Health data interpolation and optimization method combining flag bit judgment and physique constraint mechanism

Abstract

The invention relates to the field of health information management, in particular to a health data interpolation and optimization method combining flag bit judgment and a physique constraint mechanism. The method comprises the following steps of: analyzing an added flag bit, determining an abnormal data fragment, performing data interpolation on the abnormal fragment based on a GM (1, 1) model and multiple linear regression, then determining the constitution and the corresponding range thereof, and correcting the interpolation data by taking the constitution range as a constraint condition. The method specifically comprises the following steps: analyzing four flag bits to determine abnormal data segments; processing the relevance between single-dimensional data interpolation and multiple linear regression processing dimensions based on a GM (1, 1) model; and determining the physique and the corresponding data range according to the data, and correcting the interpolation data by taking the physique range as a constraint condition. The invention aims to solve the problem of how to improve the accuracy of a health data interpolation result.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of health information management, and particularly relates to a health data interpolation and optimization method combining flag bit judgment and physical constraint mechanism. BACKGROUND

[0002] In the field of smart medical treatment, individual health management, etc., continuous collection and high-quality storage of health data is of great importance. With the development of wearable devices, physiological sensors and mobile health platforms, massive multi-dimensional, time-series health data can be collected and analyzed, providing a rich foundation for personalized health management. However, due to device precision limitations, communication abnormalities or sudden changes in human body state, there are often problems such as data missing, outliers, nonlinear mutations, etc., so it is of great significance to interpolate and optimize health data.

[0003] In the field of health monitoring, physiological parameter data collected by wearable devices often has small sample characteristics. For such small sample time-series data, the grey prediction model has become a commonly used interpolation method due to its low dependence on data volume ("Research on Missing Data Interpolation of Detector Using Grey Theory", Xu Chengyong, pp. 1-83, June 2011). However, the grey model has obvious limitations: its single variable modeling mechanism cannot handle the dynamic correlation between multi-dimensional physiological parameters. For example, when there are missing data in heart rate, blood pressure and blood oxygen data at the same time, the grey model can only interpolate each parameter independently, completely ignoring the known physiological coupling relationship between them - such as the cardiovascular pressure reflex mechanism that a decrease in blood pressure is usually accompanied by a compensatory increase in heart rate. This processing method will lead to interpolation results that violate the coordinated action rules of the physiological system.

[0004] In recent years, some studies have attempted to use deep learning methods to solve the problem of correlation modeling of multi-dimensional health data ("Representation Learning for Wearable-Based Applications in the Case of Missing Data", Janosch Jungo et al., arXiv, pp. 1-9, January 2024). Although these methods can capture complex relationships between multi-dimensional parameters through attention mechanisms, they generally have a key flaw: they ignore the boundary constraints of physiological parameters due to medical constitution differences. For example, when interpolating the resting heart rate of patients with yin deficiency constitution and phlegm-dampness constitution, existing deep learning models may generate the same numerical range, while in fact the Chinese medical constitution classification standard clearly defines the difference in normal heart rate range for these two types of constitution, which can be up to 10-15 times / minute ("Classification and Determination of Chinese Medical Constitution and Establishment of Its Methodological System [J]", Wang Qi et al., Chinese Journal of Traditional Chinese Medicine, 2011, 26(1): 1-5). More seriously, when the model interpolates blood glucose values for diabetic patients, if the individualized blood glucose control target is not considered, it may generate incorrect values with clinical risks.

[0005] These two defects highlight the shortcomings of existing methods: gray models are suitable for small samples but cannot handle dimensional correlations, and deep learning methods can model complex relationships but ignore constitution differences. Therefore, there is an urgent need to develop a new interpolation method that combines small sample modeling capability, multi-dimensional correlation analysis, and medical constitution constraints to improve the reliability and medical interpretability of health data.

[0006] The above content is only used to assist in understanding the technical solutions of the present application and does not represent an acknowledgment that the above content is prior art. SUMMARY

[0007] The main purpose of the present application is to provide a health data interpolation and optimization method combining flag bit determination and constitution constraint mechanism, aiming to solve the problem of how to improve the accuracy of health data interpolation results.

[0008] To achieve the above-mentioned purpose, the present application provides a health data interpolation and optimization method combining flag bit determination and constitution constraint mechanism, as shown in Figure 1 The method comprises the following steps:

[0009] S10, the health data is expanded in the same dimension, and four flag bits are added, and whether the data is abnormal is judged according to the flag bit;

[0010] S20, interpolating the abnormal data, using GM(1, 1) model to interpolate single-dimensional data, using multivariate linear regression (MLP) to process the correlation between dimensions, and weighting the results of the two models to obtain the interpolated data;

[0011] S30, the original data is classified according to the constitution, the constitution is determined, the data range of the constitution is obtained according to the national standard, and the interpolation data is modified as a constraint condition.

[0012] Optionally, in the S10, the step of expanding the health data in the same dimension, adding a flag bit and judging the abnormality specifically includes:

[0013] S11, the n-time, N-dimensional health data X is expanded in the same dimension to obtain n-time, 2N-dimensional health data X1;

[0014] S12, four flag bits are added in the health data X1, which are original data flag bit (used to distinguish whether the data comes from the wearable device), emergency flag bit (marking whether the value is not in the normal range), time sequence flag bit (detecting the continuity of time sequence), and warning flag bit (detecting the data missing condition).

[0015] S13, the abnormality detection of health data is realized based on four key flag bits. When the data value exceeds the preset normal range, the emergency flag bit will be set to yes; if the data is not collected directly by the wearable device, the original data flag bit will be set to yes; if the time sequence appears discontinuity or discontinuity, the time sequence flag bit will be set to no; if the data missing is detected, the warning flag bit will be set to yes. When the original flag bit is yes, or the emergency flag bit is yes, or the time sequence flag bit is no, or the warning flag bit is yes, the health data is judged as abnormal data.

[0016] Optionally, in the S20, the step of interpolating the abnormal data and using the GM (1, 1) model to process the single-dimensional data interpolation specifically includes:

[0017] S21, input health data X={x (0) (1),x (0) (2),...,x (0) (n)}∈R n×N , accumulate the data once to generate the accumulated sequence X (1) ={x (1) (1),x (1) (2),...,x (1) (n)}, wherein x (1) (k) is specifically represented as follows:

[0018] x (0) (i) is the original sequence, and x (1) (k) is the accumulated sequence;

[0019] S22, a first-order gray differential equation is constructed: x (0) (k)+az (1) (k)=b In the formula, a is the development coefficient, b is the grey action quantity, and z (1) (k) is a sequence x (1) The nearest mean sequence of (k):

[0020] S23, Solve for parameters a and b using the least squares method:

[0021] In the formula, B and Y are the key matrix and vector used for parameter estimation. Matrix B is composed of the nearest mean z. (1) (k) constitutes the vector Y, which is composed of the original sequence, and its form is:

[0022] S24, Establish the time response function to predict the cumulative sequence at time t+1. The predicted value at time t+1 is obtained. The expression is as follows:

[0023] The accumulated predicted values ​​are restored to the original sequence predicted values. The expression is as follows:

[0024] Obtain the prediction vector of the GM model

[0025] Optionally, in step S20, the step of using multiple linear regression (MLP) to handle the correlation between dimensions specifically includes:

[0026] S25, for a time segment x of health data X (0) (t)∈R 1×N For time segment x (0) (t)={u1(t),u2(t),...,u N Each dimension u in (t)} i (t) A regression equation is established to handle the correlation between dimensions, and the expression is as follows: u i (t)=β0+β1u1(t)+...+β i-1 u i-1 (t)+β i+1 u i+1 (t)+...+β Nu N (t)+∈ where u1(t),...,u N (t) is the value of other dimensions at time t, β is the regression coefficient, and ∈ is the error term.

[0027] S26, the regression equation of the parameter is established, and the relationship between the current dimension value and the value of other dimensions is: y=Aβ+∈

[0028] The dimension weight is estimated by the least square method: β=(A T A) -1 A T y

[0029] S27, based on the weight parameter β obtained by calculation, the value at time t+1 is predicted The expression is as follows: where is the dimension value predicted by the GM model at time t+1, and the prediction vector of the MLR at time t+1 is

[0030] Optionally, after the step S27, the method further comprises:

[0031] S28, a GM-MLR hybrid prediction model is established, the single variable prediction result of the GM(1,1) model is fused with the multivariable prediction result of the MLR, and the expression is as follows:

[0032] where α is a weight parameter, and the value range is [0, 1], the prediction data is filled into the health data X1 ∈ R n×2N , the first N dimensions of the health data X1 are original data, and the last N dimensions are prediction data, and the health data X1 is expanded according to the original dimension to obtain the health data X2 ∈ R n×3N .

[0033] Optionally, in the S30, the step of classifying the original data according to the constitution and determining the belonging constitution to obtain the data range of the belonging constitution comprises:

[0034] S31, input health data X ∈ R n×N , divide the data into v={1, 2,.., Q} categories according to the national constitution standard, and construct a weight matrix W ∈ R n×Q , each column of W qThe weight vector corresponding to the q-th class is calculated, and the score S of each class is calculated q , and the expression is as follows: S q = X T W q In the formula, S q = [S1, S2,..., S Q ] is the score vector of each class;

[0035] S32, convert the score to a classification probability P(v=q|x):

[0036] Output the probabilities of all classes [P(v=1), P(v=2),..., P(v=Q)];

[0037] S33, select the class with the maximum probability as the classification result:

[0038] S34, determine the numerical range [y min , y max ] of the corresponding constitution according to the national standard.

[0039] Optionally, in the S30, the step of correcting the interpolation data by taking the obtained constitution data range as a constraint includes:

[0040] S35, constrain and correct the initial prediction data X by the range [y min , y max ] corresponding to the constitution category, and the expression is as follows: Subject y min ≤ y * ≤ y max In the formula, keep the correction result close to the initial prediction, make the correction result consistent with the original data distribution, and λ is a weight parameter, with a value range of [0, 1], y min ≤ y * ≤ y max ensure that the prediction data is within the national standard range, and obtain the corrected data y * , fill the corrected data y * ∈ R n×N into the health data X2∈ R n×3NThe middle N dimensions of the health data X2 are the original data, the front N dimensions are the prediction data, and the last N dimensions are the correction data, so as to more intuitively compare the differences between different data.

[0041] The present application has at least the following beneficial effects:

[0042] 1. To solve the problem of insufficient correlation modeling of multi-dimensional physiological parameters in health data interpolation, a hybrid processing method combining gray theory and multiple linear regression is proposed. By combining the small sample advantage of gray model and the multi-dimensional correlation analysis capability of multiple linear regression, the defect that traditional single-dimensional modeling method cannot capture the collaborative change relationship between small sample parameters is effectively solved.

[0043] 2. To solve the problem that existing interpolation methods ignore the difference in medical constitution, the health data is divided according to the constitution classification standard, and the preliminary prediction result is corrected based on the reasonable range of physiological parameters corresponding to each constitution, so as to ensure that the interpolation value meets the clinical characteristics of a specific constitution and improve the accuracy and medical reasonableness of the result. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The schematic diagram of the step flow of the present application;

[0045] Figure 2 The schematic diagram of the embodiment of the present application; DETAILED DESCRIPTION

[0046] The present application provides a health data interpolation and optimization method combining flag bit judgment and constitution constraint mechanism, which is suitable for missing processing and abnormal correction of multi-dimensional physiological parameters (such as heart rate, blood pressure, blood oxygen, body temperature, etc.) collected by wearable devices.

[0047] As shown in Figure 2 , the system includes the following modules:

[0048] Flag bit detection module (101): Real-time monitoring of four key data quality flags;

[0049] Interpolation calculation module (102): Using GM(1,1) gray model and multiple linear regression for data completion;

[0050] Constitution discrimination module (103): Based on user historical data analysis individual health characteristics;

[0051] Medical correction module (104): Combined with clinical medical rules for final correction.

[0052] In specific implementation, algorithm parameters or medical rules can be adjusted according to actual needs, but the core method is still based on the above architecture. Those skilled in the art can understand the technical details in combination with the drawings, and the protection scope of the present application is not limited to the illustrated examples.

[0053] First embodiment

[0054] Reference Figure 2 In the present embodiment, the health data interpolation correction method based on multi-flag detection and constitution constraint comprises the following steps:

[0055] S10, the health data is expanded in the same dimension, and four flags are added, and whether the data is abnormal is judged according to the flags;

[0056] In the present embodiment, the health data is expanded in the same dimension, and four flags (original data flag, emergency flag, time sequence flag, and warning flag) are added after the expanded health data. Whether the data is abnormal is judged according to the four flags.

[0057] S20, the abnormal data is interpolated, the GM(1, 1) model is used to interpolate the single-dimensional data, the multi-element linear regression (MLP) is used to process the correlation between dimensions, and the weighted average of the results of the two models is obtained. The interpolated data is obtained;

[0058] In the present embodiment, the abnormal data is interpolated, the GM(1, 1) model is used to predict the single-dimensional data, the multi-element linear regression is used to obtain the correlation weight between dimensions, and the MLP predicted data is obtained according to the correlation weight and the GM(1, 1) predicted data. The weighted average of the results of the two models is obtained. The interpolated data is obtained.

[0059] S30, the original data is classified according to constitution, the belonging constitution is determined, the data range of the belonging constitution is obtained according to the national standard, and the interpolated data is corrected as a constraint condition.

[0060] In the present embodiment, the belonging constitution type of the data is judged by the standard constitution questionnaire scoring model, the numerical range of the belonging constitution is determined according to the national standard, and the interpolated data is corrected as a constraint condition.

[0061] In the technical scheme provided in the present embodiment, by judging the four flags set, the abnormal data segment in the data stream is identified, the data segment that needs to be interpolated is determined, the GM(1, 1) model is used to interpolate the data in a single dimension, the multi-element linear regression is used to process the correlation between dimensions, and the weighted average of the results of the two models is obtained. The prediction result is obtained. In addition, the health data is classified according to constitution, the normal range of the data is determined, and the prediction result is corrected as a constraint to improve the medical interpretability of the data.

[0062] Second embodiment

[0063] In the embodiment, when the data has numerical value missing, time sequence confusion and numerical value abnormality, the data is subjected to an interpolation operation, and the embodiment proposes a method for judging data abnormality by using a flag bit. Specifically, in step S10, the step of determining an abnormal data segment specifically includes:

[0064] S11, the health data X of n time points and N dimensions is subjected to same-dimension expansion to obtain health data X1 of n time points and 2N dimensions;

[0065] S12, four flag bits are added in the health data X1, which are an original data flag bit (used for distinguishing whether the data comes from a wearable device), an emergency flag bit (used for marking whether the numerical value is out of a normal range), a time sequence flag bit (used for detecting the continuity of a time sequence) and a warning flag bit (used for detecting data missing).

[0066] S13, the abnormality detection of the health data is realized based on the four key flag bits. When the data value is out of a preset normal range, the emergency flag bit is set to yes; if the data is not directly collected by a wearable device, the original data flag bit is set to yes; if the time sequence has discontinuity or discontinuity, the time sequence flag bit is set to no; if data missing is detected, the warning flag bit is set to yes. When the original flag bit is yes, or the emergency flag bit is yes, or the time sequence flag bit is no, or the warning flag bit is yes, the health data is determined as abnormal data.

[0067] In the embodiment, in view of the problem that the grey model is suitable for small samples but cannot handle the difficulty of dimension correlation, the embodiment proposes a data interpolation method fusing a GM(1,1) model and a multiple linear regression. The specific steps include:

[0068] S21, input health data X={x (0) (1),x (0) (2),...,x (0) (n)}∈R n×N , perform once accumulation on the data to generate an accumulated sequence X (1) ={x (1) (1),x (1) (2),...,x (1) (n)}, wherein x (1) (k) is specifically represented as follows: x (0) (i) is an original sequence, and x (1) (k) is an accumulated sequence;

[0069] S22, a first-order grey differential equation is constructed: x (0)(k) + az (1) (k) = b where a is the development coefficient, b is the grey action amount, and z (1) (k) is the original sequence x (1) (k) is the immediately adjacent mean sequence of x

[0070] S23, parameters a and b are solved by least square method:

[0071] where B and Y are the key matrix and vector for parameter estimation, the matrix B is composed of the immediately adjacent mean z (1) (k), and the vector Y is composed of the original sequence x

[0072] S24, a time response function is established to predict the accumulated sequence x at t+1 The expression of the predicted value x at t+1 is as follows:

[0073] The accumulated predicted value is restored to the original sequence predicted value x The expression of x The predicted vector of the GM model is obtained

[0074] In this embodiment, the health data X of n time and N dimensions is predicted in time sequence based on the GM(1, 1) model, and the predicted data of a single dimension at n time is obtained

[0075] Further, after step S24, the step of processing the correlation between dimensions by multiple linear regression (MLP) includes:

[0076] S25, for a time segment x (0) (t) ∈ R 1×N , the regression equation is established for each dimension u (0) (t) = {u1(t), u2(t),..., u N (t)} in the time segment x i (t) to process the correlation between dimensions, and the expression is as follows: u i (t) = β0+ β1u1(t) +... + β i-1 u i-1 (t) + βi+1 u i+1 (t)+...+β N u N (t)+∈

[0077] In the formula, u1(t),...,u N (t) is the value of other dimensions at time t, β is the regression coefficient, and ∈ is the error term.

[0078] S26, a regression equation of the parameters is established, and the relationship between the current dimension value and the other dimension values is: y=Aβ+∈

[0079] The dimension weight is estimated by the least square method: β=(A T A) -1 A T y

[0080] S27, based on the weight parameter β obtained by calculation, the value at time t+1 is predicted The expression is as follows: In the formula, is the dimension value predicted by the GM model at time t+1, and the prediction vector of the MLR at time t+1 is

[0081] In the embodiment, by constructing the correlation relationship y=Aβ+∈ between the dimensions, the weight parameter β between different dimensions is calculated, and the data predicted by the MLP is obtained by weighting the data predicted by the GM(1,1) according to the weight parameter β and the data predicted by the GM(1,1) Further, after the step S27, the method further includes:

[0082] S28, a GM-MLR hybrid prediction model is established, and the single-variable prediction result of the GM(1,1) model and the multi-variable prediction result of the MLR are fused, and the expression is as follows: In the formula, α is a weight parameter, and the value range is [0, 1], and the value is generally 0.6. The prediction data is filled into the health data X1∈R n×2N , the first N dimensions of the health data X1 are original data, and the last N dimensions are prediction data. The health data X1 is expanded according to the original dimension to obtain the health data X2∈R n×3N .

[0083] In this embodiment, considering both the temporal and dimensional correlation of health data, abnormal data is predicted. GM(1,1) is used to predict health data from a temporal perspective. Multiple linear regression for dimensional correlation prediction data The results from the two datasets are then weighted to obtain predictive data that includes both temporal and dimensional correlations.

[0084] Third Embodiment

[0085] In this embodiment, addressing the issue that while deep learning methods can model complex relationships, they often overlook physical differences, a data correction method based on physical constraints is proposed. Specific steps include:

[0086] S31, Input health data X∈R n×N The data is divided into classes v = {1, 2, ..., Q} according to national physical fitness standards, and a weight matrix W ∈ R is constructed. n×Q Each column W q For the weight vector corresponding to the q-th class, calculate the score S for each class. q The expression is as follows: S q =X T W q In the formula, S q =[S1,S2,..., S Q] is the score vector for each class;

[0087] S32, convert the score into a classification probability P(v=q|x): Output the probabilities of all categories [P(v=1), P(v=2), ..., P(v=Q)];

[0088] S33, Select the category with the highest probability as the classification result:

[0089] S34, the numerical range of the corresponding physical constitution is determined according to national standards [y min ,y max ].

[0090] In this embodiment, the weight matrix W for determining whether data belongs to different body types is determined using a standard body constitution questionnaire scoring model, and the score vector S for each body type is calculated. q Based on the classification probability, the category with the highest probability is selected as the classification result, and the numerical range of the corresponding physical constitution is determined according to national standards. min ,y max ].

[0091] Further, the obtained physical data range [y min ,y max ] is used as a constraint condition to modify the interpolation data, and the specific steps include:

[0092] S35, the initial prediction data X is modified according to the range [y min ,y max ] corresponding to the physical category, and the expression is as follows: Subject y min ≤y * ≤y max

[0093] In the formula, the modified result is close to the initial prediction, the modified result is consistent with the original data distribution, and λ is a weight parameter, the value range is [0, 1], and the general value is 0.5, y min ≤y * ≤y max ensure that the prediction data is within the national standard range, and obtain the modified data y * , the modified data y * ∈R n×N is filled into the health data X2∈R n×3N , the first N dimensions of the health data X2 are the original data, the middle N dimensions are the prediction data, and the last N dimensions are the modified data, so as to more intuitively compare the differences between different data.

[0094] In this embodiment, the physical data range [y min ,y max ] is used as a constraint condition to modify the prediction result , if the data is not in the physical data range, the prediction result is modified, and the final result y * is obtained.

[0095] As an implementation solution, in a specific embodiment, the health data segment in the embodiment is for daily collection of 3 types of indexes (heart rate, systolic pressure, diastolic pressure), and the sampling frequency is 5 minutes per day. Four flag bits are added, i.e., an original data flag bit, an emergency flag bit, a time sequence flag bit, and a warning flag bit, to determine whether the health data deviates from the normal data range, and if so, the emergency flag bit is set to yes, whether the health data is directly from the wearable device, and if not, the original data flag bit is set to yes, whether the health data is arranged in time sequence, and if not, the time sequence flag bit is set to no, and finally whether the health data is missing, and if so, the warning flag bit is set to yes. The health data original flag bit is yes, or the emergency flag bit is yes, or the time sequence flag bit is no, or the warning flag bit is yes, which is marked as abnormal data (0 represents no and 1 represents yes), as shown in Table 1, the data segment at 08:10 is abnormal, and the heart rate and diastolic pressure data are missing.

[0096] Table 1 Health data segment

[0097] The GM(1,1) model is used according to steps S21-S24 to generate data, for example, according to the heart rate data X= {75, 76, 78, 81} to generate the accumulated sequence X (1) {75, 151, 229, 310}, the close mean sequence z (1) (k) is constructed. {113, 190, 269.5}, a first-order grey differential equation is constructed: x (0) (k) + az (1) (k) = b In the formula, a is a development coefficient, and b is a grey action amount, and the parameters a and b are solved by the least square method:

[0098] B and Y are key matrices and vectors for parameter estimation, and the matrix B and the vector Y are:

[0099] The parameters a and b are calculated by the least square method as a = -0.0442 and b = 73.34. A time response function is established to predict the accumulated sequence at t+1 The predicted value at t+1 is obtained as The expression of the predicted value is as follows: The accumulated predicted value is restored to the original sequence predicted value, The expression is as follows:

[0100] The predicted vector of the GM model is obtained As shown in Table 2.

[0101] Table 2 GM model predicted vector

[0102] The correlation between dimensions is processed using multivariate linear regression according to steps S25-S27, for a time segment of health data X For a time segment x (0) Each dimension u in (t) = {u1(t), u2(t), u3(t)} i The regression equation is established to process the correlation between dimensions, assuming that systolic blood pressure (SBP) is the dependent variable, and heart rate (HR) and diastolic blood pressure (DBP) are the independent variables, to predict systolic blood pressure using heart rate and diastolic blood pressure. u3(t) = β0 + β1u1(t) + β2u2(t)

[0103] The regression model is fitted by least squares, and the intercept β0 = 62.45, the heart rate weight β1 = 0.39, and the diastolic blood pressure weight β2 = 0.91 are obtained, and the regression equation is: u3(t) = 62.45 + 0.39u1(t) + 0.91u2(t)

[0104] The value of t+1 predicted by the MLR model is calculated according to the obtained weight β The expression is as follows: In the formula, The dimension value predicted by the GM model at t+1 is u(t+1), and the predicted vector of the MLR at t+1 is As shown in Table 3.

[0105] Table 3 Multivariate linear regression predicted vector

[0106] The GM-MLR hybrid prediction model is established according to step S28, the univariate prediction result of the GM(1,1) model is fused with the multivariate prediction result of the MLR, and the expression is as follows: In the formula, α is a weight parameter, and the value range is [0, 1], and the general value is 0.6, and the predicted value is As shown in Table 3.

[0107] Table 3 Fused predicted value Timestamp Heart rate Systolic blood pressure Diastolic blood pressure Fused heart rate Fused systolic blood pressure Fused diastolic blood pressure 2024 / 7 / 107:50 75 120 80 74.6 137.8 85.7 2024 / 7 / 107:55 76 118 79 75.0 136.2 84.8 2024 / 7 / 108:00 78 126 82 77.8 143.2 88.0 2024 / 7 / 108:05 81 124 78 79.3 140.4 84.5 2024 / 7 / 108:10 NaN 122 NaN 80.5 138.5 82.9

[0108] The body constitution and data range are determined according to steps S31-S34. According to the national standard, the maximum probability of body constitution is yang hyperactivity due to yin deficiency, and the numerical ranges of the three indexes are heart rate [75, 81], systolic pressure [120, 139], and diastolic pressure [80, 89]. The obtained body constitution data range [y min ,y max ] is used as a constraint condition to modify the interpolation data, and the initial prediction data The original data X is modified according to the range [y min ,y max ] corresponding to the body constitution category, and the expression is as follows, y * The results are shown in Table 4. Subject y min ≤y * ≤y max

[0109] Table 4 Modification results table

[0110] Although the preferred embodiments of the present application have been described, those skilled in the art who understand the basic inventive concept can make additional changes and modifications to the embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0111] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A health data imputation and optimization method combining flag bit determination and physical constraint mechanism, characterized in that, The method comprises the following steps: S10, health data is expanded in the same dimension, a flag bit is added, and it is judged whether the data is abnormal; S20, the abnormal data is interpolated, the GM(1,1) model is used to interpolate the single-dimensional data, the multi-linear regression (MLP) is used to process the correlation between dimensions, and the weighted average of the model results is obtained. S30, the original data is classified according to the constitution, the belonging constitution is determined, the data range of the constitution is obtained according to the national standard, and the interpolation data is corrected as a constraint condition.

2. The method of claim 1, wherein the data exception judgment of the flag bit is performed by: In the S10, the steps of expanding the health data in the same dimension, adding a flag bit and judging the abnormality specifically comprise: S11, the health data X of n moments and N dimensions is expanded in the same dimension to obtain the health data X1 of n moments and 2N dimensions; S12, four flag bits are added in the health data X1, which are an original data flag bit (used to distinguish whether the data comes from a wearable device), an emergency flag bit (used to mark whether the value is out of the normal range), a time sequence flag bit (used to detect the continuity of the time sequence) and a warning flag bit (used to detect the data loss condition). S13, the abnormality detection of the health data is realized based on the four key flag bits. When the data value is out of the preset normal range, the emergency flag bit is set to yes; if the data is not collected by the wearable device directly, the original data flag bit is set to yes; if the time sequence is discontinuous, the time sequence flag bit is set to no; if the data loss is detected, the warning flag bit is set to yes. When the original flag bit is yes, or the emergency flag bit is yes, or the time sequence flag bit is no, or the warning flag bit is yes, the health data is determined as abnormal data.

3. The method for interpolating abnormal data segments based on GM (1, 1) model and multiple linear regression according to claim 1, characterized in that, In the S20, the steps of interpolating the abnormal data and using the GM(1,1) model to process the single-dimensional data interpolation specifically comprise: S21, input health data The data is accumulated once to generate an accumulated sequence X (1) = {x (1) (1), x (1) (2),..., x (1) (n)}, where x (1) (k) is specifically represented as follows: x (0) (i) is the original sequence, x (1) (k) is the cumulative sequence; S22, a first-order grey differential equation is constructed: x (0) (k) + az (1) (k) = b In the formula, a is a development coefficient, b is a gray action amount, z (1) (k) is a sequence of x (1) (k) is a sequence of x S23, the parameters a and b are solved by the least square method: where B and Y are the key matrix and vector for parameter estimation, the matrix B is formed by the elements immediately adjacent to the mean z (1) (k) and the vector Y is formed by the elements of the original sequence, in the form S24, establish time response function, predict accumulated sequence at t+1 obtain predicted value at t+1 The expression is as follows: reducing the accumulated prediction values to original sequence prediction values, The expression is as follows: obtaining a prediction vector of the GM model 4. The method for interpolating abnormal data segments based on GM (1, 1) model and multiple linear regression according to claim 3, characterized in that, In the S20, the steps of using the multi-linear regression (MLP) to process the correlation between dimensions specifically comprise: S25, for a time slice of health data X for time slice x (0) (t) = {u1(t), u2(t),..., u N (t) for each dimension u i (t) establishes a regression equation to handle the correlation between dimensions, expressed as follows: u i (t) = β0+ β1u1(t) +... + β i-1 u i-1 (t) + β i+1 u i+1 (t) +... + β N u N (t) + ∈ where u1(t),...,u N (t) is the value of other dimensions at time t, β is the regression coefficient, and ∈ is the error term. S26, a regression equation of the parameters is established, and the relationship between the current dimension value and other dimension values is: y=Aβ+∈ The dimension weight is estimated by the least square method: β = (A T A) -1 A T y S27, based on the calculated weight parameter β, predict the value at time t+1 The expression is as follows: In the formula, is the dimension value predicted by the GM model at time t+1, and the prediction vector of the MLR at time t+1 is Optionally, after the step S27, the method further comprises: S28, a GM-MLR hybrid prediction model is established, the single-variable prediction result of the GM(1,1) model and the multi-variable prediction result of the MLR are fused, and the expression is as follows: In the formula, a is a weight parameter, and a value range is [0, 1], and the predicted data Fill in the health data N dimensions, the health data X1 is original data, and the last N dimensions are predicted data, and the health data X1 is expanded according to the original dimension, and the health data 5. The method for correcting the interpolated data based on the physical constraint according to claim 1, wherein, In the S30, the steps of classifying the original data according to the constitution, determining the belonging constitution and obtaining the range of the constitution data specifically comprise: S31, input health data The data is divided into v = {1, 2,..., Q} classes according to the national physical standard, and a weight matrix is constructed Each column W q The weight vector corresponding to the qth class is calculated, and the score S of each class is calculated q The expression is as follows: S q = X T W q where S = [S1, S2,..., S q Q is the score vector for each class.​ S32, the score is converted into a classification probability P(v=q|x): The probabilities of all categories [P(v=1), P(v=2),..., P(v=Q)] are outputted; S33, the category with the maximum probability is selected as the classification result: S34, determine the numerical range [y min , y max ] of the body constitution according to the national standard. S34, determine the numerical range [y min , y max ] of the body constitution according to S35, the initial prediction data The original data X is corrected by constraining the range [y min , y max ] corresponding to the physical category, and the expression is as follows: In the formula, Keep the corrected result close to the initial prediction. To ensure the corrected data distribution matches the original data distribution, λ is a weighting parameter with a value range of [0,1], and y min ≤y * ≤y max Ensure that the predicted data is within the national standard range to obtain the corrected data y. * The data will be corrected. Enter health data The data consists of N dimensions, with the first N dimensions representing the original data, the middle N dimensions representing the predicted data, and the last N dimensions representing the corrected data. This allows for a more intuitive comparison of the differences between the different data.

Images