Household noninductive blood glucose intelligent early warning and management method and system for old people

Abstract

The invention relates to the technical field of medical information and physical sign monitoring, and discloses a home noninductive blood glucose intelligent early warning and management method and system for old people. The method comprises the following steps: acquiring wrist-worn photoelectric volume pulse, skin temperature, skin electricity, activity intensity and environment context, completing term and sampling and clock registration, time alignment, artifact suppression and data restoration, and constructing a feature vector and an individual baseline; correlation estimation, trend discrimination and continuous calibration are carried out based on the features and the individual baselines, and an individual state structure is formed; carrying out grading factor selection, weight initialization, threshold value self-adaption and score aggregation, outputting an early warning event and an evidence tracing packet, and compiling a list; executing path matching, notification generation, task issuing and receipt collection, and completing parameter and strategy updating and archive storage. The method is suitable for the household noninvasive scene, can effectively improve the abnormity capture stability and cross-situation consistency, and strengthens the tracing and collaborative management.

Description

Technical Field

[0001] This invention relates to the field of medical information and vital sign monitoring technology, and in particular to a method and system for intelligent early warning and management of blood glucose at home for the elderly without their awareness. Background Technology

[0002] In the field of medical information and vital sign monitoring technology, existing solutions for wrist-worn photoplethysmography (PPG), skin temperature, conductance of skin, activity intensity, and environmental context monitoring in home settings for the elderly typically revolve around a single channel and fixed threshold rules. These solutions suffer from limitations such as insufficient time alignment, missing links in artifact suppression and data restoration processes, and a lack of updates to individual baseline definitions. Existing methods often rely on simple threshold comparisons and manually registered user profiles to generate alerts. Under the constraints of changing home environments and situations, these methods are prone to fluctuations in blood glucose correlation estimations and short-term trend discrimination biases, making it difficult to achieve stable generation and management of warning events. Regarding the joint processing of the individual state structure, the contextual label sequence, and the user profile, existing technologies generally suffer from fragmented processes in the selection of grading factors, weight initialization, adaptive threshold calculation, score aggregation, and evidence collection. This makes it difficult to establish a continuous process of data collection—time alignment—state determination—path matching—data storage in home monitoring applications, resulting in delays in the generation and retrieval of warning events and evidence tracing packages. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method for intelligent, non-intrusive blood glucose early warning and management for elderly individuals at home, comprising:

[0004] The system acquires wrist-worn photoplethysmography pulse, skin temperature, skin conductance, activity intensity, environmental context, and user profile; performs terminology registration and sampling period registration, clock reference registration and time alignment, artifact suppression and data repair, feature registration and window boundary registration, and field mapping registration processing to generate individual baseline definitions.

[0005] Based on the individual baseline definition, blood glucose correlation estimation, short-term trend discrimination and continuous calibration, window update and state determination, consistency verification and field aggregation processing are performed to generate the individual state structure.

[0006] Based on the individual state structure, the grading factor selection and weight initialization are performed, and the threshold adaptive calculation is performed. The grading strategy set includes fixed grading strategy, sliding grading strategy and situational grading strategy. The boundary adjudication unit is used to prune the strategy according to the strategy priority, the current situation number and the file constraint items to generate a grading threshold candidate set. The score aggregation is performed. For each candidate factor, the factor value of the current window is read and the corresponding grading threshold is matched to complete the grading level matching to generate the score component and aggregate it into a window-level score. The graded organization and evidence collection and trajectory backlink processing are performed to generate a list of early warning events and an evidence tracing package structure.

[0007] The system acquires the list of early warning events and the evidence tracing package structure, performs path matching and notification generation, task issuance and receipt collection, parameter update and strategy entry update, and file storage operations, and generates disposal instructions, disposal receipts, and model status version tag structures. The list of early warning events includes event categories and classification levels, a list of triggering factors and scenario numbers, and source paths and time paths.

[0008] Furthermore, the wrist-worn photoplethysmography (PPG) pulse, skin temperature, conductance of skin, activity intensity, environmental context, and user profile include:

[0009] The wrist-worn photoplethysmography (PPG) is a surface volume pulsation signal measured by an optical channel, used to reflect the waveform information of peripheral blood flow changes with heartbeat; the skin temperature is the output of the skin-touch or near-skin temperature channel, reflecting the surface temperature and rhythm; the skin conductance is the skin conductance activity signal, used to reflect the skin conductance response caused by sympathetic nerve drive; the activity intensity includes acceleration vector magnitude, gait frequency, and body position determination; the environmental context includes indoor temperature and humidity, illuminance, noise level, and geographical location time period labels; the user profile includes basic information, chronic disease history, medication records, diet records, and daily routine.

[0010] Furthermore, the process of generating individual state structures also includes:

[0011] The individual state structure is input and aligned in both the time-series and field dimensions to construct a decision-making buffer. The decision-making buffer records the start and end times of the window, the list of participating factors, the initial weight vector, the scenario number, and the file constraint entries, providing a unified entry point for tiered factor selection, adaptive threshold calculation, and score aggregation.

[0012] Furthermore, the process of selecting tiering factors also includes:

[0013] The feature factors that can participate in the classification are retrieved from the evidence list in the individual state structure. The feature factors include pulse waveform features, heart interval features, skin temperature fluctuation features, skin conductance response features, and activity intensity features. The source path is traced back to the specific channel and specific window.

[0014] Read the scene attributes, posture attributes, and time period attributes from the context label sequence, establish the association mapping between factors and contexts, and record them in the factor mapping table;

[0015] Read user profiles to generate profile constraint entries. The profile constraint entries are marked with three states: disabled factors, reduced weight factors, and enhanced factors. Disabled factors are directly removed from the current round of classification. Reduced weight factors are added to the weight reduction entry during the weight initialization stage, and enhanced factors are added to the weight increase entry, thus obtaining a candidate factor list.

[0016] Furthermore, the weight initialization includes:

[0017] Read the set of quality weights in the individual state structure and map the channel quality level to the factor weights;

[0018] Based on the pose and scene attributes in the context label sequence, adjustment entries are applied to factors strongly correlated with pose and factors strongly correlated with scene.

[0019] By combining chronic disease history and medication records in user profiles, profile adjustment items are formed. During periods when drug and dietary interventions have an impact, corresponding factors are suppressed or enhanced to obtain an initial weight vector.

[0020] Furthermore, the process of adaptive threshold calculation also includes:

[0021] The threshold adaptive calculation is performed using a set of tiered strategies, which includes fixed-tier strategies, sliding-tier strategies, and context-based strategies.

[0022] The fixed-file strategy targets long-term stable factors and generates initial, intermediate, and high-level boundaries from the long-term target range given in the user profile.

[0023] The sliding boundary strategy generates sliding boundaries from the version trajectory of the individual state structure for factors with short-term drift.

[0024] The context-based profile strategy targets factors that are significantly affected by the context, and the profile switching is driven by the context attributes of the context tag sequence.

[0025] After each of the three types of strategies outputs a candidate boundary, it enters the boundary adjudication unit. The boundary adjudication unit performs pruning according to the strategy priority, the current situation number, and the file constraint entries to generate a tiered threshold candidate set.

[0026] Furthermore, the process of performing score aggregation also includes:

[0027] For each candidate factor, read the factor value and corresponding tier threshold of the current window, and generate a score component after completing the tier matching. The score component records the factor name, tier name, matching deviation, and weight contribution.

[0028] Perform component aggregation within the window and output a window-level score, which includes the scenario number and time path.

[0029] Read the status name and level name from the individual status structure, merge them with the window-level score to form event candidate entries, and output the warning event when the triggering condition is met by the boundary of the grading threshold candidate set, the file constraint entries and the situation number.

[0030] Furthermore, the process of hierarchical organization also includes:

[0031] Read the time path and source path in the event trajectory log, merge consecutive similar events according to their time adjacency, and introduce a grace period in the merging strategy when a short interruption occurs.

[0032] The merged events are sorted hierarchically. The sorting rules read the key focus items in the user profile and the scenario priority items in the context tag sequence to form a hierarchy, generating a list of warning events arranged in chronological order.

[0033] Furthermore, the process of hierarchical organization, evidence collection, and trajectory backlink processing also includes:

[0034] Read back the source path and time path, trace back along the source path to the cleaning sequence and alignment index structure, and obtain the corresponding original fragment, cleaned fragment and alignment timestamp;

[0035] Extract the feature factor records and state determination records that are directly related to the early warning event from the evidence list of individual state structures, and encapsulate them together with the original fragments and cleaned fragments;

[0036] Read the scene attributes and posture attributes from the context tag sequence and inject the scene switching points and posture change points into the evidence timeline;

[0037] The reference window range before and after the event window is determined based on the alignment index structure, and the trajectory points of the retrieved blood glucose correlation estimation results, short-term trend discrimination results and individual baseline models are registered.

[0038] Read the weight adjustment logs during the weight initialization phase and the boundary adjudication records during the threshold adaptive calculation phase, mark the weight changes and boundary changes on the event timeline, and generate the evidence tracing package structure.

[0039] Furthermore, a home-based, non-intrusive intelligent blood glucose early warning and management system for the elderly, applied to any of the methods described above, includes:

[0040] The terminology registration and acquisition configuration module is used to register terms, sampling periods, and clock references, generate a terminology dictionary and acquisition configuration definition table, and provide the terminology dictionary and acquisition configuration definition table to the multi-source data acquisition and cleaning alignment module and the feature registration and window mapping module for calling;

[0041] The multi-source data acquisition and cleaning alignment module is used to acquire wrist-worn photoplethysmography pulse, skin temperature, skin conductance, activity intensity, environmental context and user profile, perform time alignment, artifact suppression, data repair, generate cleaning sequence, alignment index structure and context label sequence, and provide the cleaning sequence, alignment index structure and context label sequence to the feature registration and window mapping module for call;

[0042] The feature registration and window mapping module is used to perform feature item registration, window boundary registration and field mapping registration under the constraints of the cleaned sequence, aligned index structure and context label sequence, generate feature vector definition set and individual baseline definition, and provide the feature vector definition set and individual baseline definition to the baseline modeling and correlation determination module for calling;

[0043] The baseline modeling and correlation determination module is used to perform blood glucose correlation estimation, short-term trend discrimination, continuous calibration, window update, state determination, consistency check and field aggregation based on the feature vector definition set and individual baseline definition, and generate blood glucose correlation estimation results, short-term trend discrimination results, individual baseline model and individual state structure, and provide the individual state structure to the threshold adaptation and score aggregation module for calling;

[0044] The threshold adaptation and score aggregation module is used to read the individual status structure and context label sequence and user profile to perform classification factor selection and weight initialization, threshold adaptation calculation and score aggregation, output warning events and classification threshold candidate set, and provide the warning events to the hierarchical organization and evidence tracing module for call;

[0045] The hierarchical organization and evidence tracing module is used to process the early warning event set, perform hierarchical organization and evidence collection and trajectory backlink, generate an early warning event list and evidence tracing package structure, and provide the early warning event list and evidence tracing package structure to the path matching and notification distribution module for calling;

[0046] The path matching and notification distribution module is used to read the list of early warning events and the structure of the evidence tracing package and the template of the early warning handling strategy library, execute path matching and notification generation and task distribution, output handling instructions, and provide the handling instructions to the receipt collection and version update module for calling;

[0047] The receipt aggregation and version update module is used to collect and process receipt execution parameter updates and strategy entry updates and file storage, generate model status version tag structure and processing receipts, and write the model status version tag structure back to the terminology registration and collection configuration module, baseline modeling and correlation determination module and threshold adaptation and scoring aggregation module for invocation.

[0048] The key innovations of this invention include:

[0049] (1) Based on the individual state structure, the context label sequence and the user profile, construct an integrated link of grading factor selection and weight initialization, threshold adaptive calculation and score aggregation, generate a grading threshold candidate set in the same window and trigger an early warning event, forming an event-oriented threshold generation and adjudication and joint scoring mechanism.

[0050] (2) Run hierarchical organization, evidence collection and trajectory backlink on the early warning event set, read back the cleaning sequence and alignment index structure and feature item registration and window boundary registration along the source path and time path, construct the early warning event list and evidence tracing package structure, and realize evidence organization and time sequence backlink for early warning events.

[0051] (3) For the structure of the early warning event list and evidence tracing package, perform path matching and notification generation, task issuance and receipt collection, parameter update and strategy item update and file storage, construct disposal instructions and disposal receipts and model status version label structure, and complete the closed loop of version update of collection configuration definition table and individual baseline model and threshold and weight.

[0052] The following are its main beneficial effects:

[0053] (1) Under the constraints of changes in family environment and situation, the joint processing of the individual state structure, the situation label sequence and the user profile forms a candidate set of grading thresholds and outputs early warning events through scoring aggregation. Compared with the existing scheme that relies on fixed thresholds and single-channel comparison, it improves the triggering consistency of blood glucose correlation estimation and short-term trend judgment in multiple situations.

[0054] (2) Regarding the evidence organization and time-series backlink of early warning events, the cleaning sequence and alignment index structure, feature item registration and window boundary registration are uniformly encapsulated into an evidence tracing package structure. Compared with the existing scheme that only retains summary records, this improves the traceability and verification convenience of early warning events in the source data and processing path.

[0055] (3) Driven by path matching and notification generation, the model status version label structure is generated by collecting receipts, updating parameters and strategy entries, and storing files on disk. Compared with the existing schemes that rely on manual recording and discrete configuration, a continuous update link is formed between the collection configuration definition table and the individual baseline model and thresholds and weights, which is suitable for the long-term operation needs of home scenarios. Attached Figure Description

[0056] Figure 1 A flowchart illustrating a method for intelligent, unobtrusive blood glucose early warning and management for the elderly at home, provided as an embodiment of this application;

[0057] Figure 2 This is a structural block diagram of a home-based intelligent blood glucose early warning and management system for the elderly, provided as an embodiment of this application. Detailed Implementation

[0058] Example 1: Refer to Figure 1 This is a flowchart illustrating a method for intelligent, unobtrusive blood glucose early warning and management for the elderly at home, provided by an embodiment of the present invention. The process may include at least steps S100-S400:

[0059] S100: Acquire wrist-worn photoplethysmography pulse, skin temperature, skin conductance, activity intensity, environmental context, and user profile; perform terminology registration and sampling period registration, clock reference registration and time alignment, artifact suppression and data repair, feature registration and window boundary registration, and field mapping registration processing to generate individual baseline definitions;

[0060] S200. Based on the individual baseline definition, perform blood glucose correlation estimation, short-term trend discrimination and continuous calibration, window update and state determination, consistency check and field aggregation processing to generate an individual state structure.

[0061] S300. Based on the individual state structure, perform grading factor selection and weight initialization, perform threshold adaptive calculation, adopt a grading strategy set including fixed grading strategy, sliding grading strategy and situational grading strategy, and use the boundary adjudication unit to prune according to strategy priority, current situation number and file constraint items to generate a grading threshold candidate set, perform score aggregation, read the factor value of the current window and the corresponding grading threshold for each candidate factor to complete grading matching, generate score components and aggregate them into window-level scores, perform hierarchical organization and evidence collection, trajectory backlink processing, and generate a list of early warning events and evidence tracing package structure;

[0062] S400: Obtain the list of early warning events and the evidence tracing package structure; perform path matching and notification generation, task issuance and receipt collection, parameter update and strategy entry update, and file storage operation; generate disposal instructions, disposal receipts, and model status version tag structures; the list of early warning events includes event categories and classification levels, trigger factor list and scenario number, source path and time path.

[0063] Step S100 includes at least steps S110-S130:

[0064] S110: Acquire wrist-worn photoplethysmography pulse, skin temperature, skin conductance and activity intensity, environmental context and user profile, perform terminology registration, sampling period registration and clock reference registration processing, and obtain terminology dictionary and acquisition configuration definition table;

[0065] The wrist-worn photoplethysmography (PPG) is a surface volume pulsation signal measured by an optical channel, used to reflect the waveform information of peripheral blood flow changes with heartbeat; the skin temperature is the output of the skin-touch or near-skin temperature channel, reflecting body surface temperature and rhythm; the skin conductance (EDA) is the skin conductance activity signal, used to reflect the skin conductance response caused by sympathetic nerve drive; the activity intensity comes from inertial measurement and upright posture recognition, including derivative quantities such as acceleration vector magnitude, gait frequency, and body position determination; the environmental context includes indoor temperature and humidity, illuminance, noise level, and geographical location time label; the user profile includes basic information, chronic disease history, medication records, diet records, and daily routine. Specifically, the wrist-worn photoplethysmography (PPG), skin temperature, skin conductance, activity intensity, environmental context, and user profile are used as inputs to establish a source-by-source access channel, a timestamp acquisition channel, and a quality marker channel. This generates a raw data buffer, and each channel is labeled with operational tags such as source identifier, unit identifier, calibration identifier, packet loss identifier, and connection failure identifier, forming a traceable input starting point. Further, the subject's side-worn wearing information, skin sensitivity information, device model information, and sensor firmware information are read from the user profile as acquisition-side adaptation conditions to guide subsequent sampling cycle registration, artifact suppression configuration, and missing measurement repair strategy selection. Simultaneously, the home network latency level and timing source availability are read from the environmental context as a reference for clock reference registration. Understandably, a unified timestamp strategy is first established for the data stream after access is completed. Network time synchronization adopts Network Time Protocol (NTP) or Precision Time Protocol (PTP), and location time synchronization adopts Global Positioning System (GPS) or mobile network base station time. This forms a mapping relationship between device clocks and system clocks, and records the priority of time synchronization sources, switching thresholds, and drift estimates to ensure that subsequent time alignment actions have a reliable upper-level reference.

[0066] Specifically, during terminology registration, based on the established terminology system of this invention, terminology entries are created for each input channel and derived channel. These entries include attributes such as Chinese and English names and abbreviations, measurement objects and units of measurement, sampling methods and placement locations, data types and value ranges, default values ​​and null representations, reliability levels, and version numbers, forming the terminology dictionary. During sampling period registration, the device-side sampling frequency, system-side processing frequency, suggested window step size, and estimated relative delay between channels are read and uniformly registered in the acquisition configuration definition table as sampling period entries, window step size entries, and channel delay entries. When different channels have different native sampling periods or experience dynamic changes, the registration strategy includes resampling mode, interpolation mode, downsampling mode, and sample loss tolerance threshold, serving as control parameters for subsequent time alignment and data repair. During clock reference registration, a master-slave relationship is established between the device clock and the system clock, and the time point, drift slope estimate, and jitter distribution of each successful time synchronization are recorded, generating clock reference entries and writing them into the acquisition configuration definition table. During the aforementioned registration process, all registration entries and operation logs are synchronously written to the audit buffer. The audit buffer records the differences before and after registration in chronological order, which is used for time reconstruction of the downstream evidence tracing package. After this process is completed, two types of structured products are formed: the terminology dictionary and the acquisition configuration definition table. Both are used in this step for subsequent time alignment, artifact suppression, data repair, and feature registration actions. They are also called by the terminology dictionary and acquisition configuration definition table in step S120 and referenced by the model state version label structure in the closed-loop write-back path of S430.

[0067] S120. Extract the sampling period, clock reference, and attitude recognition switch from the terminology dictionary and acquisition configuration definition table, perform time alignment, artifact suppression, and data repair, and generate a cleaning sequence, alignment index structure, and context label sequence.

[0068] Specifically, during time alignment, the difference between the timestamp of each channel and the system clock is calculated based on the channel delay entries and timing records in the acquisition configuration definition table. A segmented drift model is used to correct the offset caused by long-term operation, and an alignment timestamp is generated in the data buffer. During the alignment process, an alignment error estimate and alignment mark are added to each data entry, and the mark is used for subsequent quality screening. During artifact suppression, quality gating rules are established for the four types of channels: wrist-worn photoplethysmography (PVP) pulse, electrodermal activity (EDA), skin temperature, and activity intensity. The quality gating rules use instantaneous acceleration, attitude flip markers, signal saturation markers, and optical channel occlusion markers in the activity intensity as trigger conditions. After triggering, quality grading and noise suppression are performed on the affected pulse waveform segments. For pulse waveforms, waveform morphology screening, extreme value anomaly removal, and pulsation band stability screening are performed, and the quality level and quality score are output. For EDA, slow drift removal and transient fluctuation detection are performed, and the response segment identifier is output. For skin temperature, environmental interference stripping and diurnal rhythm proportion estimation are performed, and the stable segment identifier is output. During data repair, segments with low-quality or missing markers are repaired based on homologous neighbor segment matching and cross-channel constraints. Neighbor segment matching is constrained by similar postures, similar activity intensities, and similar environmental contexts, while cross-channel constraints are constrained by the time dependencies between pulse waveforms, skin temperature, and skin conductance. Repaired segments are output when the repair strategy trigger threshold is met and the repair confidence reaches the registered value. Repair source records, repair confidence records, and repair strategy records are established for each repaired segment. The aforementioned time alignment, artifact suppression, and data repair actions collectively generate a unified sequence set, including the cleaning sequence, the structure describing the mapping relationship from multiple channels to a unified time stack, and the alignment index structure. Simultaneously, a context label sequence with time-period, posture, and scene attributes is generated from the environmental context, activity intensity, and user profile's schedule. This context label sequence participates in the window boundary alignment decision during subsequent feature registration and window boundary registration processes and is invoked by the individual state structure in step S310 as a participation factor in selection.

[0069] Furthermore, during feature registration, under the constraints of the cleaning sequence, the alignment index structure, and the context label sequence, the rise time, pulse width, rhythm stability, and waveform microstructure position of the pulse waveform are defined; the heartbeat interval is defined and uniformly referred to as the heartbeat interval feature; the hourly fluctuation amplitude, diurnal fluctuation amplitude, and short-term fall amplitude of skin temperature are defined; the number of skin conductance responses, fall duration, and peak amplitude of skin conductance are defined; and the postural probability, gait frequency estimation, and resting segment proportion of activity intensity are defined. Each feature has a name, unit, and dimension description in the terminology dictionary. The work of this registration section is to define the extraction method, window length, window step size, boundary closure strategy, and missing measurement compensation strategy at the window level, and record the mapping relationship between the feature and the alignment index structure. When registering window boundaries, a basic window and an extended window are established for each type of feature item based on the suggested window step size given by the sampling period registration item and the offset given by the channel delay item. The basic window is used to generate stable features, and the extended window is used to generate statistical trend-related features. For context label segments with rapidly changing action states, the window boundaries are aligned with the context label slices, and cross-segment boundary markers are generated. These cross-segment boundary markers are subsequently used as trigger conditions for short-term trend discrimination. When registering field mapping, the path, name, and index position of the feature item in the data structure are written into the field mapping table. The field mapping table is used to support subsequent feature vector construction and individual baseline initialization. After completing the above three registrations, a feature vector definition set and an individual baseline definition are generated. The feature vector definition set describes the arrangement order of feature items in the vector, data type, unit conversion strategy, missing test imputation strategy, and quality weight strategy. The individual baseline definition describes the baseline initialization window length, rolling update step size, stability threshold, anomaly exclusion conditions, and rhythm alignment strategy. The feature vector definition set and the individual baseline definition are written into the metadata area at the end of this step and used as standardized output objects. They are called by the feature vector definition set and individual baseline definition in S210 as cross-step inputs. The cleaning sequence, the alignment index structure, and the context label sequence are used as intermediate products. They are also referenced by factor screening when the cross-step scenario is constructed by the individual state structure in S310, and are used to restore the warning evidence time chain in the evidence collection action in S410.

[0070] S130. Register feature terms, window boundaries, and field mappings for the cleaned sequence, the aligned index structure, and the context label sequence to generate a feature vector definition set and an individual baseline definition.

[0071] During operation, for anomalies such as device disconnection, low power, and sensor obstruction, a status monitor is configured to generate disconnection, obstruction, and low power flags in real time at the acquisition end. These status events are written to the operation log area of ​​the acquisition configuration definition table. Time records in the operation log area are all written as the system clock, along with the time source identifier and time quality level, for subsequent tracing and playback. For clock jumps triggered by network switching and time source switching, clock jump suppression is implemented. Upon detecting a jump, a transition buffer is activated to temporarily isolate the data before and after the jump, perform time corrections separately, and then splice them together. Boundary markers are written at the splicing boundaries, and boundary indexes are registered in the alignment index structure. These boundary indexes are used to avoid cross-boundary mismatches in subsequent window boundary registration. For individual channels exhibiting persistent low quality or persistent missing data, weight reduction or masking is selected according to the reliability level strategy in the terminology dictionary. The masking action is recorded as a masking entry, which includes the start and end times, the affected channel, and the triggering basis. Masking entries will exist as limiting conditions in subsequent individual status structures. For cases where the unit or range changes within the same channel, a unified conversion is performed based on the unit and range mappings in the terminology dictionary, and the conversion record is written into the field mapping registration. Through the above operational control, this step forms a verifiable link at the levels of data access and time unification, quality control, and intermediate product generation, ensuring that subsequent algorithm modules operate with consistent data standards.

[0072] The output location of this step needs to be clearly marked in the text. The terminology dictionary and the acquisition configuration definition table are output as configuration classes and are directly read by the terminology dictionary and acquisition configuration definition table in S120 to guide time alignment, artifact suppression, and data repair. The cleaning sequence, the alignment index structure, and the context label sequence are output as time series classes. After entering the time alignment, artifact suppression, and data repair actions in S120 within this main step, they are called by the feature item registration, window boundary registration, and field mapping registration in S130. The feature vector definition set and the individual baseline definition are output as standard outputs across main steps and are read by the feature vector definition set and individual baseline definition in S210 to drive blood glucose correlation estimation and short-term trend discrimination actions. At the same time, the context label sequence is referenced by the grading factor selection in S310 across main steps. In the closed-loop path, the model state version label structure in S430 will be written back to the acquisition configuration definition table of this step to update the timing strategy, window strategy, and quality gating strategy for the next acquisition cycle. All the above-mentioned cross-step connections are anchored by field names and location names. The field names and location names have corresponding records in the terminology dictionary and the data collection configuration definition table, and the time sequence is maintained in the running log.

[0073] In summary, the technical effects of this step are as follows: data access and terminology system construction and operation strategy implementation are completed; time scale unification, quality gating and missing test repair are achieved; multiple intermediate products covering time series, configuration and feature definitions are produced; and a stable field-level connection relationship is formed with subsequent steps.

[0074] Step S200 includes at least steps S210-S230:

[0075] S210. Obtain the feature vector definition set and individual baseline definition, perform blood glucose correlation estimation and short-term trend discrimination processing, and obtain blood glucose correlation estimation results and short-term trend discrimination results.

[0076] The feature vector definition set originates from the feature item registration, window boundary registration, and field mapping registration in the preceding main step, including the feature item arrangement order, unit conversion strategy, missing test filling strategy, and quality weight strategy. The individual baseline definition originates from the individual baseline definition in the preceding main step, including the initial window length, rolling update step size, stability threshold, anomaly exclusion conditions, and rhythm alignment strategy. Specifically, the feature vector definition set and the individual baseline definition are used as input. The cleaned sequence, alignment index structure, and context label sequence generated in the preceding main step are read and aligned in the time dimension to construct a window-level feature buffer. The buffer structure records the window start and end times, participating channel names, quality weights, missing test percentages, and boundary markers, thus forming a unified entry point for blood glucose correlation estimation and short-term trend judgment.

[0077] Specifically, the blood glucose correlation estimation operates within the window-level feature buffer. First, features are mapped to the same dimension based on the unit conversion strategy of the feature vector definition set. Then, low-quality sources are downweighted according to the quality weight strategy, and weighted entries and source entries are registered in the local estimation log one by one. Further, for the rise time, pulse width, rhythm stability, and microstructure location of the pulse waveform, a hierarchical relationship is established between the main channel of cardiac interval and the morphological sub-channel. Cross-channel alignment is performed based on the time anchor points given by the alignment index structure. Abnormal morphological segments trigger a removal strategy, and the removal reason and range are written into the estimation log. For the skin temperature channel and the electrodermal signaling channel, segments are estimated in segments based on the activity state and resting period in the context label sequence. The segmentation results are aggregated within the window and written into the estimation vector. In the above processing, the source path and time path of each estimated component are restored. The source path records the channel name and feature name, and the time path records the window position and cross-segment splicing information for subsequent evidence backlinking. Through the above steps, the blood glucose correlation estimation result is obtained and used as an output field for subsequent continuous calibration and status determination. At the same time, it is used as a cross-step input for S220 to call the blood glucose correlation estimation result.

[0078] Furthermore, short-term trend discrimination is performed within the same window-level feature buffer. First, based on the scene slices given in the window boundary markers and context label sequences, the feature trajectory is divided into several sub-segments. Monotonicity retrieval and inflection point retrieval are performed on each sub-segment. Monotonicity retrieval outputs rising, falling, and stationary markers, while inflection point retrieval outputs the turning position and duration. Subsequently, based on the rhythm alignment strategy defined in the individual baseline, diurnal rhythm-related fluctuations are registered, separating the rhythm term from the trend term to avoid misleading trend discrimination due to nighttime temperature decline and quiet skin conductance decline. The intermediate products of trend discrimination are written into a trend vector, which includes sub-segment markers, direction markers, duration markers, and confidence markers. The trend discrimination results are aggregated by window to obtain the short-term trend discrimination results, which are used as output fields for subsequent continuous calibration and state determination stages, and are also used by the S220 for short-term trend discrimination results in cross-step operations.

[0079] S220. Extract baseline offset, confidence weight, and drift factor from blood glucose correlation estimation results and short-term trend discrimination results, perform continuous calibration and window updates, and generate individual baseline models.

[0080] Continuous calibration operates around the individual baseline definition. This process reads the blood glucose correlation estimation results, short-term trend discrimination results, and dietary, medication, and sleep time markers from the context label sequence, mapping the offset and stability within the window to baseline correction values. Correction value generation follows three constraints: the first constraint pertains to reliability level, prohibiting amplified corrections for low-reliability sources, and the correction factor is written to the calibration log; the second constraint pertains to boundary markers, employing a gradual change strategy at segment boundaries, with the gradual change step size recorded in the boundary entries of the calibration log; the third constraint pertains to anomaly exclusion conditions, directly transferring anomaly entries to the state determination stage for anomaly classification, without participating in baseline correction. The correction values ​​generated by these constraints are written to the baseline trajectory buffer, and subsequently, the baseline trajectory is updated according to the rolling update step size, with update points recording the version number, trigger source, and applicable window range. After the update, an individual baseline model is formed. This model is used as an output field for the S230's individual baseline model and continues to participate in state determination within this main step.

[0081] Window updates are driven by trigger conditions. These trigger conditions include time-step triggers, scenario-switching triggers, and boundary-jump triggers. Time-step triggers are derived from the window step size parameter registered in the acquisition configuration definition table; scenario-switching triggers are derived from scenario changes in the scenario label sequence; and boundary-jump triggers are derived from clock jump markers in the alignment index structure. The window update action translates or reconstructs existing windows and reconstructs feature vector buffers and quality weight buffers for new windows. For cross-time windows, a segmented synthesis strategy is adopted, and the segmented results are recorded in the window update log, which includes the number of segments, the start and end times of each segment, and the scenario number. After the window update is completed, the process re-enters the glucose correlation estimation and short-term trend discrimination process, forming a continuous estimation sequence.

[0082] State determination is performed under the constraints of the individual baseline model and continuous calibration products. The process first constructs a candidate state set, derived from the strength labels of glucose correlation estimation results and the direction and persistence labels of short-term trend discrimination results. These three types of labels are aggregated into the state candidate vector according to a field aggregation strategy. Subsequently, the candidate set is filtered based on the stability threshold and anomaly exclusion criteria in the individual baseline model. Selected candidates are assigned state and grade names and written into the state vector. State determination also reads chronic disease history and medication records from the user's profile, constructs contextual constraints, and records these constraints in the state determination log, which includes the source field, pruning reason, and scope of impact. After outputting the state vector, the state determination process proceeds to a consistency verification process.

[0083] S230. Perform state determination, consistency verification, and field aggregation on the individual baseline model to generate the individual state structure;

[0084] Consistency verification verifies cross-channel and cross-window conclusions. Cross-channel verification reads the estimated components of the pulse waveform's main channel and morphology sub-channel. For segments with conflicting conclusions, a re-estimation strategy is triggered, which removes low-quality components within the same window and refreshes the estimation log. Cross-window verification performs trend continuity and state consistency checks on adjacent windows. If continuity is not satisfied, a neighboring window fusion strategy is triggered, outputting a fusion flag and fusion range, which is written to the verification log. For windows that pass consistency verification, the blood glucose correlation estimation results, short-term trend discrimination results, and state vectors are encapsulated into a unified result container, containing field names, source paths, and time paths.

[0085] Field aggregation runs at the unified result container layer. Following the paths provided by the field mapping registration, this process aggregates the estimated, trend, calibration, and state components into two types of objects. The first type is the individual baseline model, whose aggregated content includes the baseline trajectory, version number, applicable scope, and most recent update source. This serves as the structured output field for the individual baseline model in S230. The second type is the individual state structure, whose aggregated content includes the state name, level name, evidence list, time path, source path, and quality weight set. This serves as the structured output field for the individual state structure in S310. After field aggregation is complete, the estimation log, trend log, calibration log, state determination log, consistency verification log, and field aggregation log generated in this main step are written to the audit buffer in chronological order, providing a time-chain index for subsequent evidence tracing package construction and report compilation.

[0086] Regarding anomaly handling, if the proportion of missing tests within a window exceeds the threshold of the missing test filling strategy, glucose correlation estimation is skipped and a skipped entry is written. The skipped entry includes the proportion of missing tests and the trigger source. Subsequently, the window is handed over to the window update module for reconstruction. If consecutive clock jump markers appear in the alignment index structure, the output of on-site conclusions is paused during the state determination phase, and the output process is resumed after the adjacent window fusion strategy is completed. If consecutive low-quality markers appear in the quality weight strategy, the correction amount update is frozen during the continuous calibration phase, and the frozen entry is written to the baseline trajectory buffer. The above-mentioned abnormal paths are automatically resolved after the next round of window update and return to the standard process.

[0087] The outputs and destinations of this main step are explicitly indicated in the text. The blood glucose correlation estimation results and short-term trend discrimination results are used as output field names, directly read by the input of S220; the individual baseline model is used as an output field name, directly read by the input of S230; and the individual state structure is used as an output field name, directly read by the input of S310. All three types of outputs are simultaneously written to the cross-main step data bus, which records the field names, position names, and time indices. The time index and alignment index have the same structure. The current main step also subscribes to the model state version tag structure from S430 to refresh the version number and applicable scope in the individual baseline model, thus taking effect during the next period window update.

[0088] In actual operation, blood glucose correlation estimation, short-term trend discrimination, continuous calibration, window update, state determination, consistency verification, and field aggregation are executed sequentially within the same processing thread. The thread is triggered by the scheduler when the window step size is reached. The scheduler reads the window step size parameter and thread priority parameter from the acquisition configuration definition table. When the thread starts, it first checks the incomplete entries in the audit buffer. If there are incomplete entries, the log is filled in before entering the standard process. When the thread ends, it outputs four types of products: blood glucose correlation estimation results, short-term trend discrimination results, individual baseline model, and individual state structure. These four types of products are written to the result cache and pushed to the cross-master step data bus, generating a processing completion event, which is recorded in the operation log.

[0089] In summary, the technical effects of this step are as follows: window-level estimation, trend identification, baseline update, and state generation are completed, resulting in structured results and log links. Four types of objects are produced that can be directly read by subsequent main steps, and consistency and version records are retained at the processing boundaries to facilitate subsequent evidence backlinks and policy invocations.

[0090] Step S300 includes at least steps S310-S330:

[0091] S310. Obtain the individual state structure, context label sequence, and user profile; perform classification factor selection and weight initialization processing to obtain a candidate set of classification thresholds.

[0092] The individual state structure originates from the field aggregation results of the preceding main step, including state name, level name, evidence list, time path, source path, and quality weight set. The context label sequence is derived from the alignment index structure and cleaning sequence of the preceding main step, including time period attributes, posture attributes, and scene attributes. The user profile is referenced from the collection configuration definition table of the preceding main step, including basic information, chronic disease history, medication records, diet records, and daily routines. Specifically, the individual state structure, the context label sequence, and the user profile are used as inputs, aligned in the time and field dimensions to construct a decision buffer. The decision buffer records the start and end times of the window, the list of participating factors, the initial weight vector, the context number, and the profile constraint entries, providing a unified entry point for subsequent grading factor selection, threshold adaptive calculation, and score aggregation.

[0093] Specifically, the selection of grading factors revolves around the state names and level names in the individual's state structure. First, feature factors eligible for grading are retrieved from the evidence list. These feature factors include pulse waveform characteristics, heart interval characteristics, skin temperature fluctuation characteristics, skin conductance response characteristics, and activity intensity characteristics, and are traced back to specific channels and windows according to their source paths. Next, scene attributes, posture attributes, and time period attributes from the context label sequence are read to establish a correlation mapping between factors and contexts. This correlation mapping is recorded in a factor mapping table, which is then written into the decision buffer. Further, chronic disease history, medication records, diet records, and daily routines from the user's profile are read to generate profile constraint entries. These entries are labeled with three categories: disabled factors, reduced-weighted factors, and enhanced factors. Disabled factors are immediately excluded from this round of grading. Reduced-weighted factors are added to the weight initialization stage, and enhanced factors are added to the weight initialization stage. After completing the above processing, a candidate factor list is obtained, which then enters the weight initialization process.

[0094] The initial weight vector is established based on the candidate factor list. Specifically, the quality weight set in the individual state structure is read, channel quality levels are mapped to factor-level weights, and the source, level, and mapping method are recorded in the initialization log. Then, based on the posture and scene attributes in the context label sequence, adjustment entries are applied to factors strongly correlated with posture and factors strongly correlated with scene; these adjustment entries are written to the weight adjustment log. Further, chronic disease history and medication records in the user profile are combined to form profile adjustment entries. During periods where medication or dietary interventions have an impact, corresponding factors are suppressed or enhanced; the magnitude of suppression or enhancement is recorded in the weight adjustment log and associated with the time path. This yields the initial weight vector, which, along with the candidate factor list, enters the threshold adaptive calculation. The threshold adaptive calculation uses a tiered strategy set, which includes three types of strategies: fixed-tier strategy, sliding-tier strategy, and context-based strategy. The fixed-level strategy targets long-term stable factors, generating initial, mid-level, and high-level boundaries from the long-term target range provided by the user profile. The sliding-level strategy targets factors with short-term drift, generating sliding boundaries from the version trajectory of the individual state structure. The contextual-level strategy targets factors significantly affected by the scenario, driving level switching based on the scenario attributes of the scenario tag sequence. After each of the three strategies outputs candidate boundaries, they enter the boundary adjudication unit. The boundary adjudication unit prunes the boundaries according to strategy priority, current scenario number, and profile constraint entries. The pruned boundaries are written into the grading threshold candidate set, which serves as one of the output field names of this main step and also as the threshold source for internal score aggregation within this main step.

[0095] S320. Extract threshold boundaries, weight parameters, and scoring items from the candidate set of tiered thresholds, perform adaptive threshold calculation and score aggregation, and generate early warning events.

[0096] The scoring aggregation process reads data from the initial weight vector and the candidate set of tier thresholds, generates score entries within the current window, and triggers event determination. Specifically, for each candidate factor, the factor value and corresponding tier threshold of the current window are read. After tier matching is completed, a score component is generated. The score component records the factor name, tier name, matching deviation, and weight contribution, and is written to the scoring log. Subsequently, component aggregation is performed within the window, outputting a window-level score, which includes the scenario number and time path. Further, the status name and tier name in the individual status structure are read and merged with the window-level score to form event candidate entries. Event candidate entries include the event category, a list of triggering factors, scenario number, source path, and time path. Trigger determination is performed at the event candidate entry level. The trigger determination conditions are jointly constrained by the boundaries of the tier threshold candidate set, the file constraint entries, and the scenario number. When the trigger conditions are met, a warning event is output. The warning event is written to the result cache and recorded in the event trajectory log. The warning event is one of the output field names of this main step and is also read by the path matching and notification generation of the next main step across main steps.

[0097] The hierarchical organization process runs on the alert event set, aiming to build a structured list and handle the merging of adjacent windows. Specifically, it reads the time path and source path from the event trajectory log, merges consecutive similar events according to their temporal adjacency, and introduces a grace period when a short interruption occurs during the merging strategy. The length of the grace period is recorded in the hierarchical organization log. The merged events are then sorted hierarchically. The sorting rules are derived from the user profile's key focus items and the scenario priority items in the context tag sequence to form a hierarchy. The sorting results are written to the alert event list. The alert event list is arranged chronologically and includes event categories, hierarchical levels, a list of triggering factors, context numbers, source paths, and time paths. The alert event list is one of the output fields of this main step and will be used in path matching and notification generation for the next main step across main steps.

[0098] Evidence aggregation constructs an evidence list for each warning event. Specifically, it reads back the source path and time path in the unified result container, traces back along the source path to the cleaned sequence and alignment index structure, and obtains the corresponding original fragment, cleaned fragment, and alignment timestamp. Then, it extracts feature factor records and state judgment records directly related to the event from the evidence list of the individual state structure, and encapsulates these two types of records along with the original and cleaned fragments, writing the encapsulated entries into the evidence repository. Further, it reads the scene attributes and attitude attributes from the context label sequence, injecting scene switching points and attitude change points into the evidence timeline to form an upstream and downstream comparison of the event. After the above evidence aggregation is completed, it enters the trajectory backlink.

[0099] S330. Classify and organize early warning events, collect evidence and track back links, and generate a list of early warning events and a structure for evidence tracing.

[0100] The trajectory backlink executes along the time path dimension, aiming to establish a complete link between the event and the upstream processing. Specifically, firstly, based on the window number and boundary markers in the alignment index structure, the reference window range before and after the event window is determined. Within the reference window range, trajectory points of the blood glucose correlation estimation results, short-term trend discrimination results, and individual baseline models are retrieved one by one. These trajectory points are registered with the evidence timeline and written to the backlink log. Subsequently, the weight adjustment log from the weight initialization phase and the boundary adjudication record from the threshold adaptive calculation phase are read, and weight changes and boundary changes are marked on the event timeline, so that the evidence entries not only present the results but also the trigger path. After the backlink is completed, an evidence tracing package structure is generated. The evidence tracing package structure includes event identifiers, factor lists, source paths, time paths, weight evolution records, boundary evolution records, and reference window ranges. As one of the output field names of this main step, the evidence tracing package structure, together with the warning event list, is used in cross-main steps for path matching and notification generation in the next main step.

[0101] Regarding anomaly handling, if the candidate factor list is insufficient within a window due to missing tests or masking, a downgrade flag is output during the tiered factor selection phase. This downgrade flag triggers the fixed-tier strategy to drive adaptive threshold calculation, and a downgrade symbol is given for the score aggregation. The downgrade symbol is written to the scoring log and marked in the warning event. If the initial weight vector contains consecutive extreme values, a frozen entry is output during the weight initialization phase. The frozen entry indicates that the window should maintain the most recently available weight. The frozen entry is written to the weight adjustment log and participates in the trajectory backlink. If a conflict boundary appears in the tiered threshold candidate set, the boundary adjudication unit outputs an adjudication failure symbol. This symbol triggers the scenario-based strategy to revert to the fixed-tier strategy. The revert record is written to the boundary adjudication record and presented in the evidence tracing package structure. All the above-mentioned anomaly paths automatically resume the standard process after the next window is updated, without changing field names.

[0102] The outputs and destinations of this main step are clearly indicated in the main text. The candidate set of tiering thresholds is one of the output field names of this main step, used for internal scoring aggregation. Warning events, also one of the output field names, are written to the result cache and used for path matching and notification generation in the next main step across main steps. The warning event list and evidence tracing package structure, as output field names of this main step, are written to the data bus in chronological order and used for path matching and notification generation in the next main step across main steps. All four types of output fields are simultaneously recorded in the audit buffer. The audit buffer records the time path, source path, and processing log index, establishing a replayable relationship with the alignment index structure and cleaning sequence generated by the previous main step. On the closed-loop path, the model status version label structure output by the next main step is written back to the strategy configuration area of ​​the user profile. The reminder frequency, channel priority, and contact information entries in the strategy configuration area will serve as a reference for the next round of tiering factor selection and weight initialization, thus forming a closed-loop parameter update that runs through the entire chain.

[0103] In summary, the technical effects of this step are as follows: It constructs a decision-making link oriented towards individual status, context, and archives; completes factor selection and weight initialization, threshold adaptive calculation, score aggregation, and hierarchical organization; and forms structured evidence and time path backlinks, thereby outputting a list of early warning events and evidence tracing package structure that can be directly called by the next main step.

[0104] Step S400 includes at least steps S410-S430:

[0105] S410. Obtain the list of early warning events, the structure of the evidence tracing package, and the template of the early warning handling strategy library; perform path matching and notification generation processing to obtain handling instructions.

[0106] The evidence tracing package structure originates from the evidence aggregation and trajectory backlink results of the preceding main steps. It includes fields such as event identifier, factor list, source path, time path, weight evolution record, boundary evolution record, and reference window range, used to read back key nodes of the upstream processing chain during the handling process. The early warning handling strategy library template is a collection of handling rules and paths, containing strategy entries, path diagrams, node actions, collaborative constraints, and failure rollback relationships, supporting differentiated notification, follow-up, and return processes for different event categories, different classification levels, and different scenario numbers. Specifically, the warning event list, the evidence tracing package structure, and the warning handling strategy library template are used as inputs to establish a handling session. In the handling session, resource views, channel views, and role views are initialized. The resource view records the list of available family doctors, follow-up seats, duty seats, and emergency contacts, as well as the available time periods. The channel view records the status, rate limiting thresholds, and retry rules of SMS channels, voice outbound call channels, in-application notification channels, and email channels. The role view records the contact information, preferred channels, and time period preferences of the elderly person, their family contacts, community doctors, and platform duty personnel. The warning event list includes event categories and classification levels, a list of triggering factors and scenario numbers, source paths, and time paths.

[0107] Path matching is used to map events to policy paths. Specifically, it iterates through each event in the warning event list, mapping the event category, level, situation number, and trigger factor list to path requirements, and retrieves candidate paths that meet the path requirements and coordination constraints from the warning handling policy library template. A candidate path consists of several node actions linked together, with each node action defining the notification object, notification channel, timing conditions, receipt type, and failure fallback action. Further, it reads the source path and time path from the evidence tracing package structure, verifies the evidence coverage and time fit of the candidate paths, and if a candidate path has insufficient evidence coverage or does not meet the time fit threshold, it is downgraded to a suboptimal path or path expansion is triggered. Path expansion is accomplished by adding supplementary evidence nodes or adjusting the node order in similar policy entries. Understandably, when the resource view shows that a target role is unavailable in the current time period, path matching sets the node corresponding to that role as a delayed node and inserts a replacement node. The replacement node and the delayed node are jointly written into the time schedule of the path instance. After path matching is completed, a path instance is output. The path instance is written to the processing session and used as direct input for the generation of subsequent notifications. At the same time, a path selection log is recorded in this main step. The log records the path source entries, the pruning basis, the expansion basis, and the resource constraints.

[0108] Notification generation is used to convert path instances into sendable messages and tasks. Specifically, based on the notification object and notification channel of each node action in the path instance, a message body, attachment body, and reference body are constructed. The message body consists of an event summary, hierarchical level, triggering factor key points, and suggested measures. The event summary and triggering factor key points are directly extracted from the warning event list, and the suggested measures are loaded from the node action descriptions in the warning handling strategy library template. The attachment body consists of an evidence snapshot, evidence timeline link, and reference window range key points. The evidence snapshot, evidence timeline link, and reference window range key points are extracted from the evidence tracing package structure and presented as readable paragraphs in the message body. The reference body contains the event identifier and path instance identifier for precise association during receipt parsing. On the channel side, the first appearance of the SMS channel in this document is given as the English name Short Message Service (SMS), the first appearance of the voice outbound call channel is given as the English name Interactive Voice Response (IVR), and the first appearance of the email channel is given as the English name Electronic Mail (Email). For different channels, separate sending queues are established for notification generation. These queues include deduplication, rate limiting, and retry mechanisms. When the same role is hit by multiple path nodes within a short period, a deduplication strategy is triggered. This strategy retains high-priority messages and merges other messages into attachments or postpones their delivery. During the notification generation phase, a draft handling instruction is created. This draft, after incorporating a timetable and receipt type, is solidified into a handling instruction. The handling instruction includes fields such as event identifier, path instance identifier, node sequence, notification object, notification channel, sending time, receipt type, and failure rollback rules. The handling instruction is one of the output field names of this main step and is read by the task issuance action within this main step.

[0109] S420: Extract execution parameters and follow-up scheduling and notification channels from the disposal instructions, issue tasks and collect receipts, and generate disposal receipts;

[0110] Task issuance is used to deliver handling instructions to the corresponding people, devices, and services. Specifically, for notification recipients who are elderly individuals, both in-app notifications and SMS messages are sent simultaneously. The in-app notification includes interactive entries for confirmation and feedback buttons, which are registered as receipt triggers on the server. For notification recipients who are family contacts, the message is sent first through their preferred channel; if the preferred channel is unavailable, it falls back to the SMS or voice call channel. For notification recipients who are community doctors or platform staff on duty, a task entry is created and written to the workbench. The workbench entry includes an event summary, evidence snapshot, handling suggestions, target response time limits, and escalation rules. On the device side, when the elderly person's wrist-worn device is detected to be online and capable of message notifications, a vibration alert is sent to the device, and the device push entry is recorded. Each sending action during task issuance generates a sending record, which is written to the sending log. The sending log records the sending time, channel name, target recipient, message identifier, and receipt type. After task issuance is completed, all sending records and handling instructions are written to the result cache, and the receipt collection module is triggered to start listening for events on the corresponding channels and workbench.

[0111] Receipt collection is used to uniformly collect responses from various channels and workbenches. Specifically, SMS channel receipts consist of an operations-side receipt code and a user-side reply message. The operations-side receipt code records the delivery, failure, and unknown status, while the user-side reply message records confirmation text and free text. These two are merged into a receipt entry on the server side. Voice outbound call channel receipts consist of connection status, key selection, and duration. Connection status and key selection are aggregated into a receipt entry. In-app notification receipts are generated by confirmation and feedback buttons. Confirmation buttons generate confirmation-type receipts, while feedback buttons generate free text-type receipts. Free text can include images or short audio clips. Images and short audio clips are converted into attachments on the server side and associated with the receipt entry. Workbench receipts are submitted by community doctors or platform staff on the treatment interface. The submitted content includes whether a patient was admitted, the measures taken, follow-up suggestions, and the next step. After submission, a receipt entry is formed with a role identifier. Receipt collection merges receipt entries from different sources according to message identifiers and event identifiers, resulting in a treatment receipt. As one of the output field names of this main step, the processing receipt is read by parameter update and policy entry update actions within this main step, and is also available for subsequent file write actions to disk in cross-main steps.

[0112] Parameter updates are used to revise model-related parameters based on the handling receipt and evidence tracing package structure. Specifically, fields such as response type, response duration, whether suggestions were adopted, and whether additional feedback was added are read from the handling receipt and aligned with the weight evolution record and boundary evolution record in the evidence tracing package structure to generate a parameter update list. The parameter update list includes tiered threshold fine-tuning entries, factor weight fine-tuning entries, reminder frequency adjustment entries, and channel priority adjustment entries. Each entry includes the trigger source, time path, applicable scope, and rollback conditions. Further, for entries related to the individual baseline model, a baseline rolling window configuration update is generated, and the updated content is written to the model configuration buffer and takes effect in the next cycle. For entries related to collection and notification, a collection configuration definition table update and a channel priority table update are generated, and the updated content is written to the configuration buffer and takes effect in the next cycle. After the parameters are updated, a model status version tag structure is formed. The model status version tag structure includes fields such as version number, effective time, applicable scope, trigger source, and update summary. As one of the output field names of this main step, the model status version tag structure is referenced by the collection configuration definition table of the previous main step and the individual baseline model and grading factor selection in cross-main steps, and is used to drive the configuration refresh of the next round of collection, modeling and decision-making.

[0113] The strategy entry update is used to evolve the warning handling strategy library template at the rule layer. Specifically, it reads the descriptions of the degree of adoption of suggestions, feasibility of implementation, and follow-up effects from the handling receipts, aligns them with the reference window range and time path of the evidence tracing package structure, and identifies the situations where the path is too long or redundant under what event category, what level of classification, and what situation number. For cases of excessively long paths, they are pruned according to the node action time records. The pruning strategy prioritizes removing nodes with duplicate evidence coverage or merging evidence appendages before nodes. For cases of redundant paths, adjacent nodes of the same type are merged, and nodes with the same notification object and the same receipt type are integrated into a single node. The strategy entry update is written to the strategy library change log, which records the entry number, old value, new value, trigger source, and effective time. After the change is completed, the strategy library index is rebuilt so that the latest entry can be used for retrieval when matching paths.

[0114] S430. Update parameters and strategy entries on the processing receipt and save the file to disk, generating a model status version tag structure;

[0115] The disk filing process is used to write the key results of this round of processing back to the user profile and audit buffer. Specifically, it reads the processing receipt, processing instructions, and evidence tracing package structure to generate file writing entries. Each entry includes an event summary, classification level, evidence index, processing process overview, role participation list, follow-up suggestions, and version number. The file writing entries are written to the processing record area of ​​the user profile, and simultaneously written to the audit buffer along with the time index, source path, and log index, forming a replayable processing trajectory. For entries involving family doctor follow-ups, disk filing also creates a follow-up appointment entry in the workbench, which includes the target time, channel suggestions, and evidence entry prompts. After disk filing is complete, a completion marker is generated in the processing session, triggering session reclamation and resource release.

[0116] In terms of anomaly handling, if no available path is found during the path matching phase, the fallback path entry in the warning handling strategy library template is invoked. This fallback path entry includes the minimum notification set and the minimum follow-up set, and the fallback source is marked in the path selection log. If the corresponding channel in the channel view is found to be unavailable and there is no alternative channel during the notification generation phase, a delay node is inserted into the handling instruction, and a retry time is set. The retry time is generated based on the rate limiting threshold given in the channel view. If consecutive failures occur during the task distribution phase, a failure rollback rule is triggered. This rule replaces the notification object with an alternative contact or elevates it to a higher-level role, and the rollback level is recorded in the sending log. If no receipt is received within the agreed time during the receipt collection phase, a timeout entry is generated, and the process enters the restriction branch of the parameter update phase. The restriction branch only updates the reminder frequency and channel priority, not the factor weight and tier threshold. All the above-mentioned abnormal paths are written to the corresponding logs and synchronized to the audit buffer for subsequent evidence backlinking and report orchestration.

[0117] The outputs and destinations of this main step are clearly defined in the main text. The disposal instruction, as one of the output fields of this main step, is directly read by the task issuance action within this main step and drives the sending queue, while also being written to the result cache for easy receipt association. The disposal receipt, as one of the output fields of this main step, is directly read by the parameter update and strategy entry update actions, and is also read by the file write-to-disk action across main steps to complete the write-back of the disposal record. The model status version tag structure, as one of the output fields of this main step, is written back to the acquisition configuration definition table and individual baseline model and grading factor selection of the preceding main step across main steps, respectively driving the next cycle refresh of the sampling strategy, baseline rolling window configuration, threshold, and weights, thus forming a closed loop connecting acquisition, modeling, decision-making, and disposal.

[0118] In summary, the technical effects of this step are as follows: Under the constraints of event evidence and strategy paths, the instantiation of the handling path, the collection of multi-channel notifications and standardized receipts, and the versioning and updating of parameters and strategies are completed. The traceable handling records and time indexes are formed by archiving the files to disk. The handling instructions, handling receipts, and model status version tag structure are output, and cross-step configuration refresh is triggered.

[0119] Example 2: Figure 2 This diagram illustrates a structural block diagram of a home-based, unobtrusive intelligent blood glucose early warning and management system for the elderly, according to an embodiment of the present invention. Figure 2 As shown, the structure may include:

[0120] The terminology registration and acquisition configuration module 01 is used to register terms, sampling periods, and clock references, generate a terminology dictionary and acquisition configuration definition table, and provide these to the multi-source data acquisition and cleaning alignment module and the feature registration and window mapping module. Specifically, it receives known field descriptions and channel lists from user profiles and environmental contexts as input to complete term registration, unit of measurement registration, value range registration, and missing measurement marker registration, forming a terminology dictionary; it reads sampling frequency, window step size, channel delay, and time source records to complete sampling period registration and clock reference registration, forming an acquisition configuration definition table; it writes the time records and source records generated during the registration process into the operation log, and the operation log maintains a field correspondence with the terminology dictionary; it outputs the terminology dictionary and acquisition configuration definition table, which are used by the multi-source data acquisition and cleaning alignment module to read acquisition rules, and by the feature registration and window mapping module to read field names and window parameters; the terminology dictionary and acquisition configuration definition table are updated with version number and effective time after the model status version label structure produced by the receipt aggregation and version update module is written back, and are called by the multi-source data acquisition and cleaning alignment module in the next round of data acquisition.

[0121] The multi-source data acquisition, cleaning, and alignment module 02 is used to acquire wrist-worn photoplethysmography pulse, skin temperature, skin conductance, activity intensity, environmental context, and user profile data. It performs time alignment, artifact suppression, and data repair, generating a cleaning sequence, alignment index structure, and context label sequence. The cleaning sequence, alignment index structure, and context label sequence are then provided to the feature registration and window mapping module for invocation. Specifically, it reads the terminology dictionary and acquisition configuration definition table from the terminology registration and acquisition configuration module, accesses the wrist-worn channel and environmental context channel according to the registered sampling period, and associates them with the user profile fields, completing timestamp unification and channel delay adjustment. The system performs correction and generates an alignment index structure; it suppresses motion and occlusion interference, repairs missing data segments, and forms a cleaned sequence; it generates a context label sequence based on time period, pose state, and scene state, and the context label sequence shares the window boundary with the cleaned sequence; it outputs the cleaned sequence, alignment index structure, and context label sequence, and passes them to the feature registration and window mapping module. These three types of outputs are indexed and called when the baseline modeling and correlation determination module needs to read back the original basis; when the timing source is abnormal or the channel is disconnected, the abnormal entry is recorded and written into the quality mark of the cleaned sequence in the next sampling period.

[0122] The feature registration and window mapping module 03 is used to perform feature item registration, window boundary registration, and field mapping registration under the constraints of the cleaned sequence, aligned index structure, and context label sequence. It generates a feature vector definition set and an individual baseline definition, and provides these to the baseline modeling and correlation determination module. Specifically, it receives the cleaned sequence and aligned index structure, completes feature item registration based on the field names and unit descriptions in the terminology dictionary, registers pulse waveform related items, heart interval related items, skin temperature fluctuation related items, skin conductance response related items, and activity intensity related items, and labels each item with missing detection handling strategies and quality weight sources. Under the context label sequence constraints, it sets the window length and window step size and writes them into the window boundary registration to form a feature output window. It registers the path of the feature item in the structure to the field mapping registration, generating a searchable index. It summarizes the above registration content to construct the feature vector definition set and the individual baseline definition. It outputs the feature vector definition set and the individual baseline definition for the baseline modeling and correlation determination module to read. The field mapping registration and window boundary registration are used as time and source location criteria by the hierarchical organization and evidence tracing module during subsequent evidence aggregation and trajectory backlinking.

[0123] The baseline modeling and correlation determination module 04 is used to perform glucose correlation estimation, short-term trend discrimination, continuous calibration, window update, state determination, consistency check, and field aggregation based on the feature vector definition set and individual baseline definition. It generates glucose correlation estimation results, short-term trend discrimination results, individual baseline models, and individual state structures, and provides the individual state structure to the threshold adaptation and scoring aggregation module for invocation. Specifically, it reads the feature vector definition set and individual baseline definition and combines the cleaning sequence, alignment index structure, and context label sequence to construct a window-level feature buffer; it calculates the glucose correlation estimation results and short-term trend discrimination results in each window, and performs adjustments for drift segments. Continuous calibration is performed and window updates are triggered. The updated window re-enters estimation and discrimination. Status labels are generated and written into the status judgment results based on the stability threshold and anomaly exclusion conditions. Consistency checks are performed on adjacent windows and field aggregation results are output. Individual baseline models and individual status structures are formed and time paths and source paths are registered. Blood glucose correlation estimation results, short-term trend discrimination results, individual baseline models and individual status structures are output. The individual status structure is passed to the threshold adaptation and scoring aggregation module. After the individual baseline model is written into the model status version label structure in the receipt aggregation and version update module, it is read by the terminology registration and acquisition configuration module to refresh the sampling and window parameters.

[0124] The threshold adaptation and scoring aggregation module 05 is used to read the individual state structure, context label sequence, and user profile to perform grading factor selection, weight initialization, threshold adaptation calculation, and scoring aggregation. It outputs warning events and a grading threshold candidate set, and provides the warning events to the hierarchical organization and evidence tracing module for invocation. Specifically, it receives the individual state structure and reads back the context label sequence and user profile. Based on the state name and level name, it completes the grading factor selection, writing disabled factors, deweighted factors, and enhanced factors into the decision buffer. It constructs an initial weight vector and triggers threshold adaptation calculation, using fixed-grading, sliding-grading, and context-grading strategies to generate candidate boundaries. The candidate boundaries enter the boundary adjudication unit to form a grading threshold candidate set. Within the window, it generates score components for the candidate factors and completes scoring aggregation. After the triggering conditions are met, it generates a warning event and registers the time path and source path. It outputs the grading threshold candidate set and the warning event. The warning event is transmitted to the hierarchical organization and evidence tracing module. The grading threshold candidate set is used as a basis field when the path matching and notification distribution module needs to generate a summary.

[0125] The hierarchical organization and evidence tracing module 06 is used to process the early warning event set by performing hierarchical organization, evidence aggregation, and trajectory backlinking, generating an early warning event list and an evidence tracing package structure, and providing the early warning event list and evidence tracing package structure to the path matching and notification delivery module for invocation; specifically, it obtains the early warning events and their time paths and source paths, merges similar events according to their temporal adjacency and writes them into the hierarchical sorting results to generate an early warning event list; it reads back along the source path the cleaning sequence and alignment index structure and feature registration and window mapping records, encapsulates the original fragments and cleaned fragments and alignment timestamps and status judgment records, and constructs the evidence tracing package structure; it then hands over the early warning event list and evidence tracing package structure to the path matching and notification delivery module, and the evidence tracing package structure is read back as the basis for weight and threshold evolution when the receipt aggregation and version update module performs parameter updates.

[0126] The path matching and notification distribution module 07 is used to read the list of early warning events, the evidence tracing package structure, and the early warning handling strategy library template to perform path matching, notification generation, and task distribution, output handling instructions, and provide the handling instructions to the receipt aggregation and version update module for invocation; specifically, it receives the list of early warning events and links with the early warning handling strategy library template to search for candidate paths, trims the node order according to event category, level, and scenario number, and solidifies the path instance after verifying the source path and time path coverage of the evidence tracing package structure; it generates message content and attachment summaries for each channel and role based on the path instance, arranges the sending time and writes it into the sending queue to form a handling instruction; it outputs the handling instruction and sends it to the target channel, while recording the sending log for the receipt aggregation and version update module to read; when a candidate path is missing, a fallback path is selected and the source entry is marked, and the original field names of the handling instruction remain unchanged.

[0127] The Receipt Aggregation and Version Update Module 08 is used to collect updates to execution parameters and strategy entries for handling receipts, and to archive them to disk. It generates a model status version tag structure and handling receipts, and writes the model status version tag structure back to the terminology registration and collection configuration module, baseline modeling and correlation determination module, and threshold adaptation and scoring aggregation module for invocation. Specifically, it receives receipts generated from handling instructions in each channel, merges delivery records, confirmation records, and free text records to form handling receipts; updates the parameters of the tiered threshold candidate set and weight records based on the handling receipts and evidence tracing package structure, and updates the strategy entries of the early warning handling strategy library template. The generated file entry is written to the user profile; the above update actions are summarized, a model status version tag structure is generated, the terminology registration and collection configuration module is written back to refresh the sampling and clock strategy, the baseline modeling and correlation determination module is written back to refresh the rolling window parameters of the individual baseline model, and the threshold adaptation and scoring aggregation module is written back to refresh the grading factor selection and weight initialization constraints; the disposal receipt and model status version tag structure are output, the disposal receipt is read back by the path matching and notification delivery module for subsequent path rewriting, and the model status version tag structure drives the next round of collection and determination process closure after the terminology registration and collection configuration module takes effect.

Claims

1. A method for intelligent, unobtrusive blood glucose early warning and management for the elderly at home, characterized in that, include: The system acquires wrist-worn photoplethysmography pulse, skin temperature, skin conductance, activity intensity, environmental context, and user profile; performs terminology registration and sampling period registration, clock reference registration and time alignment, artifact suppression and data repair, feature registration and window boundary registration, and field mapping registration processing to generate individual baseline definitions. Based on the individual baseline definition, blood glucose correlation estimation, short-term trend discrimination and continuous calibration, window update and state determination, consistency verification and field aggregation processing are performed to generate the individual state structure. Based on the individual state structure, the grading factor selection and weight initialization are performed, and the threshold adaptive calculation is performed. The grading strategy set includes fixed grading strategy, sliding grading strategy and situational grading strategy. The boundary adjudication unit is used to prune the strategy according to the strategy priority, the current situation number and the file constraint items to generate a grading threshold candidate set. The score aggregation is performed. For each candidate factor, the factor value of the current window is read and the corresponding grading threshold is matched to complete the grading level matching to generate the score component and aggregate it into a window-level score. The graded organization and evidence collection and trajectory backlink processing are performed to generate a list of early warning events and an evidence tracing package structure. The system acquires the list of early warning events and the evidence tracing package structure, performs path matching and notification generation, task issuance and receipt collection, parameter update and strategy entry update, and file storage operations, and generates disposal instructions, disposal receipts, and model status version tag structures. The list of early warning events includes event categories and classification levels, a list of triggering factors and scenario numbers, and source paths and time paths.

2. The method according to claim 1, characterized in that, Wrist-worn photoplethysmography (PPG) pulse, skin temperature, conductance of skin, activity intensity, environmental context, and user profile include: The wrist-worn photoplethysmography (PPG) is a surface volume pulsation signal measured by an optical channel, used to reflect the waveform information of peripheral blood flow changes with heartbeat; the skin temperature is the output of the skin-touch or near-skin temperature channel, reflecting the surface temperature and rhythm; the skin conductance is the skin conductance activity signal, used to reflect the skin conductance response caused by sympathetic nerve drive; the activity intensity includes acceleration vector magnitude, gait frequency, and body position determination; the environmental context includes indoor temperature and humidity, illuminance, noise level, and geographical location time period labels; the user profile includes basic information, chronic disease history, medication records, diet records, and daily routine.

3. The method according to claim 1, characterized in that, The process of generating individual state structures also includes: The individual state structure is input and aligned in both the time-series and field dimensions to construct a decision-making buffer. The decision-making buffer records the start and end times of the window, the list of participating factors, the initial weight vector, the scenario number, and the file constraint entries, providing a unified entry point for tiered factor selection, adaptive threshold calculation, and score aggregation.

4. The method according to claim 1, characterized in that, The process of selecting tiering factors also includes: The feature factors that can participate in the classification are retrieved from the evidence list in the individual state structure. The feature factors include pulse waveform features, heart interval features, skin temperature fluctuation features, skin conductance response features, and activity intensity features. The source path is traced back to the specific channel and specific window. Read the scene attributes, posture attributes, and time period attributes from the context label sequence, establish the association mapping between factors and contexts, and record them in the factor mapping table; Read user profiles to generate profile constraint entries. The profile constraint entries are marked with three states: disabled factors, reduced weight factors, and enhanced factors. Disabled factors are directly removed from the current round of classification. Reduced weight factors are added to the weight reduction entry during the weight initialization stage, and enhanced factors are added to the weight increase entry, thus obtaining a candidate factor list.

5. The method according to claim 4, characterized in that, Weight initialization includes: Read the set of quality weights in the individual state structure and map the channel quality level to the factor weights; Based on the pose and scene attributes in the context label sequence, adjustment entries are applied to factors strongly correlated with pose and factors strongly correlated with scene. By combining chronic disease history and medication records in user profiles, profile adjustment items are formed. During periods when drug and dietary interventions have an impact, corresponding factors are suppressed or enhanced to obtain an initial weight vector.

6. The method according to claim 1, characterized in that, The process of adaptive threshold calculation also includes: The threshold adaptive calculation is performed using a set of tiered strategies, which includes fixed-tier strategies, sliding-tier strategies, and context-based strategies. The fixed-file strategy targets long-term stable factors and generates initial, intermediate, and high-level boundaries from the long-term target range given in the user profile. The sliding boundary strategy generates sliding boundaries from the version trajectory of the individual state structure for factors with short-term drift. The context-based profile strategy targets factors that are significantly affected by the context, and the profile switching is driven by the context attributes of the context tag sequence. After each of the three types of strategies outputs a candidate boundary, it enters the boundary adjudication unit. The boundary adjudication unit performs pruning according to the strategy priority, the current situation number, and the file constraint entries to generate a tiered threshold candidate set.

7. The method according to claim 1, characterized in that, The process of performing score aggregation also includes: For each candidate factor, read the factor value and corresponding tier threshold of the current window, and generate a score component after completing the tier matching. The score component records the factor name, tier name, matching deviation, and weight contribution. Perform component aggregation within the window and output a window-level score, which includes the scenario number and time path. Read the status name and level name from the individual status structure, merge them with the window-level score to form event candidate entries, and output the warning event when the triggering condition is met by the boundary of the grading threshold candidate set, the file constraint entries and the situation number.

8. The method according to claim 1, characterized in that, The process of hierarchical organization also includes: Read the time path and source path in the event trajectory log, merge consecutive similar events according to their time adjacency, and introduce a grace period in the merging strategy when a short interruption occurs. The merged events are sorted hierarchically. The sorting rules read the key focus items in the user profile and the scenario priority items in the context tag sequence to form a hierarchy, generating a list of warning events arranged in chronological order.

9. The method according to claim 1, characterized in that, The process of hierarchical organization, evidence collection, and trajectory backlink processing also includes: Read back the source path and time path, trace back along the source path to the cleaning sequence and alignment index structure, and obtain the corresponding original fragment, cleaned fragment and alignment timestamp; Extract the feature factor records and state determination records that are directly related to the early warning event from the evidence list of individual state structures, and encapsulate them together with the original fragments and cleaned fragments; Read the scene attributes and posture attributes from the context tag sequence and inject the scene switching points and posture change points into the evidence timeline; The reference window range before and after the event window is determined based on the alignment index structure, and the trajectory points of the retrieved blood glucose correlation estimation results, short-term trend discrimination results and individual baseline models are registered. Read the weight adjustment logs during the weight initialization phase and the boundary adjudication records during the threshold adaptive calculation phase, mark the weight changes and boundary changes on the event timeline, and generate the evidence tracing package structure.

10. A home-based, non-intrusive intelligent blood glucose early warning and management system for the elderly, applied to the method described in any one of claims 1-9, characterized in that, include: The terminology registration and acquisition configuration module is used to register terms, sampling periods, and clock references, generate a terminology dictionary and acquisition configuration definition table, and provide the terminology dictionary and acquisition configuration definition table to the multi-source data acquisition and cleaning alignment module and the feature registration and window mapping module for calling; The multi-source data acquisition and cleaning alignment module is used to acquire wrist-worn photoplethysmography pulse, skin temperature, skin conductance, activity intensity, environmental context and user profile, perform time alignment, artifact suppression, data repair, generate cleaning sequence, alignment index structure and context label sequence, and provide the cleaning sequence, alignment index structure and context label sequence to the feature registration and window mapping module for call; The feature registration and window mapping module is used to perform feature item registration, window boundary registration and field mapping registration under the constraints of the cleaned sequence, aligned index structure and context label sequence, generate feature vector definition set and individual baseline definition, and provide the feature vector definition set and individual baseline definition to the baseline modeling and correlation determination module for calling; The baseline modeling and correlation determination module is used to perform blood glucose correlation estimation, short-term trend discrimination, continuous calibration, window update, state determination, consistency check and field aggregation based on the feature vector definition set and individual baseline definition, and generate blood glucose correlation estimation results, short-term trend discrimination results, individual baseline model and individual state structure, and provide the individual state structure to the threshold adaptation and score aggregation module for calling; The threshold adaptation and score aggregation module is used to read the individual status structure and context label sequence and user profile to perform classification factor selection and weight initialization, threshold adaptation calculation and score aggregation, output warning events and classification threshold candidate set, and provide the warning events to the hierarchical organization and evidence tracing module for call; The hierarchical organization and evidence tracing module is used to process the early warning event set, perform hierarchical organization and evidence collection and trajectory backlink, generate an early warning event list and evidence tracing package structure, and provide the early warning event list and evidence tracing package structure to the path matching and notification distribution module for calling; The path matching and notification distribution module is used to read the list of early warning events and the structure of the evidence tracing package and the template of the early warning handling strategy library, execute path matching and notification generation and task distribution, output handling instructions, and provide the handling instructions to the receipt collection and version update module for calling; The receipt aggregation and version update module is used to collect and process receipt execution parameter updates and strategy entry updates and file storage, generate model status version tag structure and processing receipts, and write the model status version tag structure back to the terminology registration and collection configuration module, baseline modeling and correlation determination module and threshold adaptation and scoring aggregation module for invocation.

Images