify the type and quantity of medication taken out by the patient. By pre-calibrating the visual characteristics of different medications within the pillbox, it can determine whether the patient has taken the correct dosage as prescribed, rather than simply sensing a decrease in the amount of medication. If the amount of medication taken does not match the prescription, a "dosage error" event is generated.

[0045] The pillbox also incorporates a gyroscope and accelerometer to sense its tilt angle and movement. When a patient frequently opens and closes the pillbox outside of the preset medication period, or when the pillbox moves abnormally (e.g., it is lost or accidentally touched by someone else), this behavioral data is recorded and transmitted as a potential adherence risk signal. This enhanced data, along with the medication timestamp and confirmation signal, is uploaded to the edge computing gateway via Bluetooth 5.0.

[0046] Secondly, at the local processing level of the edge computing gateway, the feature extraction and preliminary risk assessment related to medication adherence have been enhanced.

[0047] Refined Medication Behavior Feature Extraction: In addition to the general features in Example 1, the edge gateway will calculate the following additional medication behavior features based on the enhanced smart pillbox data: Deviation rate of single medication dispensing: (actual amount of medication dispensed - prescription dosage) / prescription dosage.

[0048] Dosage interval variability: the standard deviation of the interval between the last 5 doses from the set interval.

[0049] Medicine box abnormal opening frequency: The number of times the medicine box is opened outside of the medication administration time window.

[0050] Consistency of medication posture: By analyzing the patient's hand movements when taking medication based on gyroscope data, it can be identified whether it is a habitual medication behavior, so as to screen out accidental touches or non-personal operation.

[0051] These fine-grained features, combined with physiological and environmental features, together constitute a structured temporal feature vector with expanded dimensions.

[0052] Lightweight adherence risk warning at the edge: A small predictive model based on decision tree ensemble runs on the edge gateway's NPU. This model is specifically designed to make a preliminary judgment on the "short-term medication adherence decline risk" based on medication behavior data from the most recent 24 hours (such as two consecutive medication dosage deviations exceeding 10%, or two consecutive medication cycles of missed doses), combined with the patient's daily physiological fluctuations (such as a significant decrease in heart rate variability). Once the model identifies a high risk, it will immediately trigger a local alarm and send an instant reminder to the patient via a remote interactive terminal (e.g., a pop-up window on a mobile app or a voice prompt from a smart speaker, "You seem to have forgotten to take your medication today, please check immediately"). At the same time, a data packet labeled "high-priority adherence risk" will be sent to the cloud platform to expedite cloud processing and response.

[0053] Furthermore, at the in-depth analysis level of the cloud-based intelligent decision-making platform, the representation of multimodal health status, dynamic modeling, and cross-modal attention mechanisms have been optimized to more accurately predict medication adherence.

[0054] Expanding the dimensions of multimodal health status representation: In addition to the data from Example 1, further integration is made of patient medication education completion (obtained from questionnaire feedback via remote interactive terminals) and feedback history on specific drug adverse reactions. For example, if a patient reports a serious side effect on a certain drug, this will be taken into account as an important negative factor affecting adherence.

[0055] Structural adjustments to temporal modeling (S140): In the TCN-LSTM hybrid architecture, the 128 hidden units of the LSTM layer will focus more on capturing complex patterns and long-term dependencies in medication behavior sequences. Furthermore, a separate Gated Recurrent Unit (GRU) layer is introduced to specifically handle discrete event sequences reported by the smart pillbox (such as "pillbox opening," "abnormal dosage," and "unauthorized pillbox movement"), capturing the time, frequency, and pattern of event occurrence. Its output is then concatenated with the outputs of TCN and LSTM to form a more comprehensive fusion feature. This GRU layer can effectively handle discontinuous event data, improving the model's sensitivity to sudden behavioral anomalies.

[0056] Weight adjustments and new submodules for the cross-modal attention mechanism (S150): Weight adjustment for the physiological-drug attention submodule: In this submodule, the query vector More attention will be paid to physiological characteristics directly related to medication adherence (such as sleep quality score and morning heart rate), while key vectors The focus is on refined medication behavior characteristics (such as dosage deviation rate and dosing interval variability). Through learning, the model assigns a higher weight to the impact of physiological discomfort on adherence.

[0057] A new "Behavior-Clinical" attention submodule has been added. This submodule uses medication behavior characteristics (such as long-term adherence fluctuation patterns) as query vectors and aggregates them with weighted clinical indicators (such as viral load and CD4 count trends over the past 3 months). Its goal is to assess the strength of the impact of different medication behavior patterns on long-term treatment outcomes (such as viral suppression and immune reconstitution). For example, if a patient's adherence fluctuation pattern is highly correlated with small rebounds in viral load, this behavioral characteristic will be given a higher weight.

[0058] The focus of the individualized health risk score (S160) and the refinement of the risk intervention rule base: The focus of the risk score: The risk value of 0 to 100 predicted by the multilayer perceptron will more directly reflect the "risk of declining medication adherence". For example, a risk score of 70 may mean that the patient has an 80% probability of missing a dose or making a dosage error in the next 7 days.

[0059] Refinement of the risk intervention rule base: Based on this compliance risk score, the system will generate more targeted intervention instructions: Risk score below 30 (good adherence): Maintain regular medication reminders and regularly send HIV / AIDS health information to encourage patients to maintain good adherence.

[0060] Risk score between 30 and 70 (moderate risk of non-compliance): Enhanced reminder strategy: Trigger multiple daily sound and light reminders from the smart pillbox, mobile app push notifications, and smart speaker voice reminders such as "Please check your medication time".

[0061] Personalized educational push: Based on the patient's historical data and potential reasons for adherence problems (such as forgetting, side effects, mood swings), the system automatically generates and pushes short videos or interactive Q&As on topics such as "how to deal with medication side effects" and "balancing medication with daily life".

[0062] Family collaboration mechanism: If authorized by the patient, the system can send a reminder message to the designated family member, requesting the family member's assistance in supervising the process.

[0063] Risk score above 70 (high risk of non-compliance): Emergency Intervention Notification: Automatically notifies the responsible nurse or physician and provides a detailed report of recent abnormal medication behavior data and predicted adherence risks.

[0064] One-on-one interview recommendation: It is recommended that the responsible nurse conduct a one-on-one interview with the patient via video or telephone within 24 hours to gain a deeper understanding of the specific reasons for the decline in adherence (such as psychological stress, financial difficulties, and medication accessibility issues), and provide targeted psychological counseling or coordinate resources.

[0065] Medication regimen assessment: Doctors can use the compliance analysis report provided by the system to assess whether the existing medication regimen is suitable for the patient and consider adjusting the medication or the frequency of medication.

[0066] All intervention instructions are delivered via remote interactive terminals in various formats, including text, voice, and video.

[0067] This embodiment, through in-depth collection and analysis of medication behavior data and specific optimization of the model structure, enables the system to identify medication adherence risks earlier and more accurately, and to provide multi-level, personalized intervention measures, thereby significantly improving treatment adherence and long-term treatment outcomes for HIV patients.

[0068] Example 3 This embodiment aims to address the technical problems of limited model generalization ability and difficulty in continuous optimization based on new data in remote follow-up monitoring systems for HIV / AIDS patients. By introducing an adaptive learning mechanism and a model optimization module, it enables a cloud-based intelligent decision-making platform to continuously and adaptively optimize model parameters and risk intervention strategies. Building upon Embodiments 1 and 2, this embodiment further enhances the system's intelligent evolution capabilities.

[0069] The system architecture of this embodiment is an extension of that of Embodiment 1, mainly by adding a model adaptive optimization module and a feedback closed-loop processing module to the cloud-based intelligent decision-making platform.

[0070] First, while the data acquisition and preprocessing processes of IoT sensing terminals and edge computing gateways remain unchanged, the edge gateway, in addition to uploading structured time-series feature vectors, will also upload the confidence level of its preliminary anomaly assessment. For example, if the edge model identifies a physiological anomaly, it will upload not only the event itself but also the probability value it uses to determine that the event is abnormal. Furthermore, when medical staff manually review and correct local alarms issued by the edge gateway through remote interactive terminals, this "manual review label" will also be anonymized and uploaded to the cloud.

[0071] Secondly, the core functions of the cloud-based intelligent decision-making platform have been enhanced, with a focus on enabling continuous learning and optimization of the model: Extending Multimodal Health Status Representation: In addition to the data described in Examples 1 and 2, the cloud platform will also construct a "Model Performance Feedback Dataset." This dataset includes: Historical prediction results: This includes the system's predicted values ​​for patients' individualized health risk scores over a past period, the confidence level of the predictions, and the intervention recommendations triggered by the predictions.

[0072] Actual clinical events: such as objective clinical outcome data including actual viral load rebounds, opportunistic infections, and hospitalization records due to compliance issues. This data is usually synchronized regularly through the hospital information system.

[0073] Patient and healthcare staff feedback: This includes patient symptom self-assessment results, understanding of health guidance content, acceptance of intervention measures and effectiveness evaluation, collected through remote interactive terminals, as well as healthcare staff's response time to risk warnings and evaluation of the effectiveness of intervention measures.

[0074] After being cleaned, standardized, and semantically analyzed, this feedback data is used to evaluate the predictive accuracy of the model and the effectiveness of the intervention strategy.

[0075] Online learning and incremental optimization of temporal modeling (S140) and cross-modal attention mechanism (S150): Model Adaptive Optimization Module: The cloud platform integrates a model adaptive optimization module, which is responsible for continuous online learning or incremental training of the parameters of TCN, LSTM, and cross-modal attention mechanisms.

[0076] Optimization Objectives Expanded: In addition to minimizing health risk prediction errors, optimization objectives now include maximizing intervention effects (e.g., improving patient medication adherence and reducing the incidence of high-risk events) and enhancing model generalization ability.

[0077] Dynamic model parameter adjustment: This module employs a periodic incremental training strategy. For example, every two weeks or monthly, the system performs mini-batch incremental training on the model using the latest "model performance feedback dataset." This involves: Parameter fine-tuning based on reinforcement learning: Using patient feedback on intervention strategies (such as improved compliance and symptom relief) as reward signals, the weight allocation in the attention mechanism is adjusted so that feature combinations that lead to better clinical outcomes receive higher weights.

[0078] Adversarial training or domain adaptation techniques: To address the "data drift" problem caused by differences in patient groups or disease progression, adversarial training or domain adaptation techniques are introduced, enabling the model to better generalize to new, undertrained patient data, thereby improving the model's ability to adapt to individual differences.

[0079] Federated Learning Framework: To further enhance the model's generalization ability and strictly protect patient privacy, the cloud platform adopts the concept of Federated Learning. While ensuring that all patient data undergoes k-anonymization and differential privacy processing, anonymized feedback data from patients in different hospitals or regions can collaboratively train and update the cloud model parameters without sharing the original data, thereby jointly improving the model's intelligence.

[0080] Confidence assessment and adaptive adjustment of intervention rules for the individualized health risk score (S160): Risk Prediction Confidence: In addition to outputting continuous risk values ​​from 0 to 100, the multilayer perceptron also outputs a "risk prediction confidence level" (e.g., the entropy value based on the predicted probability distribution). When this confidence level is lower than a preset threshold (e.g., 0.7), the system will mark the prediction result as "low confidence" and trigger a prompt message to medical staff, suggesting manual review or collection of more targeted data to assist in the judgment. At the same time, the system will prioritize feeding this low-confidence prediction and its related data back to the model adaptive optimization module for focused learning.

[0081] Adaptive adjustment of the risk intervention rule base: Traditional risk rule bases are static, while the rule base in this embodiment is dynamically adjustable. The model adaptive optimization module continuously evaluates the actual effects of different intervention strategies within different risk ranges.

[0082] If the system finds that a certain intervention strategy (such as "weekly video follow-up") is not effective in improving compliance or avoiding high-risk events in a specific risk range (such as patients with a risk score between 40 and 50), the system will recommend to medical staff to adjust the intervention strategy in that risk range through the recommendation system. For example, it may reduce its priority or recommend combining it with other intervention methods such as "family supervision and reminders".

[0083] For risk score thresholds (such as 30 or 70), the system will dynamically suggest fine-tuning based on historical data and model optimization results. For example, if the model finds that adjusting the high-risk threshold to 65 can identify patients requiring urgent intervention earlier, it will suggest an adjustment. The rules will take effect after healthcare professionals confirm them on a remote interactive terminal.

[0084] Strengthening the feedback loop: Patient feedback collected via remote interactive terminals (such as understanding of guidance content, symptom relief, and acceptance of interventions) is considered a key reward or penalty signal and directly input into the model's adaptive optimization module. For example, if a patient reports "not understanding" or "not helpful" regarding certain health guidance content, the model will adjust the priority of that content or recommend more easily understood content. Evaluations of the effectiveness of interventions by healthcare professionals are also used as labeled data for supervised or semi-supervised learning of the model.

[0085] Through the above adaptive learning and optimization mechanisms, the remote follow-up monitoring system for HIV patients in this embodiment can continuously optimize its prediction model and intervention strategy based on the patient's dynamic data and actual feedback, thereby overcoming the problem of insufficient generalization ability of existing systems, realizing a true closed-loop linkage from data perception to clinical action, and continuously improving the system's intelligence level and clinical applicability.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An Internet of Things-driven remote follow-up monitoring system for HIV / AIDS patients, characterized in that, It includes the following components: IoT sensing terminals are used to collect patients' physiological parameters, medication behavior, and home environment data in real time. An edge computing gateway is used to perform local preprocessing, feature extraction, and preliminary anomaly detection on the raw data uploaded by the IoT sensing terminal. The cloud-based intelligent decision-making platform is used to receive structured feature data uploaded by the edge computing gateway, and to perform multimodal health status fusion modeling, dynamic risk level assessment, and follow-up intervention strategy generation based on a deep learning model. Remote interactive terminals are used to push risk warning information, follow-up tasks and health guidance content to patients or medical staff, and to receive patient feedback data. The data security and privacy protection module is used to implement end-to-end encryption, access control and anonymization of health data transmitted and stored throughout the entire chain; The IoT sensing terminal includes wearable physiological monitoring devices, smart pillboxes, and environmental sensor nodes. The wearable physiological monitoring device uses PPG and ECG dual-mode sensing technology to continuously monitor heart rate variability and blood oxygen saturation at a sampling rate of 1Hz. The smart pillbox has a built-in pressure sensor and near-field communication module. When the patient opens the pillbox and takes out the medicine, it automatically records the medication timestamp and transmits the confirmation signal to the edge computing gateway via Bluetooth 5.0 protocol. The environmental sensor nodes are deployed in key areas of the patient's residence, collecting temperature, humidity, and PM2.5 concentration data every 5 minutes.

2. The IoT-driven remote follow-up monitoring system for HIV / AIDS patients according to claim 1, characterized in that, The edge computing gateway uses an ARM Cortex-A72 processor and a neural network acceleration unit to run a lightweight feature extraction model; the local preprocessing includes outlier removal based on Hampel filter and missing value imputation based on linear interpolation; the feature extraction adopts a sliding window strategy with a window length of 60 minutes and a step size of 10 minutes, and calculates five statistical features for physiological parameters within each window, namely mean, standard deviation, maximum value, minimum value and rate of change, to form a structured time-series feature vector with a dimension of 30.

3. The IoT-driven remote follow-up monitoring system for HIV / AIDS patients according to claim 1, characterized in that, The deep learning model deployed on the cloud-based intelligent decision-making platform includes a hybrid architecture of temporal convolutional networks and long short-term memory networks. The temporal convolutional network contains four causal dilated convolutional layers with dilation coefficients of 1, 2, 4, and 8, a kernel size of 3, and the activation function is ReLU. The long short-term memory network contains 128 hidden units. The output feature vectors of the temporal convolutional network and the long short-term memory network are initially fused through a concatenation operation.

4. The IoT-driven remote follow-up monitoring system for HIV / AIDS patients according to claim 3, characterized in that, The cloud-based intelligent decision-making platform further introduces a cross-modal attention mechanism to adaptively fuse physiological temporal features, medication behavior features, and clinical indicator features with weights; the cross-modal attention mechanism includes a physiological-medication attention submodule and a clinical-temporal attention submodule; The physiological-medication attention submodule calculates the correlation weight between heart rate variability features and medication timestamp features, expressed by the following formula:

5. Among them This is a query vector for physiological features. Let the key vector be the medication behavior. For feature dimension, Indicates the first The physiological characteristics and the first Attention weights among individual medication behavior characteristics Dot product: measures the query vector and key vector The higher the dot product, the more correlated the two features are. Scaling factor: To prevent the dot product from becoming too large and causing the exponential function to explode, it is usually divided by . Scaling / Scaling The exponential function converts the dot product result into a positive number.

6. The IoT-driven remote follow-up monitoring system for HIV / AIDS patients according to claim 4, characterized in that, The individualized health risk score is ultimately predicted using a multilayer perceptron. The input is a fusion feature vector after cross-modal attention weighting, and the output is a continuous risk value from 0 to 100. The risk intervention rule base is preset with a three-level response mechanism: when the risk score is below 30, regular monthly follow-up is maintained; when the risk score is between 30 and 70, weekly video follow-up and medication adherence reinforcement reminders are triggered; when the risk score is above 70, the responsible doctor is automatically notified and a home assessment process is initiated within 48 hours.

7. The IoT-driven remote follow-up monitoring system for HIV / AIDS patients according to claim 1, characterized in that, The data security and privacy protection module uses the national cryptographic SM4 algorithm to encrypt transmitted data, and the stored data adopts an attribute-based access control policy, which only authorizes medical staff to decrypt and access the original data in specific diagnosis and treatment scenarios; all data used for model training is k-anonymized to ensure that any individual has at least k=50 indistinguishable neighbors in the dataset.

8. A monitoring method using the Internet of Things-driven remote follow-up monitoring system for HIV patients according to any one of claims 1-6, characterized in that, Includes the following steps: Step S110: Collect multi-source health data of HIV patients in real time through IoT sensing terminals. This multi-source health data includes heart rate, body temperature, blood oxygen saturation, sleep quality, medication timestamps, medication dosage confirmation signals, indoor temperature and humidity, and air quality parameters. Step S120: Perform data cleaning, missing value imputation, temporal alignment, and sliding window feature extraction on the multi-source health data through an edge computing gateway to generate a structured temporal feature vector. Step S130: Upload the structured temporal feature vector to a cloud-based intelligent decision-making platform, combining it with patient data. Step S140: Based on the multimodal health status representation, the patient's health status is dynamically modeled using a hybrid architecture of temporal convolutional network and long short-term memory network to extract the temporal dependency between physiological indicators and medication behavior. Step S150: A cross-modal attention mechanism is introduced to adaptively fuse physiological temporal features, medication behavior features, and clinical indicator features to generate an individualized health risk score. Step S160: Based on the individualized health risk score, match the preset risk intervention rule base, dynamically generate follow-up frequency adjustment instructions, medication reminder reinforcement strategies, or emergency medical intervention suggestions, and push them through a remote interactive terminal.

9. The IoT-driven remote follow-up monitoring method for HIV / AIDS patients according to claim 7, characterized in that, The data cleaning in step S120 adopts an outlier removal method based on Hampel filter, and the missing value filling adopts a linear interpolation method; the sliding window feature extraction adopts a strategy with a window length of 60 minutes and a step size of 10 minutes, and calculates the mean, standard deviation, maximum value, minimum value and rate of change of physiological and environmental parameters in each window to form a structured time series feature vector with a dimension of 30.

10. The IoT-driven remote follow-up monitoring method for HIV / AIDS patients according to claim 7, characterized in that, In step S160, the three-level response mechanism of the risk intervention rule base is as follows: when the risk score is below 30, regular monthly follow-up is maintained; when the risk score is between 30 and 70, weekly video follow-up and medication adherence reinforcement reminders are triggered; when the risk score is above 70, the responsible doctor is automatically notified and the on-site assessment process within 48 hours is initiated; the risk score is predicted by a multilayer perceptron using the attention-weighted fusion feature vector, and the output is a continuous value ranging from 0 to 100.

Images