Health index dynamic contrast method and system based on time series compressed sensing

Abstract

The application relates to a health index dynamic contrast method and system based on time series compressed sensing. The method comprises the following steps: a terminal device performs compressed sensing on a health index original signal x by using a measurement matrix Phi to obtain a compressed vector y; a dynamic baseline signal b is acquired and compressed by using the same Phi matrix to obtain y b ; a residual vector e=y b is calculated in the compressed domain; abnormality judgment is performed based on e, the energy of e is calculated, if the energy exceeds a standard, the first K most significant sparse components are recovered by using a sparse reconstruction algorithm, and abnormality is identified according to the component characteristics. The system comprises data acquisition, baseline management, compressed domain processing, abnormality judgment and baseline updating modules. The application directly performs dynamic contrast and sparse analysis in the compressed domain, avoids full-range reconstruction of the original signal, and realizes low-power-consumption, high-efficiency, high-personalized health monitoring and abnormality early warning.

Description

Technical Field

[0001] This invention relates to the fields of medical and health information technology and signal processing technology, and in particular to a method and system for dynamic comparison of health indicators based on time-series compressed sensing. Background Technology

[0002] With the widespread adoption of wearable devices and IoT technology, the continuous collection of time-series data on users' health indicators, such as heart rate, blood pressure, blood oxygen saturation, blood glucose, and activity levels, has become a reality. This data is of great significance for chronic disease management, health risk assessment, and early disease warning.

[0003] However, existing technologies face the following prominent challenges:

[0004] 1. Data Redundancy and Transmission Power Consumption: Health monitoring devices need to continuously collect high-frequency data, resulting in a huge amount of data. Directly transmitting or storing all raw data would consume a lot of device power, communication bandwidth, and cloud storage resources, limiting the device's battery life.

[0005] 2. Real-time performance and computational pressure: Real-time analysis and anomaly detection of complete time-series data in the cloud or on mobile terminals is computationally expensive and difficult to implement on resource-constrained devices. Uploading all data to the cloud, on the other hand, poses latency and privacy risks.

[0006] 3. Challenges of Dynamic Comparison: Effective health monitoring requires not only observing indicators at individual time points but also dynamic and continuous comparison with individual historical baselines or population statistical baselines. Existing methods often use fixed time windows for comparison, which cannot adaptively adjust the granularity and frequency of comparison according to data change characteristics (such as stable periods or fluctuating periods). This results in insensitivity to slow trend changes or slow response to sudden anomalies.

[0007] 4. Individual differences: The physiological rhythms and normal fluctuation ranges of different individuals vary greatly. The general static threshold alarm method has a high false alarm rate and a poor user experience.

[0008] Therefore, there is an urgent need for a new method and system that can perform efficient and personalized dynamic comparative analysis directly under data compression. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for dynamic comparison of health indicators based on time series compressed sensing.

[0010] The core idea of ​​this method is to use compressed sensing theory at the data acquisition end (such as wearable devices) to collect non-adaptive linear measurements of health indicators (i.e., compressed data) at a rate much lower than the Nyquist sampling rate, and directly compare them with the dynamic baseline in the compressed domain, thereby achieving low power consumption, high efficiency, and highly personalized health status monitoring and abnormal early warning.

[0011] To achieve the above objectives, in a first aspect, the present invention provides a method for dynamic comparison of health indicators based on time-series compressed sensing, comprising the following steps:

[0012] Step S100: Compressed Sensing Data Acquisition. On the terminal device, a non-adaptive linear projection measurement is performed on the time-series signal x (length N) of the target health indicator to obtain a compressed measurement vector y, y = Φx, where Φ is an M×N measurement matrix (M << N). The measurement matrix Φ is preferably a Bernoulli matrix, a partial Fourier matrix, or a sparse random matrix to ensure that the finite isometry property (RIP) is satisfied.

[0013] Step S200: Dynamic baseline modeling and compression

[0014] A dynamic baseline signal b is constructed for comparison. The dynamic baseline includes:

[0015] Personal historical baseline: A desired time series representing the user's normal state, generated based on the user's historical health data and combined with time context (such as time of day or day of week).

[0016] Population statistical baseline: A reference time series generated based on statistical information of similar populations (same age, same sex, same health status). Using the same measurement matrix Φ as in step S100, the dynamic baseline signal b is compressed to obtain the compressed baseline vector y. b =Φ b .

[0017] Step S300: Dynamic comparison of compressed domains.

[0018] Calculate the compressed measurement vector y and the compressed baseline vector y directly within the compressed domain. b The residual vector e = y - y b According to the theory of compressed sensing, the residual vector e is approximately equal to the original signal residual x. b The compression measurement value, i.e., e ≈ Φ(x) b ).

[0019] Step S400: Residual analysis and anomaly detection.

[0020] The residual vector e is analyzed to determine the health status.

[0021] S410: Energy detection. Calculate the L2 norm (energy) of the residual vector e, ||e||. 2 If ||e|| 2 Exceeding the preset first threshold T energy If so, a fine-grained anomaly detection will be triggered.

[0022] S420: Sparse pattern analysis.

[0023] When energy detection is triggered, the complete signal is not reconstructed immediately; instead, the signal (x) is used. b Given prior knowledge of sparsity in a certain transform domain (such as the Fourier domain or wavelet domain), an optimization algorithm is used to recover only the K most significant sparse components and their magnitudes from the residual vector e, where K is a small integer much smaller than N. For example, using an improved Orthogonal Matching Pursuit (OMP) algorithm, stopping after K iterations, a sparse representation α is obtained. K .

[0024] Step S430: Anomaly Pattern Identification. Analyze the characteristics of the recovered K sparse components:

[0025] Frequency domain characteristics: Determine whether it appears in a specific frequency band representing pathological noise (such as a specific high-frequency component related to atrial fibrillation in a heart rate signal).

[0026] Time-domain characteristics: Determine whether its amplitude exceeds the second threshold T obtained based on individual historical residual statistics. sparse .

[0027] Persistence: Determines whether similar sparse patterns appear within multiple consecutive time windows. Based on a combination of one or more of the above characteristics, the system ultimately determines whether the current health status is abnormal and outputs the abnormality type, level, and confidence level.

[0028] Step S500: Feedback and Baseline Adaptive Updates. Based on long-term comparison results and user feedback, dynamically adjust the individual historical baseline model using machine learning models (such as online learning models) so that the baseline can follow the slow trend changes in the user's long-term lifestyle habits or health conditions.

[0029] Secondly, the present invention provides a dynamic comparison system for health indicators based on time-series compressed sensing to implement the above method, comprising:

[0030] Data acquisition module: integrated into terminal devices (such as smart bracelets, patch monitors), including sensors and compressed sensing acquisition circuits, used to execute step S100.

[0031] Baseline Management Module: Deployed on a server or terminal device, it stores and manages personal historical data and group statistics, and is used to generate and compress dynamic baselines, executing step S200.

[0032] Compression Domain Processing Engine: Deployed on the terminal device or edge gateway, it is used to perform energy detection and sparse mode analysis in step S300 and step S400. This engine is the core of the system and can run in a resource-constrained environment.

[0033] Abnormal Judgment and Early Warning Module: Receives the output result of the compression domain processing engine, performs pattern recognition and comprehensive determination in step S430, and generates early warning information when an abnormality is determined.

[0034] Baseline Adaptive Update Module: Deployed on the server, it executes step S500 and continuously optimizes the user's personal baseline model by leveraging the powerful computing capabilities of the backend.

[0035] The dynamic comparison method and system for health indicators based on time series compressive sensing provided by the present invention:

[0036] 1. High energy efficiency: The terminal device only needs to collect and transmit M measurement values (M << N), which greatly reduces the power consumption and cost of data collection, transmission, and storage, and extends the battery life of the device.

[0037] 2. Strong real-time performance and good privacy protection: Key dynamic comparison and primary anomaly detection can be completed at the terminal or edge side without uploading the original data to the cloud, with low response latency, and the user's original physiological signals do not leave the device, ensuring high privacy and security.

[0038] 3. Combination of "compression domain dynamic comparison" and "sparse mode analysis": Creatively proposes to directly perform dynamic baseline comparison within the compression domain, and further perform fine anomaly recognition by "partially reconstructing" the most informative sparse components. This method skips the cumbersome process of "complete reconstruction -> re-comparison" in traditional solutions, and while maintaining high accuracy, reduces the computational complexity from the O(N 3 ) magnitude (typical complexity of reconstruction algorithms) to the O(KMN) magnitude (K is very small), achieving a qualitative leap in analysis efficiency and solving the long-term technical problem of achieving refined health monitoring in a resource-constrained environment.

[0039] 4. Personalization and adaptability: Through dynamic personal historical baselines and a machine learning-based baseline update mechanism, the system can learn the unique patterns of users, significantly reduce the false alarm rate, and achieve true personalized health management. Description of the Drawings

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0041] Figure 1 A general flowchart of the method provided in the embodiments of the present invention;

[0042] Figure 2 This is a schematic diagram of the system architecture provided in an embodiment of the present invention;

[0043] Figure 3 A detailed flowchart of dynamic comparison and sparse pattern analysis of the compressed domain provided in the embodiments of the present invention.

[0044] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0046] In this embodiment of the invention, reference is made to Figures 1 to 3 This invention provides a method for dynamic comparison of health indicators based on time-series compressed sensing, comprising the following steps:

[0047] Step S100: Compressed sensing data acquisition step: On the terminal device, the health indicator time series signal x is subjected to compressed sensing measurement using the measurement matrix Φ to obtain the compressed measurement vector y.

[0048] Technical Principle: Compressed sensing theory states that if a signal is sparse or compressible in a certain transform domain, then a small number of linear projections of the signal (i.e., compressed measurements) can be acquired at a frequency much lower than the Nyquist sampling rate using a measurement matrix uncorrelated with the transform basis, and the original signal can then be reconstructed with high probability through an optimization algorithm. This invention utilizes the first part of this theory at the acquisition end, namely, acquiring only compressed measurements, and basing subsequent analysis on these compressed values.

[0049] Specific implementation:

[0050] 1. Signal Buffering: The photoplethysmography (PPG) sensor on the smart bracelet collects the user's heart rate data at a frequency of 1 second (1 Hz), forming a discrete time-series signal x. The system sets a processing window, for example, N = 3600, representing 1 hour of data.

[0051] 2. Measurement Matrix Φ: A measurement matrix Φ of size M×N is pre-stored in the wristband's firmware. To balance performance and computational complexity, we choose M = 900, i.e., a compression rate of 75% (only 25% of the data is retained). Φ can be a Bernoulli matrix (elements are ±1) because they are all integers, which facilitates fast addition and subtraction operations in embedded systems, and has been proven to satisfy the Finite Isometry Property (RIP) with a high probability.

[0052] 3. Compressed Measurement: After the buffer contains 3600 data points, the data acquisition module performs matrix-vector multiplication: y = Φ x The result of this operation is a vector y of length 900, which is a compressed measurement vector. This y vector contains global feature information of the original 3600 heart rate data points, but it is no longer possible to directly read the specific heart rate value.

[0053] 4. Data Transmission: The band then sends these 900 y values, instead of the original 3600 x values, to the paired smartphone via Bluetooth. This step directly reduces wireless data transmission by 71%, significantly lowering communication power consumption.

[0054] Step S200: Dynamic baseline compression step

[0055] Acquire the dynamic baseline signal b, and compress the dynamic baseline signal b using the same measurement matrix Φ as in the compressed sensing data acquisition step to obtain the compressed baseline vector y. b ;

[0056] Technical Principle: For meaningful comparisons to be made within the compressed domain, the reference signals to be compared must also lie in the same compressed space. The key to this step is compressing the baseline signals using the exact same measurement matrix Φ, ensuring that yi and yi... b The difference e between them is mathematically exact equal to the difference (x) between the original signals. b The compression measurement value of ), i.e., e = Φ(x) b This laid the mathematical foundation for directly analyzing signal deviations in the compressed domain.

[0057] Specific implementation:

[0058] 1. Generation of dynamic baseline b: On a smartphone or in the cloud, the baseline management module maintains the user's personal historical heart rate model. For example, for the time period of "10 pm to 11 pm on weekdays", the system will aggregate all heart rate data for this time period over the past month (excluding obviously abnormal dates), calculate its rolling average and standard deviation, and generate a smooth expected heart rate curve b representing the user's normal state, with a length of N=3600.

[0059] 2. Baseline Compression: When comparison is required (e.g., upon receiving a new y signal), the baseline management module uses the same Φ matrix as the wristband to compress the baseline signal b: y b =Φ b The compressed baseline vector y is obtained. b .

[0060] 3. Baseline distribution: y is computed in the cloud. b It can be distributed to smartphones for subsequent local comparison, reducing cloud interaction latency.

[0061] Step S300: Dynamic Comparison of Compressed Domains

[0062] Calculate the compressed measurement vector y and the compressed baseline vector y b The residual vector e between them.

[0063] Technical Principle: Based on the mathematical foundation established in steps S100 and S200, this step directly operates on the compression vector. Since y and y b Φ represents the projections of x and b onto the same observation matrix Φ, and their residual e directly reflects the degree to which the original signal deviates from its individual normal baseline. This comparison is performed in a low-dimensional space (M-dimensional), which is far more computationally efficient than reconstructing first and then comparing in the original high-dimensional space (N-dimensional).

[0064] Implementation details: Within a smartphone app (or a fitness tracker, if its computing power is sufficient), the compressed domain processing engine performs a simple vector subtraction operation: e = y - y b This residual vector e, with a length of 900, is the core input for subsequent anomaly analysis.

[0065] Step S400: Residual Analysis and Anomaly Detection Steps

[0066] Based on the residual vector e, it is determined whether the health indicator is abnormal.

[0067] Technical principle: A two-stage analysis strategy of "coarse to fine" is adopted. First, a rapid coarse screening is performed by calculating a simple vector norm. If anomalies are found, a fine analysis process with controllable computational load and based on sparse priors of the signal is initiated to extract the most critical abnormal pattern features from the compressed data, rather than reconstructing the entire signal.

[0068] Specific implementation:

[0069] Step S410: Energy detection sub-step

[0070] Principle: If a user's heart rate x is roughly consistent with its baseline b, the magnitude of the residual vector e will be small. Conversely, if x shows a significant abnormality (such as sustained tachycardia or sudden arrhythmia), the magnitude (energy) of e will increase. L2 norm ||e|| 2 It is a classic and easy-to-calculate metric for measuring the total energy of a vector.

[0071] Implementation: Calculation This energy value is compared with a preset first threshold T. energy The comparison can be performed. This threshold can be customized based on the statistical distribution of the user's historical residual energy (e.g., mean + 3 standard deviations). If ||e|| 2 ≤ T energy If the heart rate is within the normal range during this time period, the process ends; if it exceeds the range, the next step of detailed analysis is triggered.

[0072] Step S420: Sparse Mode Analysis Sub-step

[0073] Principle: Many physiological signal abnormalities are sparse in specific domains. For example, an occasional premature atrial contraction manifests as an isolated, morphologically abnormal pulse in the time domain; while the characteristics of atrial fibrillation manifest as specific high-frequency oscillations in the frequency domain. This invention utilizes this prior knowledge to reconstruct (x) from e without completely reconstructing it. b Instead of retrieving the baseline, we only recover the K most significant sparse components. This is equivalent to focusing only on the parts that deviate most from the baseline, greatly reducing computational complexity.

[0074] Taking the detection of occasional premature beats as an example, we assume an abnormality (x b It is sparse in the time domain (i.e., only a few points have large values).

[0075] 1. The orthogonal matching pursuit algorithm is adopted. The input of this algorithm is: measurement vector e, measurement matrix Φ, and sparse basis (in this case, identity matrix I, because we are analyzing in the time domain).

[0076] 2. The algorithm begins to iterate. In each iteration, it selects the atom (column vector) of Φ that is most relevant to the current residual and adds it to the support set.

[0077] 3. After K iterations (e.g., K=5), the algorithm stops. At this point, we obtain a sparse solution containing K non-zero values, which indicate the 5 specific time points in the original 1-hour signal where anomalies are most likely to occur and their anomalous amplitudes.

[0078] Step S430: Anomaly Pattern Recognition Sub-step

[0079] Principle: The computational burden of performing pattern recognition on the small number of sparse components recovered in step S420 is much smaller than that of analyzing the complete signal.

[0080] Implementation: Analyze the five recovered sparse components.

[0081] Time-domain amplitude: If the amplitude of a certain component (representing the degree of heart rate variability) exceeds the second threshold T set based on personal historical data... sparse If so, it is marked as suspicious.

[0082] Continuity: Check if multiple points among these 5 points appear consecutively or periodically in time.

[0083] Comprehensive Judgment: For example, a system rule could be defined as: "If more than 3 amplitude values ​​exceed T within 15 minutes..." sparse Isolated pulse points are identified as 'frequent premature ventricular contractions'. Based on this rule, the system outputs the final abnormality type and alarm level.

[0084] Furthermore, the dynamic baseline signal b is an individual historical baseline signal and / or a group statistical baseline signal; the individual historical baseline signal is generated based on the user's historical health data and combined with time context information.

[0085] Technical principle:

[0086] Dynamic baselines are central to achieving personalized and precise health monitoring. Static thresholds cannot reflect the diurnal rhythms of human physiological indicators, weekday / weekend differences, and individual-specific fluctuation patterns. This invention constructs a multi-dimensional, three-dimensional comparative benchmark by integrating an individual's historical baseline with temporal context and a population statistical baseline as a broad-spectrum reference, significantly improving the targeting and accuracy of anomaly detection.

[0087] Specific implementation:

[0088] 1. Generation of personal historical baseline signals

[0089] Data Preparation: The system continuously collects and stores users' historical health indicator time-series data (such as heart rate and blood pressure). Each data point has a precise timestamp, from which rich temporal context information can be extracted, including:

[0090] Intraday time (e.g., 0-23 hours)

[0091] A certain day of the week (Monday to Sunday)

[0092] Workday and holiday signs

[0093] Seasonal Information

[0094] Modeling and computation: We use a “time binning aggregation” model to generate baselines.

[0095] Binning: Divide a week into multiple finer "time boxes" according to the above context. For example, a week can be divided into... The time slots can be divided into "weekday daytime", "weekday nighttime", "weekend daytime", "weekend nighttime", etc.

[0096] Aggregation: For each time bin (e.g., the bin "Wednesdays from 10 a.m. to 11 a.m."), aggregate all original signal fragments (or signals reconstructed from compressed data) that fell into this bin over the past few weeks (e.g., 4 weeks) for that user.

[0097] Statistical Calculation: For the data within each time bin, robust statistics (such as the median and interquartile range (IQR)) are calculated to resist the interference of random outliers. The median of that time bin constitutes the value of the personal historical baseline signal b at that time point. Concatenating the values ​​of all time bins in chronological order forms a complete, periodic personal historical baseline signal b. personal .

[0098] 2. Generation of population statistical baseline signals

[0099] Data source: On cloud servers, the system stores a large amount of de-identified health data of similar users (same age group, same gender, same health status).

[0100] Generation Method: A time-binding method similar to that used for individual historical baselines is employed, but within each time bin, population statistics (such as mean and standard deviation) are calculated. The resulting baseline signal is denoted as b. population .

[0101] 3. Synthesis of dynamic baseline signal b

[0102] The final dynamic baseline signal b can be:

[0103] Pure individual baseline: b = b personal Suitable for experienced users who already have sufficient historical data.

[0104] Weighted fusion of individual and group data: For new users or situations with sparse data, The weight α is dynamically adjusted from 0 to 1 as the amount of user data increases.

[0105] Dual baseline comparison: In step S300, the baseline can be calculated separately for each individual baseline b. personal and population baseline b population The residuals can be used to determine whether the anomaly is related to changes in personal habits or has deviated from the normal range of the group, thus providing richer context for the anomaly.

[0106] Furthermore, the residual analysis and anomaly detection steps include:

[0107] Energy detection sub-step: Calculate the norm of the residual vector e; when the norm exceeds a first preset threshold, trigger a fine anomaly determination.

[0108] Sparse pattern analysis sub-step: In response to the trigger, based on the prior knowledge of the sparsity of the signal residual in the transform domain, recover the top K most significant sparse components from the residual vector e, where K is a preset integer greater than zero, and K is much smaller than the original signal length N. Instead of computationally intensive full-signal reconstruction, partial reconstruction is performed, i.e., only the K most significant components that best explain the residual energy are recovered. This is equivalent to directly extracting the anomaly's "skeleton" or "fingerprint," reducing the computational complexity from O(N^2) to O(N^2). 3 ) down to This makes it possible to perform detailed analysis on resource-constrained terminals.

[0109] In this embodiment, the sparse pattern analysis sub-step employs an orthogonal matching pursuit algorithm or a compressed sampling matching pursuit algorithm, and stops after K iterations to recover the K sparse components.

[0110] Anomaly pattern recognition sub-step: Based on the frequency domain characteristics, time domain amplitude and / or persistence of the K sparse components, anomaly determination is performed, resulting in extremely high decision-making efficiency.

[0111] Furthermore, the dynamic comparison method for health indicators based on time-series compressed sensing also includes:

[0112] Baseline adaptive update step: Based on long-term anomaly detection results and user feedback, the personal historical baseline signal model is dynamically updated using a machine learning model.

[0113] In order for the system to adapt to long-term changes in users’ lifestyles or slow changes in health status (such as an overall decrease in resting heart rate after exercise), a feedback mechanism needs to be established so that the baseline model can continuously learn and evolve.

[0114] Specific implementation:

[0115] In the cloud, the baseline adaptive update module collects long-term monitoring results.

[0116] If a certain pattern is initially identified as "abnormal" by the system, but the user repeatedly marks it as "normal" in the app feedback (e.g., a transient increase in heart rate after drinking coffee), or the system continuously observes a slow, trending decrease in the user's heart rate baseline.

[0117] At this point, the system will launch an online learning algorithm (e.g., using recursive least squares or a simple exponential smoothing model) to fine-tune the parameters in the individual's historical baseline model with new, confirmed normal data at a small learning rate.

[0118] After the update, in the future, regarding the situation of "increased heart rate after drinking coffee", (x b The differences will be smaller, thus avoiding continuous false alarms and achieving true personalization.

[0119] Another aspect of the present invention provides a dynamic comparison system for health indicators based on time-series compressed sensing, comprising:

[0120] The data acquisition module is used to perform the compressed sensing data acquisition step and is integrated into the hardware circuit and embedded software of the terminal device (such as a smart bracelet or smart patch).

[0121] This module connects directly to physical sensors such as heart rate, blood oxygen, and acceleration sensors. Its core is a compressed sensing acquisition circuit / firmware. Internally, it stores a fixed measurement matrix Φ that satisfies finite isometric properties (e.g., an optimized sparse Bernoulli matrix consisting of ±1s). The module buffers the raw time-series signal x (of length N) from the sensor at a fixed frequency (e.g., 1 Hz). When the buffer is full, it does not directly upload these N data points, but instead performs a matrix-vector multiplication y = Φ. x This generates a compressed measurement vector y containing only M measurements (M << N). It then transmits y via Bluetooth or a Low Energy Wide Area Network. This step is fundamental to the system's low-power operation because it fundamentally reduces the amount of data that needs to be transmitted wirelessly.

[0122] The baseline management module is primarily deployed on cloud servers to leverage their powerful computing and storage capabilities. To improve response speed, endpoints or edge gateways can cache short-term baselines.

[0123] This module maintains each user's personal historical baseline model and group statistical baseline model.

[0124] Personal historical baseline generation: It receives and stores compressed historical data (or reconstructed data from the cloud) from users over a long period. By analyzing long-term user data, it establishes an expected value for each point in time within a typical cycle (such as a day or a week). For example, for heart rate at 9:00 AM, it might calculate a robust mean (excluding outliers) of heart rate data at the same time over the past 30 working days as the baseline value, thus forming a smooth, periodic baseline signal b.

[0125] Key operation: When comparison is required, this module generates or calls the baseline signal b, and calculates the compressed baseline vector y using the same measurement matrix Φ as the terminal. b =Φ b This is a prerequisite for achieving "compressed domain comparison," ensuring that y and y b Within the same mathematical space.

[0126] Connection: This module receives long-term historical data from the data acquisition module and provides it to the compression domain processing engine. b .

[0127] A compression domain processing engine is used to perform the dynamic comparison step of the compression domain and at least part of the residual analysis and anomaly detection steps; it is preferably deployed on a terminal device or home edge gateway / smartphone. This is crucial for achieving real-time performance and privacy protection. It receives y from the data acquisition module and y from the baseline management module. b It outputs the analysis results (such as energy values, the positions and amplitudes of the recovered K sparse components) to the anomaly detection and early warning module.

[0128] Dynamic comparison of the compressed domain: It performs a simple vector subtraction e = y - y b Based on the linearity of compressed sensing, we have e = Φ x - Φ b = Φ(x b This means that the residual vector e is mathematically equivalent to the difference (x) between the original signal and the baseline signal. b The compressed measurement values ​​of x and b. This principle allows us to analyze the differences between them directly without reconstructing the complete x and b.

[0129] Residual analysis:

[0130] Energy detection: Calculating ||e|| 2 If (x b If (x) is very small (i.e., in a normal state), then the energy of e will also be very small. b If a significant anomaly occurs (whether a sudden spike or a sustained shift), its energy will be reflected in ||e||. 2 This increases the value. This is a coarse screening step with extremely low computational cost.

[0131] Sparse Pattern Analysis (Creative Core): When energy exceeds the limit, the engine does not rebuild a complete (x) b Instead, it utilizes (x) b Given a sparsity prior in a certain transform domain (e.g., the time domain is sparse for sudden spikes, or the frequency domain is sparse for specific oscillations), an optimization algorithm (e.g., Orthogonal Matching Pursuit - OMP) is used to recover the K most significant sparse components (e.g., K=5) from e. The OMP algorithm iterates K times, finding the basis vector that contributes the most to interpreting the residual vector e each time, and recording its position and coefficients. The ingenuity of this step lies in reducing the computational complexity from O(N) for complete reconstruction to O(N). 3 The value is reduced to O(KMN), making it possible to perform fine-grained anomaly analysis on resource-constrained terminals.

[0132] The anomaly detection and early warning module, based on the output of the compressed domain processing engine, performs the final anomaly determination and generates early warning information. It can be deployed at the edge (for real-time emergency alarms) or in the cloud (for comprehensive analysis). This module receives the output results from the compressed domain processing engine. It maintains a rule base for different anomaly patterns. For example:

[0133] Rule 1: If the recovered sparse component appears as an isolated, high-amplitude pulse in the time domain, and its amplitude exceeds the threshold based on personal historical statistics, it is judged as "occasional premature beats".

[0134] Rule 2: If the recovered sparse components are concentrated in the atrial fibrillation-related frequency band in the frequency domain (which can be analyzed by performing short-time Fourier transform on a small number of recovered time-domain points) and are persistent, then it is judged as "atrial fibrillation risk".

[0135] Ultimately, it integrates these features to make a confident anomaly judgment and triggers corresponding warnings (such as screen vibration, push notifications, and contacting emergency contacts).

[0136] The baseline adaptive update module, deployed on the server, executes the baseline adaptive update steps. This module is a feedback loop. It collects long-term monitoring results and user feedback on false alarms. For example, if the system repeatedly detects an "abnormal" pattern, but the user marks it as "normal" (e.g., after exercise), the module will initiate an online learning algorithm to gradually adjust the user's personal historical baseline model, accepting this pattern as a new "normal" state. This gives the system the ability to continuously learn and evolve personalized.

[0137] Specific implementation case: Nighttime sleep apnea screening based on smart bracelets

[0138] Background: Sleep apnea is characterized by a periodic decrease in blood oxygen saturation (SpO2) during sleep. Traditional methods require recording high-sampling SpO2 data throughout the night, which is power-intensive and generates large amounts of data.

[0139] Implementation details:

[0140] 1. Initialization and parameter settings:

[0141] Terminal device: Smart bracelet equipped with SpO2 sensor.

[0142] Signal: SpO2 time series x, sampling rate 1 / 60 Hz (one point per minute), raw data length N=480 (8 hours).

[0143] Compressed sensing: The measurement matrix Φ is a 120×480 Gaussian random matrix with a compression rate of 75% (only 120 measurements are transmitted).

[0144] Baseline: Personal historical baseline b is a smoothed curve formed by the average SpO2 of the user during the same period over the past week (approximately a flat line of 95%-100%).

[0145] 2. Nighttime monitoring procedure:

[0146] S100: Data Acquisition. The wristband collects SpO2 data every minute throughout the night and caches it internally. Every 4 hours (N=240), the compressed sensing data acquisition module uses Φ to compress the cached data, obtaining y, and transmits it at low speed via Bluetooth to the edge gateway at the bedside (acting as the compression domain processing engine).

[0147] S200: Baseline Compression. The cloud-based baseline management module pre-compresses the baseline vector y for the current nighttime period. b It is then sent to the edge gateway.

[0148] 300: Comparison. The gateway calculates e = y - y b .

[0149] S400: Analysis:

[0150] S410: Gateway Computation ||e|| 2 When the user's blood oxygen levels are normal, energy levels remain low. When a sleep apnea event occurs, SpO2 begins to decrease (x... b A sustained negative offset occurs, causing ||e|| 2 A significant increase triggered detailed analysis.

[0151] S420: The sparse pattern analysis unit of the gateway is activated. Since the SpO2 decrease event in the time domain exhibits a sustained "ramp" or "plateau" lasting several minutes, rather than an instantaneous peak, we choose the first-order difference domain as the sparse basis (because the continuous decrease / rise of blood oxygen is sparse in the difference domain). Using the OMP algorithm (K=3), (x) is recovered from e. b The three most significant sparse components in the difference domain. These components precisely indicate the critical moments when blood oxygen begins to decrease, reaches its lowest point, and begins to recover.

[0152] S430: The anomaly detection module analyzes the recovered components. It identifies a typical "decline-sustainment-recovery" pattern, with the entire event lasting approximately 30 seconds (estimated based on component location) and a blood oxygen saturation decrease of approximately 4% (estimated based on component amplitude). This pattern repeats within multiple consecutive time windows. Based on this, the system classifies it as a "moderate sleep apnea-hypopnea event" and records the frequency and duration of the event.

[0153] 3. The next morning:

[0154] Users can view a detailed sleep and respiratory health report on a mobile app, instead of the original thousands of SpO2 data points.

[0155] S500: Update. If the user confirms symptoms such as snoring that night, this event data will be included in the cloud baseline adaptive update module to fine-tune the user's sensitivity threshold to SpO2 fluctuations; if the user reports good sleep, the system will slightly increase the alarm threshold to reduce the possibility of false alarms in the future.

[0156] As can be seen from the above specific examples, this invention, while greatly saving device power consumption and communication resources, has successfully achieved accurate and interpretable monitoring of the complex physiological event of sleep apnea, fully demonstrating the practical value and inventiveness of this invention.

[0157] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A dynamic comparison method for health indicators based on time-series compressed sensing, characterized in that, The method includes the following steps: Compressed sensing data acquisition steps: On the terminal device, the measurement matrix Φ is used to perform compressed sensing measurement on the health indicator time series signal x to obtain the compressed measurement vector y; Dynamic baseline compression step: Acquire dynamic baseline signal b, and compress the dynamic baseline signal b using the same measurement matrix Φ as in the compressed sensing data acquisition step to obtain compressed baseline vector y. b ; Dynamic comparison steps in the compression domain: Calculate the compression measurement vector y and the compression baseline vector y b The residual vector e between them; Residual analysis and anomaly detection steps: Based on the residual vector e, determine whether the health indicator is abnormal.

2. The method for dynamic comparison of health indicators based on time-series compressed sensing as described in claim 1, characterized in that, The dynamic baseline signal b is a personal historical baseline signal and / or a group statistical baseline signal; the personal historical baseline signal is generated based on the user's historical health data and combined with time context information.

3. The dynamic comparison method for health indicators based on time-series compressed sensing as described in claim 1, characterized in that, The residual analysis and anomaly detection steps include: Energy detection sub-step: Calculate the norm of the residual vector e; when the norm exceeds a first preset threshold, trigger a fine anomaly determination. Sparse pattern analysis sub-step: In response to the trigger, based on the prior sparsity of the signal residual in the transform domain, recover the top K most significant sparse components from the residual vector e, where K is a preset integer greater than zero and K is much smaller than the original signal length N. Anomaly pattern recognition sub-step: Based on the frequency domain characteristics, time domain amplitude, and / or persistence of the K sparse components, anomaly determination is performed.

4. The dynamic comparison method for health indicators based on time-series compressed sensing as described in claim 3, characterized in that, The sparse pattern analysis sub-step employs an orthogonal matching pursuit algorithm or a compressed sampling matching pursuit algorithm, and stops after K iterations to recover the K sparse components.

5. The dynamic comparison method for health indicators based on time-series compressed sensing as described in claim 1, characterized in that, Also includes: Baseline adaptive update step: Based on long-term anomaly detection results and user feedback, the personal historical baseline signal model is dynamically updated using a machine learning model.

6. A dynamic comparison system for health indicators based on time-series compressed sensing for implementing the method as described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to perform the compressed sensing data acquisition step; The baseline management module is used to perform the dynamic baseline compression step; A compressed domain processing engine is used to perform the compressed domain dynamic comparison step and at least part of the residual analysis and anomaly detection steps; The anomaly detection and early warning module is used to perform the final anomaly detection and generate early warning information based on the output of the compression domain processing engine.

7. The dynamic comparison system for health indicators based on time-series compressed sensing according to claim 6, characterized in that, The data acquisition module and the compression domain processing engine are deployed on the terminal device; the baseline management module is deployed on the server or the terminal device.

8. The dynamic comparison system for health indicators based on time-series compressed sensing according to claim 6, characterized in that, The compressed domain processing engine is configured to perform energy detection and sparse pattern analysis.

9. The dynamic comparison system for health indicators based on time-series compressed sensing according to claim 6, characterized in that, Also includes: The baseline adaptive update module, deployed on the server, is used to perform the baseline adaptive update steps.

10. A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images