Health data collection and analysis platform for elderly chronic disease management

Abstract

The present application relates to the technical field of health care informatics, and particularly relates to a health data collection and analysis platform for chronic disease management of the elderly, which comprises a semantic perception edge node, a space-time alignment transmission middleware, a causal reasoning analysis engine and a clinically interpretable intervention generation layer, and realizes end-to-end closed-loop optimization through sensing end context perception, transmission layer dynamic semantic encapsulation, analysis layer state transition graph neural network modeling, and output layer clinical path driven intervention instruction generation. The present application significantly improves intervention accuracy and timeliness, enhances user compliance and reduces system power consumption.

Description

Technical Field

[0001] This invention relates to the field of healthcare informatics technology, and in particular to a health data collection and analysis platform for chronic disease management in the elderly. Background Technology

[0002] Against the backdrop of a continuously aging population and a rising incidence of chronic diseases, the chronic disease management system for the elderly is gradually evolving from the traditional in-hospital treatment model to a data-driven, full-cycle health intervention model.

[0003] Current mainstream technologies mostly rely on a three-layer architecture consisting of wearable physiological monitoring devices, smart home terminals, and cloud data centers. They use Bluetooth or Wi-Fi protocols to periodically upload vital signs data and use rule engines or shallow machine learning models to trigger threshold-based alarms for abnormal indicators.

[0004] As related technologies continue to develop, while pursuing the breadth of data collection and transmission efficiency, the existing architecture has failed to effectively solve the problems of spatiotemporal alignment and causal association modeling of heterogeneous data at the semantic level, resulting in the analysis results lacking sufficient explanatory power and intervention guidance value at the level of clinical decision support. Summary of the Invention

[0005] The inventors discovered through research that the evolution of health status in elderly patients with chronic diseases often exhibits complex characteristics of nonlinearity and multi-factor coupling. Fluctuations in a single-dimensional physiological indicator may be driven by multiple factors such as decreased medication adherence, changes in environmental temperature and humidity, deterioration of sleep quality, or psychological stress events. The currently widely used linear processing paradigm of "sensor data → cloud aggregation → threshold judgment" is essentially a decontextualized data compression process. Its sampling frequency reduction and data packet encapsulation strategies implemented at the transmission layer to reduce bandwidth consumption inevitably lose key environmental context and behavioral timing information.

[0006] The purpose of this invention is to provide a health data collection and analysis platform for the management of chronic diseases in the elderly, in order to solve the technical problems of high false alarm rate, delayed intervention and poor user compliance caused by the existing technology due to data semantic fragmentation, inaccurate timing, lack of causality and weak explanatory power.

[0007] This invention provides a health data collection and analysis platform for chronic disease management in the elderly, including:

[0008] Semantic-aware edge node layer, spatiotemporal alignment transmission middleware layer, causal reasoning analysis engine layer, and clinically interpretable intervention generation layer;

[0009] The semantic perception edge node layer is deployed in the daily activity space of elderly users and includes physiological sensing units, environmental sensing units, behavior recognition units and local semantic fusion processors.

[0010] The spatiotemporal aligned transmission middleware layer is deployed within the home gateway device and includes a semantic data receiving buffer, a dynamic bandwidth allocator, a semantic priority scheduler, and an adaptive compression encoder.

[0011] The causal reasoning analysis engine layer is deployed in a cloud-based distributed computing cluster and includes a multimodal data lake, a state transition graph neural network model, a dynamic causal discovery module, and an individualized baseline modeler.

[0012] The clinically interpretable intervention generation layer includes an intervention strategy knowledge base, a causal path interpreter, and a multi-channel instruction dispatcher.

[0013] In some embodiments, the physiological sensing unit consists of a photoplethysmography pulse wave sensor, an impedance respiratory rate sensor, a thermopile body surface temperature sensor, and a triaxial accelerometer, and is connected to the local semantic fusion processor via an I²C bus.

[0014] The environmental sensing unit consists of a digital temperature and humidity sensor, an atmospheric pressure sensor, a light intensity sensor, and a carbon dioxide concentration sensor, and is connected to the local semantic fusion processor via an SPI bus.

[0015] The behavior recognition unit consists of a millimeter-wave radar module and a low-resolution infrared thermal imaging module, and is connected to the local semantic fusion processor through a parallel LVDS interface.

[0016] In some embodiments, the semantic priority scheduler divides data packets into four priority queues based on the behavioral semantic tags and abnormal physiological parameter flags embedded in the data packets:

[0017] The first priority is data packages that include the "medication" behavior tag and have blood glucose levels below 3.9 mmol / L or above 13.9 mmol / L;

[0018] The second priority is data packets that include the "toilet use" behavior tag and have a heart rate variability of less than 20ms;

[0019] The third priority is data packets that include the "unknown behavior" label and have been continuously cached for more than 3 cycles;

[0020] The fourth priority is given to other regular monitoring data packets; each queue uses a weighted round-robin scheduling algorithm with weight coefficients of 8, 4, 2, and 1 respectively.

[0021] The adaptive compression encoder employs a hybrid compression strategy, using an improved differential pulse code modulation algorithm for physiological signal data, Huffman coding for environmental data, and run-length encoding for behavioral semantic tags.

[0022] In some embodiments, the multimodal data lake adopts a columnar storage architecture, with data tables sharded by user ID. Each shard includes a physiological data table, an environmental data table, a behavioral data table, and a semantic association table. Each table is associated with a foreign key through a 64-bit globally unique timestamp, with a timestamp alignment precision of 1ms.

[0023] In some embodiments, the dynamic causal discovery module adopts a framework that combines a constraint-based PC algorithm with a score-based GES algorithm. The input is a sequence of individualized health status vectors for 7 consecutive days, and the output is a causal structure represented by a directed acyclic graph.

[0024] The GES algorithm performs a greedy equivalence class search on the skeleton graph output by the PC algorithm, and the scoring function adopts the Bayesian information criterion with a search step size of 1000 iterations.

[0025] The individualized baseline modeler is based on a causal graph structure and uses a structural equation model to fit the functional relationship between variables. The model parameters are solved by maximum likelihood estimation, with a convergence threshold of 1e-6. It is automatically updated at 2:00 AM every day, excluding data points marked as anomalous in the past 24 hours during the update.

[0026] In some embodiments, the intervention strategy knowledge base is constructed using the OWL ontology language and includes disease entities, symptom entities, drug entities, behavior entities, and intervention rule entities; the intervention rule entities are defined using the SWRL rule language, with the rule antecedent being a logical combination of clinical risk scores and causal graph structures, and the rule consequent being a specific intervention action;

[0027] The causal path interpreter automatically generates natural language explanation text after the intervention rule is triggered;

[0028] The multi-channel command distributor selects smart speaker voice broadcast, smartphone APP push, family WeChat messages, and community doctor workstation pop-up as distribution channels according to the user's device online status and interaction preferences. The distribution priority decreases in that order, and the response timeout thresholds are 30 seconds, 2 minutes, 5 minutes, and 10 minutes, respectively.

[0029] Compared with existing technologies, this invention has the following beneficial effects: Through a multi-level, multi-dimensional, and highly collaborative technical architecture, this invention achieves an intelligent upgrade of the entire chain of elderly chronic disease management, from data collection to clinical intervention, and solves the fundamental contradictions in existing technologies such as semantic breaks, causal gaps, weak explanations, and unsustainable systems, providing a complete technical solution for building an efficient, accurate, and humanized smart health service system. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the overall system architecture of the present invention;

[0032] Figure 2 This is a schematic diagram of the hardware composition of the semantic awareness edge node layer and the internal architecture of the local semantic fusion processor in the platform of this invention. Detailed Implementation

[0033] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example

[0034] This invention presents a health data collection and analysis platform for chronic disease management in the elderly. At the engineering implementation level, it constructs a four-layer collaborative architecture: a semantic-aware edge node layer, a spatiotemporally aligned transmission middleware layer, a causal reasoning analysis engine layer, and a clinically interpretable intervention generation layer. This architecture achieves end-to-end closed-loop management from raw physiological signal acquisition to clinical decision support. The following details the hardware configuration, software algorithms, data processing flow, system parameter settings, and collaborative working mechanisms of each module.

[0035] The semantic perception edge node layer is deployed in the daily activity space of elderly users. It consists of physiological sensing units, environmental sensing units, behavior recognition units, and a local semantic fusion processor. Each unit is interconnected through a low-power asynchronous bus to form a distributed sensing array. The physiological sensing units include a photoplethysmography pulse wave sensor, an impedance respiratory rate sensor, a thermopile body surface temperature sensor, and a triaxial accelerometer. The sampling frequencies are 125Hz, 50Hz, 1Hz, and 200Hz, respectively. The data output format is a 16-bit signed integer, which is connected to the local semantic fusion processor through an I²C bus.

[0036] The environmental sensing unit consists of a digital temperature and humidity sensor, an atmospheric pressure sensor, a light intensity sensor, and a carbon dioxide concentration sensor. The sampling frequency is uniformly 0.1Hz, the data output format is 32-bit floating point, and it is connected to the local semantic fusion processor via the SPI bus.

[0037] The behavior recognition unit consists of a millimeter-wave radar module and a low-resolution infrared thermal imaging module. The millimeter-wave radar operates at a frequency of 60 GHz, with a transmit power of 10 dBm, a receive sensitivity of -90 dBm, a spatial resolution of 5 cm, and a frame rate of 10 fps. The infrared thermal imaging module has a detection wavelength range of 8 μm to 14 μm, a thermal sensitivity of 50 mK, and a frame rate of 5 fps. Both modules output raw point cloud data and a heat map matrix, which are connected to the local semantic fusion processor through a parallel LVDS interface.

[0038] The local semantic fusion processor adopts a dual-core heterogeneous architecture. The real-time core is an ARM Cortex-M7 core with a main frequency of 480MHz, responsible for sensor data acquisition, preprocessing, and local caching. The inference core is an NPU coprocessor with a computing power of 4TOPS, supporting INT8 quantized inference, used to execute a lightweight behavioral semantic annotation model. The behavioral semantic annotation model is a lightweight neural network based on spatiotemporal convolution and attention mechanisms. Its input is a three-dimensional tensor after aligning millimeter-wave point cloud sequences and infrared thermal image sequences in the time dimension. The output is a predefined behavior category label, including six categories: "sitting", "standing", "walking", "lying down", "using the toilet", and "taking medication". The classification confidence threshold is set to 0.85. Samples below this threshold are marked as "unknown behavior" and trigger the local caching mechanism. After receiving the raw data from each sensor, the local semantic fusion processor first performs timestamp synchronization. It inserts a 64-bit UTC timestamp generated by a local high-stability crystal oscillator at the beginning of each sensor data frame using a hardware trigger, with a timestamp accuracy of 1μs. Then, it performs data preprocessing, including 50Hz power frequency notch filtering for physiological signals, sliding median filtering for environmental data, and background subtraction and contour extraction for behavioral data. Finally, it encapsulates the preprocessed data and behavioral semantic tags into a structured semantic data packet. The data packet format follows a custom binary protocol, including a header identifier, device ID, timestamp, data type identifier, data length, data payload, and CRC32 checksum.

[0039] The spatiotemporal aligned transmission middleware layer is deployed within the home gateway device and includes a semantic data receive buffer, a dynamic bandwidth allocator, a semantic priority scheduler, and an adaptive compression encoder. The semantic data receive buffer adopts a circular queue structure with a queue depth of 1024 data packets. Each data packet has a maximum length of 512 bytes. The buffer read and write pointers are managed by a hardware DMA controller to ensure zero-copy data transmission.

[0040] The dynamic bandwidth allocator calculates the currently available bandwidth based on channel quality metrics output by the real-time network status monitoring module, including RSSI, SNR, packet loss rate, and round-trip time. The window length is 5 seconds, and the update frequency is 1Hz. The semantic priority scheduler divides data packets into four priority queues based on embedded behavioral semantic tags and abnormal physiological parameter flags.

[0041] The first priority is data packages that include the "medication" behavior tag and have blood glucose levels below 3.9 mmol / L or above 13.9 mmol / L;

[0042] The second priority is data packets that include the "toilet use" behavior tag and have a heart rate variability of less than 20ms;

[0043] The third priority is data packets that include the "unknown behavior" label and have been continuously cached for more than 3 cycles;

[0044] The fourth priority is given to other routine monitoring data packets. Each queue uses a weighted round-robin scheduling algorithm with weight coefficients of 8, 4, 2, and 1, respectively, to ensure that high-clinical-risk data is transmitted first.

[0045] The adaptive compression encoder employs a hybrid compression strategy. It uses an improved differential pulse code modulation (DPCM) algorithm for physiological signal data, Huffman coding for environmental data, and run-length encoding for behavioral semantic tags. The improved DPCM algorithm introduces an adaptive prediction coefficient update mechanism based on traditional DPCM. The prediction coefficients are dynamically adjusted according to the slope of the first five sampling points, and the quantization step size is adaptively scaled according to the signal's local variance, dynamically maintaining the compression ratio between 4:1 and 8:1. The Huffman coding code table is dynamically generated based on the historical distribution frequency of environmental parameters and updated every 24 hours. Run-length encoding compresses the length of consecutive identical behavioral tags, triggering encoding only when the consecutive length exceeds 3. The compressed data packets are transmitted to the cloud analysis node via a Wi-Fi 6 protocol stack, using UDP over IPv6. The maximum data packet transmission unit is set to 1400 bytes, and the timeout retransmission mechanism is only enabled for first-priority data packets, with a maximum of 3 retransmissions.

[0046] The causal reasoning analysis engine layer is deployed in a cloud-based distributed computing cluster, comprising a multimodal data lake, a state transition graph neural network model, a dynamic causal discovery module, and an individualized baseline modeler. The multimodal data lake adopts a columnar storage architecture, with data tables sharded by user ID. Each shard includes a physiological data table, an environmental data table, a behavioral data table, and a semantic association table. Each table is linked by a 64-bit globally unique timestamp with a timestamp alignment precision of 1ms. The physiological data table stores fields including heart rate, systolic blood pressure, diastolic blood pressure, blood oxygen saturation, respiratory rate, body surface temperature, blood glucose level, and the corresponding device ID and sampling frequency. The environmental data table stores fields including indoor temperature, relative humidity, atmospheric pressure, light intensity, CO2 concentration, and the corresponding sensor ID. The behavioral data table stores fields including behavioral category, duration, spatial coordinates, and the corresponding radar ID and thermal imager ID. The semantic association table stores fields including data packet ID, semantic tag, anomaly flag, compression ratio, and transmission latency.

[0047] The state transition graph neural network model consists of a spatiotemporal graph convolutional network and a gated recurrent unit cascaded together. The input is a user-centric heterogeneous time series graph extracted from a multimodal data lake. In the graph, nodes represent observation events of different modalities, and edges represent the spatiotemporal proximity and semantic association strength between events.

[0048] The process of constructing a heterogeneous timing graph is as follows:

[0049] First, extract all sensor data points within a 5-minute time window;

[0050] Secondly, each data point is mapped to a graph node, and the node feature vector includes the original value, first-order difference, second-order difference, and the mode identifier.

[0051] Next, calculate the spatiotemporal distance between any two nodes. If the time difference is less than 300 seconds and the spatial distance is less than 2 meters, then establish an undirected edge with the edge weight being the similarity calculated by the Gaussian kernel function.

[0052] Finally, behavioral semantic labels are injected as global graph attributes. The spatiotemporal graph convolutional network consists of three graph convolutional modules, each employing a multi-head attention mechanism to aggregate neighbor node information, with 8 heads and a hidden layer dimension of 256. The gated recurrent unit (GRU) consists of two layers of GRUs, each with 512 hidden units, used to capture state evolution patterns across time windows. The model output is a health state vector for the current time window, with a dimension of 128, which is mapped to a clinical risk score through a fully connected layer. The score ranges from 0 to 100, with a threshold set at 75. Exceeding this threshold triggers the intervention generation process.

[0053] The dynamic causal discovery module employs a fusion framework of a constraint-based PC algorithm and a score-based GES algorithm. The input is a 7-day sequence of individualized health status vectors, and the output is a causal structure represented by a directed acyclic graph. The PC algorithm first performs a conditional independence test using a partial correlation coefficient as the test statistic, with a significance level set to 0.01 and a maximum condition set size of 5. The GES algorithm then performs a greedy equivalence class search on the skeleton graph output by the PC algorithm, using the Bayesian information criterion as the scoring function, with a search step size of 1000 iterations. In the final output causal graph, nodes represent physiological parameters or behavioral events, edges represent the direction of causal influence, and edge weights are standardized causal effect sizes. The individualized baseline modeler, based on the causal graph structure, uses a structural equation model to fit the functional relationships between variables. The model parameters are solved using maximum likelihood estimation, with a convergence threshold of 1e-6. The baseline model is automatically updated daily at 2 AM, excluding data points marked as outliers within the past 24 hours during the update.

[0054] The clinically interpretable intervention generation layer includes an intervention strategy knowledge base, a causal path interpreter, and a multi-channel instruction distributor. The intervention strategy knowledge base is constructed using the OWL ontology language and includes disease entities, symptom entities, drug entities, behavior entities, and intervention rule entities. Relationships between entities are categorized into four types: "cause," "relief," "contraindication," and "recommendation." Intervention rule entities are defined using the SWRL rule language. The antecedent of a rule is a logical combination of a clinical risk score and a causal graph structure, while the consequent is the specific intervention action. For example, the rule "IF Risk score > 85 AND 'Elevated blood glucose → Increased heart rate' causal edge exists AND 'Medication' behavior is missing THEN Push insulin injection reminder" is a valid intervention rule. After triggering an intervention rule, the causal path interpreter automatically generates natural language explanation text. The explanation text structure is: "[Abnormal parameter] detected, possibly caused by [upstream cause], it is recommended to execute [intervention action] to block [downstream consequence]."

[0055] The multi-channel command distributor selects the optimal distribution channel based on the user's device online status and interaction preferences, including smart speaker voice broadcast, smartphone APP push, family WeChat messages, and community doctor workstation pop-ups. The distribution priority decreases in that order, and the response timeout thresholds are 30 seconds, 2 minutes, 5 minutes, and 10 minutes, respectively.

[0056] At the system level, an adaptive power consumption control mechanism is implemented. This mechanism is achieved collaboratively by the edge node power budget allocator and the cloud computing load balancer. The edge node power budget allocator dynamically adjusts the operating mode of each sensor based on the remaining battery power and expected usage time. When the battery power is below 20%, the millimeter-wave radar module and infrared thermal imaging module are shut down, the sampling frequency of the physiological sensors is reduced to 50% of its original frequency, and the sampling frequency of the environmental sensors is reduced to 0.01Hz. When the battery power is below 10%, only the heart rate and blood oxygen sensors operate at a frequency of 1Hz, while the other sensors enter a sleep state. The cloud computing load balancer dynamically adjusts the inference batch size and parallelism of the state transition graph neural network model based on the cluster CPU utilization and memory usage. When the CPU utilization exceeds 80%, the batch size is reduced from 32 to 16, and the parallelism is reduced from 8 to 4. When the memory usage exceeds 90%, model parameter quantization is enabled to FP16 precision, and the causal discovery module computing tasks for non-urgent users are shut down.

[0057] At the data security level, end-to-end encryption and differential privacy protection mechanisms are implemented. End-to-end encryption uses the AES-256-GCM algorithm, with the key generated jointly by the user's biometrics and the device's hardware fingerprint, and the key update cycle is 7 days. Differential privacy protection is implemented during the individualized baseline modeling phase, adding Laplacian noise to the gradient values ​​used for parameter estimation. The noise scale is dynamically calculated based on sensitivity and privacy budget ε, with ε set to 0.5 to ensure that the impact of individual data points on model parameters is effectively masked. All data access operations are logged, with log entries including operator ID, operation time, operation type, accessed data range, and operation result. The logs are stored in an independent security domain and retained for 180 days.

[0058] At the user interaction level, a progressive guidance and negative feedback suppression mechanism is implemented. The progressive guidance mechanism is initiated upon the user's first use, gradually activating the sensor module in three stages:

[0059] The first phase involves only physiological sensors and lasts for 3 days.

[0060] The second phase involves adding environmental sensors and lasts for two days.

[0061] The third phase involves activating the behavior recognition module, which will last for one day.

[0062] At the end of each phase, the system automatically generates a usage report and provides feedback to the user via voice broadcast. The report includes data collection completeness rate, number of abnormal events, and estimated device battery life. A negative feedback suppression mechanism is triggered when the system detects that the user has not worn the device for three consecutive days or has manually turned off the sensors more than five times. This automatically reduces the data collection frequency to the minimum guaranteed level, pushes a simplified health report, and sends a device usage anomaly alert to family members. Once the user resumes wearing the device and uses it normally for more than seven consecutive days, the system automatically reverts to standard data collection mode.

[0063] The clinical validation layer incorporates a built-in efficacy assessment module. After each intervention instruction is executed, this module marks the intervention time point and extracts health status vector sequences for 72 hours before and after the intervention. It then calculates the differences in changes in key physiological parameters between the intervention group and the control group. Control group data comes from a similar user group that did not receive the same intervention; similarity is measured by cosine similarity with a threshold of 0.85. Assessment indicators include the magnitude of systolic blood pressure reduction, the rate of reduction in blood glucose variability, and the reduction in nighttime awakenings, with a statistical significance level set at 0.05. The assessment results automatically generate a clinical research report, including a description of the intervention, sample size, effect size, confidence interval, and p-value, and are pushed to the attending physician's end.

[0064] The system expansion layer supports a modular plug-in architecture, allowing third-party developers to connect new sensor types or analysis algorithms through standardized API interfaces. These standardized API interfaces include a data access interface, a model registration interface, and a result callback interface. The data access interface requires new sensors to provide data format description files and drivers; the data format description files must declare the sampling frequency, data type, unit, and precision. The model registration interface requires new algorithms to provide model weight files and inference scripts; the inference scripts must implement predefined input / output specifications. The result callback interface is used to receive analysis results, and the callback function must return an acknowledgment response within 500ms. All plug-ins must pass sandbox environment testing before integration. Testing includes resource utilization, inference latency, output stability, and security scanning; a pass rate of 99.9% is required for deployment.

[0065] In terms of deployment, it supports three modes: private cloud, public cloud, and hybrid cloud. In private cloud mode, all computing nodes are deployed in the medical institution's local data center, ensuring data stays within the hospital and requiring network latency of less than 50ms. In public cloud mode, computing nodes are deployed on a carrier-grade cloud platform, supporting elastic scaling with a minimum of 3 instances and a maximum of 100 instances. In hybrid cloud mode, edge nodes and transmission middleware are deployed locally, while the analysis engine and intervention generation layer are deployed in the cloud, connected via a VPN dedicated line with a guaranteed bandwidth of no less than 100Mbps. System performance metrics remain consistent across different deployment modes, including end-to-end latency of less than 2 seconds, data loss rate of less than 0.1%, and intervention command accuracy of more than 95%.

[0066] At the operation and maintenance monitoring level, a full-link health assessment mechanism is implemented, encompassing four dimensions: edge node health, transmission link health, analytics engine health, and intervention channel health. Edge node health is calculated using a weighted score based on battery power, sensor online rate, and local cache utilization, with weights of 0.4, 0.3, and 0.3, respectively. Transmission link health is calculated using a weighted score based on packet loss rate, latency jitter, and bandwidth utilization, with weights of 0.5, 0.3, and 0.2, respectively. Analytics engine health is calculated using a weighted score based on CPU utilization, memory utilization, and model inference accuracy, with weights of 0.4, 0.3, and 0.3, respectively. Intervention channel health is calculated using a weighted score based on command delivery rate, user confirmation rate, and family member response rate, with weights of 0.5, 0.3, and 0.2, respectively. A yellow alert is triggered when the score in any dimension falls below 80, and a red alert is triggered when the score falls below 60. Alert information is pushed to the operation and maintenance console in real time, and a fault diagnosis report is automatically generated.

[0067] To control costs, a dynamic resource recycling mechanism is implemented. This mechanism automatically releases the cloud computing instance and storage space allocated to a user when no valid data is uploaded for 30 consecutive days, retaining only metadata and historical reports. When a user reactivates their device, the system rebuilds the computing environment and restores access to historical data within two hours. For inactive users, the system automatically reduces their data sampling frequency to 25% of the standard value and schedules analysis tasks to a low-priority computing queue, ensuring that the service quality for highly active users is not affected.

[0068] To enhance anti-interference capabilities, a multi-source data cross-validation mechanism is implemented. When an anomaly is detected in data from a single sensor, this mechanism automatically retrieves data from other modalities for corroboration. For example, when the heart rate sensor reports a heart rate exceeding 120 bpm, the system automatically checks accelerometer data within the same time window. If the accelerometer indicates the user is "walking" or "standing," it is determined to be physiological tachycardia, and no alarm is triggered. If the accelerometer indicates the user is "sitting" or "lying down," it combines this with ambient temperature data. If the ambient temperature is above 30°C, it is determined to be an environmental stress response, generating a hydration reminder. If the ambient temperature is normal, a high-priority alarm is triggered. The cross-validation rule base includes 128 predefined rules, covering common physiological parameter anomaly scenarios.

[0069] In the long-term adaptive phase, a dynamic evolution mechanism for user profiles is implemented. This mechanism updates the user's health feature vector monthly. The feature vector has a dimension of 64 and includes the mean of basic physiological parameters, environmental tolerance threshold, behavioral pattern entropy, and drug response slope. The feature vector update employs an incremental learning algorithm with a learning rate set to 0.01 to ensure the model can adapt to the slow drift of the user's physiological state. When the monthly change in the feature vector exceeds a preset threshold (Euclidean distance greater than 2.0), the system automatically initiates a baseline model recalibration process and sends a notification of significant changes in health status to the attending physician.

[0070] At the emergency response level, a multi-level circuit breaker mechanism is implemented, including edge-level, transmission-level, and analysis-level circuit breakers. Edge-level circuit breakers are triggered when the local semantic fusion processor detects five consecutive data packet CRC check failures, automatically restarting the sensor driver and switching to a backup communication channel. Transmission-level circuit breakers are triggered when the dynamic bandwidth allocator detects ten consecutive data packet transmission timeouts and retransmission failures, automatically reducing the data compression ratio and enabling data fragmentation transmission. Analysis-level circuit breakers are triggered when the causal inference analysis engine detects three consecutive model inference timeouts or abnormal outputs, automatically rolling back to the previous stable model version and enabling the rule engine as a temporary alternative. The status of each circuit breaker level is synchronized to the operations and maintenance console in real time and automatically recovers after the fault is resolved.

[0071] At the human-machine collaboration level, a closed-loop doctor feedback mechanism is implemented. This mechanism allows attending physicians to confirm, modify, or reject intervention instructions generated by the system. All operations are recorded in a blockchain-based evidence storage system, with a block generation cycle of 10 minutes and the consensus algorithm using PBFT. Modified intervention instructions from physicians are automatically injected back into the intervention strategy knowledge base to optimize subsequent rule generation. The system compiles monthly statistics on physician adoption rate, modification rate, and rejection rate, generating a model optimization suggestion report to guide the algorithm team in adjusting model parameters and rule thresholds.

[0072] A federated learning architecture is implemented at the cross-institutional collaboration level. This architecture achieves collaborative optimization of model parameters while protecting the data privacy of each medical institution. Each participating institution trains its state transition graph neural network model locally, encrypting only the model gradients before uploading them to the central aggregation server. The aggregation algorithm employs a secure multi-party computation protocol to ensure that gradient information cannot be back-engineered to the original data. The central server distributes the aggregated global model every 24 hours, and each institution fine-tunes its model locally with a learning rate set to 0.001. Participating institutions in the federated learning process must meet the following criteria: a dataset of more than 1000 cases and a labeling accuracy rate higher than 90%.

[0073] To ensure sustainability, a green computing strategy is implemented. This strategy employs dynamic voltage and frequency adjustment technology at the cloud analytics engine layer, adjusting the CPU operating frequency based on real-time load, with a frequency adjustment range of 50% to 100% of the base frequency. At the edge node layer, a solar-assisted power supply module is used. This module consists of a flexible photovoltaic thin film and a supercapacitor. The photovoltaic thin film has a conversion efficiency of 22%, and the supercapacitor has a capacity of 500F, providing an additional 4 hours of power per day under standard lighting conditions. The system automatically records daily energy consumption data, generates a carbon footprint report, and pushes it to the organization's management.

[0074] At the traceability level, a full lifecycle data lineage tracking mechanism is implemented. This mechanism records the complete processing path from raw sensor data to the final intervention command, including data acquisition time, preprocessing steps, transmission nodes, analysis model version, intervention rule ID, and execution results. Lineage information is stored in a graph database, supporting multi-dimensional queries by time, user, device, parameter, etc. Any data anomaly can be traced back to the specific processing stage, facilitating fault location and responsibility determination.

[0075] At the standardization level, the system adheres to the HL7 FHIR R4 medical data exchange standard and the IEEE 11073-10400 series of personal health device communication protocols. All external interfaces provide RESTful APIs compliant with the FHIR standard, with resource types including Patient, Observation, Device, and CarePlan. Sensor data formats comply with the IEEE 11073-10400 specification, ensuring interoperability with third-party medical devices. The system undergoes regular compliance audits by third-party certification bodies, covering areas such as data security, privacy protection, clinical effectiveness, and system stability.

[0076] A multi-layered defense system is implemented to mitigate attacks. This system includes network-layer DDoS protection, application-layer API authentication, data-layer encrypted storage, and model-layer adversarial sample detection. DDoS protection employs traffic scrubbing and rate limiting strategies, with a scrubbing threshold of 1000 QPS. API authentication uses the OAuth 2.0 protocol, with tokens valid for one hour. Encrypted storage uses the AES-256 algorithm, with keys managed by a hardware security module. Adversarial sample detection is performed before model inference, using a gradient mask-based input preprocessing algorithm. The detection threshold is set to reject inference if the L2 norm of the input perturbation is greater than 0.1. All security events are logged in an independent log system and retained for 365 days.

[0077] At the user experience level, an adaptive interface rendering mechanism is implemented, dynamically adjusting the interactive interface based on the user's vision, hearing, and operational abilities. For users with declining vision, the font size is automatically enlarged to 24pt, and a high-contrast color scheme is enabled; for users with hearing loss, the voice broadcast volume is increased to 85dB, and vibration alerts are enabled; for users with declining operational abilities, interface elements are simplified, and voice control and gesture recognition are enabled. Interface adaptation parameters are automatically updated based on the user's monthly health assessment report to ensure that the interaction method matches their physiological capabilities.

[0078] At the research support level, an anonymized data open interface is implemented. After obtaining user authorization, this interface provides desensitized health datasets for research institutions. Data desensitization employs a combination of k-anonymity and l-diversity algorithms, with k set to 50 and l set to 10, ensuring that individuals cannot be re-identified. The open dataset includes time-aligned multimodal observation sequences and labeled clinical events, supporting downloads in both CSV and Parquet formats. Research institutions are required to sign a data usage agreement, committing to use the data solely for non-commercial research purposes.

[0079] At the technological evolution level, a quantum-safe communication interface is reserved. This interface supports post-quantum cryptography algorithms, including the lattice-based CRYSTALS-Kyber key encapsulation mechanism and the hash-based SPHINCS+ signature scheme. The interface is in an inactive state; when a quantum computing threat is detected to the traditional cryptographic system, the system will switch algorithms within 72 hours to ensure long-term data security. The quantum-safe module has passed the third round of evaluation by the NIST post-quantum cryptography standardization project and meets Level 3 security requirements.

[0080] A tiered early warning and response mechanism is implemented at the community level, categorizing intervention instructions into community, street, and district / county levels based on risk level. Community-level warnings are handled by family doctors with a response time of 2 hours; street-level warnings are handled by community health service centers with a response time of 1 hour; and district / county-level warnings are handled by regional medical centers with a response time of 30 minutes. Warning information is synchronized to all levels of response units via the government intranet to ensure rapid coordination. The system evaluates response timeliness and effectiveness monthly, optimizing the early warning tiering standards accordingly.

[0081] In end-of-life care, a comfort-first mode is implemented. This mode is automatically activated when a user enters the terminal stage of their illness, reducing data collection frequency to the minimum necessary level, disabling non-emergency alarms, and prioritizing the user's rest and dignity. Intervention strategies shift towards symptom relief and psychological support, including pain management reminders, access to palliative care resources, and suggestions for family companionship. Activation of the mode requires confirmation from both the attending physician and the family to ensure it aligns with the patient's wishes.

[0082] The historical data utilization module implements a longitudinal trend analysis module, which retrospectively analyzes users' health data over five years to identify chronic disease progression patterns, treatment response trajectories, and the cumulative effects of risk factors. The analysis employs time series decomposition and survival analysis methods to output personalized disease prediction curves and intervention window recommendations. Access to historical data requires secondary authorization from the user to ensure informed consent. The analysis results assist physicians in developing long-term management plans and improve treatment foresight.

[0083] In terms of technology accessibility, low-power wide-area network (LPWAN) support is implemented. LPWAN supports NB-IoT and LoRaWAN protocols, suitable for remote areas and older communities. The NB-IoT module operates at 800MHz with a transmit power of 23dBm and standby power consumption of less than 1μA; the LoRaWAN module operates at 470MHz with a spreading factor of 12 and a communication range of up to 10 kilometers. The system automatically detects the network environment and prioritizes the LPWAN protocol with the best signal quality. Data transmission employs an integrated compression and encryption scheme to ensure security and efficiency under low bandwidth conditions.

[0084] To verify the effectiveness of the present invention, Table 1 provides a set of comparative experimental data of embodiments and comparative examples. The experimental subjects were elderly patients aged 65 and above with type 2 diabetes and hypertension, with a sample size of 200 people. They were randomly divided into the implementation group (using the platform of the present invention) and the control group (using the traditional remote monitoring system), and the observation period was 6 months.

[0085] Table 1 Comparative Experimental Data

[0086] Evaluation indicators Implementation Group (This Invention) Control group (traditional system) Increase p-value Average daily data completeness 98.7% 82.3% +16.4% <0.001 Average response time for high-risk events 42 seconds 3 minutes and 18 seconds -78.9% <0.001 Average monthly device usage time per user 21.5 hours / day 14.2 hours / day +51.4% <0.01 Physician intervention instruction adoption rate 93.6% 68.4% +25.2% <0.001 System average power consumption (edge ​​node) 0.85W 2.10W -59.5% <0.001 False alarm rate (non-clinically relevant alerts) 3.2% 27.8% -88.5% <0.001 Systolic blood pressure control compliance rate (<140) 89.1% 65.3% +23.8% <0.01 HbA1c reduction (mmol / mol) -8.7 -3.2 +172% <0.05 Nocturnal hypoglycemia event incidence 0.8 times / month 3.5 times / month -77.1% <0.01 User satisfaction rating (out of 10) 9.2 6.1 +50.8% <0.001

[0087] Experimental results show that the platform of this invention significantly outperforms traditional solutions in terms of data integrity, response speed, user compliance, clinical efficacy, and system efficiency. Particularly in terms of false alarm rate control and power consumption optimization, thanks to the behavioral semantic annotation mechanism of the semantic-aware edge node layer and the priority scheduling strategy of the spatiotemporally aligned transmission middleware layer, the system significantly reduces invalid alarms and energy consumption while maintaining clinical sensitivity. The state transition graph neural network model and dynamic causal discovery module of the causal reasoning analysis engine layer effectively improve the accuracy and timeliness of risk prediction, while the natural language interpretation mechanism of the clinically explainable intervention generation layer significantly enhances users' trust in the system's decisions and their willingness to cooperate.

[0088] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0089] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A health data collection and analysis platform for elderly chronic disease management, characterized in that, The application relates to a health monitoring system for the elderly, which comprises a semantic perception edge node layer, a space-time alignment transmission middleware layer, a causal reasoning analysis engine layer and a clinically interpretable intervention generation layer. The semantic perception edge node layer is arranged in the daily activity space of an elderly user and comprises a physiological sensing unit, an environment sensing unit, a behavior recognition unit and a local semantic fusion processor. The space-time alignment transmission middleware layer is arranged in a home gateway device and comprises a semantic data receiving buffer, a dynamic bandwidth allocator, a semantic priority scheduler and an adaptive compression encoder. The causal reasoning analysis engine layer is arranged in a cloud distributed computing cluster and comprises a multi-modal data lake, a state transition graph neural network model, a dynamic causal discovery module and an individualized baseline modeler. The clinically interpretable intervention generation layer comprises an intervention strategy knowledge base, a causal path interpreter and a multi-channel instruction distributor. The physiological sensing unit is composed of an optical plethysmography sensor, an impedance respiration rate sensor, a thermocouple body temperature sensor and a three-axis accelerometer, and is connected to the local semantic fusion processor through an I2C bus. The environment sensing unit is composed of a digital temperature and humidity sensor, an atmospheric pressure sensor, an illumination intensity sensor and a carbon dioxide concentration sensor, and is connected to the local semantic fusion processor through an SPI bus. The behavior recognition unit is composed of a millimeter wave radar module and a low-resolution infrared thermal imaging module, and is connected to the local semantic fusion processor through a parallel LVDS interface. The semantic priority scheduler divides data packets into four priority queues according to the behavior semantic labels and physiological parameter abnormal flag bits embedded in the data packets. The first priority is for data packets comprising a "medicine taking" behavior label and having a blood glucose value lower than 3.9 mmol / L or higher than 13.9 mmol / L. The second priority is for data packets comprising a "defecation" behavior label and having a heart rate variability lower than 20 ms. The third priority is for data packets comprising an "unknown behavior" label and having continuous buffering for more than 3 cycles. The fourth priority is for the rest of the conventional monitoring data packets. Each queue adopts a weighted round-robin scheduling algorithm, and the weight coefficients are 8, 4, 2 and 1 respectively. The adaptive compression encoder adopts a hybrid compression strategy, adopts an improved differential pulse code modulation algorithm for physiological signal data, adopts Huffman coding for environment data and adopts run-length coding for behavior semantic labels. The dynamic causal discovery module adopts a constraint-based PC algorithm and a score-based GES algorithm fusion framework, takes a 7-day individualized health state vector sequence as input and outputs a causal structure in the form of a directed acyclic graph. The GES algorithm performs a greedy equivalence class search on the skeleton graph output by the PC algorithm, adopts a Bayesian information criterion as a score function and performs 1000 iterations as a search step. The individualized baseline modeler adopts a structural equation model to fit the functional relationship between variables based on the causal graph structure, solves model parameters through maximum likelihood estimation, takes 1e-6 as a convergence threshold, automatically updates at 2 o'clock in the morning every day and excludes data points marked as abnormal in the past 24 hours during the updating.

2. The platform of claim 1, wherein, The multi-modal data lake adopts a columnar storage architecture, and data tables are sharded by user ID, each shard including a physiological data table, an environmental data table, a behavior data table, and a semantic association table, each table being associated by a 64-bit global unique timestamp with an alignment accuracy of 1 ms.

3. The platform of claim 1, wherein, The intervention strategy knowledge base is constructed using the OWL ontology language, and includes disease entities, symptom entities, drug entities, behavior entities, and intervention rule entities; the intervention rule entities are defined using the SWRL rule language, the rule antecedent is a logical combination of clinical risk scores and causal graph structures, and the rule consequent is a specific intervention action; The causal path interpreter automatically generates natural language explanation text after triggering the intervention rule; The multi-channel instruction distributor selects a smart speaker voice broadcast, a smartphone APP push, a family WeChat message, and a community doctor workstation pop-up window as the distribution channels according to the online state and interaction preferences of the user equipment, and the distribution priority decreases in turn, and the response timeout thresholds are 30 seconds, 2 minutes, 5 minutes, and 10 minutes, respectively.

Images