Intelligent health prediction method based on wearable device data

Abstract

The invention relates to the technical field of health data analysis, in particular to an intelligent health prediction method based on wearable device data. According to the method, multiple types of physiological signals are collected through a unified timestamp, multi-scale decomposition and energy entropy calculation are carried out on the signals, internal complexity characteristics of the physiological signals are accurately extracted, correlation analysis and synchronous fluctuation interval recognition of the signal energy entropy are subsequently utilized, and a correlation combination of collaborative change of the physiological signals is selected; the defect that complex interaction of multi-source signals cannot be captured in the prior art is overcome, meanwhile, bidirectional sequence processing is introduced to conduct deep analysis on an associated physiological signal time sequence, switching of a motion state and a resting state is dynamically predicted, a time weighting function is constructed in combination with state information, weighted smoothing processing is conducted on entropy feature data, and the accuracy of the data is improved. The nonlinear variation trend of the physiological signals is effectively reflected, and finally, the recursive filtering is used for updating the physiological parameters, so that finer and more accurate dynamic prediction of the health state of the individual is realized.

Description

Technical Field

[0001] This invention relates to the field of health data analysis technology, and in particular to an intelligent health prediction method based on wearable device data. Background Technology

[0002] The field of health data analysis technology involves the collection, processing, and analysis of human physiological, biochemical, and behavioral data using computer science, sensing technology, and statistical modeling methods to achieve functions such as health status monitoring, disease risk assessment, and individual health management. This includes the acquisition of multi-source health data, feature extraction and modeling, trend analysis of time-series health parameters, and intelligent judgment and prediction of health status.

[0003] Among them, intelligent health prediction methods refer to using physiological signals and exercise data collected by wearable devices to make a preliminary judgment on an individual's health status through preset statistical analysis or rule matching methods. Such methods typically rely on fixed thresholds to compare data such as heart rate, blood pressure, and cadence, and achieve short-term trend prediction through time series analysis, thereby issuing a prompt when the monitored values ​​exceed the standard range.

[0004] Existing technologies mainly rely on preset statistical analysis or rule matching methods to compare physiological data such as heart rate, blood pressure, and cadence using fixed thresholds, thereby making preliminary judgments and short-term trend predictions of an individual's health status. This operating mode has significant limitations. Its fixed threshold setting fails to fully consider the physiological differences and dynamic changes that are common among individuals, resulting in low differentiation between normal fluctuations and abnormal changes in physiological signals, leading to false alarms or missed alarms during monitoring and early warning. In addition, based solely on preliminary judgments and short-term trend predictions, it is difficult to deeply capture the complex nonlinear interactions between multi-source physiological signals, nor can it effectively identify the subtle evolutionary patterns of physiological states. The ability to detect and accurately assess potential health risks in the early stages is insufficient, affecting the comprehensiveness and reliability of health status assessment. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent health prediction method based on wearable device data.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent health prediction method based on wearable device data, comprising the following steps:

[0007] S1: Collect patients' physiological signal data with a unified timestamp, perform multi-scale decomposition on each type of physiological signal and calculate the corresponding signal energy entropy to obtain the initial entropy feature dataset;

[0008] S2: Analyze the correlation of the signal energy entropy corresponding to each type of physiological signal in the initial entropy feature dataset, sort each type of physiological signal according to the correlation, and select the associated physiological signal combination with the same trend to obtain the associated feature combination data;

[0009] S3: Analyze the continuous time series of each associated physiological signal combination in the associated feature combination data to predict the switching between the patient's motor state and resting state, and obtain the state prediction result;

[0010] S4: Based on the state prediction results and the initial entropy feature dataset, a time-weighted function is constructed to reflect the nonlinear change trend of physiological signals, and weighted smooth feature data is obtained;

[0011] S5: Establish a dynamic prediction relationship between the patient's physiological parameters and the state prediction results through the weighted smoothed feature data, and generate the patient's health prediction results.

[0012] As a further aspect of the present invention, the initial entropy feature dataset includes heart rate signal energy entropy, skin temperature signal energy entropy, and signal time series features; the associated feature combination data includes a group of associated features between heart rate and skin temperature, signal synchronous fluctuation intervals, and correlation ranking information; the state prediction results include an exercise state interval, a resting state interval, and a state switching time identifier; the weighted smoothing feature data includes a time-weighted coefficient sequence, smoothed heart rate features, and smoothed skin temperature features; and the health prediction results include heart rate change trends, skin temperature change trends, and a comprehensive health status assessment.

[0013] As a further aspect of the present invention, the steps for obtaining the initial entropy feature dataset are specifically as follows:

[0014] S111: Collect heart rate and skin temperature data of patients through wearable devices, and perform synchronous sampling with a unified timestamp. Perform mirror extension on the time sequence of the collected signals at the sampling boundary to establish synchronous time sequence signal data of heart rate and skin temperature signals.

[0015] S112: Perform multi-scale decomposition on the heart rate signal and skin temperature signal in the synchronous time-series signal data, extract the detail coefficients and approximation coefficients of the heart rate signal and skin temperature signal, calculate the energy of the detail coefficients and approximation coefficients in each frequency band, obtain the heart rate signal energy and skin temperature signal energy, and integrate them into a multi-scale signal energy set.

[0016] S113: Calculate the probability distribution of the proportion of heart rate signal energy and skin temperature signal energy in each frequency band of the multi-scale signal energy set, obtain the entropy of heart rate signal energy and skin temperature signal energy based on the probability distribution, and arrange them in time series to generate an initial entropy feature dataset.

[0017] As a further aspect of the present invention, the step of obtaining the associated feature combination data specifically includes:

[0018] S211: Perform interpolation correction on the time series of heart rate signal energy entropy and skin temperature signal energy entropy in the initial entropy feature dataset to supplement missing measurement points, and calculate the Pearson correlation coefficient of the two at multiple time points to represent their mutual relationship, generating a set of signal energy entropy correlation coefficients;

[0019] S212: Sort the Pearson correlation coefficients in the set of signal energy entropy correlation coefficients, and judge the changing trends of heart rate signal energy entropy and skin temperature signal energy entropy based on the consistency of the changing direction, and obtain the synchronous fluctuation range of the two types of physiological signals changing in tandem.

[0020] S213: For the synchronous fluctuation interval, select the associated physiological signal combination with consistent change trend from the initial entropy feature dataset, extract the coordinated change features of heart rate signal energy entropy and skin temperature signal energy entropy within the interval, and establish associated feature combination data.

[0021] As a further aspect of the present invention, the step of obtaining the state prediction result specifically includes:

[0022] S311: Input the continuous time series of heart rate signal energy entropy and skin temperature signal energy entropy of each associated physiological signal combination in the associated feature combination data into the bidirectional long short-term memory model. The model performs bidirectional processing on the sequence to extract the cooperative change pattern within the sequence and analyze the change trend of the sequence to establish a time state vector.

[0023] S312: Calculate the rate of change of data in the time state vector over time, and compare the calculated rate of change of the signal with the preset motion-rest switching threshold to determine the patient’s motion state and resting state switching in the corresponding time period, and obtain a single-point state switching judgment value.

[0024] S313: The state switching results represented by the single-point state switching judgment values ​​of multiple time periods are serialized and arranged in chronological order to generate state prediction results.

[0025] As a further aspect of the present invention, the step of obtaining the weighted smoothed feature data specifically includes:

[0026] S411: Based on the state prediction results, determine the distribution of the patient's motion state and resting state in each time period, and construct a time weighting function to enhance the weighting factor in the motion state and decay the weighting factor in the resting state to obtain the time series weighting factor.

[0027] S412: Call the time series weighting factor, and combine it with the heart rate signal energy entropy and skin temperature signal energy entropy in the initial entropy feature dataset. Perform a weighted operation on the weighting factor and the energy entropy corresponding to each time point to obtain weighted entropy sequence data.

[0028] S413: Perform a smoothing operation on the weighted heart rate signal energy entropy and skin temperature signal energy entropy sequences within the weighted entropy sequence data to reflect the state change characteristics and obtain weighted smoothed feature data.

[0029] As a further aspect of the present invention, the steps for obtaining the patient health prediction results specifically include:

[0030] S511: Based on the weighted smoothed feature data and the state prediction results, establish a dynamic prediction relationship of the patient's physiological parameters, call the Kalman filter function to update and estimate the predicted state at the current moment, and input the filtered output of heart rate and skin temperature as the state estimate value to the next moment for calculation to obtain the updated state quantity;

[0031] S512: Generate health prediction output data based on the updated state variables, and classify and judge the change range of each type of physiological parameter in the health prediction output data to establish classified health prediction data.

[0032] S513: Analyze the consistency of the classified health prediction data over consecutive time periods to maintain the temporal continuity of the patient's health status and generate the patient health prediction results.

[0033] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0034] This invention collects multiple types of physiological signals using a unified timestamp and performs multi-scale decomposition and energy entropy calculation on the signals to accurately extract the inherent complexity features of the physiological signals. Subsequently, by using correlation analysis of signal energy entropy and identification of synchronous fluctuation intervals, it selects associated combinations of coordinated changes in physiological signals, overcoming the shortcomings of existing technologies that fail to capture the complex interactions of multi-source signals. At the same time, it introduces bidirectional sequence processing to perform in-depth analysis of the time series of associated physiological signals, dynamically predicting the switching between exercise and resting states. By combining state information to construct a time-weighted function, it performs weighted smoothing processing on entropy feature data, effectively reflecting the nonlinear change trend of physiological signals. Finally, it uses recursive filtering to update physiological parameters, achieving a more refined and accurate dynamic prediction of individual health status, significantly improving the sensitivity of health risk identification and the comprehensiveness of assessment. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the main steps of the present invention;

[0036] Figure 2 This is a flowchart of step S1 of the present invention;

[0037] Figure 3 This is a flowchart of step S2 of the present invention;

[0038] Figure 4 This is a flowchart of step S3 of the present invention;

[0039] Figure 5 This is a flowchart of step S4 of the present invention;

[0040] Figure 6 This is a flowchart of step S5 of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] Please see Figure 1 This invention provides a technical solution: an intelligent health prediction method based on wearable device data, comprising the following steps:

[0043] S1: Collect patients' physiological signal data with a unified timestamp, perform multi-scale decomposition on each type of physiological signal and calculate the corresponding signal energy entropy to obtain the initial entropy feature dataset;

[0044] S2: Analyze the correlation of the signal energy entropy corresponding to each type of physiological signal in the initial entropy feature dataset, sort each type of physiological signal according to the correlation, and select the associated physiological signal combination with the same trend to obtain the associated feature combination data;

[0045] S3: Analyze the continuous time series of each associated physiological signal combination in the associated feature combination data to predict the switching between the patient's motor state and resting state, and obtain the state prediction results;

[0046] S4: Based on the state prediction results and the initial entropy feature dataset, a time-weighted function is constructed to reflect the nonlinear change trend of physiological signals, and weighted smooth feature data is obtained;

[0047] S5: Establish a dynamic predictive relationship between patient physiological parameters by using weighted smoothed feature data and state prediction results to generate patient health prediction results.

[0048] The initial entropy feature dataset includes heart rate signal energy entropy, skin temperature signal energy entropy, and signal time series features; the associated feature combination data includes the associated feature group of heart rate and skin temperature, the synchronous fluctuation interval of the signal, and the correlation ranking information; the state prediction results include the exercise state interval, the resting state interval, and the state switching time marker; the weighted smoothing feature data includes the time weighting coefficient sequence, the smoothed heart rate feature, and the smoothed skin temperature feature; the health prediction results include the heart rate change trend, the skin temperature change trend, and the comprehensive health status assessment.

[0049] Please see Figure 2 The specific steps for obtaining the initial entropy feature dataset are as follows:

[0050] S111: Collect heart rate and skin temperature data of patients through wearable devices, and perform synchronous sampling with a unified timestamp. Perform mirror extension on the time sequence of the collected signals at the sampling boundary to establish synchronous time sequence signal data of heart rate and skin temperature signals.

[0051] The wearable device uses a built-in heart rate sensor and skin temperature sensor to continuously collect heart rate and skin temperature data from a patient in a conscious state at a preset sampling frequency. Heart rate is measured in beats per minute, and skin temperature in degrees Celsius. Each collected data point is appended with a timestamp accurate to the millisecond level to ensure strict temporal synchronization of the two physiological signals. For example, at a specific time point, both heart rate and skin temperature data are collected simultaneously. All collected data points are arranged in timestamp order to form two parallel raw time series: one for heart rate and the other for skin temperature. When processing the collected signal time series data, a mirror extension operation is performed at the start and end boundaries of the sequence. Specifically, for the end portion of a signal time series, several data points are reversed in order and appended to the end of the sequence as extension data; for the beginning portion of the sequence, several data points are reversed in order and appended to the beginning of the sequence as extension data. The extension length is determined according to the needs of subsequent signal analysis. The same mirror continuation operation was performed on both the heart rate signal timing and the skin temperature signal timing. Through this continuation process, synchronized timing signal data of the heart rate signal and the skin temperature signal were obtained.

[0052] S112: Perform multi-scale decomposition on the heart rate signal and skin temperature signal in the synchronous time-series signal data, extract the detail coefficients and approximation coefficients of the heart rate signal and skin temperature signal, calculate the energy of the detail coefficients and approximation coefficients in each frequency band, obtain the heart rate signal energy and skin temperature signal energy, and integrate them into a multi-scale signal energy set.

[0053] Taking heart rate signals as an example, the heart rate signal time series is decomposed into three levels. First, the original heart rate signal is decomposed into a first-level detail coefficient sequence and a first-level approximation coefficient sequence. Then, the first-level approximation coefficient sequence is used as input for a second decomposition, generating a second-level detail coefficient sequence and a second-level approximation coefficient sequence. Next, the second-level approximation coefficient sequence is used as input for a third decomposition, generating a third-level detail coefficient sequence and a third-level approximation coefficient sequence. After three-level decomposition, the original heart rate signal is effectively decomposed into three detail coefficient sequences in three different frequency bands and an approximation coefficient sequence representing the low-frequency components of the signal. The same multi-scale decomposition process is performed on the skin temperature signal, similarly obtaining its corresponding detail coefficient sequence and approximation coefficient sequence. Next, the energy of these coefficients in each frequency band is calculated. The energy is calculated by squaring the values ​​of all coefficients within a specific frequency band, then summing all the squared results to obtain a total value, which is the energy of the corresponding frequency band. The detail coefficient energy and approximation coefficient energy of the heart rate signal in all frequency bands are calculated to obtain a set of heart rate signal energy values. The same energy calculation was performed on the skin temperature signal, resulting in a set of skin temperature signal energy values. Finally, the heart rate signal energy and the skin temperature signal energy were integrated to form a multi-scale signal energy set, which contains the energy distribution information of both signals across all decomposed frequency bands.

[0054] S113: Calculate the probability distribution of the proportion of heart rate signal energy and skin temperature signal energy in each frequency band of the multi-scale signal energy set, obtain the entropy of heart rate signal energy and skin temperature signal energy based on the probability distribution, and arrange them in time series to generate an initial entropy feature dataset.

[0055] Calculate the probability distribution of the energy proportion of the heart rate signal and skin temperature signal in each frequency band. This calculation involves dividing the energy value of the specified signal in a single frequency band by the sum of its energy across all frequency bands. This calculation yields the probability of the signal's energy proportion in each frequency band. This process is repeated to calculate the probability proportions for all frequency bands, thus obtaining the probability distribution of the heart rate signal. The same calculation is performed on the skin temperature signal to obtain its probability distribution. Based on these probability distributions, the energy entropy of the heart rate signal and the skin temperature signal is calculated. The energy entropy is calculated using the following formula:

[0056] ;

[0057] in, The calculated signal energy entropy is denoted as , where Represents entropy, subscript This represents the signal type, corresponding to heart rate and skin temperature signals respectively. It is a dimensionless value used to quantify the complexity of the signal energy distribution. For signal In the The energy proportion probability of the first frequency band is determined by the first frequency band. The energy of each frequency band is divided by the total energy of the signal. This parameter is dimensionless and ranges from 0 to 1. (Subscript...) Refers to a specific signal type, subscript Refers to the first Each frequency band The total number of frequency bands obtained after multi-scale decomposition is a dimensionless integer. For example, a three-level decomposition will produce four frequency bands. It is a frequency band index, from 1 to... , For the first The weighting coefficient for each frequency band is a dimensionless value. The basis for setting this weighting coefficient is that the more important a frequency band is in representing changes in physiological state, the greater its weight. By analyzing historical physiological data, the frequency bands most strongly correlated with specific physiological states (such as stress or fatigue) are identified. For example, if the energy change of a certain high-frequency band in the heart rate signal is highly correlated with sympathetic nerve excitability, then the weight of that frequency band... The weights will be set relatively high, and all weight coefficients will be normalized after setting so that their sum is 1. This is a dimensionless smoothing adjustment parameter, set to a positive number close to zero. The purpose of introducing this parameter is to handle the energy proportion of a certain frequency band. The special case of zero is used to avoid taking the logarithm of a term with a value of zero, which could lead to calculation interruption and ensure the stability and continuity of the calculation.

[0058] Taking heart rate signals as an example, assuming they are obtained through three-level decomposition... One frequency band, 1. Parameter settings: signal For heart rate signals, the total number of frequency bands Energy percentage probability of each frequency band They are respectively: , , , Weighting coefficients for each frequency band They are respectively: , , , Smoothly adjust parameters .

[0059] Substitute into the formula to calculate:

[0060] First, calculate the smoothed probability term for each frequency band: :

[0061] ;

[0062] ;

[0063] ;

[0064] .

[0065] Calculate the entropy contribution for each frequency band. :

[0066] ;

[0067] ;

[0068] ;

[0069] .

[0070] Calculate total energy entropy :

[0071] Finally, the energy entropy of the heart rate signal within this time window is obtained. It is approximately 0.4033.

[0072] Please see Figure 3 The specific steps for obtaining associated feature combination data are as follows:

[0073] S211: Perform interpolation correction on the time series of heart rate signal energy entropy and skin temperature signal energy entropy in the initial entropy feature dataset to supplement missing measurement points, and calculate the Pearson correlation coefficient of the two at multiple time points to represent their mutual relationship, generating a set of signal energy entropy correlation coefficients;

[0074] First, check the sequence for missing data points caused by signal acquisition or transmission interruptions. If data is missing at a certain time point, linear interpolation is used for correction and supplementation. Specifically, locate the two nearest valid data points before and after the missing data point, calculate the arithmetic mean of these two data points, and use the result as the supplementary value for the missing data point. After supplementing all missing data points, calculate the Pearson correlation coefficient of the two energy entropy time series at multiple time points. During calculation, a fixed-length time window is selected, and within the window, the average values ​​of the heart rate signal energy entropy sequence segment and the skin temperature signal energy entropy sequence segment are calculated respectively. Then, for each time point within the window, the value at that point is subtracted from the average value of the corresponding sequence segment to obtain the respective deviation. Multiply the deviations of each pair of corresponding time points, and sum all the products within the window. Simultaneously, calculate the sum of squared deviations for each of the two sequence segments. Finally, divide the sum of the deviation products by the square root of the product of the sums of squared deviations of the two sequences. Record the calculated correlation coefficient values, and slide the time window forward by one time point along the time axis. Repeat the above calculation process to generate a set of signal energy entropy correlation coefficients arranged in chronological order.

[0075] S212: Sort the Pearson correlation coefficients in the set of signal energy entropy correlation coefficients, and judge the changing trends of heart rate signal energy entropy and skin temperature signal energy entropy based on the consistency of the changing direction, and obtain the synchronous fluctuation range of the two types of physiological signals changing in tandem.

[0076] At two consecutive time points, the values ​​of heart rate signal energy entropy and skin temperature signal energy entropy are acquired. If the heart rate signal energy entropy at the later time point is greater than that at the earlier time point, and the skin temperature signal energy entropy at the later time point is also greater than that at the earlier time point, the trend between these two points is determined to be "synchronous increase." Conversely, if both the heart rate signal energy entropy and skin temperature signal energy entropy at the later time point are less than their respective values ​​at the earlier time point, the trend is determined to be "synchronous decrease." All other combinations of changes besides these two cases, such as one signal energy entropy increasing while the other decreases, or one remaining unchanged while the other changes, are determined to be "asynchronous" trends. Next, the synchronous fluctuation intervals of the coordinated changes of the two types of physiological signals are obtained. The criterion for determining a synchronous fluctuation interval is: in the time series, the same type of synchronous trend appears continuously for a preset minimum length (3 to 10 time points). The basis for setting this minimum length is: through analysis of historical data, it was found that the consistency of random trends is usually short-lived, while real coordinated changes often have a certain degree of persistence. By observing the stability of trend judgment and the accuracy of subsequent predictions over different durations, the minimum number of consecutive time points that can effectively filter out instantaneous noise interference is determined, thereby ensuring that the identified synchronized fluctuation intervals have physiological stability. By performing such trend judgment and continuity screening on the entire time series, all synchronous fluctuation intervals that meet the conditions are finally obtained.

[0077] S213: For the synchronous fluctuation interval, select the combination of related physiological signals with consistent changing trends from the initial entropy feature dataset, extract the coordinated change features of heart rate signal energy entropy and skin temperature signal energy entropy within the interval, and establish the related feature combination data.

[0078] For each identified synchronous fluctuation interval, its start and end times are recorded. Then, based on this time information, heart rate signal energy entropy sequence fragments and skin temperature signal energy entropy sequence fragments for the corresponding time period are extracted from the initial entropy feature dataset. These two extracted sequence fragments together constitute a composite of associated physiological signals. For example, if a synchronous fluctuation interval is determined to be from N minutes to M minutes, and the co-change trend of this interval is "synchronous decline," then the sequence of all heart rate signal energy entropy values ​​within these M-N minutes, and the sequence of all skin temperature signal energy entropy values ​​within these M-N minutes, are extracted. These two sequences together form a co-change feature characterizing the "synchronous decline" state. The same sequence fragment extraction operation is performed on all identified synchronous fluctuation intervals, regardless of whether the trend is "synchronous rise" or "synchronous decline." All heart rate signal energy entropy sequence fragments and skin temperature signal energy entropy sequence fragments extracted from all synchronous fluctuation intervals are aggregated to establish composite feature data.

[0079] Please see Figure 4 The specific steps for obtaining the state prediction results are as follows:

[0080] S311: Input the continuous time series of heart rate signal energy entropy and skin temperature signal energy entropy of each associated physiological signal combination in the associated feature combination data into the bidirectional long short-term memory model. The model performs bidirectional processing on the sequence to extract the cooperative change pattern within the sequence and analyze the change trend of the sequence to establish a time state vector.

[0081] The continuous time series of heart rate and skin temperature energy entropy for each associated physiological signal combination in the established associated feature combination data are input into a bidirectional sequence processing structure. This structure consists of two independent sequence processing units: a forward processing unit and a backward processing unit, operating in parallel. The forward processing unit inputs data sequentially from the beginning to the end of the sequence. At each time point, based on the input signal and the previous hidden state, it dynamically updates its internal memory unit through gating logic, thereby capturing the historical dependencies from the start to the current time. The backward processing unit operates in the opposite time direction, inputting data sequentially from the end of the sequence backward, using the same gating mechanism to establish information dependencies from future time points to the current time point, thus recording the changing characteristics of future segments of the sequence. Both units contain input regulation gates, forgetting regulation gates, and output regulation gates to control the retention ratio of historical information and the update ratio of new input information. To ensure model stability and sensitivity to subtle signal changes, the input signals are normalized before being fed into the model, keeping their values ​​between 0 and 1. Each processing unit has a hidden layer size of 64 units, with standard time-step updates between layers. The time-step length is consistent with the signal sampling interval, e.g., one update per second. During model training, 10,000 labeled samples of heart rate and skin temperature co-changes are used, with a batch size of 128 sequences. The training runs for 30 epochs, with independent forward and backward updates in each epoch. The model is initialized with random weights in the initial stage and converges gradually through error backpropagation. To prevent overfitting, a 20% random dropout mechanism is implemented after each hidden unit, and an early stopping strategy is introduced to monitor the stability of the verification error. After bidirectional processing, at each time point, the hidden state vectors output by the forward and backward processing units are concatenated into a composite vector. This composite vector contains contextual feature information from the past and future, and has a length of 128 dimensions. The model uses this composite vector to establish a temporal state vector sequence, reflecting the co-change trend and dynamic equilibrium characteristics of the signal over time. The entire processing is performed using a sliding window approach, with each window covering 60 time steps, corresponding to a continuous signal interval of approximately one minute. This ensures that the model can detect short-term fluctuations while maintaining its ability to track long-term trends. Ultimately, the model generates a sequence of time-state vectors for each combination of associated physiological signals from the input.

[0082] S312: Calculate the rate of change of data in the time state vector over time, and compare the calculated rate of change of the signal with the preset motion-rest switching threshold to determine the patient's motion state and resting state switching in the corresponding time period and obtain the single-point state switching judgment value.

[0083] The process of calculating the rate of change of data within the time state vector over time, comparing the calculated rate of change of the signal with a preset motion-resting state switching threshold, and determining the patient's motion and resting state switching within the corresponding time period is as follows:

[0084] Extract the state data corresponding to each time point in the time state vector in sequence, calculate the numerical difference between the state data at the current time point and the state data at the previous adjacent time point, and use the absolute value of the numerical difference as the instantaneous change quantity representing the signal change between the two time points.

[0085] The rate of change of the signal is obtained by averaging all instantaneous changes within a preset time window.

[0086] The signal change rate is compared with the preset motion-rest switching threshold. When the signal change rate exceeds the preset motion-rest switching threshold, the single-point state switching judgment value is set as the motion state indicator. When the signal change rate does not exceed the preset motion-rest switching threshold, the single-point state switching judgment value is set as the rest state indicator.

[0087] First, state data corresponding to each time point within the time state vector sequence is extracted sequentially; the state data is a numerical vector. The norm of this vector, such as the Euclidean norm, is calculated to obtain a scalar value as the state data value for that point. Then, the numerical difference between the current state data value and the state data value of the previous adjacent time point is calculated, and the absolute value of this difference is used as the instantaneous change in signal between the two time points. Next, a preset time window is set, the length of which is determined experimentally. The rate of signal change is calculated under different window lengths, and its sensitivity and specificity in distinguishing between motion and resting states are evaluated. A window length that effectively smooths instantaneous fluctuations while promptly reflecting state transitions is selected. The arithmetic mean of all instantaneous changes within this time window is calculated to obtain the rate of signal change. Then, the calculated rate of signal change is numerically compared with a preset motion-resting state transition threshold. The motion-resting state transition threshold is set based on the collection of a large amount of physiological signal data from patients known to be in clearly defined motion and resting states. Then, the rate of signal change for these data under different states is calculated according to the method described in this step. Statistical analysis is performed on these signal change rates, such as calculating the mean and standard deviation of the signal change rate distribution under different states. A threshold is set to a value that can best distinguish the signal change rate distribution between motion and resting states; for example, selecting the intersection of the two distributions or a certain percentile to optimize the accuracy of state classification. Specifically, this threshold is set to 0.12, a value that can effectively distinguish the signal change rate between the two states. When the calculated signal change rate exceeds the preset motion-resting state switching threshold, the single-point state switching judgment value for the corresponding time period is set as a motion state identifier; when the signal change rate does not exceed the preset motion-resting state switching threshold, the single-point state switching judgment value is set as a resting state identifier.

[0088] S313: Serialize and arrange the state transition results represented by the single-point state transition judgment values ​​of multiple time periods in chronological order to generate state prediction results;

[0089] Each judgment value, namely a "motor state identifier" or "resting state identifier," represents the patient's main physiological activity state within the corresponding time period. This arrangement process combines a series of discrete judgment results into a continuous time series. For example, by processing the data of the entire monitoring period, a sequence of state identifiers is obtained. This sequence may present multiple resting state identifiers appearing consecutively, then changing to multiple motor state identifiers appearing consecutively, and then switching back to the resting state identifier pattern. This complete sequence is directly output as the state prediction result. This result, in the form of discrete state labels, clearly depicts the dynamic switching of the patient's activity state over time throughout the entire monitoring period, forming a timeline of the patient's macroscopic activity state.

[0090] Please see Figure 5 The specific steps for obtaining weighted smoothed feature data are as follows:

[0091] S411: Based on the state prediction results, determine the distribution of the patient's motion state and resting state in each time period, and construct a time weighting function to enhance the weighting factor in the motion state and decay the weighting factor in the resting state to obtain the time series weighting factor.

[0092] Based on the state distribution, a time-weighted function is constructed to generate time-series weighting factors. The specific logic of this function is to set a basic weight adjustment factor for both the motion and resting states. The setting of these adjustment factors is based on experimental verification on historical datasets to examine the impact of different weight factor combinations on the accuracy of the final health prediction results. The experimental design includes setting the enhancement factor range between 1.1 and 1.5 (for the motion state) and the decay factor range between 0.6 and 0.9 (for the resting state), with a combination test using a step size of 0.05. By evaluating the mean squared error and prediction consistency index of the health prediction model under each combination, it was finally determined that when the enhancement factor is 1.3 and the decay factor is 0.8, the model output prediction error is minimized and the discrimination between the motion and resting states is highest. This result, validated with multiple batches of data, stably reflects the enhancement effect of signal characteristics under motion states and the mitigation characteristics of signal changes under resting states. In practice, the state prediction result sequence is traversed. If the state at a certain time point is identified as "movement state", the weight factor for that time point is set to 1.3; if the state is identified as "resting state", the weight factor for that time point is set to 0.8. By performing this operation on all time points, a time series weight factor synchronized with the time series is obtained.

[0093] S412: Call the time series weighting factor, and combine it with the heart rate signal energy entropy and skin temperature signal energy entropy in the initial entropy feature dataset. Perform a weighted operation on the weighting factor and the energy entropy corresponding to each time point to obtain the weighted entropy sequence data.

[0094] At each time point, the time series weighting factor, along with the heart rate signal energy entropy value and skin temperature signal energy entropy value, are extracted. The time series weighting factor is then multiplied by the heart rate signal energy entropy value and the skin temperature signal energy entropy value, respectively. For example, at a certain time point, if the state prediction result is "exercise state," the corresponding enhancement factor is used as the weight; if the heart rate signal energy entropy is one value and the skin temperature signal energy entropy is another value, these two energy entropy values ​​are multiplied by the enhancement factor to obtain the weighted heart rate signal energy entropy and the weighted skin temperature signal energy entropy. If at another time point the state is "resting state," the corresponding decay factor is used as the weight, and the same multiplication operation is performed on the heart rate signal energy entropy and the skin temperature signal energy entropy at that time. This weighting operation is performed for each time point in the initial entropy feature dataset, ultimately obtaining two new time series, namely the weighted entropy sequence data.

[0095] S413: Perform a smoothing operation on the weighted heart rate signal energy entropy and skin temperature signal energy entropy sequences within the weighted entropy sequence data to reflect the state change characteristics and obtain weighted smoothed feature data;

[0096] The purpose of smoothing is to filter out short-term, sharp fluctuations in a sequence caused by random factors, thereby revealing the long-term trend reflecting changes in state more clearly. Here, the moving average method is used for smoothing. A smoothing window of fixed length is set. The length of the smoothing window is determined experimentally. Smoothing is performed at multiple different window lengths, and the balance between preserving trend information and filtering out instantaneous fluctuations in the smoothed sequence is evaluated. A window length that can effectively filter out high-frequency noise without excessively losing the signal reflecting changes in physiological state is selected. The specific operation process is as follows: the smoothing window is placed at the beginning of the sequence, the arithmetic mean of all data points within the window is calculated, and this average is taken as the first point of the smoothed sequence. Then, the window is moved one data point backward along the time axis, and the arithmetic mean of all data points within the new window is calculated, which is taken as the second point of the smoothed sequence. This process is repeated until the window slides to the end of the sequence. After performing this smoothing operation on both the weighted heart rate signal energy entropy sequence and the weighted skin temperature signal energy entropy sequence, two new time series with smoother fluctuations are obtained, which are the weighted smoothed feature data.

[0097] Please see Figure 6 The specific steps for obtaining patient health prediction results are as follows:

[0098] S511: Based on the weighted smooth feature data and state prediction results, establish a dynamic prediction relationship of the patient's physiological parameters, call the Kalman filter function to update the predicted state at the current moment and estimate the error, and use the filtered output of heart rate and skin temperature as the state estimate value to be input into the next moment for calculation to obtain the updated state quantity;

[0099] Based on weighted smoothed feature data and state prediction results, a dynamic prediction relationship for patient physiological parameters is established. In this step, a recursive process is used to update the predicted state and estimate the error at the current moment. The core update step of this process is executed according to the following formula:

[0100] ;

[0101] in, Let be the updated state vector at time t. This vector contains the optimal estimates of heart rate and skin temperature at that time, and its physical dimensions are consistent with the estimated physiological parameters (e.g., beats / minute, degrees Celsius). The subscripts are... Represents the current point in time. Let be the prior state prediction vector at time t, where " The symbol "" indicates that this is an estimate or prediction. This vector is obtained based on the optimal state estimate at time t-1 and the prediction from the system dynamic model. It represents the prediction before receiving the current measurement, and its physical dimensions are the same as... same, Let be the measurement vector at time t. The content of this vector comes from the weighted smoothed feature data of step S413, namely the weighted smoothed heart rate signal energy entropy and skin temperature signal energy entropy. Since entropy is a dimensionless quantity, this vector is dimensionless. The subscript... Represents the current point in time. As a measurement matrix, it establishes a mapping relationship between physiological parameters (such as heart rate and skin temperature) in the state space and entropy features in the measurement space. The elements of the measurement matrix M are established through system identification and physiological modeling. Specifically, by performing multiple linear regression analysis or nonlinear fitting on a large amount of physiological data and entropy feature data collected under controlled conditions, a quantitative relationship between physiological parameters and entropy features is established. This process ensures that the measurement matrix M can accurately map physiological parameters in the state space to entropy features in the measurement space. The dimensional configuration of the matrix elements is designed to ensure... Dimensions and The dimensions are consistent. Let be the Kalman gain matrix at time t. The value of this matrix is ​​dynamically calculated at each time step and is used to balance the confidence level between prior predictions and new measurements. Its setting is based on the principle that when measurement noise is low but prediction model uncertainty is high, the Kalman gain... A larger value indicates greater trust in the current measurement; conversely, a smaller value indicates greater confidence in the current measurement when measurement noise is high but the prediction model is accurate. A smaller value indicates greater trust in the model's predictions; the dimensions of the Kalman gain matrix must be guaranteed. The results and The dimensions are consistent.

[0102] Let the state vector be 2-dimensional, representing heart rate and skin temperature respectively. 1. Parameter setting: the prior state prediction vector at time t. (Predicted heart rate 75 beats / minute, skin temperature 36.8 degrees Celsius), measurement vector at time t. (The measured weighted smoothed heart rate entropy is 0.41, and the skin temperature entropy is 0.55), measurement matrix (Mapping heart rate and skin temperature to their respective entropy values), the Kalman gain matrix at time t. (This indicates that at the current moment, there is a high degree of confidence in the measured heart rate and the predicted skin temperature.)

[0103] Substitute into the formula to calculate:

[0104] First, calculate the predicted measurement value. :

[0105] ;

[0106] Next, the measurement residuals are calculated. :

[0107] ;

[0108] Then, calculate the correction term. :

[0109] ;

[0110] Finally, the updated state vector is calculated. :

[0111] Finally, the updated state at time t is: heart rate 75.021 beats / minute, skin temperature 36.7992 degrees Celsius.

[0112] S512: Generate health prediction output data based on the updated state variables, and classify and judge the change range of each type of physiological parameter in the health prediction output data to establish classified health prediction data.

[0113] The process of classifying and judging the variation range of each type of physiological parameter in the health prediction output data is as follows:

[0114] Calculate the rate of change of the centrality of the updated state variables and the rate of change of skin temperature over consecutive time points;

[0115] The calculated rate of change in heart rate and rate of change in skin temperature are compared with preset stable change thresholds and abnormal change thresholds, respectively.

[0116] Based on the comparison results, the intervals with a rate of change lower than the stable change threshold are marked as stable intervals, the intervals with a rate of change between the stable change threshold and the abnormal change threshold are marked as fluctuating intervals, and the intervals with a rate of change higher than the abnormal change threshold are marked as abnormal intervals.

[0117] First, the rate of change of the updated state variables, heart rate and skin temperature, is calculated over consecutive time points. The specific calculation method is to divide the difference between the current value and the previous value by the previous value, and then multiply by 100%. Then, the calculated rates of change of heart rate and skin temperature are compared with preset stable and abnormal change thresholds, respectively. The stable change threshold is set based on: collecting data on the rate of change of heart rate and skin temperature from a large number of healthy individuals during daily physiological activities (especially at rest), statistically analyzing these rates of change, calculating the mean and standard deviation of their distribution, and setting the stable change threshold as an upper limit that can cover the normal physiological fluctuations of most healthy individuals; for example, it can be determined by calculating the mean plus 2 or 3 times the standard deviation. The abnormal change threshold is set based on: collecting and analyzing physiological data under different health risk states (e.g., early disease warnings or acute physiological stress) to identify the critical points where the rate of change of heart rate and skin temperature significantly deviates from the normal range. This threshold is set so that when the rate of change exceeds this value, it may indicate potential health problems or physiological abnormalities. This threshold setting method aims to balance the sensitivity and specificity of monitoring, ensuring timely detection of anomalies while reducing false alarms. Interval labeling is performed based on the comparison results: if the calculated rate of change at a time point is lower than the corresponding stable change threshold, that time point is labeled as a stable interval; if the rate of change is between the stable change threshold and the abnormal change threshold, it is labeled as a fluctuating interval; if the rate of change is higher than the abnormal change threshold, it is labeled as an abnormal interval. This classification is applied to all time points of both heart rate and skin temperature sequences to establish categorical health prediction data.

[0118] S513: Analyze the consistency of categorized health prediction data over consecutive time periods, maintain the temporal continuity of patients' health status, and generate patient health prediction results.

[0119] By examining the temporal consistency of state labels, isolated misjudgments that may be caused by transient noise or interference from a single data point are corrected. The specific analysis process employs a rule-based post-processing logic. A fixed-length time window is set and slides along the time axis. Within each window, it is checked whether there are isolated labels inconsistent with the preceding and following state labels. For example, in a sequence segment, if the labels immediately preceding and following an abnormal interval label are all stable intervals, then the abnormal interval label is considered an isolated label. In this case, the isolated abnormal interval label is corrected to a fluctuating interval to reduce false alarms caused by single-point mutations. Conversely, if a non-stable state label (such as a fluctuating interval or an abnormal interval) appears consecutively, reaching or exceeding a preset minimum duration, it indicates that the state change is persistent and not accidental; therefore, the original judgment is maintained, and it is not modified. The time window length was set based on the following: analysis of continuous physiological signal data from multiple subjects was conducted to compare the smoothness and accuracy of the corrected label sequences under different window lengths. The results showed that a window length of 5 time steps (approximately 5 seconds) effectively detected and corrected isolated labels caused by transient noise without introducing misjudgments due to over-smoothing. The minimum duration length was set based on the statistical analysis of the duration of real-state change events, finding that most stable state transitions lasted no less than 15 time steps (approximately 15 seconds). Therefore, the minimum duration length was set to 15 time steps. This parameter combination demonstrated the best balance in experimental verification, ensuring both the suppression of transient noise by the correction logic and maintaining sensitivity to real-state changes. By performing this consistency analysis and correction on the entire classification health prediction data sequence, patient health prediction results were generated.

[0120] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A smart health prediction method based on wearable device data, characterized in that, Includes the following steps: S1: Collect patients' physiological signal data with a unified timestamp, perform multi-scale decomposition on each type of physiological signal and calculate the corresponding signal energy entropy to obtain the initial entropy feature dataset; S2: Analyze the correlation of the signal energy entropy corresponding to each type of physiological signal in the initial entropy feature dataset, sort each type of physiological signal according to the correlation, and select the associated physiological signal combination with the same trend to obtain the associated feature combination data; S3: Analyze the continuous time series of each associated physiological signal combination in the associated feature combination data to predict the switching between the patient's motor state and resting state, and obtain the state prediction result; S4: Based on the state prediction results and the initial entropy feature dataset, a time-weighted function is constructed to reflect the nonlinear change trend of physiological signals, and weighted smooth feature data is obtained; S5: Establish a dynamic prediction relationship between the patient's physiological parameters and the state prediction results through the weighted smoothed feature data, and generate the patient's health prediction results.

2. The intelligent health prediction method based on wearable device data according to claim 1, characterized in that, The initial entropy feature dataset includes heart rate signal energy entropy, skin temperature signal energy entropy, and signal time series features; the associated feature combination data includes the associated feature group of heart rate and skin temperature, signal synchronous fluctuation interval, and correlation ranking information; the state prediction results include the exercise state interval, the resting state interval, and the state switching time identifier; the weighted smoothing feature data includes the time weighting coefficient sequence, the smoothed heart rate feature, and the smoothed skin temperature feature; the health prediction results include the heart rate change trend, the skin temperature change trend, and the comprehensive health status assessment.

3. The intelligent health prediction method based on wearable device data according to claim 2, characterized in that, The specific steps for obtaining the initial entropy feature dataset are as follows: S111: Collect heart rate and skin temperature data of patients through wearable devices, and perform synchronous sampling with a unified timestamp. Perform mirror extension on the time sequence of the collected signals at the sampling boundary to establish synchronous time sequence signal data of heart rate and skin temperature signals. S112: Perform multi-scale decomposition on the heart rate signal and skin temperature signal in the synchronous time-series signal data, extract the detail coefficients and approximation coefficients of the heart rate signal and skin temperature signal, calculate the energy of the detail coefficients and approximation coefficients in each frequency band, obtain the heart rate signal energy and skin temperature signal energy, and integrate them into a multi-scale signal energy set. S113: Calculate the probability distribution of the proportion of heart rate signal energy and skin temperature signal energy in each frequency band of the multi-scale signal energy set, obtain the entropy of heart rate signal energy and skin temperature signal energy based on the probability distribution, and arrange them in time series to generate an initial entropy feature dataset.

4. The intelligent health prediction method based on wearable device data according to claim 3, characterized in that, For the energy entropy of heart rate signals and skin temperature signals, the formula is used: ； in, The calculated signal energy entropy, These represent signal types, corresponding to heart rate and skin temperature signals, respectively. For signal In the The probability of energy proportion in each frequency band This represents the total number of frequency bands obtained after multi-scale decomposition. It is a frequency band index, from 1 to... , For the first Weighting coefficients for each frequency band To smoothly adjust parameters.

5. The intelligent health prediction method based on wearable device data according to claim 3, characterized in that, The specific steps for obtaining the associated feature combination data are as follows: S211: Perform interpolation correction on the time series of heart rate signal energy entropy and skin temperature signal energy entropy in the initial entropy feature dataset to supplement missing measurement points, and calculate the Pearson correlation coefficient of the two at multiple time points to represent their mutual relationship, generating a set of signal energy entropy correlation coefficients; S212: Sort the Pearson correlation coefficients in the set of signal energy entropy correlation coefficients, and judge the changing trends of heart rate signal energy entropy and skin temperature signal energy entropy based on the consistency of the changing direction, and obtain the synchronous fluctuation range of the two types of physiological signals changing in tandem. S213: For the synchronous fluctuation interval, select the associated physiological signal combination with consistent change trend from the initial entropy feature dataset, extract the coordinated change features of heart rate signal energy entropy and skin temperature signal energy entropy within the interval, and establish associated feature combination data.

6. The intelligent health prediction method based on wearable device data according to claim 5, characterized in that, The specific steps for obtaining the state prediction result are as follows: S311: Input the continuous time series of heart rate signal energy entropy and skin temperature signal energy entropy of each associated physiological signal combination in the associated feature combination data into the bidirectional long short-term memory model. The model performs bidirectional processing on the sequence to extract the cooperative change pattern within the sequence and analyze the change trend of the sequence to establish a time state vector. S312: Calculate the rate of change of data in the time state vector over time, and compare the calculated rate of change of the signal with the preset motion-rest switching threshold to determine the patient’s motion state and resting state switching in the corresponding time period, and obtain a single-point state switching judgment value. S313: The state switching results represented by the single-point state switching judgment values ​​of multiple time periods are serialized and arranged in chronological order to generate state prediction results.

7. The intelligent health prediction method based on wearable device data according to claim 6, characterized in that, The process of calculating the rate of change of data within the time state vector over time, comparing the calculated rate of change of the signal with a preset motion-resting state switching threshold, and determining the patient's motion state and resting state switching status within the corresponding time period is as follows: Extract the state data corresponding to each time point in the time state vector in sequence, calculate the numerical difference between the state data at the current time point and the state data at the previous adjacent time point, and use the absolute value of the numerical difference as the instantaneous change quantity representing the signal change between the two time points. The rate of change of the signal is obtained by averaging all instantaneous changes within a preset time window. The signal change rate is compared with a preset motion-rest switching threshold. When the signal change rate exceeds the preset motion-rest switching threshold, the single-point state switching judgment value is set as the motion state indicator. When the signal change rate does not exceed the preset motion-rest switching threshold, the single-point state switching judgment value is set as the rest state indicator.

8. The intelligent health prediction method based on wearable device data according to claim 7, characterized in that, The specific steps for obtaining the weighted smoothed feature data are as follows: S411: Based on the state prediction results, determine the distribution of the patient's motion state and resting state in each time period, and construct a time weighting function to enhance the weighting factor in the motion state and decay the weighting factor in the resting state to obtain the time series weighting factor. S412: Call the time series weighting factor, and combine it with the heart rate signal energy entropy and skin temperature signal energy entropy in the initial entropy feature dataset. Perform a weighted operation on the weighting factor and the energy entropy corresponding to each time point to obtain weighted entropy sequence data. S413: Perform a smoothing operation on the weighted heart rate signal energy entropy and skin temperature signal energy entropy sequences within the weighted entropy sequence data to reflect the state change characteristics and obtain weighted smoothed feature data.

9. The intelligent health prediction method based on wearable device data according to claim 8, characterized in that, The specific steps for obtaining the patient health prediction results are as follows: S511: Based on the weighted smoothed feature data and the state prediction results, establish a dynamic prediction relationship of the patient's physiological parameters, call the Kalman filter function to update and estimate the predicted state at the current moment, and input the filtered output of heart rate and skin temperature as the state estimate value to the next moment for calculation to obtain the updated state quantity; S512: Generate health prediction output data based on the updated state variables, and classify and judge the change range of each type of physiological parameter in the health prediction output data to establish classified health prediction data. S513: Analyze the consistency of the classified health prediction data over consecutive time periods to maintain the temporal continuity of the patient's health status and generate the patient health prediction results.

10. The intelligent health prediction method based on wearable device data according to claim 9, characterized in that, The process of classifying and judging the variation range of each type of physiological parameter in the health prediction output data is as follows: Calculate the rate of change of the centrality of the updated state variables and the rate of change of skin temperature over consecutive time points; The calculated rate of change in heart rate and rate of change in skin temperature are compared with preset stable change thresholds and abnormal change thresholds, respectively. Based on the comparison results, the intervals with a rate of change lower than the stable change threshold are marked as stable intervals, the intervals with a rate of change between the stable change threshold and the abnormal change threshold are marked as fluctuating intervals, and the intervals with a rate of change higher than the abnormal change threshold are marked as abnormal intervals.

Images