Artificial intelligence-based chronic disease health management system and method

Abstract

The present application relates to the technical field of artificial intelligence health management, in particular to a chronic disease health management system and method based on artificial intelligence, comprising: receiving health monitoring streams of heterogeneous sensors and user input, generating spatiotemporal aligned multi-modal health data streams through analysis and alignment; performing asynchronous feature fusion to obtain individual health feature tensors, inputting dynamic decision forest for iterative evaluation of chronic disease state trajectory and outputting evaluation index; activating personalized monitoring protocol according to the evaluation index, adjusting sensor acquisition mode to obtain directional enhancement data, performing hierarchical anomaly decomposition and progressive pattern matching to generate a comprehensive risk assessment spectrum with time window annotation, and generating a health management execution list containing specific instructions after analysis. The method can realize accurate fusion of multi-modal health data and dynamic evaluation of chronic disease state, adapt to personalized monitoring, accurately identify health deviation, and output implementable health management instructions.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence health management technology, and in particular to an artificial intelligence-based chronic disease health management system and method. Background Technology

[0002] Current chronic disease health management relies on collecting physiological data from a single type of sensor, combining it with basic user behavior input to form health monitoring data, processing the data through conventional feature extraction methods, using static assessment models to determine chronic disease health status, acquiring data with fixed sensor collection parameters, identifying health abnormalities based on single-layer anomaly detection rules, completing health risk assessment by simple comparison with historical data, and finally outputting routine health management recommendations.

[0003] Current technologies cannot perform spatiotemporal alignment processing of data collected by heterogeneous sensors with unstructured physiological data and behavioral record sequences input by users. Multimodal health data can only be simply stitched together, failing to adapt to the asynchronous characteristics of heterogeneous sensors with different sampling rates. This results in delays or loss of information fusion, and asynchronous feature fusion processing cannot be completed. Existing assessment models are mostly static in structure, with fixed model weights, unable to be dynamically adjusted based on the data quality of individual patients and the model's recent performance. They lack adaptability to individual differences and cannot perform iterative trajectory assessment for chronic disease states. At the anomaly identification level, existing methods are mostly performed on a single time scale, making it difficult to simultaneously and effectively capture instantaneous abnormal events and long-term trend shifts. Furthermore, the lack of cross-level anomaly verification mechanisms leads to a high false alarm rate. In addition, health risk assessment results are mostly static risk probability values, lacking prediction of the time dimension of risk occurrence. They cannot clearly indicate when the risk may increase, making it difficult to grasp the timing of intervention. Health management measures often lag behind actual changes in health status. The perception link in the entire management closed loop is rigid. The sensor acquisition mode remains fixed and cannot be adaptively adjusted based on the health assessment results. Health abnormalities can only be detected at a single layer and cannot be deconstructed at different levels. The offset pattern matching has no progressive logic and cannot generate risk assessment results with time window annotations. Ultimately, it is impossible to generate health management execution content containing specific instructions.

[0004] To address the issues that multimodal health data cannot achieve spatiotemporal alignment and asynchronous feature fusion, and cannot use dynamic models to iteratively assess the trajectory of chronic disease status, as well as the issues that cannot activate personalized monitoring protocols to adjust the acquisition mode based on assessment results, and cannot generate a comprehensive risk assessment spectrum with time window annotations through hierarchical anomaly deconstruction and progressive pattern matching. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an artificial intelligence-based chronic disease health management system and method.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a chronic disease health management method based on artificial intelligence, comprising:

[0007] The system receives health monitoring streams from heterogeneous sensors and user input, the health monitoring streams including unstructured physiological data and behavioral record sequences, parses and aligns the health monitoring streams, and generates a spatiotemporally aligned multimodal health data stream.

[0008] Asynchronous feature fusion processing is performed on the spatiotemporally aligned multimodal health data stream to generate a fused individual health feature tensor. The individual health feature tensor is then input into a dynamic decision forest to iteratively evaluate the chronic disease state trajectory and output the state trajectory evaluation index.

[0009] Based on the state trajectory evaluation index, the personalized monitoring protocol is activated to guide the heterogeneous sensors to adjust their acquisition modes and obtain directional enhanced monitoring data streams;

[0010] The targeted enhanced monitoring data stream is subjected to hierarchical anomaly deconstruction to identify potential health shift patterns at different levels, forming a hierarchical shift pattern set;

[0011] The hierarchical offset pattern set is progressively matched with the historical offset pattern library to generate a comprehensive risk assessment spectrum with time window annotations.

[0012] The comprehensive risk assessment spectrum is analyzed to generate a health management execution checklist containing specific instructions.

[0013] As a further aspect of the present invention, a health monitoring stream is received from heterogeneous sensors and user input, the health monitoring stream including unstructured physiological data and behavioral recording sequences. The health monitoring stream is parsed and aligned to generate a spatiotemporally aligned multimodal health data stream, including:

[0014] Establish a continuous data connection channel with smart wearable devices, home medical devices, and mobile terminal applications, and receive continuously incoming raw data packets through the data connection channel to form the health monitoring stream;

[0015] The health monitoring stream is analyzed in real time to distinguish data sources from different devices and applications, and the unstructured physiological data and the behavior record sequence are extracted. The unstructured physiological data includes time-series waveform fragments, discrete numerical readings, and text log summaries. The behavior record sequence includes dietary events, exercise events, medication confirmation events triggered by the user and their corresponding timestamps.

[0016] Maintain a local logical clock for each independent data source and establish a mapping relationship with the global standard time. Use the mapping relationship to perform timestamp alignment correction on all extracted data points.

[0017] After completing the timestamp alignment correction, the unstructured physiological data belonging to the same short time domain window are spatially associated with the behavioral record sequence, and a context label is attached to each associated data point to form the spatiotemporally aligned multimodal health data stream.

[0018] As a further aspect of the present invention, asynchronous feature fusion processing is performed on the spatiotemporally aligned multimodal health data stream to generate a fused individual health feature tensor, including:

[0019] Independent feature extraction pipelines are assigned to different types of data in the spatiotemporally aligned multimodal health data stream, and the feature extraction pipelines include waveform feature extraction pipelines, numerical sequence feature extraction pipelines, and behavioral event feature extraction pipelines.

[0020] In the waveform feature extraction pipeline, multi-scale convolution kernels are applied to time-series waveform segments to perform feature scanning and capture waveform morphology features;

[0021] In the numerical sequence feature extraction pipeline, the sliding window statistics of discrete numerical readings are calculated, and the slope of their changing trend is tracked to form numerical statistical features;

[0022] In the behavioral event feature extraction pipeline, events in the behavioral record sequence are encoded, and the density and regularity of event occurrence are calculated to form behavioral pattern features;

[0023] An asynchronous feature fusion engine is set up, which allows the waveform feature extraction pipeline, the numerical sequence feature extraction pipeline, and the behavioral event feature extraction pipeline to run at different processing speeds, and submits the output waveform morphology features, numerical statistical features, and behavioral pattern features to a shared feature buffer after each has completed feature extraction.

[0024] When the number of feature vectors accumulated in the shared feature buffer reaches the preset batch size, the asynchronous feature fusion engine performs concatenation and normalization operations on the feature dimensions of the feature vectors to generate the fused individual health feature tensor.

[0025] As a further aspect of the present invention, the individual health feature tensor is input into a dynamic decision forest to iteratively evaluate the chronic disease state trajectory and output a state trajectory evaluation index, including:

[0026] The dynamic decision forest is constructed, which consists of multiple growable decision trees. The node splitting rules of each decision tree are based on chronic disease state transition patterns learned from historical health data.

[0027] The individual health feature tensor is used as input and fed into each decision tree in the dynamic decision forest in parallel.

[0028] Within each decision tree, the individual health feature tensor starts from the root node and is passed down the branches of the tree according to the node splitting rules until it reaches a leaf node. Each leaf node is associated with a predefined state trajectory score.

[0029] Collect the state trajectory scores output by all decision tree leaf nodes, and perform weighted aggregation on the state trajectory scores. The aggregation weight is dynamically adjusted according to the confidence level of each decision tree in the recent evaluation.

[0030] The weighted aggregated values ​​are mapped to a standardized discrete index range, and the state trajectory evaluation index is output.

[0031] As a further aspect of the present invention, based on the state trajectory evaluation index, a personalized monitoring protocol is activated to guide heterogeneous sensors to adjust their acquisition modes and obtain directional enhanced monitoring data streams, including:

[0032] Multiple monitoring protocol templates are predefined, each bound to a specific state trajectory evaluation index interval, and the combination of sensor type, acquisition frequency, and duration is specified;

[0033] The state trajectory evaluation index is matched with a predefined index range to determine the active monitoring protocol template;

[0034] Based on the determined monitoring protocol template, a protocol execution script containing specific control instructions is generated, and the protocol execution script is sent to the corresponding smart wearable device and home medical device through the data connection channel;

[0035] The smart wearable device and home medical device execute scripts according to the received protocol, adjust the sampling frequency and working cycle of their built-in sensors or enable additional monitoring functions, and operate according to the new acquisition mode.

[0036] During the execution of the agreement, more intensive, focused, or multi-dimensional monitoring data is received from the adjusted smart wearable devices and home medical devices, forming the targeted enhanced monitoring data stream.

[0037] As a further aspect of the present invention, the targeted augmentation monitoring data stream is subjected to hierarchical anomaly deconstruction to identify potential health shift patterns at different levels, forming a hierarchical shift pattern set, including:

[0038] Define multiple analysis levels, including the immediate vital signs level, the daily fluctuation level, and the long-term trend level;

[0039] At the immediate vital signs level, high-frequency physiological signals in the targeted enhanced monitoring data stream are scanned in real time to detect instantaneous abnormal peaks or valleys that exceed the individualized adaptive threshold, and their amplitude, duration and morphology are recorded as immediate abnormal segments.

[0040] At the diurnal fluctuation level, the index curves aggregated by day in the targeted enhanced monitoring data stream are analyzed to identify morphological distortions, rhythm disorders, or phase shifts that occur compared with the individual historical baseline daily fluctuation curves, which are then used as diurnal fluctuation patterns.

[0041] At the long-term trend level, the indicator sequence spanning multiple days in the targeted enhanced monitoring data stream is fitted with a trend, the slope and curvature changes of the fitted curve are calculated, and it is compared with the expected healthy evolution trend line to identify deviations from the trend as long-term offset trends.

[0042] The instantaneous abnormal fragments, the intraday fluctuation patterns, and the long-term offset trends are structured and encapsulated to form the hierarchical offset pattern set.

[0043] As a further aspect of the present invention, the hierarchical migration pattern set is progressively matched with the historical migration pattern library to generate a comprehensive risk assessment spectrum with time window annotations, including:

[0044] The historical migration pattern library stores typical migration patterns mined and labeled from group historical data, along with their corresponding risk levels and evolution paths.

[0045] A progressive matching process is set up to sequentially calculate the similarity between the real-time vital signs level pattern, the daily fluctuation level pattern, and the long-term trend level pattern in the hierarchical offset pattern set and the typical offset patterns of the corresponding level in the historical offset pattern library.

[0046] For each level, several typical offset patterns with similarity exceeding the matching threshold are selected as candidate matching patterns.

[0047] Analyze the risk events and their time intervals that subsequently occurred in historical cases for the candidate matching patterns, and statistically obtain the risk probability distribution and typical time window distribution;

[0048] The matching results from the three levels are fused, the risk probabilities are superimposed and calibrated, and the intersection of the time windows is calculated to generate a multi-dimensional vector. The multi-dimensional vector describes the probability of different risk levels occurring in different time windows in the future, which is the comprehensive risk assessment spectrum with time window annotation.

[0049] As a further aspect of the present invention, the comprehensive risk assessment spectrum is analyzed to generate a health management execution list containing specific instructions, including:

[0050] Establish an intervention knowledge graph, where nodes represent different health risk states and situations, and edges represent feasible intervention actions and their expected impacts;

[0051] Using the highest priority risk indicated in the comprehensive risk assessment spectrum and its corresponding time window as the starting point for querying, a graph traversal is performed in the intervention measure knowledge graph.

[0052] During the traversal, an optimal sequence of intervention actions is selected based on the risk status nodes, the correlation strength of available intervention actions, the execution cost of the actions, and the expected effect path.

[0053] Each action in the optimal intervention sequence is transformed into a specific instruction that can be executed by the user or the linked device. The specific instructions include generating medication adjustment suggestions, performing specific rehabilitation exercises, scheduling a specialist appointment, and contacting a health manager. An execution time point and triggering conditions are assigned to each instruction.

[0054] All the generated specific instructions are arranged in chronological order of execution to form the health management execution list.

[0055] As a further aspect of the present invention, the method further includes:

[0056] The health management execution checklist is pushed to the user terminal and displayed visually in a timeline view on the user terminal interface;

[0057] Receive user confirmation of execution, modification feedback, or reports of execution difficulties for instructions in the list through the user terminal;

[0058] When the preset triggering conditions of the instruction are met or the execution time is reached, a reminder notification is triggered through the user terminal.

[0059] As a further aspect of the present invention, the present invention also includes an artificial intelligence-based chronic disease health management system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the artificial intelligence-based chronic disease health management method described above.

[0060] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0061] The system performs parsing and alignment operations on health monitoring streams from heterogeneous sensors and user input to form a spatiotemporally aligned multimodal health data stream. Asynchronous feature fusion processing is then performed on this data stream to generate individual health feature tensors. These individual health feature tensors are input into a dynamic decision forest, which iteratively completes the assessment of chronic disease state trajectories and outputs a state trajectory assessment index. This approach can adapt to the temporal feature differences between unstructured physiological data and behavioral record sequences, achieving accurate feature integration of multimodal health data. The dynamic decision forest can continuously iterate and calculate in accordance with changes in the input data, conforming to the dynamic evolution of chronic disease states. The output state trajectory assessment index can fully reflect the temporal change characteristics of chronic disease states.

[0062] Based on the state trajectory assessment index, a personalized monitoring protocol is activated to guide heterogeneous sensors to adjust their acquisition modes and acquire directional augmentation monitoring data streams. The directional augmentation monitoring data streams are then subjected to hierarchical anomaly deconstruction to identify potential health deviation patterns at different levels and form a hierarchical deviation pattern set. This hierarchical deviation pattern set is then progressively matched with a historical deviation pattern library to generate a comprehensive risk assessment spectrum with time window annotations. After parsing the comprehensive risk assessment spectrum, a health management execution list containing specific instructions is generated. Sensor acquisition logic can be adjusted according to different assessment results. The directional augmentation monitoring data stream can focus on collecting data from key health dimensions. Hierarchical anomaly deconstruction can break down health deviation features at different levels. Progressive pattern matching can complete the deviation feature comparison layer by layer. Time window annotations can clearly define the specific time period corresponding to the risk. The generated execution list can directly form actionable health management instructions. Attached Figure Description

[0063] Figure 1 This is a flowchart of the artificial intelligence-based chronic disease health management method described in this invention;

[0064] Figure 2 A flowchart for receiving health monitoring streams and generating spatiotemporally aligned multimodal health data streams;

[0065] Figure 3 This is a flowchart for inputting individual health characteristic tensors into a dynamic decision forest to output a state trajectory evaluation index. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0067] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0068] See Figure 1 This invention provides an artificial intelligence-based method for chronic disease health management. The method includes: Executed by a computing system, the system receives health monitoring streams from heterogeneous sensors and user input. These streams include unstructured physiological data and behavioral record sequences. The system parses and aligns the health monitoring streams to generate a spatiotemporally aligned multimodal health data stream. Asynchronous feature fusion processing is performed on the spatiotemporally aligned multimodal health data stream to generate a fused individual health feature tensor. This individual health feature tensor is input into a dynamic decision forest to iteratively evaluate the chronic disease state trajectory and output a state trajectory evaluation index. Based on the state trajectory evaluation index, a personalized monitoring protocol is activated to guide the heterogeneous sensors to adjust their acquisition modes and acquire a targeted augmented monitoring data stream. The targeted augmented monitoring data stream undergoes hierarchical anomaly deconstruction to identify potential health deviation patterns at different levels, forming a hierarchical deviation pattern set. The hierarchical deviation pattern set is progressively matched with a historical deviation pattern library to generate a comprehensive risk assessment spectrum with time window annotations. This comprehensive risk assessment spectrum is parsed to generate a health management execution list containing specific instructions.

[0069] In one embodiment of the present invention, the specific process of receiving health monitoring streams from heterogeneous sensors and user input, and generating a spatiotemporally aligned multimodal health data stream includes, as described in [reference] Figure 2The system establishes a continuous data connection channel with smart wearable devices, home medical devices, and mobile terminal applications. Through this channel, it receives continuously incoming raw data packets, which constitute a health monitoring stream. The health monitoring stream is analyzed in real time to distinguish data sources from different devices and applications, extracting unstructured physiological data and behavioral record sequences. The unstructured physiological data includes time-series waveform fragments, discrete numerical readings, and text log summaries. The behavioral record sequences include user-triggered dietary events, exercise events, medication confirmation events, and their corresponding timestamps. The system maintains a local logical clock for each independent data source and establishes a mapping relationship with the global standard time. This mapping relationship is used to perform timestamp alignment correction on all extracted data points. After timestamp alignment correction, unstructured physiological data and behavioral record sequences belonging to the same short time domain window are spatially correlated, and context labels are attached to each correlated data point, ultimately forming a spatiotemporally aligned multimodal health data stream.

[0070] In practice, the process of receiving health monitoring streams from heterogeneous sensors and user input, and parsing and aligning these streams to generate a spatiotemporally aligned multimodal health data stream, is implemented through a health management platform module deployed on a cloud server. This health management platform module establishes continuous data connection channels with smart wearable devices, home medical devices, and mobile terminal applications. These continuous data connection channels are based on secure network communication protocols, such as a combination of authenticated WebSocket long connections and HTTPS polling, to maintain data links with various terminals. The health monitoring stream consists of continuously incoming raw data packets received through the continuous data connection channels. These raw data packets include device identifiers, raw sensor readings, user interface interaction event logs, and corresponding timestamps. In one example scenario, raw byte streams of ECG waveforms from a smart bracelet, systolic and diastolic blood pressure values ​​from a home blood pressure monitor, and a user-recorded lunch carbohydrate intake log from a mobile terminal application are simultaneously received as different raw data packets and merged into the health monitoring stream.

[0071] In practical implementation, the health management platform module performs real-time parsing of the health monitoring stream. The real-time parsing process activates a dedicated stream processing engine to differentiate data sources from different devices and applications. Based on the device identifier and application identifier in the header of the original data packet, the real-time parsing distributes the mixed data stream to the corresponding parser instance. The parser instance then extracts unstructured physiological data and behavioral record sequences from the payload of the original data packet according to a predefined data pattern. In one data comparison scenario, the unstructured physiological data extracted from the smart bracelet data packet is a 5-second time-series electrocardiogram waveform segment; the unstructured physiological data extracted from the home blood pressure monitor data packet is discrete numerical readings containing systolic blood pressure, diastolic blood pressure, and heart rate; and the unstructured physiological data extracted from the mobile terminal application logs might be a text log summary, such as "The user reported feeling slightly dizzy in the afternoon." The extraction of behavioral record sequences focuses on identifying structured events, such as identifying user-triggered events like "record breakfast" (classified as a diet event), "start walking training" (classified as an exercise event), and "confirm taking antihypertensive medication" (classified as a medication confirmation event) from the interaction logs of mobile terminal applications, and extracting the corresponding precise timestamp for each event.

[0072] In some embodiments, maintaining a local logical clock for each independent data source and establishing a mapping relationship with the global standard time is key to solving the time synchronization problem of heterogeneous data sources. The local logical clock is a timer maintained internally by the data source device or application. The global standard time is synchronized from an authoritative time server by the health management platform module. The process of establishing the mapping relationship involves the platform sending a time synchronization request packet containing the platform's current global standard time to each data source when it first connects to the platform. Upon receiving the request, the data source records its current local logical clock value and immediately replies. The platform calculates a time offset based on the time difference between sending and receiving the request, thus establishing a mapping relationship of "data source local logical clock value + time offset = platform global standard time". Using this mapping relationship, timestamp alignment correction is performed on all extracted data points, meaning that the original local logical timestamps carried by the data points are converted into a unified global standard timestamp using the above mapping formula. The mapping relationship can be expressed as the formula:

[0073]

[0074] in: This represents the corrected global standard timestamp. Indicates from the The original local logical timestamps of each data source, Represented as the first The time offset constant calculated from each data source This represents an estimated compensation value for network transmission latency, and is a configurable, small correction. It can be understood that after this correction, the timestamps of events recorded by different devices at the same physical moment will be highly consistent within the system.

[0075] In some embodiments, after completing timestamp alignment correction, the spatiotemporal alignment module spatially correlates unstructured physiological data and behavioral recording sequences belonging to the same short time-domain window. The short time-domain window is a configurable time interval, such as 5 minutes. The duration of the short time-domain window is designed to balance the ability to capture acute physiological changes with data processing efficiency. The window length needs to be short enough to correlate nearly simultaneous multimodal data, such as motion events and immediate heart rate changes, while not being too short to avoid invalid correlations due to minor time synchronization errors between devices or transient signal fluctuations. The spatial correlation operation, based on time alignment, clusters and links all data points from different data sources within the same time window according to their data themes. A context label is attached to each correlated data point, including the source device type, the associated physiological parameter type, the associated user behavior event type, and the time interval identifier of the correlation window. In one example, a user triggers a "start jogging" exercise event within the time window of "14:00-14:05". Within the same time window, heart rate and accelerometer waveforms collected by the smart bracelet, along with ambient temperature and humidity readings recorded by home environment sensors, are clustered and labeled with contextual tags such as "exercise period - ECG", "exercise period - exercise intensity", and "exercise period - environment". Ultimately, the ordered set of all data points, after timestamp correction, spatial association, and the addition of contextual tags, forms a spatiotemporally aligned multimodal health data stream. This spatiotemporally aligned multimodal health data stream provides standardized input that is precisely aligned in both time and semantic dimensions for subsequent feature extraction.

[0076] In one embodiment of the present invention, the specific process of performing asynchronous feature fusion processing on a spatiotemporally aligned multimodal health data stream to generate a fused individual health feature tensor includes assigning independent feature extraction pipelines to different types of data in the spatiotemporally aligned multimodal health data stream. These pipelines include a waveform feature extraction pipeline, a numerical sequence feature extraction pipeline, and a behavioral event feature extraction pipeline. In the waveform feature extraction pipeline, multi-scale convolutional kernels are applied to time-series waveform segments for feature scanning to capture waveform morphological features. The width of the multi-scale convolutional kernels is designed to cover morphological features at different time scales in the target physiological waveform. Narrower convolutional kernels help capture subtle local morphological features of the waveform, while wider convolutional kernels help perceive the overall rhythm and trend changes of the waveform. In the numerical sequence feature extraction pipeline, sliding window statistics are calculated for discrete numerical readings, and the slope of their changing trends is tracked to form numerical statistical features. In the behavioral event feature extraction pipeline, events in the behavioral recording sequence are encoded, and the density and regularity measures of event occurrence are calculated to form behavioral pattern features. The system employs an asynchronous feature fusion engine. This engine allows waveform feature extraction pipelines, numerical sequence feature extraction pipelines, and behavioral event feature extraction pipelines to run at different processing speeds. After each pipeline completes its feature extraction, it submits the output waveform morphology features, numerical statistical features, and behavioral pattern features to a shared feature buffer. When the number of feature vectors accumulated in the shared feature buffer reaches a preset batch size, the asynchronous feature fusion engine concatenates and normalizes these feature vectors along the feature dimension to generate a fused individual health feature tensor.

[0077] In practical implementation, the process of asynchronous feature fusion processing on spatiotemporally aligned multimodal health data streams to generate fused individual health feature tensors is completed by the feature computation service module. This module allocates independent feature extraction pipelines to different types of data in the spatiotemporally aligned multimodal health data streams. The waveform feature extraction pipeline, numerical sequence feature extraction pipeline, and behavioral event feature extraction pipeline are deployed as independent microservices, receiving data slices of the corresponding types from the data stream in parallel. In one example scenario, the waveform feature extraction pipeline receives ECG waveform segments and photoplethysmography (PPG) waveform segments from the spatiotemporally aligned multimodal health data stream; the numerical sequence feature extraction pipeline receives discrete numerical reading sequences such as systolic blood pressure, diastolic blood pressure, heart rate, and blood glucose; and the behavioral event feature extraction pipeline receives medication confirmation event sequences, dietary event sequences, and exercise event sequences. Each feature extraction pipeline has independent computing resources and processing logic, and they do not block each other.

[0078] In practical implementation, the waveform feature extraction pipeline applies multi-scale convolutional kernels to time-series waveform segments for feature scanning. Multi-scale convolutional kernels refer to a set of one-dimensional convolutional kernels with different widths, used to capture waveform morphological features at different time scales. In one example, for a 5-second ECG waveform, the waveform feature extraction pipeline simultaneously applies three one-dimensional convolutional kernels with widths of 30, 100, and 300 sampling points. The output of each convolutional kernel, after pooling, generates a set of feature vectors representing the local morphology and rhythm information of the waveform; these feature vectors together constitute the waveform morphological features. The calculation process of waveform morphological features can be formally described as follows: for a waveform signal... , No. The output features of each convolution kernel Given by the formula:

[0079]

[0080] in: It is the first The width of each convolutional kernel, It is the first The convolutional kernel at the _th ... Weight parameters for each position, The pooling operation represents the maximum value within a specified step window. The pooling results from all convolutional kernels are concatenated into a waveform morphology feature vector. The numerical sequence feature extraction pipeline calculates sliding window statistics for discrete numerical readings. These statistics include the mean, standard deviation, minimum, maximum, and quartiles within a preset time window, and track the slope of their trend. The slope coefficient is obtained by linearly fitting the numerical sequence within the window. The resulting numerical statistical feature vector contains a combination of statistics and trend information. The behavioral event feature extraction pipeline performs one-hot encoding or embedding encoding on events in the behavioral record sequence and calculates the density of events occurring within a preset period, such as the number of events per hour, as well as a measure of the regularity of event occurrences. This regularity measure is obtained by calculating the coefficient of variation of the interval between consecutive events. The final behavioral pattern feature vector encodes the intensity and pattern information of the behavioral events.

[0081] In some embodiments, the system employs an asynchronous feature fusion engine, which acts as a central coordinating component. This engine allows waveform feature extraction pipelines, numerical sequence feature extraction pipelines, and behavioral event feature extraction pipelines to operate at their respective inherent processing speeds. The waveform feature extraction pipeline, due to its involvement in convolution operations, may be slower than the other pipelines. After extracting features from a batch of input data, each feature extraction pipeline submits the output waveform morphology features, numerical statistical features, behavioral pattern feature vectors, and their corresponding time signatures to a shared feature buffer. This shared feature buffer is a queue structure maintained in memory, temporarily storing and aligning feature vectors from different pipelines according to time windows. It is understood that the asynchronous feature fusion engine does not force all pipelines to complete processing synchronously; instead, it allows the results from faster pipelines to wait in the shared feature buffer for the results from slower pipelines. In some embodiments, when the number of feature vectors accumulated in the shared feature buffer reaches a preset batch size (e.g., a set of features corresponding to 100 aligned time segments), the asynchronous feature fusion engine triggers a fusion operation. The preset batch size is set to provide a stable statistical basis for subsequent batch operations such as normalization, while ensuring the timeliness of feature fusion output. Too small a batch size may lead to unstable feature distribution, while too large a batch size may introduce unnecessary processing delays. The asynchronous feature fusion engine retrieves a batch of time-aligned feature vectors from the shared feature buffer and concatenates these feature vectors along the feature dimension. Concatenation refers to joining waveform morphology feature vectors, numerical statistical feature vectors, and behavioral pattern feature vectors from the same time point end-to-end into a longer one-dimensional feature vector. Subsequently, the asynchronous feature fusion engine performs a normalization operation on the concatenated long feature vector using the Z-score standardization method, ensuring that the mean of each feature dimension is 0 and the variance is 1. The system maintains a user's individual baseline mean and baseline dispersion for each feature dimension. This baseline value comes from the user's historical stable health data. The normalization process involves subtracting the individual baseline mean from the current feature value and then dividing by the individual baseline dispersion. The individual baseline data is updated periodically to adapt to the long-term, slow changes in the user's physiological parameters. After concatenation and normalization, a batch of feature vectors is organized into a three-dimensional tensor, whose dimensions represent the sample index, time step, and feature dimension, respectively. This three-dimensional tensor is the fused individual health feature tensor. Optionally, the shared feature buffer has a timeout mechanism. If the feature vectors of a certain time window have not received all the inputs from the pipelines within the buffer after waiting for the maximum time limit, the asynchronous feature fusion engine can fill the buffer with historical features or default values ​​based on a preset strategy to ensure the continuity of the processing flow. It can be understood that through asynchronous feature fusion processing, the system efficiently integrates multi-rate, multi-modal features, generating a unified format individual health feature tensor for use by downstream models.

[0082] In one embodiment of the present invention, the specific process of inputting individual health feature tensors into a dynamic decision forest, iteratively evaluating chronic disease state trajectories, and outputting state trajectory evaluation indexes includes, as described in [reference needed]. Figure 3 A dynamic decision forest is constructed, consisting of multiple growable decision trees. The node splitting rules of each decision tree are based on chronic disease state transition patterns learned from historical health data. Individual health feature tensors are fed into each decision tree in the dynamic decision forest in parallel. The growability of the dynamic decision forest is achieved through an incremental learning mechanism. The system establishes a buffer storage area for newly incoming labeled historical health data. When the buffer data accumulates to a certain scale, it triggers local updates and structural adjustments to some decision trees in the forest, rather than retraining the entire forest. This process includes growing new decision tree branches and optimizing existing node splitting rules. Simultaneously, the system imposes constraints on the overall size of the decision trees and introduces regularization strategies in model updates to prevent the forgetting of existing knowledge. This allows the forest to continuously adapt to new patterns in chronic disease state trajectories and maintain the stability of the evaluation. Within each decision tree, the individual health feature tensor starts from the root node and is passed down the branches of the tree according to the node splitting rules until it reaches a leaf node. Each leaf node is associated with a predefined state trajectory score. The system collects state trajectory scores output by all leaf nodes of the decision trees, and performs weighted aggregation on these scores. The aggregation weights are dynamically adjusted based on the confidence level of each decision tree in recent evaluations. The weighted aggregation values ​​are mapped to a standardized discrete index interval, outputting a state trajectory evaluation index. Based on the state trajectory evaluation index, a personalized monitoring protocol is activated to guide heterogeneous sensors to adjust their acquisition modes and obtain targeted enhanced monitoring data streams. The specific process includes: the system predefines multiple monitoring protocol templates, each bound to a state trajectory evaluation index interval, specifying combinations of sensor type, acquisition frequency, and duration. The state trajectory evaluation index is matched with the predefined index interval to determine the activated monitoring protocol template. Based on the determined monitoring protocol template, a protocol execution script containing specific control instructions is generated and sent to the corresponding smart wearable devices and home medical devices via a data connection channel. The smart wearable devices and home medical devices adjust the sampling frequency, working cycle, or enable additional monitoring functions of their built-in sensors according to the received protocol execution script, operating according to the new acquisition mode. During the implementation of the agreement, more intensive, focused, or multi-dimensional monitoring data will be received from the adjusted smart wearable devices and home medical devices, forming a targeted and enhanced monitoring data stream.

[0083] In practice, the process of inputting individual health feature tensors into a dynamic decision forest and iteratively evaluating chronic disease state trajectories to output a state trajectory evaluation index is executed by the state evaluation module. During the initialization phase, the state evaluation module constructs the dynamic decision forest, which consists of multiple growable decision trees. The node splitting rules for each decision tree are based on chronic disease state transition patterns learned from historical health data. In one example, the growth process of the decision tree is determined by analyzing the correlation between individual health feature tensors in historical health data and clinical outcomes over a subsequent period. The node splitting rules are selected based on feature dimensions and thresholds that best differentiate between different prognostic risks. For example, a node splitting rule might be "whether the standard deviation of the mean heart rate variability over the past 24 hours is less than 20 milliseconds," a rule learned from historical data that is a key differentiating factor related to readmission risk in heart failure patients. After deployment, each decision tree can still be structurally adjusted based on newly incoming labeled data, such as generating new branches or adjusting leaf node scores, thus achieving "growthability."

[0084] In its implementation, the state assessment module takes the individual health feature tensor as input and feeds it in parallel into each decision tree in the dynamic decision forest. Within each decision tree, the individual health feature tensor, starting from the root node, is passed down the tree branches according to node splitting rules. These rules assess specific feature dimensions within the individual health feature tensor. For example, for an individual health feature tensor representing a diabetic patient, the root node rule of the first decision tree might check the "coefficient of variation of fasting blood glucose over the past 7 days," guiding the individual health feature tensor to the left or right subtree based on the comparison result. Subsequent nodes might further check features such as "total duration of moderate-intensity exercise in the past 3 days," until the individual health feature tensor reaches a leaf node. Each leaf node is associated with a predefined state trajectory score, a real value reflecting the risk of chronic disease deterioration for the sample reaching that leaf node in historical data; a higher score indicates a more pessimistic risk trend.

[0085] In some embodiments, the state evaluation module collects state trajectory scores output by all leaf nodes of the decision tree, and performs weighted aggregation on these scores. The aggregation weights are dynamically adjusted based on the confidence level of each decision tree in recent evaluations. The confidence level of each decision tree in recent evaluations is calculated by its accuracy in evaluating validation set samples over a past period, such as using weighted accuracy or F1 score, and is recalculated in each update cycle. The state evaluation module maps the weighted aggregation values ​​to a standardized discrete index range, outputting a state trajectory evaluation index. The state trajectory evaluation index is an integer, for example, ranging from 1 to 5, where 1 represents a very stable state and 5 represents a high risk of significant state deterioration. The weighted aggregation and mapping process can be expressed by the following formula:

[0086]

[0087] in: Indicates the final state trajectory evaluation index. It is a scaling and rounding factor that maps continuous ratings to discrete intervals. This represents the total number of decision trees in a dynamic decision forest. Indicates the first The current dynamic adjustment weights of the decision trees Indicates the first The decision tree outputs the leaf node state trajectory score after evaluating the current individual health characteristic tensor. This indicates the floor function.

[0088] In some embodiments, the process of activating a personalized monitoring protocol based on the state trajectory evaluation index and guiding heterogeneous sensors to adjust their acquisition modes is executed by the protocol management module. The protocol management module predefines multiple monitoring protocol templates, each bound to a state trajectory evaluation index interval, and specifies a combination of sensor type, acquisition frequency, and duration. See Table 1 for a definition of a monitoring protocol template.

[0089] Table 1: Monitoring Protocol Template

[0090]

[0091] The protocol management module matches the calculated state trajectory evaluation index with predefined index intervals to determine the activated monitoring protocol template. For example, if the state trajectory evaluation index is 4, falling into the "high-risk" interval, then the high-intensity monitoring protocol template bound to it is activated. Based on the determined monitoring protocol template, the protocol management module generates a protocol execution script containing specific control instructions. The protocol execution script is a structured instruction set, such as "Device ID: Bracelet_001, Set heart rate sampling interval: 60 seconds, Enable continuous blood oxygen monitoring, Effective duration: 72 hours". The protocol execution script is then distributed to the corresponding smart wearable devices and home medical devices via a data connection channel.

[0092] In some embodiments, the internal command parsing agent of the smart wearable device and home medical device adjusts the sampling frequency, duty cycle, or enables additional monitoring functions of its built-in sensors according to the received protocol execution script, operating according to the new acquisition mode. For example, a smart bracelet that originally measured heart rate every 30 minutes, after receiving the new protocol execution script, adjusts the heart rate sampling frequency to once every 5 minutes and simultaneously activates its single-lead electrocardiogram recording function. During protocol execution, more intensive, focused, or multi-dimensional monitoring data is received from the adjusted smart wearable device and home medical device, forming a targeted enhanced monitoring data stream. It can be understood that this targeted enhanced monitoring data stream, compared to the baseline monitoring mode, is enhanced in terms of data temporal resolution, signal dimension, or specific indicators, providing a data foundation for subsequent detailed analysis. Optionally, the protocol execution script can include conditional logic, such as "if the heart rate exceeds the threshold three times consecutively at night, then automatically start an electrocardiogram recording," enabling the monitoring behavior to have a certain degree of adaptive and intelligent response capabilities. It is understandable that the entire closed-loop process of "assessment-adjustment monitoring-reassessment" enables health management to allocate and focus resources in real time based on the dynamic changes in an individual's condition.

[0093] In one embodiment of the present invention, the specific process of performing hierarchical anomaly deconstruction on the targeted augmentation monitoring data stream to identify potential health deviation patterns at different levels and form a hierarchical deviation pattern set includes: the system defines multiple analysis levels, including an immediate vital signs level, a diurnal fluctuation level, and a long-term trend level. At the immediate vital signs level, high-frequency physiological signals in the targeted augmentation monitoring data stream are scanned in real time to detect instantaneous abnormal peaks or troughs exceeding individualized adaptive thresholds, and their amplitude, duration, and morphology are recorded as immediate abnormal segments. At the diurnal fluctuation level, the index curves aggregated daily in the targeted augmentation monitoring data stream are analyzed to identify morphological distortions, rhythmic disturbances, or phase shifts compared to the individual's historical baseline daily fluctuation curve, as diurnal fluctuation patterns. At the long-term trend level, the index sequences spanning multiple days in the targeted augmentation monitoring data stream are trend-fitted, and the slope and curvature changes of the fitted curves are calculated and compared with the expected health evolution trend line to identify deviation trends, as long-term deviation trends. The immediate abnormal segments, diurnal fluctuation patterns, and long-term deviation trends are then structurally encapsulated to form a hierarchical deviation pattern set. The specific process of generating a comprehensive risk assessment spectrum with time window annotations by progressively matching the hierarchical offset pattern set with the historical offset pattern library includes the following steps: The historical offset pattern library stores typical offset patterns mined and annotated from historical population data, along with their corresponding risk levels and evolution paths. The system sets up a progressive matching process, sequentially calculating the similarity between the immediate vital sign level patterns, diurnal fluctuation level patterns, and long-term trend level patterns in the hierarchical offset pattern set and the corresponding typical offset patterns in the historical offset pattern library. For each level, several typical offset patterns with similarity exceeding the matching threshold are selected as candidate matching patterns. The risk events and their time intervals that subsequently occurred with the candidate matching patterns in historical cases are analyzed, and the risk probability distribution and typical time window distribution are statistically obtained. The matching results from the three levels are fused, the risk probabilities are superimposed and calibrated, and the intersection of the time windows is calculated to generate a multi-dimensional vector. This multi-dimensional vector describes the probability of different risk levels occurring within different future time windows, i.e., the comprehensive risk assessment spectrum with time window annotations.

[0094] In practice, the process of hierarchically deconstructing the targeted augmentation monitoring data stream to identify potential health deviation patterns at different levels and form a hierarchical deviation pattern set is executed by a hierarchical analysis engine. This engine defines multiple analysis levels, including the immediate vital signs level, the diurnal fluctuation level, and the long-term trend level. The immediate vital signs level focuses on high-frequency physiological signal changes at the second or minute level; the diurnal fluctuation level focuses on physiological rhythms and pattern changes over a 24-hour period; and the long-term trend level analyzes the direction of indicator evolution spanning several days to several weeks. In an example scenario, for the targeted augmentation monitoring data stream of a heart failure patient, the hierarchical analysis engine analyzes heart rate and respiratory waveforms every 5 seconds at the immediate vital signs level, analyzes the diurnal variation curve of daily systolic blood pressure at the diurnal fluctuation level, and analyzes the sequence of daily average weight changes over the past two weeks at the long-term trend level.

[0095] In practical implementation, at the real-time vital signs level, the hierarchical analysis engine performs real-time scanning of high-frequency physiological signals in the targeted enhanced monitoring data stream. These high-frequency physiological signals include continuous heart rate, respiratory rate, and blood oxygen saturation waveforms. It detects instantaneous abnormal peaks or troughs exceeding a personalized adaptive threshold. This threshold is dynamically calculated based on the statistical distribution of the user's recent historical data; for example, the abnormal heart rate peak threshold is set above three standard deviations of the individual's recent resting heart rate mean. The personalized adaptive threshold is calculated based on the user's historical health data. The system first filters physiological data segments representing the user's resting state based on the user's activity status data (such as signals from the accelerometer) and time period. Using these resting state data segments as a basis, it statistically analyzes the central tendency and dispersion of the heart rate data. The adaptive threshold is then calculated and set based on these central tendency and dispersion values. This threshold is periodically recalculated and updated using the user's latest resting state data to adapt to drift in the user's physiological baseline. The system records the amplitude, duration, and shape of instantaneous abnormal peaks or troughs. For example, it records the amplitude of an abnormal atrial premature beat waveform, the duration of the abnormal waveform in milliseconds, and the specific waveform morphological characteristics. These records constitute instantaneous abnormal segments. At the diurnal fluctuation level, the hierarchical analysis engine analyzes the daily aggregated indicator curves in the targeted augmentation monitoring data stream. Indicator curves, such as those formed by connecting the average blood pressure values ​​for each half-hour over a 24-hour period, are examples. The engine identifies morphological distortions, rhythm disturbances, or phase shifts in the indicator curves compared to the individual's historical baseline daily fluctuation curve, which is obtained by calculating the average of daily curves over a stable period in the user's past history. For example, if the daily blood pressure curve loses its "dipper" shape due to the physiological decline at night and becomes "non-dipper," this morphological distortion is recorded as a diurnal fluctuation pattern. At the long-term trend level, the hierarchical analysis engine performs trend fitting on indicator sequences spanning multiple days in the targeted augmentation monitoring data stream. For example, it performs linear or polynomial fitting on the daily morning fasting blood glucose value sequence over the past ten days, calculating the slope and curvature changes of the fitted curve. The fitted curve is compared with the expected health evolution trend line, which is the ideal trajectory of change based on the user's treatment goals. Deviations are identified, such as a continuously rising slope in the fitted weight curve, while the expected trend line is flat or declining. This deviation is recorded as a long-term offset trend. Immediate abnormal segments, diurnal fluctuation patterns, and long-term offset trends are then structured and encapsulated. This structured encapsulation involves assigning a standardized data format to each identified pattern, including pattern type, start time, end time, quantified feature vector, and confidence score, ultimately forming a hierarchical offset pattern set.

[0096] In some embodiments, the process of progressively matching a hierarchical offset pattern set with a historical offset pattern library to generate a comprehensive risk assessment spectrum with time window annotations is completed by the pattern matching and risk assessment module. The historical offset pattern library stores typical offset patterns mined and annotated from historical population data, along with their corresponding risk levels and evolution paths. Typical offset patterns are clinically validated pattern templates statistically associated with specific adverse health events. The pattern matching and risk assessment module sets up a progressive matching process, which sequentially calculates the similarity between the immediate vital sign level patterns, diurnal fluctuation level patterns, and long-term trend level patterns in the hierarchical offset pattern set and the corresponding typical offset patterns in the historical offset pattern library. Similarity calculation can employ methods such as dynamic time warping, morphological distance, or cosine similarity of feature vectors. For each level, several typical offset patterns with similarity exceeding a matching threshold are selected as candidate matching patterns. Table 2 shows the intermediate results of a progressive matching process.

[0097] Table 2: Candidate Results of Hierarchical Pattern Matching

[0098]

[0099] This analysis examines the subsequent risk events and their time intervals in historical cases related to candidate matching patterns, such as hospitalization due to acute exacerbation of heart failure. Statistical analysis yields the risk probability distribution and typical time window distribution. For example, in historical data, matching pattern P001 has a 30% probability of a moderate-risk event occurring within the next 7 days and a 10% probability of a high-risk event occurring within the next 3 days. The typical time window distribution describes the temporal probability density of the event. It can be understood that the risk probability distribution and typical time window distribution are statistical learning results based on historical data.

[0100] In some embodiments, the pattern matching and risk assessment module fuses matching results from three levels, performs superimposed calibration on the risk probabilities, and calculates the intersection of time windows. The superimposed calibration is not a simple addition, but rather uses a fusion model, such as a Bayesian network or weighted logistic regression, to consider the interactions between patterns at different levels to correct the overall risk probability. The intersection calculation of time windows involves taking the intersection of the predicted risk event occurrence time windows from each level to obtain a more accurate time estimate. Finally, a multidimensional vector is generated, describing the probability of different risk levels occurring within different future time windows—that is, a comprehensive risk assessment spectrum with time window annotations. The generation of the comprehensive risk assessment spectrum can be formally expressed as: for a future time interval... Calculate the risk level within this interval. probability

[0101]

[0102] in: These represent the risk logic values ​​contributed by the matching patterns at the immediate, intraday, and long-term levels, respectively. It is the hierarchical weight coefficient. It is an sigmoid function that maps weighted sums to probabilities. Optionally, the comprehensive risk assessment spectrum can be represented as a matrix, where rows represent different risk levels, columns represent different future time windows, and the elements in the matrix are the probabilities of occurrence of the corresponding risk level within the corresponding time window. In essence, a comprehensive risk assessment spectrum with time window annotations provides a quantitative description of future risks in both time and severity dimensions.

[0103] In one embodiment of the present invention, the specific process of parsing the comprehensive risk assessment spectrum and generating a health management execution list containing specific instructions includes: the system establishes an intervention measure knowledge graph, where nodes represent different health risk states and situations, and edges represent feasible intervention actions and their expected impacts. Using the highest priority risk indicated in the comprehensive risk assessment spectrum and its corresponding time window as the query starting point, the system traverses the intervention measure knowledge graph. During the traversal, an optimal sequence of intervention actions is selected based on the risk state nodes, the correlation strength of available intervention actions, the execution cost of the actions, and the expected effect path. Each action in the optimal sequence of intervention actions is transformed into a specific instruction that can be executed by the user or linked devices. Specific instructions include generating medication adjustment suggestions, performing specific rehabilitation exercises, scheduling a specialist appointment, and contacting a health manager. Each instruction is assigned an execution time point and triggering conditions. All generated specific instructions are arranged in chronological order to form a health management execution list. The health management execution list is pushed to the user terminal and visualized in a timeline view on the user terminal interface. The user terminal receives user confirmation of execution of instructions in the list, modification feedback, or reports of execution difficulties. When the preset triggering conditions of the instruction are met or the execution time is reached, a reminder notification is triggered through the user terminal.

[0104] In practice, the process of parsing the comprehensive risk assessment spectrum and generating a health management execution list containing specific instructions is executed by the intervention planning module. This module establishes an intervention knowledge graph, a graph-structured database. Nodes in the intervention knowledge graph represent different health risk states and situations, such as "high risk of exacerbation of acute heart failure within the next 48 hours." Situation nodes can include attributes such as specific physiological parameter ranges, time windows, and risk levels. Edges in the intervention knowledge graph represent feasible intervention actions and their expected effects. Edges include attributes such as intervention action description, execution cost, expected effect intensity, and contraindications. These edges connect risk state nodes to subsequent state nodes after the intervention. In an example scenario, the risk state node "Nighttime systolic blood pressure consistently 20 mmHg above baseline" is connected to the intervention action "Temporarily increase diuretic dosage once (within the doctor's remote authorization range)" via an edge. This edge indicates the expected effect as "potentially reducing cardiac preload within 6-12 hours," with an execution cost of "low." This intervention action node is then connected to the subsequent possible state node "Nighttime systolic blood pressure returns to near baseline."

[0105] In practice, the intervention planning module uses the highest priority risk indicated in the comprehensive risk assessment spectrum and its corresponding time window as the starting point for querying, and performs graph traversal in the intervention measure knowledge graph. The highest priority risk refers to the risk with the highest severity among all risk levels, such as "the risk of arrhythmia due to hyperkalemia within the next 72 hours is high risk." Graph traversal begins with the node representing this risk state and time window combination, using a graph search algorithm to explore connected intervention action edges and subsequent state nodes. During the traversal, an optimal sequence of intervention actions is selected based on the risk state nodes, the correlation strength of available intervention actions, the execution cost of the actions, and the expected effect path. The correlation strength is derived from statistical data of historical successful intervention cases; the execution cost includes quantitative indicators such as economic cost, user compliance cost, and operational complexity; and the expected effect path refers to the probability and magnitude of risk reduction after executing the action. The selection process can be implemented using an evaluation function, such as the function:

[0106]

[0107] in: This represents the total utility of the intervention sequence. It is the number of actions in the sequence. It is the first The expected effect strength of the intervention action It is the first The strength of the correlation between each intervention action and the current risk status. It is the first The execution cost of each intervention action, and the goal of the traversal are to find the one that maximizes total utility. The maximum sequence of intervention actions.

[0108] In some embodiments, the intervention planning module transforms each action in the optimal sequence of intervention actions into a specific instruction that can be executed by the user or linked devices. These specific instructions include generating medication adjustment suggestions, performing specific rehabilitation exercises, scheduling a specialist appointment, and contacting a health manager. For example, the intervention action "temporarily increase the diuretic dose" is transformed into a specific instruction sent to the user's smart pillbox: "At 18:00 today, open compartment C of the glibenclamide pillbox," with the text prompt "Please take the extra dose of diuretic as prescribed." The intervention action "perform 30 minutes of seated breathing training" is transformed into a specific instruction sent to the user's mobile application: "This afternoon, start the 'seated breathing training' initiation program." The intervention planning module assigns an execution time and triggering condition to each instruction. The execution time is determined based on the risk assessment time window and the time required for the intervention to take effect. The triggering condition can be a specific physiological threshold or the completion status of a preceding instruction. All generated specific instructions are arranged in chronological order to form a health management execution list. This health management execution list is a structured list, where each instruction includes the action content, the executing entity, the execution time, the triggering condition, and the status field.

[0109] In some embodiments, the system pushes a health management execution checklist to a user terminal, which can be a user's smartphone, tablet, or smartwatch. The checklist is visualized on the user terminal interface using a timeline view, arranging the instructions in the checklist chronologically along a horizontal or vertical timeline. Each instruction is presented as a graphical card, displaying a summary, time, and status. The system receives user confirmations, modification feedback, or reports of execution difficulties for the instructions on the checklist via the user terminal. Users can perform actions such as "Confirm Completion," "Request Modification," or "Report Difficulty" by clicking buttons on the instruction cards. When a preset trigger condition for an instruction is met or the execution time is reached, the system triggers a notification via the user terminal. This notification can be a push notification on the phone, an SMS message, or a vibration on the smartwatch. For example, if the trigger condition for an instruction is "If the blood glucose reading is still higher than 10 mmol / L at 10 AM," the system automatically triggers a "Contact a Health Manager" notification when the user terminal receives data meeting the condition from the connected blood glucose meter. Optionally, for instructions requiring device linkage, such as generating medication adjustment suggestions, the system will send control commands directly to the corresponding smart home or medical devices while triggering the notification. It is understandable that the generation and delivery of the health management implementation checklist constitutes a closed loop from risk assessment to concrete action, making risk management operable. It is also understandable that through user terminal interaction and feedback, the system can update the instruction status in real time and collect user compliance data, providing a basis for subsequent personalized adjustments.

[0110] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications 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 protection scope of the present invention.

Claims

1. A chronic disease health management method based on artificial intelligence, characterized in that, The method includes: The system receives health monitoring streams from heterogeneous sensors and user input, the health monitoring streams including unstructured physiological data and behavioral recording sequences. It then parses and aligns the health monitoring streams to generate a spatiotemporally aligned multimodal health data stream, including: A continuous data connection channel is established with smart wearable devices, home medical devices, and mobile terminal applications. Raw data packets are continuously received through this channel, forming the health monitoring stream. The health monitoring stream is analyzed in real time to distinguish data sources from different devices and applications, extracting unstructured physiological data and behavioral record sequences. The unstructured physiological data includes time-series waveform fragments, discrete numerical readings, and text log summaries. The behavioral record sequences include user-triggered dietary events, exercise events, medication confirmation events, and their corresponding timestamps. A local logical clock is maintained for each independent data source, and a mapping relationship with the global standard time is established. Using this mapping relationship, timestamp alignment correction is performed on all extracted data points. After timestamp alignment correction, the unstructured physiological data and behavioral record sequences within the same short time domain window are spatially correlated, and context labels are added to each correlated data point to form the spatiotemporally aligned multimodal health data stream. Asynchronous feature fusion processing is performed on the spatiotemporally aligned multimodal health data stream to generate a fused individual health feature tensor. The individual health feature tensor is then input into a dynamic decision forest to iteratively evaluate the chronic disease state trajectory and output the state trajectory evaluation index. Based on the state trajectory evaluation index, the personalized monitoring protocol is activated to guide the heterogeneous sensors to adjust their acquisition modes and obtain directional enhanced monitoring data streams; The targeted enhanced monitoring data stream is subjected to hierarchical anomaly deconstruction to identify potential health shift patterns at different levels, forming a hierarchical shift pattern set; The hierarchical offset pattern set is progressively matched with the historical offset pattern library to generate a comprehensive risk assessment spectrum with time window annotations. The comprehensive risk assessment spectrum is analyzed to generate a health management execution checklist containing specific instructions.

2. The chronic disease health management method based on artificial intelligence according to claim 1, characterized in that, Asynchronous feature fusion processing is performed on the spatiotemporally aligned multimodal health data stream to generate a fused individual health feature tensor, including: Independent feature extraction pipelines are assigned to different types of data in the spatiotemporally aligned multimodal health data stream, and the feature extraction pipelines include waveform feature extraction pipelines, numerical sequence feature extraction pipelines, and behavioral event feature extraction pipelines. In the waveform feature extraction pipeline, multi-scale convolution kernels are applied to time-series waveform segments to perform feature scanning and capture waveform morphology features; In the numerical sequence feature extraction pipeline, the sliding window statistics of discrete numerical readings are calculated, and the slope of their changing trend is tracked to form numerical statistical features; In the behavioral event feature extraction pipeline, events in the behavioral record sequence are encoded, and the density and regularity of event occurrence are calculated to form behavioral pattern features; An asynchronous feature fusion engine is set up, which allows the waveform feature extraction pipeline, the numerical sequence feature extraction pipeline, and the behavioral event feature extraction pipeline to run at different processing speeds, and submits the output waveform morphology features, numerical statistical features, and behavioral pattern features to a shared feature buffer after each has completed feature extraction. When the number of feature vectors accumulated in the shared feature buffer reaches the preset batch size, the asynchronous feature fusion engine performs concatenation and normalization operations on the feature dimensions of the feature vectors to generate the fused individual health feature tensor.

3. The method for chronic disease health management based on artificial intelligence according to claim 1, characterized in that, The individual health feature tensor is input into a dynamic decision forest to iteratively evaluate the chronic disease state trajectory and output the state trajectory evaluation index, including: The dynamic decision forest is constructed, which consists of multiple growable decision trees. The node splitting rules of each decision tree are based on chronic disease state transition patterns learned from historical health data. The individual health feature tensor is used as input and fed into each decision tree in the dynamic decision forest in parallel. Within each decision tree, the individual health feature tensor starts from the root node and is passed down the branches of the tree according to the node splitting rules until it reaches a leaf node. Each leaf node is associated with a predefined state trajectory score. Collect the state trajectory scores output by all decision tree leaf nodes, and perform weighted aggregation on the state trajectory scores. The aggregation weight is dynamically adjusted according to the confidence level of each decision tree in the recent evaluation. The weighted aggregated values ​​are mapped to a standardized discrete index range, and the state trajectory evaluation index is output.

4. The chronic disease health management method based on artificial intelligence according to claim 3, characterized in that, Based on the aforementioned state trajectory evaluation index, a personalized monitoring protocol is activated to guide heterogeneous sensors to adjust their acquisition modes and obtain directional enhanced monitoring data streams, including: Multiple monitoring protocol templates are predefined, each bound to a specific state trajectory evaluation index interval, and the combination of sensor type, acquisition frequency, and duration is specified; The state trajectory evaluation index is matched with a predefined index range to determine the active monitoring protocol template; Based on the determined monitoring protocol template, a protocol execution script containing specific control instructions is generated, and the protocol execution script is sent to the corresponding smart wearable device and home medical device through the data connection channel; The smart wearable device and home medical device execute scripts according to the received protocol, adjust the sampling frequency and working cycle of their built-in sensors or enable additional monitoring functions, and operate according to the new acquisition mode. During the execution of the agreement, more intensive, focused, or multi-dimensional monitoring data is received from the adjusted smart wearable devices and home medical devices, forming the targeted enhanced monitoring data stream.

5. The method for chronic disease health management based on artificial intelligence according to claim 1, characterized in that, The targeted augmentation monitoring data stream is subjected to hierarchical anomaly deconstruction to identify potential health shift patterns at different levels, forming a hierarchical shift pattern set, including: Define multiple analysis levels, including the immediate vital signs level, the daily fluctuation level, and the long-term trend level; At the immediate vital signs level, high-frequency physiological signals in the targeted enhanced monitoring data stream are scanned in real time to detect instantaneous abnormal peaks or valleys that exceed the individualized adaptive threshold, and their amplitude, duration and morphology are recorded as immediate abnormal segments. At the diurnal fluctuation level, the index curves aggregated by day in the targeted enhanced monitoring data stream are analyzed to identify morphological distortions, rhythm disorders, or phase shifts that occur compared with the individual historical baseline daily fluctuation curves, which are then used as diurnal fluctuation patterns. At the long-term trend level, the indicator sequence spanning multiple days in the targeted enhanced monitoring data stream is fitted with a trend, the slope and curvature changes of the fitted curve are calculated, and it is compared with the expected healthy evolution trend line to identify deviations from the trend as long-term offset trends. The instantaneous abnormal fragments, the intraday fluctuation patterns, and the long-term offset trends are structured and encapsulated to form the hierarchical offset pattern set.

6. The method for chronic disease health management based on artificial intelligence according to claim 5, characterized in that, The hierarchical migration pattern set is progressively matched with the historical migration pattern library to generate a comprehensive risk assessment spectrum with time window annotations, including: The historical migration pattern library stores typical migration patterns mined and labeled from group historical data, along with their corresponding risk levels and evolution paths. A progressive matching process is set up to sequentially calculate the similarity between the real-time vital signs level pattern, the daily fluctuation level pattern, and the long-term trend level pattern in the hierarchical offset pattern set and the typical offset patterns of the corresponding level in the historical offset pattern library. For each level, several typical offset patterns with similarity exceeding the matching threshold are selected as candidate matching patterns. Analyze the risk events and their time intervals that subsequently occurred in historical cases for the candidate matching patterns, and statistically obtain the risk probability distribution and typical time window distribution; The matching results from the three levels are fused, the risk probabilities are superimposed and calibrated, and the intersection of the time windows is calculated to generate a multi-dimensional vector. The multi-dimensional vector describes the probability of different risk levels occurring in different time windows in the future, which is the comprehensive risk assessment spectrum with time window annotation.

7. The method for chronic disease health management based on artificial intelligence according to claim 6, characterized in that, The comprehensive risk assessment spectrum is analyzed to generate a health management execution checklist containing specific instructions, including: Establish an intervention knowledge graph, where nodes represent different health risk states and situations, and edges represent feasible intervention actions and their expected impacts; Using the highest priority risk indicated in the comprehensive risk assessment spectrum and its corresponding time window as the starting point for querying, a graph traversal is performed in the intervention measure knowledge graph. During the traversal, an optimal sequence of intervention actions is selected based on the risk status nodes, the correlation strength of available intervention actions, the execution cost of the actions, and the expected effect path. Each action in the optimal intervention sequence is transformed into a specific instruction that can be executed by the user or the linked device. The specific instructions include generating medication adjustment suggestions, performing specific rehabilitation exercises, scheduling a specialist appointment, and contacting a health manager. An execution time point and triggering conditions are assigned to each instruction. All the generated specific instructions are arranged in chronological order of execution to form the health management execution list.

8. The method for chronic disease health management based on artificial intelligence according to claim 7, characterized in that, The method further includes: The health management execution checklist is pushed to the user terminal and displayed visually in a timeline view on the user terminal interface; Receive user confirmation of execution, modification feedback, or reports of execution difficulties for instructions in the list through the user terminal; When the preset triggering conditions of the instruction are met or the execution time is reached, a reminder notification is triggered through the user terminal.

9. An artificial intelligence-based chronic disease health management system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based chronic disease health management method according to any one of claims 1 to 8.

Images