Time-series data driven health state dynamic prediction system and method

Abstract

The application discloses a time series data driven health state dynamic prediction system and method, belongs to the technical field of health big data management and time series database, and aims to solve the problems of poor adaptability of fixed reference individuals, redundancy of fixed cycle collection and misjudgment and missing judgment of health risks of existing health monitoring technologies.A whole-cycle health time series database is constructed, health risk preliminary screening is carried out, a health risk assessment data set is constructed for triggered risks, individual health evolution characteristics are extracted and mined, an individual health dynamic baseline is constructed, health risk comprehensive judgment is executed, a real risk triggered health state evolution trend is predicted, and an individual exclusive evolution path, a self-adaptive health data collection sequence and a targeted collection strategy are generated.The application realizes accurate identification and whole life cycle closed loop control of health risks, significantly improves health prediction accuracy and data processing operation efficiency, and is suitable for long cycle, non-steady state and multi-sampling frequency health monitoring scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for time-series data, and more specifically to a time-series data-driven dynamic prediction system and method for health status. Background Technology

[0002] With the increasing demand for dynamic monitoring of healthy individuals throughout their entire life cycle, health status monitoring based on time-series data has become a core technical means for health status management. In actual health monitoring scenarios, the health data of individuals under health monitoring generally exhibit non-stationary characteristics such as multi-source heterogeneity, non-fixed collection cycles, non-fixed collection indicators, and multiple sampling frequencies. Traditional health monitoring methods struggle to achieve efficient collection and standardized processing of such complex health data.

[0003] Most existing health anomaly monitoring technologies use industry-standard fixed thresholds as judgment criteria, which cannot adapt to the differences in characteristics of different individuals under health monitoring, resulting in poor adaptability and accuracy in health status determination. Furthermore, anomaly determination often relies on localized judgments based on single batches of monitoring data, making it difficult to achieve comprehensive identification by combining full-cycle time-series evolution characteristics. This leads to insufficient ability to detect gradual and hidden health anomalies, resulting in a high probability of false positives and false negatives. In addition, existing monitoring methods often employ fixed-cycle data collection and monitoring patterns, which easily miss key nodes in anomaly evolution, generate a large amount of invalid monitoring data, and cause redundant consumption of computing and storage resources. Moreover, they lack the ability to predict and regulate health status based on anomaly evolution patterns, failing to meet the practical needs of high-precision, high-efficiency, and highly adaptable dynamic health monitoring. Therefore, to overcome these limitations, this invention proposes a time-series data-driven dynamic health status prediction system and method. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a time-series data-driven dynamic prediction system and method for health status, which solves the technical problems of improving the accuracy of abnormal health status monitoring and prediction, adapting to individual characteristics of health monitoring, and reducing false positives and false negatives and ineffective monitoring overhead in health monitoring scenarios with multi-source heterogeneous health time-series data and non-fixed collection rules.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Time-series data-driven methods for dynamic prediction of health status include:

[0007] Based on the monitoring data of individuals, a full-cycle health time-series database is constructed, and health risk anomalies are screened in real-time monitoring data to locate abnormal collection indicators and map abnormal health status risks, triggering the health risk assessment process.

[0008] For each abnormal health status risk that triggers the health risk assessment process, historical valid monitoring data of its associated abnormal collection indicators are extracted to construct a health risk assessment dataset. Adaptive multi-scale time series feature extraction is performed on the historical valid monitoring data of the health risk assessment dataset to construct an individual health dynamic baseline for the health monitoring individual.

[0009] The active health risk assessment dataset is dynamically supplemented and updated. Combined with the individual health dynamic baseline, abnormal fluctuation characteristics are identified, the health risk assessment score of the abnormal health status risk corresponding to each collection batch is quantified, and a time series sequence of health risk assessment scores is constructed. This is used to perform a comprehensive health risk existence determination of abnormal health status risk, so as to close the update channel of the health risk assessment dataset where abnormal health status risk does not exist, and trigger the prediction of health status evolution trend.

[0010] When the prediction of health status evolution trend is triggered, an individual health status evolution path is constructed based on the standard health status evolution path, thereby generating an adaptive monitoring time point sequence of abnormal health status risk to guide the targeted collection of health data.

[0011] Specifically, the steps for constructing an individual health dynamic baseline for health monitoring include:

[0012] Based on the pre-defined correlation mapping between the collection indicators and the risk of abnormal health status, the associated collection indicators corresponding to each abnormal health status risk are extracted to form a collection indicator set.

[0013] From the full-cycle health time-series database, retrieve all historical valid monitoring data corresponding to all collected indicators in the collection indicator set to construct the basic data source for the health risk assessment dataset.

[0014] Based on the set of collected indicators and basic data sources corresponding to abnormal health status risks, a health risk assessment dataset is initially constructed.

[0015] Extract the sampling frequency of each collection indicator in the health risk assessment dataset and classify each collection indicator hierarchically; perform adaptive multi-scale sliding window partitioning for historical effective monitoring data of different hierarchical classifications, and perform adaptive multi-scale time series feature extraction to generate a multi-scale time series feature set for each associated collection indicator, and then construct an individual time series feature set;

[0016] Based on the individual time-series feature set, the feature benchmark interval and adaptive deviation threshold corresponding to each associated collection indicator are calculated to construct the individual health dynamic baseline.

[0017] Specifically, the steps for constructing an individual time-series feature set include:

[0018] Based on a multi-scale sliding window, basic feature extraction operations are performed on the historical valid monitoring data of each associated collection indicator to generate a single sliding window basic feature set corresponding to each sliding window; the basic features include time-domain statistical features, frequency-domain features and nonlinear dynamic features.

[0019] For each associated collection indicator, a weighted weight is set for each sliding window of the basic feature set of all single sliding windows. The basic feature sets under the same collection indicator are then fused through weighted fusion to generate a multi-scale time series feature set.

[0020] Based on the historical collection timestamps of all associated collection indicators in the multi-scale temporal feature set, a temporal alignment correlation matrix is ​​established to mine the accompanying change relationship, causal relationship, and temporal coupling relationship among the multi-scale temporal features of different collection indicators, and to generate a cross-indicator correlation feature set.

[0021] For all associated collected indicators, corresponding multi-scale time-series feature sets and cross-indicator associated feature sets, a standardized integration operation of all features is performed to generate individual time-series feature sets.

[0022] Specifically, the steps for performing adaptive multi-scale sliding window partitioning on historical valid monitoring data for different classification levels include:

[0023] Using the minimum or average sampling interval of each sampling index under each sampling frequency level as the basic sliding window unit, multiple rounds of expansion are performed upwards according to a preset multiple gradient to generate multiple sliding windows with different time spans, forming a multi-scale sliding window group.

[0024] Based on the basic sliding window unit of the corresponding level, and according to the time-series fluctuation amplitude of the historical valid monitoring data of each collected indicator within the level, a preset mapping relationship between the time-series fluctuation amplitude and the step size adjustment coefficient is established, and the step size of each sliding window corresponding to each collected indicator within the level is set respectively; and the overlap rate of the sliding window is adaptively determined according to the ratio of the step size to the time span of the corresponding sliding window, thus forming an adaptive multi-scale sliding window for each collected indicator.

[0025] Specifically, the steps for locating abnormal data collection indicators, mapping abnormal health status risks, and triggering the health risk assessment process include:

[0026] Based on the unique identifier of the health monitoring individual, an initial full-cycle health time-series database of the health monitoring individual is established. The monitoring data collected in each batch is classified, archived, and preprocessed. The valid monitoring data collected in each batch is stored in the full-cycle health time-series database.

[0027] The real-time values ​​of each collected indicator in a single batch of real-time effective monitoring data are compared one by one with the corresponding safety benchmark interval to identify abnormal collected indicators; if at least one abnormal collected indicator is generated in this batch, the process of associating abnormal health status risk with mapping is initiated.

[0028] The system retrieves the pre-defined mapping relationship between the collection indicators and the abnormal health status risks. For each abnormal collection indicator, it matches and retrieves all associated abnormal health status risks, generates the set of abnormal health status risks corresponding to this batch, and verifies one by one whether each abnormal health status risk has a corresponding active assessment process.

[0029] If there is no active assessment process for the abnormal health status risk, a unique assessment process identifier is generated to trigger the health risk assessment process.

[0030] Specifically, the steps for quantifying the health risk assessment score for the abnormal health status corresponding to each collection batch include:

[0031] For the health risk assessment dataset that has been dynamically updated, newly supplemented real-time effective monitoring data is extracted. For each collected indicator associated with the abnormal risk of this health status, adaptive multi-scale time series feature extraction is performed to generate a real-time multi-scale time series feature set.

[0032] For each feature dimension in the real-time multi-scale time series feature set, it is compared one by one with the feature benchmark interval of the corresponding feature dimension in the individual health dynamic baseline. The relative deviation between the real-time feature value of each feature dimension and the feature benchmark interval is calculated. Combined with the adaptive deviation threshold of the corresponding feature dimension, it is determined whether the deviation of each feature dimension exceeds the normal fluctuation range.

[0033] If the fluctuation exceeds the normal range, the feature dimension is marked as an abnormal fluctuation feature and included in the abnormal feature candidate set; otherwise, the feature dimension is marked as a normal fluctuation feature and included in the normal feature set.

[0034] For each collected indicator, the proportion of the number of feature dimensions included in the abnormal feature candidate set, the maximum relative deviation value of a single feature, and the weighted average of the relative deviation values ​​of all feature dimensions are calculated to generate the single indicator abnormality quantification value corresponding to the collected indicator.

[0035] The abnormal quantitative values ​​of each collected indicator associated with abnormal health status risk are standardized and normalized; and a weighted summation average is calculated based on the preset associated weights of each collected indicator to generate the health risk assessment score for the abnormal health status risk corresponding to this batch of data collection.

[0036] Specifically, the steps for conducting a comprehensive health risk assessment to determine the presence of abnormal health status include:

[0037] The health risk assessment scores calculated for each abnormal health status are sorted to construct a time series sequence of health risk assessment scores corresponding to the abnormal health status risk. It is then determined whether the actual time span of the current health risk assessment score time series sequence is less than the preset minimum assessment period. If so, the current comprehensive health risk existence determination for the corresponding abnormal health status risk is terminated; otherwise, the comprehensive health risk existence determination operation for the corresponding abnormal health status risk is initiated.

[0038] If the health risk assessment score for each health status abnormality risk corresponding to this batch of data collection is greater than or equal to the single abnormality threshold, it is marked as meeting the single abnormality significance standard, and the abnormality risk of that health status is determined to exist; otherwise, it is marked as not meeting the single abnormality significance standard.

[0039] For health status abnormality risks where the significance of a single abnormality does not meet the standard, the corresponding health risk assessment score time series is subjected to standardized preprocessing, and time series evolution features are extracted to construct a full-cycle time series evolution feature set; the time series evolution features refer to the core quantitative feature set that characterizes the trend of changes in the abnormality risk status and the abnormality persistence characteristics of the corresponding health monitoring individual.

[0040] If the preset continuous anomaly determination requirements are retrieved and the full-cycle time-series evolution feature set does not meet the continuous anomaly determination requirements, then the abnormal risk of the health state is determined to be non-existent. The life cycle termination status of the health risk assessment dataset bound to the abnormal risk of the health state is set, and the time-series data update channel of the health risk assessment dataset is closed. Otherwise, the abnormal risk of the health state is determined to exist, and the health state evolution trend prediction is triggered.

[0041] Specifically, the standard health status evolution path refers to a standardized time-series path that uses the health risk assessment score of the corresponding health status abnormal risk as a quantitative indicator to characterize the universal evolution law of the abnormal risk of such health status throughout its entire cycle; it is used to specify the standard evolution stage of abnormal risk of each health status, as well as the standard range of health risk assessment score, the standard range of stage duration, the standard evolution rate, and the standard core abnormal indicator set for each standard evolution stage, while clearly distinguishing the standard abnormal critical threshold between normal and abnormal states.

[0042] Specifically, the steps for constructing the evolutionary path of an individual's health monitoring state include:

[0043] The standard health status evolution path for various health status abnormal risks is pre-constructed. For the health status abnormal risks that are identified as existing, the time series sequence of the health risk assessment score corresponding to the health status abnormal risk is retrieved to extract the current health risk assessment score and historical evolution rate of the health status abnormal risk, and the current core abnormal indicator set of the health status abnormal risk is obtained.

[0044] The current health risk assessment score is matched one by one with the standard range of health risk assessment scores for each standard evolution stage in the standard health status evolution path to determine the current anchoring time node and corresponding standard evolution stage of the health monitoring individual in the standard health status evolution path.

[0045] The ratio of the historical abnormal evolution rate of an individual under health monitoring to the standard evolution rate of the corresponding standard evolution stage is used as the individual rate scaling factor to linearly scale the standard range of the stage duration of the standard health state evolution path.

[0046] Specifically, the steps for constructing the evolutionary path of an individual's health monitoring state also include:

[0047] Using the final determined anchor time node as the prediction starting point, and the duration of each standard evolution stage after linear scaling as the time scale benchmark, combined with the standard evolution rate of the corresponding standard evolution stage and the standard range of health risk assessment scores, the initial state evolution path of the health monitoring individual is generated.

[0048] The difference between the individual single abnormality threshold set for the health monitoring individual as the corresponding abnormality risk and the standard abnormality threshold in the standard health state evolution path is used as the core translation correction amount. The standard interval of health risk assessment score in each standard evolution stage in the initial state evolution path is linearly translated with equal offset to obtain the state evolution path of the health monitoring individual.

[0049] A time-series data-driven dynamic prediction system for health status includes an anomaly screening module, a baseline generation module, a comprehensive assessment module, and a trend prediction module.

[0050] The anomaly screening module is used to construct a full-cycle health time-series database based on the monitoring data of individuals under health monitoring, and to screen for health risk anomalies in real-time monitoring data, locate abnormal collection indicators and map abnormal health status risks, and trigger the health risk assessment process.

[0051] The baseline generation module is used to extract historical valid monitoring data of associated abnormal collection indicators for each health status abnormal risk that triggers the health risk assessment process, construct a health risk assessment dataset, and construct an individual health dynamic baseline for the health monitoring individual by adaptively extracting multi-scale temporal features from the historical valid monitoring data of the health risk assessment dataset.

[0052] The comprehensive assessment module is used to dynamically supplement and update the health risk assessment dataset that is in an active state. Combined with the individual health dynamic baseline, it identifies abnormal fluctuation characteristics, quantifies the health risk assessment score of the abnormal health status risk corresponding to each collection batch, and constructs a time series sequence of health risk assessment scores. This is used to perform a comprehensive health risk existence determination of abnormal health status risk, so as to close the update channel of the health risk assessment dataset where abnormal health status risk does not exist, and trigger the prediction of health status evolution trend.

[0053] The trend prediction module is used to construct an individual health status evolution path based on the standard health status evolution path when the health status evolution trend prediction is triggered, and then generate an adaptive monitoring time point sequence of abnormal health status risk to guide the targeted collection of health data.

[0054] The beneficial effects of this invention are:

[0055] This application constructs a full-cycle health time-series database for individuals under health monitoring and performs anomaly screening on real-time health monitoring data, triggering a health status anomaly risk assessment process on demand. This effectively avoids invalid assessment calculations in the absence of anomaly data and reduces redundant computing power consumption. By constructing a dedicated health risk assessment dataset for each health status anomaly risk triggering the health risk assessment process, and employing an adaptive multi-scale time-series feature extraction method to mine individual evolutionary features to construct an individual health dynamic baseline for the monitored individuals, this approach can fully adapt to the characteristics of multi-source heterogeneous and non-fixed collection rules of health data and the individual differences of the monitored individuals, significantly improving the adaptability and accuracy of health status determination. Combining the single health risk assessment score with the full-cycle data... The time-series evolution characteristics are used to perform a comprehensive existence determination of abnormal health status risks. This method can overcome the shortcomings of single local determinations and effectively reduce the probability of misjudgment and missed judgment of gradual and hidden anomalies. At the same time, it can accurately close the update channel of the health risk assessment dataset without anomalies, and realize closed-loop management of the entire life cycle of the assessment process. Based on the pre-constructed standard health status evolution path, it generates a health monitoring individual-specific status evolution path, and generates an adaptive monitoring time point sequence according to the stage switching node to guide the targeted collection of health data. This can avoid the problems of missed collection of key evolution nodes and redundant invalid monitoring data in fixed-period monitoring, so that the monitoring rhythm is highly matched with the health anomaly evolution law, and further improve the efficiency and reliability of dynamic health monitoring. Attached Figure Description

[0056] Figure 1 This is a flowchart of the time-series data-driven dynamic prediction method for health status according to the present invention;

[0057] Figure 2 A flowchart for constructing an individual health dynamic baseline for health monitoring in this invention;

[0058] Figure 3 This is a flowchart illustrating the health risk assessment score for quantifying the risk of abnormal health status corresponding to each data collection batch in this invention.

[0059] Figure 4 This is a flowchart illustrating the comprehensive existence determination of health risks associated with abnormal health status in this invention. Detailed Implementation

[0060] Example 1

[0061] Please see Figure 1 This embodiment introduces a time-series data-driven dynamic prediction method for health status, including:

[0062] Step S1: Unify, clean, and standardize multi-source heterogeneous monitoring data from the entire lifespan of health monitoring individuals, including data collected over non-fixed periods, with non-fixed indicators and at different sampling frequencies. This will build a full-cycle health time-series database for subsequent anomaly screening, baseline construction, and status assessment. Furthermore, rapidly screen newly collected real-time monitoring data within this database to avoid invalid full-scale assessment calculations without anomalies. Simultaneously, locate abnormal indicators and map associated health status anomaly risks, triggering the health risk assessment process. This step aims to construct a structured full-cycle health time-series database covering the entire lifespan of health monitoring individuals by unifying, cleaning, and standardizing multi-source heterogeneous non-steady-state monitoring data. Simultaneously, it performs rapid anomaly screening on newly collected real-time monitoring data, avoiding invalid full-scale assessment calculations without anomalies and accurately triggering the health status anomaly risk assessment process with abnormal data. This provides a reliable data foundation and prerequisite triggering conditions for subsequent accurate assessment of abnormal states and prediction of evolutionary trends throughout the entire process.

[0063] In this embodiment, the health monitoring individual refers to the object through which full life cycle status information collection, dynamic feature analysis, and status evolution trend monitoring can be carried out through multi-source time-series monitoring data. This includes, but is not limited to, human subjects in the TCM health status time-series monitoring scenario for non-diagnostic and treatment purposes, as well as key power equipment and rotating machinery equipment in the industrial equipment operation and maintenance scenario. Specifically, the human subjects in the TCM health status time-series monitoring scenario include subjects for dynamic identification of TCM constitution status, subjects for time-series monitoring of sub-health status, subjects for dynamic tracking of TCM physiological status, and subjects for time-series evaluation of TCM constitution conditioning effects.

[0064] For example, when the health monitoring of individuals is for the purpose of TCM health status time-series monitoring for non-diagnostic and treatment purposes, and for the dynamic identification of the TCM constitution status of the subjects, the collected indicators include, but are not limited to: high-frequency time-series collected TCM pulse diagnosis quantitative characteristic indicators, specifically quantifiable time-domain and frequency-domain characteristic indicators such as pulse rate, pulse strength, pulse rhythm, superficiality, slowness, smoothness, and tightness of the six pulse positions (cun, guan, chi); non-fixed-period collected TCM tongue appearance quantitative characteristic indicators, specifically quantifiable visual characteristic indicators such as tongue color, tongue coating thickness, tongue coating moisture, tongue coating color, tongue shape, and the fullness and tortuosity of the sublingual veins; non-fixed-period collected... The study includes quantitative indicators for TCM facial diagnosis, specifically quantifiable visual features such as facial luster, facial color classification, lip color, and periorbital condition; time-series data on TCM physical sensations reported by subjects as needed, including the presence, severity, and duration of physical sensations such as dizziness, headache, palpitations, tinnitus, lower back and knee weakness, insomnia, and irritability; scores from a TCM constitution identification scale collected at fixed or non-fixed periods, and quantitative scores for TCM constitution characteristic elements; and supporting quantitative indicators of physiological signs at multiple sampling frequencies, specifically resting heart rate, daily systolic and diastolic blood pressure at multiple time points, fasting blood glucose, four blood lipids, and biochemical indicators related to liver and kidney function. When the health monitoring individual is the main drive gearbox of a coal mine hoist in an industrial equipment operation and maintenance scenario, the collected indicators include, but are not limited to: high-frequency continuous acquisition of vibration acceleration signals and spindle speed signals at various measuring points of the gearbox; non-fixed periodic acquisition of lubricating oil supply and return temperatures, oil contamination levels, and drive motor operating current; and periodic inspection acquisition of equipment operating noise, gear meshing clearance, and bearing clearance indicators.

[0065] In one implementation, the specific steps for locating abnormal data collection indicators and mapping them to associated health status abnormalities to trigger the health risk assessment process include:

[0066] Based on the unique identifier of each health monitoring individual, a full-cycle health time-series database for that individual is initialized. Monitoring data collected in each batch is categorized and archived according to collection timestamp, data indicator type, and sampling frequency. Data preprocessing is then performed to obtain the valid monitoring data for each batch and store it in the full-cycle health time-series database for each individual. It should be noted that data preprocessing includes: adaptive cleaning and completion of missing values, abnormal noise values, and outliers in the monitoring data; for monitoring data with different sampling frequencies, unified time granularity alignment and dimensional standardization are performed based on the collection timestamp to eliminate data heterogeneity between different sampling frequencies and indicators with different dimensions; and a dynamic field expansion mechanism is configured for the full-cycle health time-series database to adapt to the addition and deletion of collection indicators, addressing the dynamic adjustment needs of non-fixed collection indicators.

[0067] It continuously receives monitoring data reported by various collection terminals throughout the entire life cycle of the health monitoring individual. For each batch of received monitoring data, it performs classification, archiving, and data preprocessing operations consistent with the initialization phase. The generated valid monitoring data is synchronously archived into the full-cycle health time-series database of the health monitoring individual, completing the dynamic and continuous updating of the full-cycle health time-series database. At the same time, it generates a unique batch identifier and global time-series index for each batch of valid monitoring data that has been archived. It synchronously retrieves the newly collected and archived single batch of real-time valid monitoring data and enters the subsequent anomaly screening process.

[0068] The system retrieves all collected indicators corresponding to the newly collected batch of real-time valid monitoring data. Simultaneously, it retrieves the preset safety benchmark intervals corresponding to each collected indicator. The safety benchmark interval is the range of state values ​​applicable to the same type of monitored object within the industry field to which the corresponding collected indicator belongs. The system compares the real-time values ​​of each collected indicator in the batch of real-time valid monitoring data with the corresponding safety benchmark interval one by one to determine whether the real-time value exceeds the corresponding safety benchmark interval. Collected indicators that exceed the corresponding safety benchmark interval are marked as abnormal collected indicators. The batch identifier, collection timestamp, and real-time value corresponding to the abnormal collected indicator are recorded simultaneously. If no abnormal collected indicators are generated in this batch, the subsequent evaluation process of this batch of data is terminated, and only data archiving and storage are completed. If at least one abnormal collected indicator is generated in this batch, the system proceeds to the subsequent health status abnormal risk association mapping process.

[0069] For all abnormal data collection indicators marked in this batch, the pre-defined many-to-many association mapping relationship between the data collection indicators and abnormal health status risks is retrieved. This mapping relationship is pre-constructed through domain knowledge sorting, historical data association mining, and expert verification and iteration, clarifying the corresponding binding relationship between each data collection indicator and all abnormal health status risks with causal and accompanying associations. For each abnormal data collection indicator, all abnormal health status risks with associated relationships are matched and retrieved to generate a set of abnormal health status risks corresponding to this batch. The set of abnormal health status risks is then bound and stored in the full-cycle health time series database along with the corresponding batch identifier and abnormal data collection indicator information.

[0070] For each abnormal health status risk in the set of abnormal health status risks, all active assessment process information corresponding to the individual being monitored is retrieved from the full-cycle health time-series database. Each abnormal health status risk is checked to see if there is an active assessment process. If there is an active assessment process for the abnormal health status risk, the triggering operation for a new assessment process for that abnormal health status risk is terminated, and only the corresponding abnormal collection indicator information is synchronously pushed to the active assessment process for dynamic updating of the corresponding health risk assessment dataset. If there is no active assessment process for the abnormal health status risk, the subsequent trigger condition verification process is initiated.

[0071] For health status abnormal risks that do not have an activated assessment process, a unique assessment process identifier is generated for the health status abnormal risk, and the assessment process corresponding to the health status abnormal risk is launched simultaneously. At the same time, the assessment process identifier, the launch time and the corresponding abnormal collection indicator information are bound and stored in the full-cycle health time series database.

[0072] Step S2: For each abnormal health status risk that triggers the health risk assessment process, extract historical valid monitoring data of its associated abnormal collection indicators to construct a health risk assessment dataset. This provides targeted data support for subsequent individual baseline construction and status assessment, avoiding data interference between different abnormal situations. For each abnormal health status risk corresponding to the health risk assessment dataset, construct an individual health dynamic baseline for the monitored individual through adaptive multi-scale temporal feature extraction. This baseline adapts to the individual characteristics of the monitored individual and to data at different sampling frequencies, abandoning the static judgment logic of fixed thresholds, and providing accurate data for subsequent deviation analysis and status assessment. The reference benchmark is used to construct a standardized health risk assessment dataset with complete indicator dimensions and unified data sources for each abnormal health status risk triggered on demand in step S1. This avoids data cross-interference when assessing multiple abnormal scenarios in parallel. At the same time, by adapting to the adaptive multi-scale time series feature extraction of time series data with different sampling frequencies, it fully explores the individual evolution characteristics and gradual feature changes in long-term time series data, and constructs a dynamic baseline that fits the individual characteristics of health monitoring individuals. This solves the problems of poor adaptability of general fixed benchmarks, insufficient mining of individual time series features, low feature extraction accuracy of data with different sampling frequencies, and inability to capture gradual state changes in existing technologies.

[0073] In this embodiment, the health risk assessment dataset, which is one-to-one bound to the abnormal health status risk, refers to a structured dataset that contains only all the relevant collected indicators for the corresponding abnormal health status risk and covers the entire historical valid monitoring data of the individual throughout the health monitoring cycle. It is uniquely bound to the assessment process identifier generated in step S1 and has an independent lifecycle management mechanism and data update channel. For example, when the health monitoring individual is a subject undergoing dynamic identification of TCM constitution status in a non-diagnostic / treatment-oriented TCM health status time-series monitoring scenario, and the abnormal health status risk triggering the health risk assessment process is a Qi deficiency constitution state shift, the health risk assessment dataset corresponding to this abnormal health status risk contains all the collected indicators associated with the Qi deficiency constitution state shift. Historically valid monitoring data specifically includes quantitative characteristic indicators of TCM pulse diagnosis collected at high frequency, quantitative characteristic indicators of TCM tongue appearance collected at non-fixed periods, quantitative characteristic indicators of TCM facial diagnosis, time-series data of TCM body sensation status self-reported by subjects, scores of TCM constitution identification scale, quantitative scores of TCM constitution characteristic elements, and corresponding quantitative indicators of physiological signs. For this health risk assessment dataset, an adaptive multi-scale time-series feature extraction method is used to perform multi-scale feature mining on the high-frequency sampled pulse diagnosis time-series data, the non-fixed period sampled tongue and facial diagnosis indicators, and the discretely reported body sensation status data. This constructs an individual health dynamic baseline for the subject's deviation from Qi deficiency constitution, providing an individual reference benchmark for subsequent assessment of the degree of constitution deviation. When the health monitoring individual is the main drive gearbox of a coal mine hoist in an industrial equipment operation and maintenance scenario, and the health status abnormality risk that triggers the health risk assessment process is the abnormal lubrication status of the gearbox, the corresponding health risk assessment dataset contains historical valid monitoring data of all collected indicators related to the abnormal lubrication status of the gearbox. Specifically, this includes high-frequency continuously collected vibration acceleration signals and spindle speed signals at various measuring points of the gearbox, non-fixed-period collected lubricating oil supply and return temperatures, oil contamination degree, and drive motor operating current, as well as equipment operating noise, gear meshing clearance, and bearing clearance indicators collected during regular inspections. For this health risk assessment dataset, an adaptive multi-scale temporal feature extraction method is used to perform multi-scale feature mining on the high-frequency sampled vibration temporal data, non-fixed-period collected oil temperature and oil indicators, and discrete inspection clearance data, respectively, to construct an individual health dynamic baseline for the main drive gearbox of the coal mine hoist in response to the abnormal lubrication status, providing an individual reference benchmark for subsequent assessment of the degree of abnormal lubrication status.

[0074] Please see Figure 2 In one embodiment, the specific steps for constructing an individual health dynamic baseline for a health monitoring individual include:

[0075] For each health status abnormality risk that initiates the assessment process, based on the preset many-to-many association mapping relationship between the collection indicators and the health status abnormality risk, all associated collection indicators corresponding to the health status abnormality risk are extracted to form a collection indicator set for the health status abnormality risk. Then, from the full-cycle health time-series database of the health monitoring individual, all historical valid monitoring data corresponding to all collection indicators in the collection indicator set are retrieved to construct the basic data source for the health risk assessment dataset.

[0076] Based on the collection of indicators and basic data sources corresponding to abnormal health status risks, a health risk assessment dataset is initialized and constructed, which is bound one-to-one with the abnormal health status risk. A unique identifier consistent with the corresponding assessment process identifier is assigned to the health risk assessment dataset, and the binding relationship between the health risk assessment dataset and the corresponding abnormal health status risk and health monitoring individual is established. At the same time, a dynamic field expansion mechanism and time series data update channel are configured for the health risk assessment dataset to adapt to the dynamic adjustment of associated collection indicators and the continuous supplementation of subsequent effective monitoring data. Finally, the initialized health risk assessment dataset is archived and stored in the full-cycle health time series database of the health monitoring individual.

[0077] The sampling frequency of each collection indicator in the health risk assessment dataset is extracted, and each collection indicator is classified hierarchically. Adaptive multi-scale sliding window partitioning is performed on the historical valid monitoring data of different hierarchical classifications. The sampling frequency parameter is the number of valid data collections of the corresponding collection indicator per unit time. According to the order of sampling frequency from high to low, all related collection indicators are divided into multiple levels based on the preset hierarchical threshold. The hierarchical threshold is a preset sampling frequency interval threshold used to divide different sampling frequency levels. This threshold is preset based on the conventional sampling frequency distribution characteristics of various collection indicators in the corresponding application scenario. It can be dynamically adapted and adjusted according to the monitoring scenario and collection terminal characteristics of the health monitoring individual. The sampling frequency intervals corresponding to adjacent levels do not overlap and are fully covered, ensuring that all related collection indicators can be uniquely classified into the corresponding sampling frequency level.

[0078] It should be noted that adaptive multi-scale sliding window partitioning includes:

[0079] Using the minimum or average sampling interval of each collected indicator under each sampling frequency level as the basic sliding window unit, multiple rounds of expansion are performed upwards according to a preset multiple gradient to generate multiple sliding windows with different time spans. This forms a multi-scale sliding window group covering a certain proportion of the overall time span of the entire historical data from the basic sliding window unit to the benchmark indicator at that level, thus completing the basic framework for the general sliding window time span at that level. The preset multiple gradient refers to a preset multiple sequence used to expand upwards from the basic sliding window unit to generate sliding windows with different time spans. The multiples in this sequence are arranged in ascending order, with a fixed interval between adjacent multiples. The preset multiple gradient is comprehensively set based on the feature evolution time scale of abnormal health status risk in the corresponding application scenario, the sampling frequency characteristics of the corresponding level benchmark indicator, and the historical data time span of the individuals under health monitoring. It can be dynamically optimized and adjusted based on the feedback of the effectiveness of subsequent feature extraction, ensuring that the generated multi-scale sliding window group can cover the full-time-scale feature capture needs from short-term subtle feature changes to long-term trend evolution.

[0080] Based on the basic sliding window unit of the corresponding level, and according to the temporal fluctuation amplitude of the historical valid monitoring data of each collected indicator within the level, a preset mapping relationship between the temporal fluctuation amplitude and the step size adjustment coefficient is established. The step size of each sliding window corresponding to each collected indicator within the level is set respectively. Then, the overlap rate of the sliding window is adaptively determined according to the ratio of the step size to the time span of the corresponding sliding window, ensuring that there is continuous temporal data overlap between adjacent sliding windows. Finally, combined with the time span framework of the general multi-scale sliding window group of the level, an adaptive multi-scale sliding window that is exclusive to each collected indicator and adapted to its own temporal characteristics is formed. The preset mapping relationship is pre-constructed based on the historical time-series fluctuation characteristics of similar monitoring indicators in the corresponding application scenarios. The time-series fluctuation amplitude and the step size adjustment coefficient are monotonically negatively correlated: the coefficient of variation or the proportion of extreme value difference of the historical effective monitoring data of the collected indicators is used as the quantitative indicator of the time-series fluctuation amplitude and normalized to the [0,1] interval. The step size adjustment coefficient takes values ​​in the range of (0,1], which satisfies the monotonically negative correlation law that decreases as the fluctuation amplitude increases and increases as the fluctuation amplitude decreases. When the time-series fluctuation amplitude increases, a smaller step size adjustment coefficient is selected, and the step size is reduced by multiplying the adjustment coefficient by the basic sliding window unit. When the time-series fluctuation amplitude decreases, a larger step size adjustment coefficient is selected, and the step size is increased by multiplying the adjustment coefficient by the basic sliding window unit. This realizes the adaptive dynamic adjustment of the step size according to the time-series fluctuation characteristics of a single indicator.

[0081] For each collection indicator, the historical valid monitoring data is divided into multi-scale sliding windows. Adaptive multi-scale time series feature extraction is performed to generate a multi-scale time series feature set for each associated collection indicator, and then the individual time series feature set corresponding to the health risk assessment dataset is constructed.

[0082] It should be noted that the specific steps for constructing the individual time-series feature set corresponding to the health risk assessment dataset include:

[0083] For each historical valid monitoring data of the associated acquisition index after multi-scale sliding window division, basic feature extraction is performed. Basic features refer to a set of core features that comprehensively characterize the intrinsic properties and evolutionary features of time-series data from different dimensions. These features are adaptable to non-steady-state time-series data with different sampling frequencies and acquisition periods, including time-domain statistical features, frequency-domain features, and nonlinear dynamic features. A single sliding window basic feature set is generated for each sliding window, achieving full coverage extraction of basic features for time-series data within sliding windows of different scales. For example, for the time-series data within each sliding window, time-domain statistical features characterizing the data's central trend, dispersion, and distribution characteristics are first extracted. Then, frequency-domain features characterizing the data's frequency distribution and energy changes are extracted using time-frequency domain transformation methods. Finally, nonlinear dynamic features characterizing the data's temporal evolution complexity and chaotic characteristics are extracted using nonlinear dynamic analysis methods.

[0084] For each associated data collection indicator, a multi-scale feature fusion operation is performed on all single sliding window basic feature sets. First, for the basic feature sets corresponding to sliding windows with different time spans, weighted weights matching the sliding window scale are set based on the sliding window's time scale, the proportion of effective data points within the sliding window, and the discriminative power of the features within the sliding window. The sliding window's time scale refers to the time span covered by the window, and its weight is set based on the degree of matching between the corresponding time scale and the evolution cycle of abnormal health status risk features. The degree of matching is obtained by normalizing the overlap between the time scale and the evolution cycle; a higher degree of matching results in a higher weight. The proportion of effective data points within the sliding window refers to the ratio of the number of effective data points within the window to the theoretical maximum number of data points; its weight is set based on this ratio; a higher ratio results in a higher weight. The discriminative power of the features within the sliding window refers to the sensitivity of the features to changes in the target object's state, obtained by normalizing between-class variance, within-class variance, or mutual information; a higher degree of sensitivity results in a higher weight. The weighted fusion method is then used to fuse the basic feature sets of all sliding windows under the same data collection indicator to generate a multi-scale time-series feature set corresponding to that data collection indicator.

[0085] For the multi-scale time-series feature sets corresponding to all associated collection indicators, cross-indicator time-series correlation feature mining is performed. First, a time-series alignment correlation matrix for all indicators is established based on the historical collection timestamps of all associated collection indicators. Then, based on the time-series alignment correlation matrix, the accompanying change relationships, causal relationships, and time-series coupling relationships among the multi-scale time-series features of different collection indicators are mined, generating a cross-indicator correlation feature set that characterizes the degree of feature correlation between different collection indicators. Accompanying change relationships refer to the synchronous and unidirectional or synchronous and inverse linkage change characteristics of the multi-scale time-series features of different collection indicators within the same time interval, used to characterize the degree of synchronous correlation between different indicator features. The multi-scale time-series features of different collection indicators are mapped to a unified time-series baseline interval through the time-series alignment correlation matrix. The time-series correlation coefficient between the feature sequences of different indicators is calculated. Combining the synchronous time difference and the matching degree of the change amplitude of feature changes, the degree of correlation of accompanying change relationships is quantified. Causal correlation refers to the temporal driving and driven relationship between multi-scale time-series features of different collected indicators. That is, a change in the feature of one indicator will trigger a corresponding change in the feature of another indicator in a subsequent time interval. This is used to characterize the strength and significance of the causal influence between different indicator features. Temporal lag analysis is performed on the multi-scale time-series feature sequences of different collected indicators using a temporal alignment correlation matrix. Combined with sliding window temporal traversal, the causal driving effect of different indicator feature sequences under different lag orders is verified. The significance and strength of the causal correlation between different indicators are calculated, quantifying the degree of correlation. Temporal coupling refers to the nonlinear co-evolutionary characteristics of multi-scale time-series features of different collected indicators during long-term time-series evolution. This is used to characterize the long-term evolutionary matching degree and mutual constraint relationship of different indicator features across the entire time-series dimension. Full-cycle nonlinear coupling analysis is performed on the multi-scale time-series feature sequences of different collected indicators using a temporal alignment correlation matrix. The nonlinear mutual information value and coupling synchronization coefficient between different indicator feature sequences are calculated, quantifying the degree of correlation of the temporal coupling relationship.

[0086] For all associated collection indicators corresponding to multi-scale time-series feature sets and cross-indicator associated feature sets, a full-scale feature standardization and integration operation is performed. First, a uniform dimensionless standardization process is performed on all feature dimensions to eliminate the dimensional differences between different feature dimensions, generating a full set of individual time-series features covering single-indicator multi-scale time-series features and cross-indicator associated features. At the same time, a unique feature identifier is set for each feature dimension in the individual time-series feature set, and a mapping relationship is established between feature dimensions and corresponding collection indicators and corresponding sliding window scales, completing the final construction of the individual time-series feature set.

[0087] Based on the individual time-series feature set corresponding to the health risk assessment dataset, the feature benchmark interval and adaptive deviation threshold corresponding to each associated collection indicator are calculated. An individual health dynamic baseline bound to the abnormal risk of the health status is constructed for the health monitoring individual. At the same time, the generated individual health dynamic baseline is strongly bound to the unique data identifier of the corresponding health risk assessment dataset and the corresponding assessment process identifier, and archived and stored in the full-cycle health time-series database of the health monitoring individual. In addition, a dynamic adjustment mechanism is configured for the individual health dynamic baseline to be updated synchronously with the health risk assessment dataset.

[0088] It should be noted that the feature baseline interval refers to the range of fluctuations in the state values ​​of each feature dimension in the corresponding individual time-series feature set, generated based on the individual's own historical valid monitoring data. By fitting the probability distribution of the full historical feature values ​​of each feature dimension in the individual time-series feature set, the distribution confidence interval of the corresponding feature dimension is determined, and the individual feature baseline interval corresponding to each feature dimension is obtained.

[0089] An adaptive deviation threshold is a critical threshold used to determine whether the deviation between an individual's real-time characteristics and their dynamic health baseline exceeds a certain range. This critical threshold can be adaptively adjusted according to the temporal fluctuation characteristics of the feature dimension, the data volume, and the historical deviation distribution characteristics. By statistically analyzing the historical deviation values ​​of each feature dimension in the individual's temporal feature set, and combining the correlation weight between the corresponding feature dimension and the risk of abnormal health status (i.e., based on the correlation degree between the collected indicators of the feature dimension and the risk of abnormal health status), the basic threshold obtained from the historical deviation statistics is scaled with correlation sensitivity to calculate the initial deviation threshold for each feature dimension. Then, based on the feature deviation distribution corresponding to the newly added valid monitoring data, the initial deviation threshold is dynamically iteratively optimized. That is, by statistically analyzing the newly added deviation distribution through a sliding window, when the distribution difference exceeds the initial deviation threshold, it is updated by weighting the historical weights and the new distribution statistics, so that the initial deviation threshold can be adaptively adjusted according to the individual's status changes, and finally, the adaptive deviation threshold corresponding to each feature dimension is obtained.

[0090] Step S3: For each health risk assessment dataset corresponding to an active health status abnormality risk, perform dynamic supplementation and update of newly collected valid monitoring data. Combined with the corresponding bound individual health dynamic baseline, quantify the health risk assessment score of the health status abnormality risk corresponding to each collection batch. At the same time, construct a time series sequence of health risk assessment scores since the initial establishment of the health risk assessment dataset. Combining the evolution characteristics of single real-time assessment results and full-cycle score time series sequence, perform a comprehensive health risk existence determination of the health status abnormality risk. For health risk assessment datasets determined to have no abnormalities, terminate the update and corresponding assessment process. For health status abnormalities determined to be real, maintain the activation state of the assessment process and synchronously iteratively optimize the corresponding individual health dynamic baseline.

[0091] This step aims to achieve continuous dynamic tracking and assessment of triggered abnormal health status risks. On the one hand, it enables real-time iteration of the health risk assessment dataset and accurate quantification of individual states through the dynamic supplementation of newly collected data. On the other hand, it avoids local misjudgments and omissions caused by accidental fluctuations in single-collection data by constructing a time series of health risk assessment scores. Combined with the state evolution trend within the complete assessment cycle, it completes the comprehensive determination of the existence of non-local health risks. At the same time, it achieves precise closed-loop management of the life cycle of the health risk assessment dataset, and promptly terminates the assessment process without abnormalities to avoid ineffective computing power consumption. This provides a reliable basis for subsequent abnormal trend prediction and collection strategy optimization, realizing full-cycle, highly reliable dynamic closed-loop management of the individual's health status.

[0092] In this embodiment, the health risk assessment score time sequence refers to the standardized time sequence formed by arranging the health risk assessment scores corresponding to each valid collection batch according to the collection timestamps since the initial establishment of the health risk assessment dataset corresponding to the abnormal health status risk and the formal start of the assessment process. The comprehensive existence determination of health risk must simultaneously meet the abnormal determination conditions of a single real-time assessment and the continuous abnormal determination conditions of the whole cycle time sequence evolution, abandoning the local determination logic based only on a single collection of data or a single assessment result, and ensuring the accuracy and anti-interference of abnormal determination. For example, when the health monitoring individual is a subject undergoing dynamic identification of TCM constitution status in a TCM health status time-series monitoring scenario for non-diagnostic and treatment purposes, and the abnormal health status risk is a deviation from Qi deficiency constitution status, the corresponding health risk assessment score time-series sequence is formed by sorting the health risk assessment scores corresponding to each batch of data collected since the start of the Qi deficiency constitution status deviation assessment process. A comprehensive judgment requires simultaneously verifying whether the current score exceeds the abnormal threshold and whether the sequence shows a continuous trend of constitution status deviation evolution. Similarly, when the health monitoring individual is the main drive gearbox of a coal mine hoist in an industrial equipment operation and maintenance scenario, and the abnormal health status risk is an abnormal gearbox lubrication status, the corresponding health risk assessment score time-series sequence is formed by sorting the health risk assessment scores corresponding to each batch of data collected since the start of the lubrication status abnormality assessment process. A comprehensive judgment requires simultaneously verifying whether the current score exceeds the abnormal threshold and whether the sequence shows a continuous trend of lubrication performance deterioration evolution.

[0093] Please see Figure 3 In one embodiment, the specific steps for quantifying the health risk assessment score corresponding to the abnormal health status risk of each collection batch include:

[0094] For the health risk assessment dataset that has been dynamically updated, the newly added real-time effective monitoring data is extracted. For each collected indicator associated with the abnormal risk of this health status, the adaptive multi-scale sliding window partitioning rule and multi-scale temporal feature extraction method, which are consistent with the individual health dynamic baseline construction stage in step S2, are used to generate a real-time multi-scale temporal feature set that matches the feature dimension and feature caliber of the corresponding individual health dynamic baseline. This ensures the comparability between real-time features and individual baseline features and provides a unified feature benchmark for subsequent deviation quantification analysis.

[0095] For the generated real-time multi-scale time series feature set, an individual baseline deviation quantification assessment is performed. Each feature dimension in the real-time multi-scale time series feature set is compared one by one with the feature benchmark interval of the corresponding feature dimension in the individual health dynamic baseline. The relative deviation value between the real-time feature value of each feature dimension and the feature benchmark interval is calculated. That is, based on the center and width of the feature benchmark interval, the deviation degree between the real-time feature value and the interval is normalized to the [0,1] interval. Then, combined with the adaptive deviation threshold of the corresponding feature dimension, it is determined whether the deviation of each feature dimension exceeds the normal fluctuation range. That is, when the relative deviation value exceeds the adaptive deviation threshold, it is determined to be abnormal fluctuation; otherwise, it is normal fluctuation.

[0096] If the fluctuation exceeds the normal range, the feature dimension is marked as an abnormal fluctuation feature and included in the abnormal feature candidate set; otherwise, the feature dimension is marked as a normal fluctuation feature and included in the normal feature set.

[0097] For each data collection indicator, firstly, the abnormal fluctuation features and normal fluctuation features corresponding to all feature dimensions under that indicator are collected. Then, the proportion of feature dimensions included in the abnormal feature candidate set, the maximum relative deviation value of a single feature, and the weighted average of the relative deviation values ​​of all feature dimensions are calculated. This generates the single-indicator abnormality quantification value, completing the pure quantitative representation of the abnormality degree of a single data collection indicator. The weight of the weighted average is the historical correlation contribution of each feature dimension to the corresponding abnormal health status risk, i.e., the quantification value obtained based on historical abnormal cases, representing the effectiveness and causal correlation strength of the feature dimension in abnormal early warning. Specifically, the calculation of the single-indicator abnormality quantification value includes: for each feature dimension under that data collection indicator, a 0-1 interval standardized mapping is performed using the ratio of the relative deviation value of the real-time feature value to the corresponding feature baseline interval and the adaptive deviation threshold of that feature dimension, obtaining the single-feature standardized deviation value. Then, the proportion of feature dimensions included in the abnormal feature candidate set is calculated. The maximum value of the standardized deviation of a single feature among all feature dimensions is extracted. The weighted average deviation value is calculated by weighting the contribution of each feature dimension to the historical correlation of the corresponding abnormal health status risk. The three statistics are weighted and fused by a preset fusion coefficient that sums to 1 to obtain the single indicator abnormality quantification value. The fusion coefficient is preset according to the domain monitoring needs. For example, in high-sensitivity scenarios, extreme deviations can be taken as (0.2, 0.5, 0.3), in multi-feature linkage scenarios, abnormal proportions can be taken as (0.4, 0.3, 0.3), and in general scenarios, a balanced value of (0.3, 0.3, 0.4) is taken.

[0098] For the individual abnormality quantification values ​​of each collected indicator associated with abnormal health status risk, a standardized normalization process is performed using a 0-1 interval standardization normalization method adapted to individual baseline extreme values. Specifically, for each collected indicator's individual abnormality quantification value, the extreme value interval of the historical normal individual abnormality quantification value corresponding to the full historical valid monitoring data during the construction of the individual health dynamic baseline is retrieved. The upper limit of the extreme value interval of the historical normal individual abnormality quantification value is the maximum individual abnormality quantification value of the collected indicator under the historical normal state, and the lower limit is the minimum individual abnormality quantification value of the collected indicator under the historical normal state. Using this extreme value interval as the normalization benchmark, a linear normalization mapping is performed on the individual abnormality quantification values ​​of this collection batch, uniformly mapping the individual abnormality quantification values ​​of all collected indicators to the 0-1 interval, eliminating the incomparability caused by differences in data volume and feature dimensions between different collected indicators, and obtaining the normalized individual abnormality quantification value corresponding to each collected indicator.

[0099] After normalizing all collected indicators, a weighted average of the abnormal quantification values ​​of all normalized single indicators is calculated based on the preset association weights corresponding to each indicator to generate the health risk assessment score for this batch of data collection. Specifically, the normalized single indicator abnormal quantification value of each collected indicator is multiplied by its corresponding preset association weight to obtain the weighted abnormal quantification value of each indicator. Then, the weighted abnormal quantification values ​​of all collected indicators are summed and divided by the sum of the preset association weights of all collected indicators to obtain the final health risk assessment score for this batch of data collection. The health risk assessment score ranges from 0 to 1, and the score is positively correlated with the degree of abnormality of the corresponding health risk for the individual being monitored. The closer the score is to 1, the more significant the degree of abnormality of the corresponding health risk. The preset association weight refers to a preset weight coefficient used to quantify the driving contribution of a single collected indicator to the occurrence and evolution of the corresponding abnormal health status risk. The weight coefficient ranges from 0 to 1, and its value is positively correlated with the intensity of the causal influence of the collected indicator on the abnormal health status risk, the tightness of the temporal association, and the driving contribution in historical abnormal events. Among them, the intensity of causal influence represents the direct driving degree of the indicator on the abnormality, the tightness of the temporal association represents the temporal synchronicity between the indicator abnormality and the abnormal event, and the driving contribution in historical abnormal events represents the actual deviation ratio of the indicator in historical abnormalities. The three are obtained by weighted summation and normalization through preset scenario adaptation coefficients. The configuration is based on the preset many-to-many association mapping relationship between the collected indicator and the abnormal health status risk, the statistical results of the abnormal driving contribution of the collected indicator in the historical confirmed abnormal events of the health monitoring individual, and the historical sensitivity of the temporal characteristics of the collected indicator to changes in the abnormal health status risk.

[0100] Simultaneously, the health risk assessment score, the single-indicator abnormal quantitative value and normalization result of each collected indicator, the preset correlation weight, the abnormal feature candidate set, the corresponding collection timestamp, the batch identifier, and the correlation assessment process identifier are strongly bound and archived into the full-cycle health time-series database of the health monitoring individual, providing complete and unified data support for the subsequent comprehensive existence determination of full-cycle health risks.

[0101] Please see Figure 4 In one embodiment, the specific steps for performing a comprehensive health risk existence determination of abnormal health status include:

[0102] Based on the chronological order of data collection timestamps, the health risk assessment scores calculated for all batches within the assessment period corresponding to each abnormal health status risk are sorted to construct a time-series sequence of health risk assessment scores corresponding to that abnormal health status risk. A preset minimum assessment period is retrieved to determine if the actual time span of the current health risk assessment score time-series sequence is less than the minimum assessment period. If it is less, the current comprehensive health risk existence determination for the corresponding abnormal health status risk is terminated, and the corresponding assessment process remains active until subsequent batches of data are supplemented, at which point the determination is executed again. If it is greater than or equal to the minimum assessment period requirement, the comprehensive health risk existence determination operation for the corresponding abnormal health status risk is initiated. The minimum assessment period is preset based on the minimum evolution period of abnormal health status risks in the corresponding application scenario and the sampling frequency characteristics of the collected indicators, ensuring that the time-series sequence within the assessment period can fully represent the trend of health status changes in the monitored individuals.

[0103] It should be noted that the comprehensive assessment of the existence of health risks specifically includes:

[0104] The health risk assessment score for each abnormal health status risk corresponding to this batch of data is compared with the single-abnormality threshold. If the health risk assessment score for the abnormal health status risk is greater than or equal to the single-abnormality threshold, it is marked as meeting the single-abnormality significance standard. This indicates that the abnormal health status risk in the current batch has shown a significant deviation exceeding the individual's normal fluctuation range, with clear abnormal characteristics and high certainty of occurrence, and the existence of the abnormal health status risk is directly determined. Otherwise, it is marked as not meeting the single-abnormality significance standard, and enters the full-cycle time-series evolution feature verification operation. The single-abnormality threshold is pre-set based on the historical normal state score distribution characteristics of the corresponding abnormal health status risk in the individual's health dynamic baseline. The value range is 0 to 1 and can be synchronously optimized with the iteration of the individual's health dynamic baseline. It is used to define the abnormal significance of a single batch of data.

[0105] For health status abnormality risks where the significance of a single abnormality does not meet the standard, standardized preprocessing is performed on the time series of the corresponding health risk assessment scores. This includes adaptive completion of missing values ​​corresponding to batches with interrupted sequences and smoothing filtering of abnormal jump values ​​caused by data acquisition errors, to ensure the continuity and effectiveness of the time series.

[0106] The temporal evolution characteristics of the preprocessed health risk assessment score time series are extracted to construct a full-cycle temporal evolution feature set. The temporal evolution characteristics refer to the core quantitative feature set that characterizes the trend and persistence of abnormal health status of individuals under health monitoring. These features include the overall trend evolution slope of the sequence, the cumulative total duration of health risk assessment scores exceeding the preset single abnormal threshold, and the proportion of abnormal batches with health risk assessment scores exceeding the single abnormal threshold.

[0107] Retrieve the preset continuous anomaly judgment requirements. The continuous anomaly judgment requirements refer to the anomaly judgment criteria preset for time-series evolution characteristics, including trend evolution judgment items, duration judgment items, and anomaly batch proportion judgment items.

[0108] For example, the requirements for determining persistent anomalies include:

[0109] Trend evolution judgment item: The overall trend evolution slope of the health risk assessment score time series is greater than the preset positive growth threshold. The preset positive growth threshold is based on the trend slope of historical progressive abnormal cases and the scene sensitivity requirements. A smaller value is taken for high-risk scenarios to identify progressive abnormalities earlier. That is, the overall sequence shows a continuous upward abnormal development trend without a continuous decline and convergence feature exceeding the preset duration. The preset duration is taken as the maximum duration of short-term fluctuations under normal working conditions of the target object or 10% of the abnormal evolution cycle, which is used to verify whether the abnormal state of the individual under health monitoring shows a progressive and hidden development.

[0110] Duration determination item: The proportion of the cumulative total duration of health risk assessment scores exceeding the single abnormal threshold to the total duration of the assessment period is greater than the preset proportion threshold. This preset proportion threshold is based on the historical abnormality persistence characteristics and scenario continuity requirements. A higher proportion is taken for high continuity scenarios to verify the degree of continuous coverage of abnormal states.

[0111] Abnormal Batch Proportion Determination Item: The proportion of abnormal batches whose health risk assessment scores exceed the single abnormal threshold within the assessment period is greater than the preset abnormal proportion threshold of the total number of valid batches in the entire period. This preset abnormal proportion threshold is based on the historical abnormal frequency and the scenario false alarm tolerance. A lower proportion is taken for high-frequency monitoring scenarios to verify the frequency of abnormal states.

[0112] If the entire time-series evolution feature set does not meet the above-mentioned continuous anomaly determination requirements, then the risk of an abnormal health state is determined to be non-existent; otherwise, the risk of an abnormal health state is determined to be present.

[0113] If it is determined that there is no abnormal health status risk, the health risk assessment dataset bound to the abnormal health status risk is set to lifecycle termination status, the execution permission of the corresponding assessment process is terminated simultaneously, the time series data update channel of the health risk assessment dataset is closed, and the subsequent continuous assessment operation of the abnormal health status risk is stopped. Only the full historical data of the health risk assessment dataset, the time series sequence of health risk assessment scores, the full process assessment records and the final judgment results are archived and retained for subsequent traceability and query. If it is determined that there is an abnormal health status risk, the corresponding health risk assessment dataset and assessment process are kept active, and the prediction of health status evolution trend is triggered.

[0114] Step S4: For the health status abnormality risks finally identified in Step S3, trigger the prediction of health status evolution trends. Based on the pre-constructed standard health status evolution path corresponding to the health status abnormality risk, and combined with the individual temporal evolution characteristics of the health monitoring individual, the full-cycle abnormal status data, and the corresponding bound individual health dynamic baseline, construct a state evolution path exclusive to the health monitoring individual. Based on the predicted switching nodes of each standard evolution stage in the state evolution path, generate the subsequent adaptive monitoring time point sequence of the health status abnormality risk, and push it synchronously to the corresponding data acquisition terminal to guide the targeted collection of health data. Based on the actual monitoring data collected subsequently, continuously optimize and correct the stage division and switching node prediction results of the state evolution path, and synchronously and dynamically adjust the adaptive monitoring time point sequence to form a closed-loop management and control throughout the entire process.

[0115] This step aims to build upon the anomaly determination results of step S3 and address the core issues of existing fixed-cycle monitoring models, such as missed data collection at key stage switching nodes, redundant invalid monitoring data, and an imbalance between monitoring accuracy and computational cost. By using the standard evolution stage switching of the state evolution path as the core anchor point, it differentiates the monitoring density and accurately matches the stage characteristics of anomaly evolution. While ensuring the complete capture of the anomaly development stage switching process, it significantly reduces the computational and storage costs of invalid monitoring. At the same time, through continuous feedback of actual monitoring data, it achieves dynamic iterative optimization of the state evolution path and monitoring strategy, providing accurate and efficient monitoring support for the full-cycle continuous tracking of health state anomaly risks.

[0116] In one embodiment, the specific steps for constructing the state evolution path include:

[0117] For each type of abnormal health status risk, based on the recognized abnormal evolution mechanism in the relevant industry field and the full-cycle historical data of a large sample of similar abnormal events, a standard health status evolution path corresponding to that type of abnormal health status risk is pre-constructed.

[0118] It should be noted that the standard health status evolution path refers to a standardized time-series path that uses the health risk assessment score of the corresponding health status abnormality risk as a quantitative indicator, and takes the universally recognized full-cycle evolution mechanism of the relevant industry field as the underlying logic. It represents the universal evolution law of the full-cycle abnormality risk of this type of health status. It is a benchmark evolution template for the abnormality risk of the same type of health status under the condition of natural development without external intervention. It is used to provide a unified benchmark for the division of evolution stages, the quantitative reference of status, the anchoring rules of core indicators, and the benchmark for individual path adjustment for all individuals under the same type of health monitoring. Its components include the standard evolution stages of the abnormality risk of this type of health status, as well as the standard range of health risk assessment score, the standard range of stage duration, the standard evolution rate, and the standard set of core abnormal indicators corresponding to each standard evolution stage. At the same time, it clearly distinguishes the standard abnormality threshold between normal and abnormal states.

[0119] Specifically, the standard evolution stage refers to a continuous evolutionary stage, based on the universally recognized evolutionary mechanism of the occurrence, development, and outcome of anomalies within the relevant industry sector. This stage, which is non-overlapping, fully covered, and possesses clear business connotations and progressive logic, divides the complete lifecycle of abnormal health status risks of the same type. It includes at least three stages: the anomaly initiation stage, the anomaly continuous development stage, and the anomaly aggravation stage. The health risk assessment score standard range refers to the upper and lower threshold ranges for health risk assessment scores set for each standard evolution stage, based on the statistical distribution characteristics of large-sample, full-cycle data of similar potential abnormal events. The upper and lower threshold ranges refer to the typical normal fluctuation range of health risk assessment scores for similar potential abnormal events within their corresponding evolutionary stages. The lower limit is the initial characteristic level of this stage, and the upper limit is the critical level for the evolution from this stage to subsequent stages. These ranges are set through a combination of statistical distribution of large-sample historical data and constraints imposed by the domain's business context. For example, the 5% to 95th percentiles of historical samples for each stage are extracted and fine-tuned based on business context, covering 90% of normal fluctuations. The standard range for stage duration refers to the statistical distribution of the duration of each evolutionary stage based on a large sample of similar potential abnormal events, defining each standard evolutionary stage as... The set of duration fluctuation ranges under natural development conditions without external intervention; the duration fluctuation range refers to the typical distribution range of the natural duration of the corresponding evolutionary stage of the same type of potential abnormal event under natural development conditions without external intervention, representing the typical time span from the initiation of the stage to the subsequent stage, and is set by statistical analysis of a large sample of natural evolution duration and filtering samples with artificial intervention; the standard evolution rate refers to the standard change range of the health risk assessment score per unit time for each standard evolutionary stage under natural development conditions without external intervention; the standard core abnormality indicator set refers to the set of indicators based on the corresponding health status abnormality risk. The intrinsic evolutionary mechanism of abnormal health status risk is a set of collection indicators set for each standard evolution stage, which plays a core driving role in the abnormal evolution of that stage, can accurately characterize the core business characteristics of the stage, and is strongly correlated with the progress of abnormal development. The standard abnormality critical threshold is a general critical score set for the corresponding abnormal health status risk based on the statistical characteristics of the boundary between normal and abnormal states of a large sample of similar potential abnormal events. It distinguishes between normal state and abnormal budding state and is the lower limit of the abnormal budding stage in the standard evolution stage. It is also a general abnormal trigger judgment benchmark for the same type of abnormal health status risk, with a value range of 0 to 1.

[0120] For any identified abnormal health status risk, the time series sequence of the health risk assessment score corresponding to the abnormal health status risk is retrieved to extract the current health risk assessment score and historical evolution rate of the abnormal health status risk, and to obtain the current set of core abnormal indicators for the abnormal health status risk. Among them, the historical evolution rate refers to the actual change in the health risk assessment score of the health monitoring individual per unit time during the entire historical abnormal period from the first triggering of the health risk assessment process and the start of the abnormal development cycle to the current assessment batch.

[0121] The current health risk assessment score is matched one by one with the standard range of health risk assessment scores for each standard evolution stage in the standard health state evolution path to determine the current anchor time node and corresponding standard evolution stage of the health monitoring individual on the standard health state evolution path; whereby, the anchor time node refers to the time position of the health monitoring individual on the full-cycle time axis of the standard health state evolution path for the corresponding health state abnormal risk, and is the core anchor point that characterizes the degree of matching between the individual's abnormal development progress and the standard general evolution law;

[0122] Specifically, the steps for obtaining the anchor time node include:

[0123] The current health risk assessment score is compared one by one with the standard range of health risk assessment scores for each standard evolution stage of the standard health state evolution path. The standard evolution stage to which the current score belongs is initially determined. Within the standard time axis of this stage, the time point with the smallest absolute value of the difference between the health risk assessment score and the current score in the standard health state evolution path is located and determined as the initial matching time point.

[0124] Retrieve the current set of core abnormal indicators for the individual under health monitoring and match them with the set of standard core abnormal indicators for the standard evolution stage corresponding to the initial matching time point in the standard health status evolution path. Calculate the indicator feature matching degree.

[0125] If the indicator feature matching degree reaches the preset matching threshold, the initial matching time point is determined as the final anchoring time node. If the indicator feature matching degree does not reach the preset matching threshold, within the time axis range of the corresponding standard evolution stage, the process is traversed back and forth centered on the initial matching time point. The time point with the highest comprehensive matching degree is selected by combining the comprehensive weighted calculation result of the score matching degree and the indicator matching degree, and determined as the final anchoring time node. The preset matching threshold is a critical judgment value used to determine whether the feature matching degree between the current core abnormal indicator set of an individual under health monitoring and the standard core abnormal indicator set of the corresponding stage of the standard health status evolution path meets the requirements for direct anchoring. The value range is fixed from 0 to 1, and it is comprehensively set by considering the statistical distribution characteristics of the core indicator matching degree of similar closed-loop abnormal events in a large sample of the same monitoring scenario, the risk level of the corresponding health status abnormality risk, and the business monitoring accuracy requirements.

[0126] The ratio of the historical abnormal evolution rate of an individual under health monitoring to the standard evolution rate of the corresponding standard evolution stage is used as the individual rate scaling factor. The standard range of the stage duration of the standard health state evolution path is linearly and inversely scaled to adapt to the inherent personalized characteristics of individual abnormal development that are too fast or too slow, and to eliminate the difference in evolution rate between the individual and the standard path.

[0127] Using the final determined anchor time node as the prediction starting point and the duration of each standard evolution stage after linear scaling as the time scale benchmark, combined with the standard evolution rate of the corresponding standard evolution stage and the standard range of health risk assessment scores, the initial state evolution path of the health monitoring individual is generated. Simultaneously, the predicted switching time nodes of each subsequent standard evolution stage in the path, the predicted sequence of health risk assessment scores for the corresponding evolution period, and the evolution law of the core abnormal indicators corresponding to each stage are identified.

[0128] The difference between the individual single-abnormality threshold set for the health monitoring individual as the corresponding abnormal risk of health status and the standard abnormality threshold in the standard health status evolution path is used as the core translation correction amount. The standard range of health risk assessment scores in each standard evolution stage in the initial state evolution path is linearly translated with equal offset to eliminate the difference between the general standard of large sample and the individual normal state benchmark. Finally, a state evolution path that is unique to the health monitoring individual and adapted to the individual characteristics is obtained. At the same time, the final generated state evolution path is strongly bound with the corresponding abnormal health risk, the unique identifier of the health monitoring individual, and the assessment process identifier, and archived and stored in the full-cycle health time series database of the health monitoring individual.

[0129] Based on the health risk assessment score prediction sequence and standard evolution stage division rules of the individual's state evolution path, the critical nodes for stage switching between adjacent standard evolution stages are identified, and the prediction time, critical health risk assessment score, and critical characteristics of core abnormal indicators corresponding to the nodes are clarified to generate an adaptive monitoring time point sequence. Furthermore, a set of standard core abnormal indicators corresponding to the standard evolution stage is matched for each monitoring time point to form a targeted collection instruction.

[0130] The specific generation process is as follows:

[0131] For the identified critical nodes of phase switching, encrypted monitoring time points are set at equal intervals within the preset transition period before and after the critical nodes of phase switching. The time interval of encrypted monitoring does not exceed half of the regular monitoring interval of the corresponding phase, so as to ensure complete capture of the status change data during the phase switching process. The preset transition period is preset based on the average duration of the phase switching of the corresponding abnormal health status risk and the sampling frequency characteristics of the collected indicators.

[0132] For the non-switching periods of each standard evolution stage in the state evolution path, differentiated regular monitoring intervals are set based on the standard evolution rate and abnormal risk level of the corresponding standard evolution stage, generating regular monitoring time points within the stage; a maximum allowable monitoring interval is set for the entire cycle. If the regular monitoring interval within the standard evolution stage exceeds this maximum allowable value, a minimum monitoring time point is added according to the maximum allowable monitoring interval to ensure that there are no monitoring blind spots throughout the entire monitoring cycle.

[0133] This embodiment introduces a time-series data-driven dynamic prediction system for health status, including an anomaly screening module, a baseline generation module, a comprehensive assessment module, and a trend prediction module.

[0134] The anomaly screening module receives multi-source heterogeneous monitoring data from individuals throughout their entire health monitoring lifecycle. It initializes and dynamically maintains a corresponding full-cycle health time-series database based on the individual's unique identifier. It completes the classification, archiving, standardized preprocessing, and effective data archiving and storage of all monitoring data. It also retrieves newly archived single-batch real-time effective monitoring data from the individual's full-cycle health time-series database to perform rapid anomaly screening of real-time monitoring data. It locates anomaly collection indicators and maps them to the corresponding health status anomaly risks. It verifies the activated assessment process status of health status anomaly risks. For health status anomaly risks without a corresponding activated assessment process, it generates a unique assessment process identifier and initiates the corresponding assessment process. This achieves the pre-emptive avoidance of invalid assessment calculations in the absence of anomaly data and the precise on-demand triggering of the health status anomaly risk assessment process in the presence of anomaly data.

[0135] For each triggered and initiated health status abnormality risk assessment process, the baseline generation module constructs a dedicated health risk assessment dataset that is one-to-one bound to that health status abnormality risk. It completes the dynamic updating and full lifecycle management of the health risk assessment dataset. Adaptive multi-scale time-series feature extraction is performed on the effective monitoring data within the health risk assessment dataset to construct an individual time-series feature set corresponding to the health status abnormality risk of the health monitoring individual. Based on the individual time-series feature set, an individual health dynamic baseline strongly bound to the health status abnormality risk is generated, providing a dynamic reference benchmark that fits the individual characteristics of the health monitoring individual for subsequent abnormal status deviation analysis and accurate assessment.

[0136] The comprehensive assessment module dynamically supplements the health risk assessment dataset corresponding to each active health status abnormality risk, calculates and generates the health risk assessment score for each collection batch, constructs a time series sequence of full-cycle health risk assessment scores, and performs a comprehensive health risk existence determination of the health status abnormality risk by combining the single real-time assessment result with the full-cycle time series evolution characteristics. Based on the final determination result, it completes the life cycle closed-loop management of the corresponding health risk assessment dataset and assessment process, and simultaneously iteratively optimizes the corresponding individual health dynamic baseline for the determined real health status abnormality risks.

[0137] The trend prediction module triggers health status evolution trend prediction for health status risks that are ultimately determined to be real. Based on the pre-constructed standard health status evolution path corresponding to the health status risk, combined with the individual time-series evolution characteristics of the health monitoring individual, the full-cycle abnormal status data and the corresponding bound individual health dynamic baseline, it constructs a state evolution path exclusive to the health monitoring individual. Based on the standard evolution stage division and stage switching critical nodes of the state evolution path, it generates the subsequent adaptive monitoring time point sequence of the health status risk and the corresponding targeted collection instructions, which are simultaneously pushed to the corresponding data collection terminal to guide the targeted collection of health data. Based on the actual monitoring data collected later, it continuously optimizes and corrects the state evolution path and adaptive monitoring strategy to form a closed-loop management and control throughout the entire process.

[0138] Working principle and its effects:

[0139] This invention focuses on multi-source heterogeneous time-series data of individuals for health monitoring, and constructs a closed-loop health monitoring system for the entire process. By adapting to individual characteristics, mining time-series features, and optimizing monitoring strategies, it solves the pain points of existing health monitoring technologies such as poor individual adaptability, high false negative and false negative rates, and redundant computing resources, thereby achieving accurate and efficient intelligent monitoring of health status and reducing ineffective computing power consumption.

[0140] This invention first constructs a full-cycle health time-series database based on multi-source health data throughout an individual's life cycle and completes data preprocessing. It then rapidly screens real-time health data for anomalies, identifies abnormal health indicators, maps abnormal health status risks, and triggers health risk assessment processes as needed, achieving standardized management of health data and avoiding ineffective calculations. Next, it constructs a dedicated health risk assessment dataset for each triggered health status abnormality risk. Through adaptive multi-scale time-series feature extraction, it mines individual health evolution characteristics and constructs an individual health dynamic baseline, addressing the problem of insufficient individual adaptation and improving the accuracy of health status determination. Then, it dynamically updates the health risk assessment dataset, quantifies health risk assessment scores based on the individual health dynamic baseline, and constructs a time-series sequence. Through comprehensive judgment, it reduces the rate of false positives and false negatives, achieving closed-loop control of the assessment process. Finally, for real-world health status abnormalities, it constructs an individual-specific evolutionary path based on the standard health status evolution path, generating an adaptive monitoring sequence to guide targeted health data collection, avoiding missed data collection at key nodes and data redundancy, and improving the efficiency of health monitoring and prediction.

[0141] In summary, this invention organically combines multi-source heterogeneous health time-series data processing, individual dynamic adaptation, accurate determination of health anomalies, and prediction of health status evolution. It achieves efficient processing of multi-source health information and accurate identification of health anomalies. By optimizing monitoring strategies to reduce ineffective resource consumption, it significantly improves the accuracy of health monitoring, individual adaptability, and operational efficiency. It meets the needs of dynamic health monitoring and prediction of individuals throughout their entire life cycle in multiple scenarios and has strong practicality and application value.

[0142] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A time-series data driven health state dynamic prediction method, characterized in that, include: Based on the monitoring data of individuals, a full-cycle health time-series database is constructed, and health risk anomalies are screened in real-time monitoring data to locate abnormal collection indicators and map abnormal health status risks, triggering the health risk assessment process. For each abnormal health status risk that triggers the health risk assessment process, historical valid monitoring data of its associated abnormal collection indicators are extracted to construct a health risk assessment dataset. Adaptive multi-scale time series feature extraction is performed on the historical valid monitoring data of the health risk assessment dataset to construct an individual health dynamic baseline for the health monitoring individual. The active health risk assessment dataset is dynamically supplemented and updated. Combined with the individual health dynamic baseline, abnormal fluctuation characteristics are identified, the health risk assessment score of the abnormal health status risk corresponding to each collection batch is quantified, and a time series sequence of health risk assessment scores is constructed. This is used to perform a comprehensive health risk existence determination of abnormal health status risk, so as to close the update channel of the health risk assessment dataset where abnormal health status risk does not exist, and trigger the prediction of health status evolution trend. When the prediction of health status evolution trend is triggered, an individual health status evolution path is constructed based on the standard health status evolution path, thereby generating an adaptive monitoring time point sequence of abnormal health status risk to guide the targeted collection of health data.

2. The time-series data driven health state dynamic prediction method of claim 1, wherein, The steps for constructing an individual's dynamic health baseline for health monitoring include: Based on the pre-defined correlation mapping between the collection indicators and the risk of abnormal health status, the associated collection indicators corresponding to each abnormal health status risk are extracted to form a collection indicator set. From the full-cycle health time-series database, retrieve all historical valid monitoring data corresponding to all collected indicators in the collection indicator set to construct the basic data source for the health risk assessment dataset. Based on the set of collected indicators and basic data sources corresponding to abnormal health status risks, a health risk assessment dataset is initially constructed. Extract the sampling frequency of each collection indicator in the health risk assessment dataset and classify each collection indicator hierarchically; perform adaptive multi-scale sliding window partitioning for historical effective monitoring data of different hierarchical classifications, and perform adaptive multi-scale time series feature extraction to generate a multi-scale time series feature set for each associated collection indicator, and then construct an individual time series feature set; Based on the individual time-series feature set, the feature benchmark interval and adaptive deviation threshold corresponding to each associated collection indicator are calculated to construct the individual health dynamic baseline.

3. The time-series data driven health state dynamic prediction method of claim 2, wherein, The steps for constructing the individual temporal feature set include: Based on a multi-scale sliding window, basic feature extraction operations are performed on the historical valid monitoring data of each associated collection indicator to generate a single sliding window basic feature set corresponding to each sliding window; the basic features include time-domain statistical features, frequency-domain features and nonlinear dynamic features. For each associated collection indicator, a weighted weight is set for each sliding window of the basic feature set of all single sliding windows. The basic feature sets under the same collection indicator are then fused through weighted fusion to generate a multi-scale time series feature set. Based on the historical collection timestamps of all associated collection indicators in the multi-scale temporal feature set, a temporal alignment correlation matrix is ​​established to mine the accompanying change relationship, causal relationship, and temporal coupling relationship among the multi-scale temporal features of different collection indicators, and to generate a cross-indicator correlation feature set. For all associated collected indicators, corresponding multi-scale time-series feature sets and cross-indicator associated feature sets, a standardized integration operation of all features is performed to generate individual time-series feature sets.

4. The time-series data driven health state dynamic prediction method of claim 2, wherein, The steps for performing adaptive multi-scale sliding window partitioning on historical valid monitoring data for different classification levels include: Using the minimum or average sampling interval of each sampling index under each sampling frequency level as the basic sliding window unit, multiple rounds of expansion are performed upwards according to a preset multiple gradient to generate multiple sliding windows with different time spans, forming a multi-scale sliding window group. Based on the basic sliding window unit of the corresponding level, and according to the time-series fluctuation amplitude of the historical valid monitoring data of each collected indicator within the level, a preset mapping relationship between the time-series fluctuation amplitude and the step size adjustment coefficient is established, and the step size of each sliding window corresponding to each collected indicator within the level is set respectively; and the overlap rate of the sliding window is adaptively determined according to the ratio of the step size to the time span of the corresponding sliding window, thus forming an adaptive multi-scale sliding window for each collected indicator.

5. The time-series data driven health state dynamic prediction method of claim 1, wherein, The steps involved in locating abnormal indicators and mapping them to abnormal health status risks to trigger the health risk assessment process include: Based on the unique identifier of the health monitoring individual, an initial full-cycle health time-series database of the health monitoring individual is established. The monitoring data collected in each batch is classified, archived, and preprocessed. The valid monitoring data collected in each batch is stored in the full-cycle health time-series database. The real-time values ​​of each collected indicator in a single batch of real-time effective monitoring data are compared one by one with the corresponding safety benchmark interval to identify abnormal collected indicators; if at least one abnormal collected indicator is generated in this batch, the process of associating abnormal health status risk with mapping is initiated. The system retrieves the pre-defined mapping relationship between the collection indicators and the abnormal health status risks. For each abnormal collection indicator, it matches and retrieves all associated abnormal health status risks, generates the set of abnormal health status risks corresponding to this batch, and verifies one by one whether each abnormal health status risk has a corresponding active assessment process. If there is no active assessment process for the abnormal health status risk, a unique assessment process identifier is generated to trigger the health risk assessment process.

6. The time-series data driven health state dynamic prediction method of claim 1, wherein, The steps for quantifying the health risk assessment score corresponding to the abnormal health status risk of each data collection batch include: For the health risk assessment dataset that has been dynamically updated, newly supplemented real-time effective monitoring data is extracted. For each collected indicator associated with the abnormal risk of this health status, adaptive multi-scale time series feature extraction is performed to generate a real-time multi-scale time series feature set. For each feature dimension in the real-time multi-scale time series feature set, it is compared one by one with the feature benchmark interval of the corresponding feature dimension in the individual health dynamic baseline. The relative deviation between the real-time feature value of each feature dimension and the feature benchmark interval is calculated. Combined with the adaptive deviation threshold of the corresponding feature dimension, it is determined whether the deviation of each feature dimension exceeds the normal fluctuation range. If the fluctuation exceeds the normal range, the feature dimension is marked as an abnormal fluctuation feature and included in the abnormal feature candidate set; otherwise, the feature dimension is marked as a normal fluctuation feature and included in the normal feature set. For each collected indicator, the proportion of the number of feature dimensions included in the abnormal feature candidate set, the maximum relative deviation value of a single feature, and the weighted average of the relative deviation values ​​of all feature dimensions are calculated to generate the single indicator abnormality quantification value corresponding to the collected indicator. The abnormal quantitative values ​​of each collected indicator associated with abnormal health status risk are standardized and normalized; and a weighted summation average is calculated based on the preset associated weights of each collected indicator to generate the health risk assessment score for the abnormal health status risk corresponding to this batch of data collection.

7. The time-series data driven health state dynamic prediction method of claim 1, wherein, The steps for determining the comprehensive existence of health risks associated with abnormal health status include: The health risk assessment scores calculated for each abnormal health status are sorted to construct a time series sequence of health risk assessment scores corresponding to the abnormal health status risk. It is then determined whether the actual time span of the current health risk assessment score time series sequence is less than the preset minimum assessment period. If so, the current comprehensive health risk existence determination for the corresponding abnormal health status risk is terminated; otherwise, the comprehensive health risk existence determination operation for the corresponding abnormal health status risk is initiated. If the health risk assessment score for each health status abnormality risk corresponding to this batch of data collection is greater than or equal to the single abnormality threshold, it is marked as meeting the single abnormality significance standard, and the abnormality risk of that health status is determined to exist; otherwise, it is marked as not meeting the single abnormality significance standard. For health status abnormality risks where the significance of a single abnormality does not meet the standard, the corresponding health risk assessment score time series is subjected to standardized preprocessing, and time series evolution features are extracted to construct a full-cycle time series evolution feature set; the time series evolution features refer to the core quantitative feature set that characterizes the trend of changes in the abnormality risk status and the abnormality persistence characteristics of the corresponding health monitoring individual. If the preset continuous anomaly determination requirements are retrieved and the full-cycle time-series evolution feature set does not meet the continuous anomaly determination requirements, then the abnormal risk of the health state is determined to be non-existent. The life cycle termination status of the health risk assessment dataset bound to the abnormal risk of the health state is set, and the time-series data update channel of the health risk assessment dataset is closed. Otherwise, the abnormal risk of the health state is determined to exist, and the health state evolution trend prediction is triggered.

8. The time-series data driven health state dynamic prediction method of claim 1, wherein, The standard health status evolution path refers to a standardized time-series path that uses the health risk assessment score of the corresponding health status abnormal risk as a quantitative indicator to characterize the universal evolution law of the entire cycle of such health status abnormal risk. It is used to define the standard evolution stages of abnormal risks in various health states, as well as the standard range of health risk assessment scores, the standard range of stage duration, the standard evolution rate, and the standard set of core abnormal indicators for each standard evolution stage, while clearly distinguishing the standard abnormal threshold between normal and abnormal states.

9. The time-series data-driven dynamic prediction method for health status as described in claim 7 or 8, characterized in that, The steps for constructing the evolutionary path of an individual's health monitoring state include: The standard health status evolution path for various health status abnormal risks is pre-constructed. For the health status abnormal risks that are identified as existing, the time series sequence of the health risk assessment score corresponding to the health status abnormal risk is retrieved to extract the current health risk assessment score and historical evolution rate of the health status abnormal risk, and the current core abnormal indicator set of the health status abnormal risk is obtained. The current health risk assessment score is matched one by one with the standard range of health risk assessment scores for each standard evolution stage in the standard health status evolution path to determine the current anchoring time node and corresponding standard evolution stage of the health monitoring individual in the standard health status evolution path. The ratio of the historical abnormal evolution rate of an individual under health monitoring to the standard evolution rate of the corresponding standard evolution stage is used as the individual rate scaling factor to linearly scale the standard range of the stage duration of the standard health state evolution path.

10. The time-series data-driven dynamic prediction method for health status as described in claim 9, characterized in that, The steps for constructing the evolutionary path of an individual's health monitoring state also include: Using the final determined anchor time node as the prediction starting point, and the duration of each standard evolution stage after linear scaling as the time scale benchmark, combined with the standard evolution rate of the corresponding standard evolution stage and the standard range of health risk assessment scores, the initial state evolution path of the health monitoring individual is generated. The difference between the individual single abnormality threshold set for the health monitoring individual as the corresponding abnormality risk and the standard abnormality threshold in the standard health state evolution path is used as the core translation correction amount. The standard interval of health risk assessment score in each standard evolution stage in the initial state evolution path is linearly translated with equal offset to obtain the state evolution path of the health monitoring individual.

11. A time-series data-driven dynamic health status prediction system, used to implement the time-series data-driven dynamic health status prediction method according to any one of claims 1 to 10, characterized in that, It includes an anomaly screening module, a baseline generation module, a comprehensive evaluation module, and a trend prediction module; The anomaly screening module is used to construct a full-cycle health time-series database based on the monitoring data of individuals under health monitoring, and to screen for health risk anomalies in real-time monitoring data, locate abnormal collection indicators and map abnormal health status risks, and trigger the health risk assessment process. The baseline generation module is used to extract historical valid monitoring data of associated abnormal collection indicators for each health status abnormal risk that triggers the health risk assessment process, construct a health risk assessment dataset, and construct an individual health dynamic baseline for the health monitoring individual by adaptively extracting multi-scale temporal features from the historical valid monitoring data of the health risk assessment dataset. The comprehensive assessment module is used to dynamically supplement and update the health risk assessment dataset that is in an active state. Combined with the individual health dynamic baseline, it identifies abnormal fluctuation characteristics, quantifies the health risk assessment score of the abnormal health status risk corresponding to each collection batch, and constructs a time series sequence of health risk assessment scores. This is used to perform a comprehensive health risk existence determination of abnormal health status risk, so as to close the update channel of the health risk assessment dataset where abnormal health status risk does not exist, and trigger the prediction of health status evolution trend. The trend prediction module is used to construct an individual health status evolution path based on the standard health status evolution path when the health status evolution trend prediction is triggered, and then generate an adaptive monitoring time point sequence of abnormal health status risk to guide the targeted collection of health data.

Images