A health data processing and alarm method for a family monitoring group

Abstract

The present application relates to the technical field of electric digital data processing, and discloses a health data processing and alarming method for a family monitoring group, comprising: establishing a dynamic baseline of individual long-term stable habits and a short-term baseline reflecting a recent mode for a multi-modal data stream of a person being monitored; and executing two processing paths in parallel, one of which generates a system steady-state entropy value based on a double baseline difference aggregation to identify a long-term system evolution trend, and the other of which performs a multi-dimensional resonance judgment based on a difference between original data points and the short-term baseline to perceive a transient system impact event, the present application integrates the trend analysis and the abnormality detection, which are independent of each other in a traditional data processing mode, into an overall processing structure with internal data cooperation and complementary functions by constructing a dual-lineage collaborative monitoring framework of slow-changing risks and acute events, thereby solving the inherent information omission problem of a single processing mode.

Description

TECHNICAL FIELD

[0001] The present application relates to a health data processing and alarm method for a family monitoring group, belonging to the technical field of electronic digital data processing. BACKGROUND

[0002] The current monitoring and analysis of time series data related to individual health is a known technical approach, which usually collects user's daily activity and sleep data streams through sensors, and pre-sets a judgment threshold or a statistical range. When the real-time data exceeds this range, the system triggers an alarm to prompt the guardian to pay attention. When this data processing method is applied to long-term health trend monitoring of the elderly at home, there is a limitation in principle, that is, it mainly focuses on the absolute value of the data at a certain moment or the deviation from the short-term mean. However, the risks that need to be identified in the long-term monitoring scenario are not always represented by dramatic instantaneous abnormalities, but rather by a slow and systematic state drift of the overall life pattern of the monitored person on a time scale of weeks or months. The drift process is submerged in the normal fluctuations of daily physiological data due to its change amplitude, making it difficult for traditional threshold judgment mechanisms to effectively respond to such risks.

[0003] To solve this technical problem, one improved idea is to use a more complex machine learning model to try to identify weak change patterns from complex data, but this path usually requires a large amount of accurately labeled medical data for training under the realistic constraints of home monitoring, which is difficult to achieve in a family environment that emphasizes privacy and non-standard data sources. At the same time, complex models also have high computational costs and the problem of difficult interpretation of results. Another approach is to set more complex multi-dimensional linked alarm rules, but this makes the system sensitive to individual differences and easily generates a large number of non-critical alarm information, which continuously interferes with the guardian. In addition to the limitations of the threshold judgment mechanism in principle, some existing monitoring devices that integrate multi-dimensional data collection functions also fail to escape this situation in the design of their core alarm logic. For example, the Chinese utility model patent with the authorization announcement number CN204950922U discloses a health monitoring call device for families. Although this device integrates various data collection devices including acceleration sensors, gyroscopes, heart rate and blood pressure collection, etc., its alarm judgment control logic still remains in a passive response event-driven mode. The triggering conditions of its alarm are limited to the old person actively calling or in the event of a fall, data indicators being temporarily ill, i.e. single reading exceeding a certain static threshold, etc. The fundamental defect of this design concept is that it completely focuses on isolated and dramatic instantaneous events, and has no perception and analysis capability for the coordinated and progressive deterioration trend of multiple physiological parameters in the long time dimension, which is more indicative of potential health risks. In other words, it can respond to the result of an event that has already occurred (such as a fall), but it cannot identify the process that leads to the result (such as a systematic decline in activity and physiological stability) that lasts for weeks or even months. This is the inherent information blind spot of traditional monitoring methods when dealing with slow-changing risks.

[0004] Specifically, the existing technology mainly has the following technical problems: 1. The triggering conditions of the alarm mechanism are difficult to match the slow-changing risk characteristics of the slow-changing risk, and cannot effectively identify the progressive evolution of the health status; 2. The conventional technical path to improve the recognition sensitivity introduces unacceptable limitations in engineering universality and operating cost. Therefore, how to design a data processing method that uses the historical data stream generated by the user itself to establish a long-term baseline for each monitoring object that represents its individual dynamic stability, and on this basis, continuously quantifies the deviation of the overall state from its historical baseline, thereby effectively filtering the background noise of daily behavior fluctuations while identifying the systematic long-term evolution trend that has indicative significance, becomes the technical problem to be solved by the present invention. SUMMARY

[0005] The application provides a health data processing and alarm method for a family monitoring group, which mainly aims to solve the problem that the existing data processing method is difficult to balance filtering daily fluctuation noise and identifying individual long-term slow change risk.

[0006] To achieve the above-mentioned purpose, the application provides a health data processing and alarm method for a family monitoring group, which comprises the following steps:

[0007] Obtaining at least two data streams representing the state of the monitored person;

[0008] For each data stream of the at least two data streams, a long-term moving average value representing the individual long-term stable habit of the data stream is established and updated, and the long-term moving average value is used as an individual dynamic steady-state baseline for subsequent judgment;

[0009] For each data stream, a short-term moving average value reflecting the recent behavior pattern of the data stream is established and updated;

[0010] A first processing path is executed, which calculates the normalized deviation of each data stream based on the difference between the short-term moving average value and the individual dynamic steady-state baseline, and aggregates the normalized deviation of each data stream into a system steady-state entropy value, and determines whether there is a continuous system evolution according to the time variation trend of the system steady-state entropy value, and sends a first type of prompt information when the determination is yes;

[0011] A second processing path is executed, which calculates the instantaneous dispersion of each data stream based on the difference between the current original data point of each data stream and the short-term moving average value, and determines whether the number of data streams whose instantaneous dispersion exceeds a preset dispersion threshold at the same time point reaches a preset resonance dimension threshold, to determine whether there is a system shock event, and sends a second type of prompt information when the determination is yes.

[0012] Preferably, the long-term moving average value is calculated using a first time period, and the short-term moving average value is calculated using a second time period smaller than the first time period.

[0013] Preferably, in the first processing path, the calculation method of the normalized deviation of each data stream is as follows: wherein, the normalized deviation of the i th data stream is the short-term moving average value of the data stream is the long-term moving average value of the data stream; and the generation method of the system steady-state entropy value is to perform weighted summation on the absolute values of the normalized deviations of all data streams.

[0014] ​Preferably, in the second processing path, before calculating the instantaneous dispersion, further comprising: for each data stream, calculating a short-term standard deviation as a statistical quantity representing its recent fluctuation amplitude based on a short-term moving average; the instantaneous dispersion is calculated by dividing the absolute value of the difference between the current raw data point of each data stream and its corresponding short-term moving average by its corresponding short-term standard deviation; the second type of prompt information is sent through a separate alarm channel.

[0015] Preferably, the method further comprises: pre-allocating the at least two data streams into at least two pre-set life domain categories; in the first processing path, while generating the system steady-state entropy value, calculating a domain entropy contribution for each life domain category based on the normalized deviation degree and the pre-allocated category; the first type of prompt information contains an indication information indicating the life domain category that contributes most to the change trend of the system steady-state entropy value.

[0016] Preferably, before sending the first type of prompt information, the method further comprises: for the life domain category that contributes most, calculating a coefficient of variation of the normalized deviation degrees of the data streams allocated within it to obtain an intra-domain consensus index; comparing the intra-domain consensus index with a pre-set consensus threshold; and according to the comparison result, adaptively adjusting the content of the first type of prompt information, when the intra-domain consensus index is higher than the pre-set consensus threshold, the content of the first type of prompt information is adjusted to contain a prompt for reliability attention to the data sources within the life domain category.

[0017] Preferably, in the first processing path, the way of judging according to the time change trend of the system steady-state entropy value comprises calculating the slope of the moving average line of the time series of the system steady-state entropy value, and comparing the slope with a pre-set trend slope threshold.

[0018] Preferably, the first type of prompt information is a non-urgent text suggestion that does not contain specific numerical guidance for the guardian to perform active emotional care.

[0019] Preferably, the method further comprises: monitoring the continuity of the data streams, and when detecting that the data interruption duration exceeds a pre-set hibernation threshold, suspending the calculation of the system steady-state entropy value; after recovering data reception from the data interruption state, entering a temporary calibration state, and calculating a temporary short-term baseline based on the newly received data; comparing the temporary short-term baseline with the long-term moving average stored before the interruption, and according to the comparison result, selectively performing one of the following operations: if the difference between the two is less than a pre-set smooth transition threshold, resuming the calculation of the system steady-state entropy value based on the long-term moving average; or, if the difference between the two is greater than or equal to the pre-set smooth transition threshold, discarding the long-term moving average and starting a brand new long-term moving average learning process.

[0020] Preferably, the method further comprises: storing the system steady state entropy values periodically generated by the first processing path to form a time series of system steady state entropy values; performing autocorrelation analysis on the time series of system steady state entropy values to obtain a physiological elasticity index representing the correlation strength of data points in the time series; and generating an assessment report on the resilience level of the health status of the person under care based on the physiological elasticity index, or dynamically adjusting the parameters used in the judgment in the first processing path.

[0021] Compared with the prior art, the present application has the following advantages:

[0022] 1. By establishing and updating long-term and short-term moving averages for each data stream and processing them in parallel, a system steady state entropy value representing the long-term evolution trend of the individual state of the person under care is obtained, as well as a judgment basis representing instantaneous systemic impact. Since the calculation of the system steady state entropy value smooths out short-term data fluctuations, it focuses on identifying persistent overall state deviations. At the same time, the calculation of instantaneous dispersion uses the aforementioned short-term moving average as a dynamic reference to identify coordinated and dramatic deviations of multiple data streams at the same time point. These two parallel processing paths share the intermediate calculation result of the short-term moving average, but respond to different time scale changes, making the previously independent and blind observation methods for trend analysis of slow-changing risks and detection of acute events into a comprehensive processing structure with internal data coordination and complementary functions, avoiding the inherent information omission problem of a single processing method.

[0023] 2. After calculating the system steady state entropy value representing the overall state deviation, instead of directly using it for alarm, the normalized deviation degrees constituting the entropy value are first classified into different life domain categories according to pre-set logic, and the domain entropy contribution of each category is calculated. This step reassociates a single abstract overall entropy value to its specific behavioral source. When a prompt message is needed, the system can retrieve the life domain category with the highest contribution and include it as part of the indication information. This way, the judgment result of the system state change is no longer a trigger signal that needs to be guessed by the receiver, but a data processing product with inherent logical explanation that contains preliminary directional guidance, improving the understanding and subsequent processing efficiency of the information receiver.

[0024] 3、In identifying the most contribution life field category, further set up an internal verification step, the step is for the normalization deviation of each data stream assigned in this category, calculate its dispersion statistics, get a domain consensus index, the high and low of the index, directly reflects the change trend of the multiple data sources under the category is consistent, on this basis, the system according to the index and the comparison result of consensus threshold, adaptively adjust the content of the final output of the indication information, when the consensus index is lower than the threshold, it is shown that the internal data consistency is high, then output the prompt indicating the change of the life field;When the index is higher than the threshold, the content of the prompt information is adjusted to contain the suggestion of the reliability of the data source in the category, this mechanism makes the behavior of information output directly linked with the quality of the internal data driving the output, avoids a misleading specific interpretation caused by the fault or isolated anomaly of a single data source, and improves the reliability of the whole data processing and alarm process.

[0025] 4、The method claimed in the application also includes a data stream continuity monitoring mechanism, when the data interruption duration exceeds the hibernation threshold, the calculation of the system steady state entropy value will be suspended, avoiding the error calculation based on the old long-term behavior baseline, after the data recovery, the system does not immediately resume the calculation, but enters a temporary calibration state first, calculates a temporary short-term baseline through the new data, and compares it with the long-term behavior baseline stored before the interruption, according to the comparison result, the system autonomously selects whether to resume the calculation based on the original baseline or to discard the old baseline and start a brand new baseline learning process, this suspension and calibration decision process makes the method adapt to the inevitable data interruption of the monitoring device in actual use, ensures that the core algorithm can maintain the effectiveness and logical rigor of its data processing results under the condition of discontinuous data. BRIEF DESCRIPTION OF DRAWINGS

[0026] Fig. 1 The health data double path parallel processing architecture of the application;

[0027] Fig. 2 The slow change risk and acute event detection performance comparison chart of the application;

[0028] Fig. 3 The double path processing core algorithm and correlation analysis module chart of the application. DETAILED DESCRIPTION

[0029] In order to make the purposes, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described in detail below, but those skilled in the art can understand that the embodiments in the specification and the features in the embodiments can be combined with each other without conflict, all of which belong to the protection scope of the present application, and the following embodiments are intended to explain the present application, and cannot be understood as a limitation on the protection scope of the present application.

[0030] The health data processing and alarm method for a family monitoring group according to the present application has a whole data processing architecture configured as a double-path parallel processing structure, which is composed of a first processing path for long-term system evolution trend analysis and a second processing path for sensing instantaneous system impact events. The two processing paths use partially same intermediate calculation results to improve processing efficiency, but respond to changes in different time scales respectively, so as to complementarily integrate trend analysis and abnormality detection in time series data processing, so as to realize digital processing of risk characteristics in different time scales in data flow representing the state of the monitored person. In an application scenario, the data processing flow of the present application starts from obtaining at least two data streams representing the state of the monitored person. These data streams can come from sensors or application program interfaces deployed in the monitoring environment for collecting non-medical health-related data. For example, one data stream can be daily step data representing activity amount, and another data stream can be night sleep duration data representing work and rest regularity. After receiving these data streams in time series form, the system establishes and continuously performs data processing for each independent data stream at the data processing server. Before any processing, the system first anonymizes the obtained original data, separates the personal identity information of the user from the data stream, and performs all subsequent processing on the de-identified data.

[0031] To effectively identify a slow and systematic state drift submerged within the normal fluctuations of daily data, the data processing system needs to establish a benchmark for each data stream that can characterize its individualized long-term stable habits. To this end, the system establishes and updates a long-term moving average for each data stream, using this long-term moving average as the individualized dynamic steady-state baseline for subsequent judgments. In one implementation, this long-term moving average is calculated using an exponential moving average algorithm. The calculation employs a first time period, the length of which is set to sufficiently smooth out most short-term and seasonal behavioral fluctuations, thus reflecting a stable long-term habit. For example, a 90-day time period can be set. The choice of this period length takes into account various factors. The technical trade-off between the pervasive impact of seasonal changes on human behavior patterns and the stability required for the data baseline reveals that too short a period cannot effectively filter out medium-term fluctuations, while too long a period makes the baseline too sluggish in responding to real, fundamental changes in lifestyle. In parallel with establishing an individualized dynamic steady-state baseline, to keenly capture recent changes in behavior patterns, the system also establishes and updates a short-term moving average reflecting its recent behavior patterns for each data stream. This short-term moving average can also be calculated using the exponential moving average algorithm, but it uses a second time period shorter than the aforementioned first time period, for example, a 15-day time period. The choice of this time period length aims to achieve effective tracking of recent behavioral changes while filtering out daily random fluctuations.

[0032] After establishing the aforementioned dual-track dynamic baseline, the system initiates the calculation of the first processing path, which aims to quantify the long-term systematic evolution of the overall state of the ward. Given that the data streams of different modalities have different physical dimensions and numerical ranges, directly comparing their absolute deviations is not meaningful in engineering practice. Therefore, the system performs a normalization step, calculating the normalized deviation of each data stream based on the difference between the short-term moving average and the individualized dynamic steady-state baseline. This calculation method is defined as a dimensionless formula: ,in, For the first Normalization deviation of each data stream This is the short-term moving average of the data stream. This is the long-term moving average of the data stream. Through this calculation, all different types of data streams are converted into a unified, directly comparable relative percentage deviation. For example, for a data stream measured in daily steps, if its long-term moving average is 8000 steps and its recent short-term moving average is 7200 steps, then its normalized deviation is calculated as (7200-8000) / 8000 = -0.10. To obtain a single indicator that can quantify the overall health stability of the ward, the system aggregates the normalized deviations of each data stream to generate a system steady-state entropy value. This value is generated by weighted summation of the absolute values ​​of the normalized deviations of all data streams. ,in, These are preset weighting coefficients, which can be offline calibrated based on the importance of different data streams to the overall state representation. For example, core behavioral data such as activity level and sleep can be assigned a higher weight, such as 0.4, while other auxiliary data can be assigned a lower weight. The sum of the weighting coefficients is set to 1. This ensures that the various data streams converge to form the system's steady-state entropy value. When, its weighting coefficient The allocation is based on objective evidence and reflects the individualized stability characteristics of each data stream. Before formal monitoring begins, the system can execute a weight calibration procedure based on initial accumulated data. This procedure uses a stable period of data, such as 90 days, to assign weights to each data stream. Calculate a coefficient of variation that characterizes its degree of dispersion. , ,in The standard deviation of the data stream during the steady-state period. Its arithmetic mean, then, the... Weight coefficients of each data stream It is set to be proportional to the reciprocal of the coefficient of variation of the data stream, and normalized to make the sum of all weight coefficients equal to 1. The calculation method is as follows: Where N is the total number of data streams, this approach assigns higher sensitivity to deviations of data streams with lower volatility when assessing overall system instability. It's important to note that the absolute value of the deviation is used here because assessing system instability focuses on the magnitude of the deviation rather than its direction of increase or decrease. Continuing the previous example, suppose there are two data streams, data stream 1 (step number)... Its weight Data stream 2 (sleep duration) Its weight The steady-state entropy of the system at this moment is calculated as follows: The final output The value is a dimensionless index, and the continuous rise of its value objectively indicates that the entire system of the person under guardianship is continuously deviating from its historical stable track.

[0033] An isolated High value may be caused by short-term disturbance, and does not necessarily constitute a risk, therefore, the method claimed in the present application is not based on the instantaneous value of , but on the time-varying trend of the steady-state entropy value of the system to determine whether there is a persistent systemic evolution; the specific determination method includes that the system stores the daily calculated value on the server side to form a time series of the value, and calculates a moving average line, such as a 30-day moving average line, for the time series to further smooth short-term fluctuations, then calculates the slope of the 30-day moving average line, and compares the slope with a preset trend slope threshold value, when it is determined that the slope is continuously greater than the threshold value, the system determines that there is a persistent systemic evolution, and sends a first type of prompt information to the terminal device of the guardian group; the first type of prompt information is a non-emergency text suggestion that does not contain specific numerical values, guiding the guardian to actively provide emotional care, for example, it is observed that the overall life pattern of the father in the past month is showing a continuous and slow change trend compared to his long-term habits, and it is suggested that you can actively communicate and care when convenient; to make up for the limitations of the first processing path that is not sensitive to acute changes, the system performs a second processing path in parallel, which is dedicated to sensing systemic shock events; the mechanism is that a real acute event is systemic, and usually at the same time, in multiple data dimensions, signals deviating from the recent normal are triggered at the same time, in order to capture this resonance phenomenon, the second processing path calculates the instantaneous dispersion of each data stream based on the difference between the current raw data point of each data stream and the aforementioned calculated short-term moving average value; in order to achieve the comparability of the instantaneous dispersion of different data streams, before calculating the instantaneous dispersion, a short-term standard deviation is also calculated for each data stream based on its short-term moving average value as a statistical quantity representing its recent fluctuation amplitude, and then the calculation method of the instantaneous dispersion is to divide the absolute value of the difference between the current raw data point of each data stream and its corresponding short-term moving average value by its corresponding short-term standard deviation, which is statistically equivalent to Z-score, representing the standard deviation multiple of the current data point deviating from the recent mean.

[0034] After obtaining the instantaneous dispersion of each data stream, the system performs a multi-dimensional resonance judgment, that is, whether the number of data streams whose instantaneous dispersion exceeds a preset dispersion threshold at the same time point reaches a preset resonance dimension threshold; wherein the dispersion threshold is set according to statistical principles, for example, it can be set to 3.0, which means that a data point falls outside 3 times the standard deviation, which is a small probability event, and the resonance dimension threshold is determined according to the total number of data streams accessed, for example, in a system that accesses 5 data streams, it can be set to 3, that is, at least 3 data streams occur at the same time Small probability deviation; when the resonance condition is met, the system determines that a systematic shock event has occurred, and sends a second type of prompt information, which is sent through an independent alarm channel with a higher priority than the first type of prompt information; in order to improve the effectiveness and interpretability of the first type of prompt information, the method can also include an information enhancement process, which first pre-allocates all data streams to different life field categories according to their behavior meanings, for example, daily steps, number of outings, etc. Data streams are classified into dynamic activity fields, and sleep duration, sitting time, etc. are classified into static rest fields; the specific engineering implementation steps of classifying each data stream into a life field category include: first, a data stream-field mapping table is pre-constructed on the data processing server side, which is a two-dimensional data structure, one column is the unique identifier of all data streams that the system can receive, and the other column is the preset field category name, such as dynamic activity or static rest; when a new data stream first accesses the system, the processing program first queries the mapping table, if the corresponding identifier is found, the normalized deviation degree calculated subsequently is automatically classified into the corresponding field entropy contribution degree accumulation; if not found, the system will place the data stream in a temporary unclassified field, and send a configuration prompt to the system maintenance terminal requesting manual calibration of its field attribution, after manual configuration is completed, the new mapping relationship of the data stream is updated to the mapping table, so that the entire classification management process has the ability to expand to future new data stream types; while calculating the total value in the first processing path, the system also calculates a field entropy contribution degree for each life field category based on the normalized deviation degree of each data stream and the preset category, that is, the weighted absolute deviation degree of all data streams in the field is summed up; when the first type of prompt information is triggered, the system will additionally compare the sizes of each field entropy contribution degree, and the life field category with the largest contribution degree will be included as an indication information in the final sent text, so that the prompt content is upgraded from the overall law being changed to the overall law being changed, which is mainly reflected in the habits related to daily activities, thereby providing preliminary directional guidance for subsequent communication of guardians.

[0035] ​​To avoid misleading interpretation of the system due to the failure of a single data source or isolated abnormal behavior, the system can also perform an internal verification step before sending the first type of prompt information containing the indication information. The step calculates a variation coefficient of the normalized deviation of each data stream allocated in the life domain category that contributes the most, obtaining an intra-domain consensus index. The variation coefficient is the ratio of the standard deviation to the average value, which is a dimensionless dispersion measure. A lower variation coefficient value indicates that the deviation of each data stream in the domain is similar, and the change has systematic consistency. A higher value indicates that the change is likely driven by extreme values of a single data stream. The system compares the calculated intra-domain consensus index with a pre-set consensus threshold, and adjusts the content of the prompt information adaptively according to the comparison result. If the consensus index is lower than the threshold, the prompt information containing the specific domain indication is sent as planned. If the consensus index is higher than the threshold, the system determines that the internal consistency of the change in this domain is questionable, and actively adjusts the content of the prompt information to include a prompt for reliability attention to the data sources in the life domain category. For example, if a certain data related to the father's daily activities is observed to have a large fluctuation recently, it is suggested that you can check whether the related device is working properly when convenient. This mechanism enables the system's output behavior to be associated with the quality status of its internal data.

[0036] Considering that in actual application, the monitoring device may cause long-term interruption of data flow due to user's outing, device damage or network problems, the method also includes a monitoring mechanism for data flow continuity. The system monitors the timestamp of each data flow. When detecting that the data interruption duration exceeds a pre-set hibernation threshold, such as 7 days, the system will suspend the calculation of the system steady-state entropy value and enter a hibernation state to prevent the risk of false calculation based on outdated and invalid long-term moving average. After recovering data reception from the data interruption state, the system does not immediately resume calculation, but enters a temporary calibration state based on a small period of data, such as 3 days, which can represent the current state, to quickly calculate a temporary short-term baseline. Then, the system compares the temporary short-term baseline with the long-term moving average stored before the interruption, and selectively performs subsequent operations according to the comparison result. If the difference between the two is less than a pre-set smooth transition threshold, indicating that the user's living habits have not changed fundamentally during the interruption period, the system determines that the old baseline is still valid, and resumes the calculation based on the long-term moving average. If the difference is greater than the threshold, indicating that the user's living habits have changed fundamentally during the interruption period, the system determines that the old baseline is no longer valid, and resumes the calculation based on the temporary short-term baseline. The calculation of the value, if the difference between the two is greater than or equal to the threshold value, the system determines that the old baseline has been invalidated, and it is abandoned, and automatically start a new long-term moving average learning process, this pause-calibration-decision process, to ensure that the core algorithm of the method in the real working conditions of the data is not continuous, still can maintain the effectiveness of its data processing results; for the daily fluctuations in the system state to dig out the deeper health information, the method can also include an evaluation of individual physiological resilience parallel analysis module; the module using the first processing path generated daily value, which is stored to form a time series of values, and in a fixed period, for example every week, on the recent period, for example 90 days of time series of self-correlation analysis, specifically, is to calculate the first-order autocorrelation coefficient, thereby obtaining a physiological resilience index representing the correlation strength of the data points before and after the time series, the internal logic of the index is that a good physiological resilience system, after being subjected to minor daily disturbances can quickly return to baseline, its value memory is weak, which is manifested as a lower autocorrelation, on the contrary, a declining system, the impact of the disturbance will continue for several days, which is value before and after the strong autocorrelation, that is, a higher index value; the obtained physiological resilience index can be used to generate a long-term assessment report on the resilience level of the health status of the person being monitored, or as a personalized parameter basis for dynamically adjusting the trend slope threshold in the first processing path.

[0037] Example 1: In a specific application deployment, the claimed method is used to process multiple non-medical grade health related data streams of a ward A, including daily step count data, nightly sleep duration data and social application usage duration data, for which individualized dynamic steady state baselines have been established at the beginning of the system operation, representing the individualized long term stable habits of the ward A. During a subsequent three-month period, the data of the ward A starts to change slowly, with the mean daily step count decreasing from about 7500 steps to about 6800 steps, and the mean nightly sleep duration decreasing from about 7 hours to about 6.2 hours. However, since the magnitude of the changes are within the normal fluctuation of each data stream, for example, the daily step count reading is still randomly distributed between 5500 steps and 8500 steps, a conventional data processing approach that relies on fixed thresholds set for each data stream independently, for example, triggering an alarm when the step count is below 4000 steps, fails to make any effective technical response during this period. For the above scenario, the claimed method operates according to the illustrated procedure, with the system utilizing the established and continuously updated long term moving average and short term moving average, and performing the calculation of the first processing path in parallel. For the daily step count data stream, the long term moving average stays around 7500 steps, while the short term moving average decreases slowly from 7450 steps to 6850 steps during the three-month period. Correspondingly, the normalized deviation calculated by the system each day also changes from close to -0.01 to about -0.09. In parallel, for the nightly sleep duration data stream, the long term moving average stays at 7 hours, while the short term moving average decreases slowly from 6.9 hours to 6.3 hours. Correspondingly, the normalized deviation also changes from -0.01 to about -0.10. The two data streams each have a small and statistically insignificant persistent deviation in a single dimension, which is integrated into a single index through the aggregate calculation of the steady state entropy value of the system, causing the value of the daily to increase from an initial baseline level of close to 0.01 to above 0.09 at the end of the period.

[0038] This processing approach resolves the contradiction between high sensitivity and high false alarm rate in data processing. Instead of directly thresholding the value of the daily , the system further performs trend analysis on the time series of the value, by calculating the 30-day moving average of the value sequence, the system obtains a line that filters out the single-day or multi-day fluctuations, and thus the value of the daily The smooth trend curve of the value contingency jump presents a clear slope that continuously and monotonously rises in the three-month period. When the slope finally exceeds the preset trend slope threshold of the system, the system determines that a continuous systematic evolution has occurred, and automatically sends a first type of prompt information to the designated guardian terminal device. The information content is a non-urgent text suggestion guiding active care. Subsequent communication triggered by the prompt information enables the guardian to learn about the objective life situation that the ward A recently faces, which leads to the gradual change in his behavior pattern. This long-term trend quantification method based on multi-dimensional data flow changes the focus of data processing from the out-of-bound judgment of instantaneous values to the quantification of the overall system instability trend, so that the slow-changing risk signal in the background noise can be effectively identified.

[0039] To further verify the technical advantages of the method of the present application in identifying such slow-changing risk signals, the following comparative examples are provided.

[0040] Comparative Example 1: This comparative example adopts a conventional data processing method based on independent channel fixed threshold judgment in the background technology, processes a group of synthetic time series data for the same purpose of testing as Example 2, to verify the technical limitations of the conventional method in dealing with slow-moving risks and systemic shock events; the most essential difference between the processing method of this comparative example and the method claimed in the present application is that it does not establish a short-term moving average, nor does it perform double-path parallel processing, but establishes a single long-term mean and its long-term standard deviation based on 90 days of historical data for each data stream, and deviates from the long-term mean by more than 3 times the long-term standard deviation of the single-day data point as the only alarm trigger condition. The three synthetic data streams used in the test have the same characteristics as in Example 2, including baseline noise, slow-moving drift signal (from day 31 to day 120, mean -0.2% per day), and instantaneous impulse signal (three data streams all deviate from the recent mean by +4 times the short-term standard deviation at a certain time); slow-moving risk identification test: during the 120-day test period, synthetic data containing slow-moving drift signals are input into the test system using the conventional data processing method, although the intrinsic mean of the three data streams has been declining since day 31, the daily change amplitude (-0.2%) is much smaller than the daily fluctuation range of the data itself (i.e. the standard deviation of the baseline Gaussian white noise), therefore, during the entire test period, no single-day data point deviates from its 90-day long-term mean by a magnitude that reaches the preset 3 times long-term standard deviation alarm threshold, and the test system log does not record any abnormal events; acute event perception test: in another test period, the test system is input with synthetic data that has a systemic shock at a certain time, although the instantaneous amplitude of the shock reaches 4 times the short-term standard deviation, since the judgment criterion adopted by the conventional method is the long-term standard deviation, which includes all fluctuations over a longer time range, its value is usually greater than the short-term standard deviation, and the calculation shows that the +4 times short-term standard deviation shock is equivalent to only about +2.2 times the long-term standard deviation after conversion, which does not reach the preset 3 times long-term standard deviation alarm threshold, therefore, the test system also fails to respond technically to this systemic shock event; in addition, when testing the baseline data containing random noise for a long time, it is found that due to the lack of multi-dimensional resonance judgment mechanism, the conventional method triggers two non-systemic false alarms in the test due to isolated noise spikes in a single data stream, which interferes with the effectiveness of the alarm, and the test results and analysis are shown in Table 1.

[0041] Table 1: Summary of test results of Comparative Example 1.

[0042]

[0043] The test results of Comparative Example 1 show that, in its technical principle, the conventional data processing method based on independent channel fixed threshold judgment cannot effectively identify the persistent slow change risk submerged in the daily fluctuation noise; at the same time, the fixed threshold set based on long-term statistical characteristics has the problem of insufficient sensitivity to systemic shock events caused by short-term state dramatic changes.

[0044] Example 2: In order to objectively verify the performance of the method claimed in the present application in identifying data evolution characteristics of different time scales, a comparative test based on synthetic time series data is designed and performed in this example, which aims to compare the performance of the method claimed in the present application with a conventional data processing method based on independent channel threshold judgment. The test platform is a computing simulation environment, in which three independent synthetic data streams simulating key data characteristics in a monitoring scenario are generated by a pre-set statistical model, namely data stream 1, data stream 2 and data stream 3. Each data stream is composed of a baseline Gaussian white noise sequence with a determined mean and standard deviation, a linearly superimposed slow change drift signal and an instantaneous pulse signal. Two processing groups are set in the test, namely a control group and a test group. The data processing method used in the control group is to monitor the three data streams independently, and when the single-day data point of any data stream deviates from its long-term historical mean by more than 3 times the standard deviation, it is determined as an abnormal event. The test group adopts the method claimed in the present application and performs data processing according to the disclosed double-path parallel processing procedure. The parameters are set as follows: the time period of long-term moving average is set to 90 days, the time period of short-term moving average is set to 15 days, the weight of each data stream in calculating the system steady-state entropy value is set to 1 / 3, the trend slope threshold of the first processing path is set according to the baseline data, the instantaneous dispersion threshold of the second processing path is set to 3.0, and the resonance dimension threshold is set to 3. In the slow change risk identification test in the first stage, a period of 120 days of synthetic data is input to the two processing systems. During the period from the 31st day to the 120th day, a linearly decreasing drift of -0.2% per day is applied to the mean of the three data streams. During the entire test period, the control group does not produce any abnormal event judgment, because although the mean of the data stream is continuously decreasing, each independent data point after superimposing the baseline noise does not deviate from the long-term historical mean by more than the pre-set 3 times standard deviation threshold. In contrast, the first processing path of the test group identifies this systemic evolution, and the key nodes of the data processing process are shown in Table 2.

[0045] Table 2: Data processing process example table of the first processing path of the test group in the slow change risk identification test.

[0046] ​​

[0047] As can be seen from Table 2, as the short-term EMA continues to go down with the tracking data mean, the absolute value of the normalized deviation of each data stream also increases steadily, which makes the aggregated daily SHE value present a one-way upward trend. In other words, the 30-day moving average of the SHE value further smooths the daily fluctuations, and its slope has continuously exceeded the preset trend slope threshold around the 90th day, thereby triggering the first type of prompt information. The test results show that by calculating the deviation of the individualized dynamic steady-state baseline from the actual mode and performing multi-dimensional aggregation, the method claimed in the present application can quantify and present the slowly varying signals with the same trend that are contained in the noise of each independent data stream; in the second phase of the acute event perception test, a synthetic data in which the values of the three data streams all deviate from their recent mean + 4 times the short-term standard deviation at the same hour of a day are input to the two groups of processing systems. The control group cannot respond because the fluctuation of the single point cannot break through the relatively wide threshold range of its judgment reference, which is the long-term mean and the long-term standard deviation; the second processing path of the test group recognizes this impact, and the core judgment basis is shown in Table 3.

[0048] Table 3: Example of data processing process of the second processing path of the test group in the acute event perception test.

[0049]

[0050] As shown in Table 3, at the moment of the impact, the instantaneous dispersion of the three data streams is all calculated as 4.0, which exceeds the preset dispersion threshold of 3.0. Since the number of data streams that exceed the threshold, i.e. 3, reaches the preset resonance dimension threshold, i.e. 3, the system determines that a systematic impact event has occurred and starts the sending process of the second type of prompt information. This resonance decision mechanism distinguishes between systematic impact and isolated noise of a single data stream by requiring multiple data streams to achieve a coordinated and dramatic deviation in a statistical sense. The test data show that the method claimed in the present application can process pattern changes of different scales in time series data through its dual-path parallel processing architecture.

[0051] Embodiment 3: This embodiment combines Figs. 1 to 3 a health data processing and alarm method for a family monitoring group, as Fig. 1As shown, it shows the complete path of data from input to output, the raw health data from sensor / application interface goes through data collection and preprocessing steps to generate de-identified data stream, which is used to build dual-track dynamic baseline on one hand, the baseline data is stored in A1: long / short-term baseline database and continuously updated and read, on the other hand, it directly enters the second processing path, the core of the architecture is two parallel processing paths, among which, the first processing path: trend analysis reads long / short-term baseline data, calculates normalized deviation, and generates system steady-state entropy value SHE, which is stored in A2: system steady-state entropy value sequence database, and based on its time series, the system evolution trend is judged, so as to obtain the evolution trend judgment result, in parallel, the second processing path: impact perception calculates instantaneous dispersion based on de-identified data stream and read short-term baseline, and performs multi-dimensional resonance judgment to obtain resonance judgment result, finally, the evolution trend judgment result and resonance judgment result of the two paths are jointly fed into the generation and sending prompt information module, which is responsible for sending first / second type prompt information to the guardian group terminal.

[0052] As shown in Fig. 2 , the horizontal axis of the chart is the method type, which is divided into traditional single threshold method and double baseline method of the present application, and the vertical axis is the detection performance%, three key performance indicators, slow risk detection rate, acute event detection rate and false positive rate are respectively shown by three different filling mode column charts, as can be seen from the figure, the slow risk detection rate of the traditional single threshold method is low, while the false positive rate is high, in contrast, the double baseline method of the present application shows performance improvement in slow risk detection rate and acute event detection rate, while the false positive rate is controlled at a low level.

[0053] As shown in Fig. 3 , the figure shows that after the multi-source data stream including daily step data, sleep duration data, social application data and other data streams are subjected to data anonymization processing and data continuity monitoring, they are used to calculate the long-term moving average with a time period of 90 days and the short-term moving average with a time period of 15 days in parallel; in the first processing path, the system performs normalized deviation calculation, system steady-state entropy value generation, 30-day moving average slope analysis and trend threshold judgment in sequence, and finally outputs the first type of prompt information, non-emergency emotional care suggestion, in the second processing path, the system performs instantaneous dispersion calculation, multi-dimensional resonance judgment, abnormal flow counting and resonance threshold judgment, and finally outputs the second type of prompt information, system impact event alarm, in addition, the figure also shows three analysis modules associated with the main processing flow, namely, life field classification field entropy contribution degree, domain consensus reliability verification, and physiological resilience index autocorrelation analysis.

[0054] Example 4: To ensure the data processing performance of the claimed method of the present application when applied to different individuals, after the system is initially deployed for a new ward B, a standardized offline calibration procedure can be performed to determine a set of core algorithm parameters suitable for the individual, which uses the historical data generated by the ward B during an initial data collection phase to configure a set of operating parameters for him / her; the calibration procedure is started after an initial data collection phase, which is set to 120 days before the system is formally put into operation, during which the system only collects and stores various data streams of the ward B, including daily steps, sleep duration and number of outings, without performing alarm judgment, after the collection phase ends, the calibration module on the server side of the system analyzes the 120 days of historical data in a background batch processing manner to determine the moving average time period, weight coefficient and judgment threshold, first, to determine the time period of the long-term moving average and the short-term moving average, the calibration module performs autocorrelation analysis on the time series of each data stream to calculate the change of the autocorrelation coefficient with time delay, the time period of the long-term moving average is set to the time delay required for the autocorrelation coefficient to first drop to a low correlation threshold of 0.2, while the time period of the short-term moving average is set to the time delay required for the autocorrelation coefficient to first drop to a higher correlation threshold of 0.6; then, the calibration module determines the weight coefficient of each data stream for calculating the system steady-state entropy value based on the 120 days of stable period data , the calibration module first calculates the coefficient of variation of each data stream in the stable period, i.e. the ratio of the standard deviation to the mean value, then sets the weight of each data stream to be inversely proportional to the reciprocal of its coefficient of variation, and performs normalization to make the sum of all weights equal to 1.

[0055] After the time period and weight are determined, the calibration module uses the calibrated parameters to perform a complete simulation calculation on the 120 days of stable period data to generate a historical time series of the system steady-state entropy value of 120 days in length, the fluctuation of which is used to represent the baseline noise level of the system, the calibration module further calculates the slope of the 30-day moving average of the sequence, and performs statistics on all values of the slope during this period to obtain its mean and standard deviation, finally, the trend slope threshold used for the first processing path judgment is set to the value of the mean slope plus 3 times the slope standard deviation; in parallel with this, to calibrate the consensus threshold for checking the reliability of the interpretation, the calibration module synchronously calculates the historical sequence of the intra-domain consensus index for each life domain category during the stable period when performing the aforementioned sequence simulation calculation, and performs statistical analysis on the value distribution of the index, finally, the consensus threshold Set as the 95th percentile of the historical distribution, when the consensus index within a domain exceeds this threshold, the system determines that the data consistency within that domain is abnormal. Through this series of calibration steps, a set of parameters based on the individual data characteristics of the ward B is determined and applied to subsequent online monitoring.

[0056] Example 5: During the long-term operation of the method claimed in this invention, in order to address the potential for continuous quality degradation in a single data stream, the system also includes a set of dynamic weight degradation and isolation procedures for data streams. In a specific scenario, the system detects that the data stream of outings belonging to the dynamic vitality domain has its corresponding sensor start outputting irregular noise data for some reason, causing the consensus index within the dynamic vitality domain to be higher than its preset consensus threshold for 5 consecutive days. In this situation, the system determines that there is a data source conflict within the domain, and further identifies the exit count data stream as the main contributor to the high domain consensus index. Subsequently, the system triggers a weight downgrade procedure, gradually reducing the weight of the exit count data stream in the calculation system's steady-state entropy value during subsequent calculation cycles. The weighting coefficient used at that time If the abnormal state of the data stream persists, its weight coefficient will eventually be reduced to 0, thereby functionally isolating the failed data source from the overall system status assessment. This procedure is used to protect core indicators. The long-term accuracy of the value calculation is not affected by persistent failures of local sensors.

[0057] Furthermore, to achieve individualized adaptive adjustment of the alarm logic, this method can also utilize the physiological resilience index calculated in the aforementioned specific implementation to dynamically modulate the judgment parameters in the first processing path. In a system deployed for over a year, the system has calculated a stable physiological resilience index for the ward C. The value is 0.7, which is higher than a baseline value set at 0.5, indicating that the system state of the ward C recovers more slowly from daily disturbances. The value fluctuates significantly on a daily basis. To avoid non-critical alerts caused by this individual characteristic, the system uses a preset adjustment logic: the adjusted trend slope threshold is equal to the initial calibration threshold multiplied by a factor. For coefficients with a positive correlation, the trend slope threshold of the first processing path was increased by 20% accordingly; conversely, if the other ward D's... If the value is calculated to be 0.3, the system will correspondingly lower its trend slope threshold by 20%, making the monitoring more sensitive to it. This parameter adaptive adjustment mechanism based on the physiological elasticity index enables the alarm logic of this method to better match the state fluctuation characteristics of different individuals.

[0058] Example 6: When the method is first deployed for a new ward, a pre-calibration procedure is performed to establish individualized parameter baselines for the ward; the procedure first enters a 90-day data accumulation period during which the system collects and stores data streams of daily step count, total nighttime sleep duration, daytime away-from-home activity duration, etc. but does not perform any alarm judgment, after the accumulation period, for each data stream , the arithmetic mean value thereof in the past 90 days is calculated as the initial long-term moving average , and the arithmetic mean value thereof in the past 15 days is calculated as the initial short-term moving average ; subsequently, the system performs retrospective calculation using the 90-day historical data accumulated to calibrate the relevant thresholds, for the first processing path, the system calculates a time series of 90 historical system steady-state entropy values (SHE), and performs sliding linear regression on the series using a 30-day time window to obtain 61 slope values; based on the statistical distribution of these slope values, the trend slope threshold is set at the 95th percentile, and its calculation formula is: wherein is the mean value of the 61 slope values, is the standard deviation; for the second processing path, the system calculates the instantaneous dispersion of all data streams in the 90-day historical data, and its calculation formula is: wherein is the original data point, is the 15-day short-term standard deviation of the corresponding data stream; the system counts the proportion of data points with values greater than 3.0, and if the proportion is less than 1%, the dispersion threshold is determined to be 3.0; at the same time, the system counts the number of data streams that exceed at the same time point , and based on the statistical result, the resonance dimension threshold is set to 3.

[0059] Taking a ward as an example, after the 90-day data accumulation period ends, the system obtains the initial parameters of the daily step count data stream as steps, steps, steps, and the initial parameters of the nighttime sleep duration data stream as hours, hours, hours, on the first day of entering formal monitoring, the system collects steps, hours, and the system performs the following calculations: first, the normalized deviation of each data stream is calculated, , ; subsequently, the system steady-state entropy value In parallel, the instantaneous dispersion is calculated , The number of data streams whose instantaneous dispersion exceeds the threshold value 3.0 is 0, and the resonance dimension threshold value 3 is not reached, so the second type of prompt information is not generated. At this point, the system completes the individualized parameter matrix calibration for the monitored person, and enters the long-term monitoring state with alarm capability.

[0060] It is apparent for a person skilled in the art that the present application is not limited to the details of the above exemplary embodiments, but can be implemented in other concrete forms without departing from the spirit or essential characteristics of the present application.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A health data processing and alarming method for a family care group, characterized in that, The method comprises: obtaining at least two data streams representing the state of the person under care; for each of the at least two data streams, establishing and updating a long-term moving average representing the individual long-term stable habit of the data stream, and taking the long-term moving average as an individual dynamic steady-state baseline for subsequent judgment; for each of the data streams, establishing and updating a short-term moving average reflecting the recent behavior pattern thereof; executing a first processing path, which calculates the normalized deviation of each data stream based on the difference between the short-term moving average and the individual dynamic steady-state baseline, and aggregates the normalized deviations of the data streams into a system steady-state entropy value, and determines whether there is a sustained systemic evolution according to the time trend of the system steady-state entropy value, and sends a first type of prompt information when the determination is positive; executing a second processing path, which calculates the instantaneous dispersion of each data stream based on the difference between the current raw data point of each data stream and the short-term moving average, and determines whether the number of data streams whose instantaneous dispersion exceeds a preset dispersion threshold at the same time point reaches a preset resonance dimension threshold, to determine whether there is a systemic shock event, and sends a second type of prompt information when the determination is positive; In the first processing path, the calculation of the normalized deviation for each data stream is as follows: wherein, the normalized deviation for the i-th data stream, is calculated as follows: is the short-term moving average of the i-th data stream, is the long-term moving average of the i-th data stream; and the system steady-state entropy is generated by weighted summing the absolute values of the normalized deviations of all data streams.

2. The health data processing and alarming method for a family care group according to claim 1, wherein, the calculation of the long-term moving average adopts a first time period, and the calculation of the short-term moving average adopts a second time period smaller than the first time period.

3. The health data processing and alarming method for a family care group according to claim 1, wherein, In the second processing path, before calculating the instantaneous dispersion, it further comprises: for each data stream, calculating a short-term standard deviation based on the short-term moving average as a statistical quantity representing the recent fluctuation amplitude thereof; the calculation method of the instantaneous dispersion is to divide the absolute value of the difference between the current raw data point of each data stream and the corresponding short-term moving average by the corresponding short-term standard deviation; the second type of prompt information is sent through an independent alarm channel.

4. The health data processing and alarming method for a family care group according to claim 1, wherein, The method further comprises: pre-allocating the at least two data streams to at least two preset life domain categories; in the first processing path, while generating the system steady-state entropy value, calculating a domain entropy contribution for each life domain category based on the normalized deviation and the pre-allocated category; the first type of prompt information contains indication information indicating the life domain category that contributes most to the change trend of the system steady-state entropy value.

5. The health data processing and alarming method for a family care group according to claim 4, wherein, Before sending the first type of prompt information, the method further comprises: for the life domain category that contributes most, calculating a coefficient of variation of the normalized deviations of the data streams allocated therein to obtain an intra-domain consensus index; comparing the intra-domain consensus index with a preset consensus threshold; and according to the comparison result, adaptively adjusting the content of the first type of prompt information, and when the intra-domain consensus index is higher than the preset consensus threshold, the content of the first type of prompt information is adjusted to contain a prompt for reliability attention to the data sources in the life domain category.

6. The health data processing and alarming method for a family care group according to claim 1, wherein, In the first processing path, the way of determining according to the time trend of the system steady-state entropy value comprises calculating the slope of the moving average line of the time series of the system steady-state entropy value, and comparing the slope with a preset trend slope threshold.

7. The health data processing and alarming method for a family care group according to claim 1, wherein, The first type of prompt information is a non-emergency text suggestion that does not contain specific numerical values and guides the guardian to provide active emotional care.

8. The health data processing and alarming method for a family care group according to claim 1, wherein, The method further comprises: monitoring continuity of the data stream, suspending calculation of the system steady-state entropy value when a duration of data interruption is detected to exceed a preset hibernation threshold; after resuming data reception from the data interruption state, entering a temporary calibration state, and calculating a temporary short-term baseline based on newly received data; comparing the temporary short-term baseline with a long-term moving average value stored before the interruption, and selectively performing one of the following operations according to the comparison result: if a difference between the two is less than a preset smooth transition threshold, resuming calculation of the system steady-state entropy value based on the long-term moving average value; or if the difference is greater than or equal to the preset smooth transition threshold, discarding the long-term moving average value and starting a brand-new long-term moving average value learning process.

9. The health data processing and alarming method for a family care group according to claim 1, wherein, The method further comprises: storing the system steady-state entropy values periodically generated by the first processing path to form a time series of system steady-state entropy values; performing autocorrelation analysis on the time series of system steady-state entropy values to obtain a physiological elasticity index representing correlation strength between data points before and after the time series; and based on the physiological elasticity index, generating an assessment report on the health state resilience level of the person being monitored, or dynamically adjusting parameters used for judgment in the first processing path.

Images