Risk assessment and early warning method and system for older pregnant women

Abstract

The present invention discloses a risk assessment and early warning method and system for older pregnant women, and relates to the technical field of data processing. The method comprises: step 1: collecting physiological index data of the target pregnant woman at each time point; step 2: using wavelet transform to perform time-frequency domain analysis on the health index vector at each time point to obtain the fused features at each time point; step 3: using the fused features at each time point as input, inputting them into a bidirectional LSTM network, and the bidirectional LSTM network outputs the forward hidden state and backward hidden state corresponding to each time point; step 4: using the backward hidden state with an abnormal score exceeding a set threshold as the forward abnormal hidden state; step 5: if the final risk probability exceeds the risk judgment threshold, a risk warning is issued. The present invention significantly improves the accuracy, real-time and dynamic performance of health risk assessment for older pregnant women.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to a method and system for risk assessment and early warning for older pregnant women. Background Art

[0002] In the process of the development of modern medical technology, health management and risk assessment for pregnant women, especially older mothers, have become a research hotspot in the field of obstetrics. Older mothers generally refer to pregnant women over the age of 35. As they age, the health risks they face during pregnancy increase significantly, and they are prone to complications such as hypertension, diabetes, placental abruption, and fetal growth restriction. Therefore, how to effectively monitor the health and warn of risks for older mothers has become a key issue to improve pregnancy safety and reduce maternal and child health risks.

[0003] Existing maternal health monitoring technologies mostly use traditional physiological indicator detection methods to assess the health status of pregnant women by regularly monitoring indicators such as blood pressure, blood sugar, and fetal heart rate. For example, during routine prenatal checkups, doctors will regularly test the blood pressure of pregnant women to assess whether there is a risk of gestational hypertension, and use blood sugar tests to determine whether pregnant women have gestational diabetes. However, this static, regular detection method has many limitations.

[0004] First, the traditional prenatal examination method is based on fixed-point detection at time intervals, which cannot continuously and dynamically monitor the health status of pregnant women. Maternal physiological indicators, such as heart rate and blood pressure, may fluctuate significantly in a short period of time, and regular examinations alone may not be able to capture these fluctuations in time. When the interval between prenatal examinations is long, potential health risks may be overlooked, especially in older mothers. Some acute health problems, such as fetal distress or sudden hypertension, may occur between two prenatal examinations, which greatly increases the health risks of mothers and infants. Secondly, most existing health monitoring methods rely on the analysis of a single physiological indicator and fail to fully consider the association and dynamic changes between multiple physiological parameters. For example, although hypertension and hyperglycemia are both common problems faced by older mothers, there may be mutual influence between the two, and monitoring a single indicator may not fully reflect the overall health status of the mother. In addition, the assessment of fetal development also mainly relies on ultrasound detection, such as biparietal diameter, fetal weight estimation and other parameters. However, these tests often have a large error range and are limited to physical growth assessment, failing to provide an overall dynamic assessment of the maternal-fetal system. Summary of the invention

[0005] The main purpose of the present invention is to provide a risk assessment and early warning method and system for older pregnant women, which significantly improves the accuracy, real-time and dynamic nature of health risk assessment for older pregnant women, and can better provide personalized health management and early risk warning for pregnant women. This not only improves the safety of pregnant women and fetuses, but also provides more scientific decision-making support for medical staff.

[0006] In order to solve the above problems, the technical solution of the present invention is achieved as follows:

[0007] A risk assessment and early warning method for older pregnant women, the method comprising:

[0008] Step 1: Collect the physiological index data of the target parturient at each time point, standardize each physiological index data, and use it as an element to construct the health index vector at each time point; use the health index vector at each time point as an element to construct the health index time series of the target parturient;

[0009] Step 2: Use wavelet transform to perform time-frequency domain analysis on the health index vector at each time point, extract the wavelet transform coefficient at each time point; calculate the maximum Lyapunov exponent of the health index vector at each time point; perform weighted feature fusion on the wavelet transform coefficient and the maximum Lyapunov exponent at each time point to obtain the fused feature at each time point;

[0010] Step 3: Take the fused features at each time point as input and input them into a bidirectional LSTM network. The bidirectional LSTM network outputs the forward hidden state and backward hidden state corresponding to each time point.

[0011] Step 4: Fit the forward hidden state of all time points through the first Gaussian mixture model to obtain the corresponding forward hidden state distribution, perform anomaly analysis on the forward hidden state distribution, calculate the anomaly score of the forward hidden state, and take the forward hidden state whose anomaly score exceeds the set threshold as the forward abnormal hidden state; fit the backward hidden state of all time points through the second Gaussian mixture model to obtain the corresponding backward hidden state distribution, perform anomaly analysis on the backward hidden state distribution, calculate the anomaly score of the backward hidden state, and take the backward hidden state whose anomaly score exceeds the set threshold as the backward abnormal hidden state;

[0012] Step 5: Take the forward hidden state distribution as the first prior knowledge, the anomaly score of the forward abnormal hidden state as the first observation data, and use the Bayesian method to calculate the first risk probability; take the backward hidden state distribution as the second prior knowledge, the anomaly score of the backward abnormal hidden state as the second observation data, and use the Bayesian method to calculate the second risk probability; and take the average of the first risk probability and the second risk probability as the final risk probability, set a risk judgment threshold, and issue a risk warning if the final risk probability exceeds the risk judgment threshold.

[0013] Furthermore, the physiological indicator data at least include: systolic blood pressure, diastolic blood pressure, heart rate, blood sugar, body temperature, blood oxygen saturation, respiratory rate, fetal heart rate, weight, fetal biparietal diameter, fetal femur length, fetal weight estimation, and amniotic fluid index.

[0014] Furthermore, in step 1, let Indicates the target maternal age at time The Physiological index data, assuming that Different physiological index data; each physiological index data at each time point is standardized: ;in, It's time Time Normalized values ​​of physiological index data; It is The mean of physiological index data; It is The standard deviation of the physiological index data; each time point All standardized physiological index data Combined into a vector, the health index vector is ;in It's time The health index vector at each moment contains all Normalized physiological index data; By transforming the health index vector at each time point Arranged in chronological order, the target maternal health index time series is formed as follows: ; Indicates the total number of time points for data collection, It means that the mother has Time series of health index at each time point; It's time The standardized value of the first physiological index data; It's time The standardized value of the second physiological index data; It's time Time Normalized values ​​of physiological index data; is the health index vector at the first time point; is the health index vector at the second time point; It is A vector of health indices at each time point.

[0015] Furthermore, in step 2, the health index vector at each time point is analyzed in the time-frequency domain using wavelet transform, and the wavelet transform coefficients at each time point are extracted as follows: ; Calculate the maximum Lyapunov exponent of the health index vector at each time point as ; The wavelet transform coefficient and the maximum Lyapunov exponent at each time point are weighted and fused by the following formula to obtain the fused feature at each time point : ;in, ; is the wavelet weight, is the Lyapunov weight, and The value range of ; represents wavelet transform, is the scale of wavelet transform; is the displacement of wavelet transform.

[0016] Furthermore, in step 3, each time point The fusion feature vector Input into the bidirectional LSTM network; the bidirectional LSTM network contains a forward LSTM network unit and a backward LSTM network unit; the fusion feature vector After inputting into the forward LSTM network unit, the forward hidden state is output ;in, For time The forward hidden state of Represents the forward LSTM network unit; fused feature vector After inputting into the backward LSTM network unit, the backward hidden state is output ; For time The backward hidden state of Represents a backward LSTM network unit.

[0017] Furthermore, the process of obtaining the forward abnormal hidden state in step 4 specifically includes: forming the forward hidden states of all time points into a forward hidden state sequence , it is fitted by the first Gaussian mixture model, where the probability density function of the first Gaussian mixture model is It is expressed as: ;in, It is The mixing coefficient of the Gaussian distribution satisfies ; It is The probability density function of a Gaussian distribution with a mean of , the covariance matrix is ; and They are The mean and covariance of a Gaussian distribution; is the number of preset Gaussian distributions; calculate the Mahalanobis distance between each forward hidden state and its corresponding Gaussian distribution: ;in, is the forward hidden state With Gaussian distribution mean The Mahalanobis distance between is the inverse of the covariance matrix; each forward hidden state is assigned a forward anomaly score based on the size of the Mahalanobis distance , the specific formula is: ;in, is the forward hidden state The forward anomaly score of yes The maximum Mahalanobis distance from the center of all Gaussian distributions; The forward hidden state exceeding the set threshold is regarded as the forward abnormal hidden state.

[0018] Furthermore, the process of obtaining the backward abnormal hidden state in step 4 specifically includes: forming the backward hidden states of all time points into a backward hidden state sequence , and fit it through the second Gaussian mixture model, where the probability density function of the second Gaussian mixture model is It is expressed as: ;in, It is The mixing coefficient of the Gaussian distribution satisfies ; It is The probability density function of a Gaussian distribution with a mean of , the covariance matrix is ; and They are The mean and covariance of a Gaussian distribution; is the number of preset Gaussian distributions; calculate the Mahalanobis distance between each forward hidden state and its corresponding Gaussian distribution: ;in, is the forward hidden state With Gaussian distribution mean The Mahalanobis distance between is the inverse of the covariance matrix; each forward hidden state is assigned a forward anomaly score based on the size of the Mahalanobis distance , the specific formula is: ;in, is the forward hidden state The forward anomaly score of yes The maximum Mahalanobis distance from the center of all Gaussian distributions; The forward hidden state exceeding the set threshold is regarded as the forward abnormal hidden state.

[0019] Furthermore, in step 5, the first risk probability is calculated by the following formula: ;in, is the first risk probability, which represents the anomaly score given the forward anomaly hidden state The mother is at risk The probability of is the first likelihood function, indicating that the mother is in a risk state When , the anomaly score of the forward anomaly hidden state The probability of is the prior probability, which indicates the prior knowledge that the mother is in a risky state; is the first marginal likelihood, representing the anomaly score of the forward hidden state The overall probability of the second risk is calculated by the following formula: ;in, is the second risk probability, representing the anomaly score given the backward anomaly hidden state The mother is at risk The probability of is the second likelihood function, indicating that the mother is in a risk state When , the anomaly score of the backward anomaly hidden state The probability of is the prior probability, which indicates the prior knowledge that the mother is in a risky state; is the second marginal likelihood, representing the abnormality score of the backward hidden state The overall probability of the final risk Calculated using the following formula: .

[0020] Furthermore, in step 5, the first likelihood function is expressed by the following formula: ;in, is the number of forward anomaly hidden states; is the subscript index; is the Gaussian kernel function; the second likelihood function is expressed by the following formula: ;in, is the number of backward anomaly hidden states; the first marginal likelihood Use the following formula to calculate: ;in, is the prior probability that the mother is not at risk; is the distribution of the forward hidden state anomaly scores in the non-risk state; Indicates not being at risk; is the number of forward anomaly hidden states; the second marginal likelihood Use the following formula to calculate: ;in, is the distribution of the backward hidden state anomaly scores in the non-risk state.

[0021] A risk assessment and early warning system for older pregnant women is used to implement the risk assessment and early warning method for older pregnant women as described above. The system includes: a data acquisition part, which is used to collect physiological index data of target pregnant women at each time point, and after data standardization of each physiological index data, as an element, construct a health index vector at each time point; the health index vector at each time point is used as an element to constitute the health index time series of the target pregnant woman; a feature extraction and fusion part, which is used to use wavelet transform to perform time-frequency domain analysis on the health index vector at each time point, and extract the wavelet transform coefficient at each time point; calculate the maximum Lyapunov exponent of the health index vector at each time point; perform weighted feature fusion on the wavelet transform coefficient and the maximum Lyapunov exponent at each time point to obtain the fused feature at each time point; a bidirectional LSTM network part, which is used to input the fused feature at each time point as input into a bidirectional LSTM network, and the bidirectional LSTM network outputs the forward hidden state and backward hidden state corresponding to each time point; a risk warning distribution part, which is used to perform time-frequency domain analysis on the health index vector at each time point, and extract the wavelet transform coefficient at each time point; calculate the maximum Lyapunov exponent of the health index vector at each time point; perform weighted feature fusion on the wavelet transform coefficient and the maximum Lyapunov exponent at each time point to obtain the fused feature at each time point; a bidirectional LSTM network part, which is used to input the fused feature at each time point into a bidirectional LSTM network, and the bidirectional LSTM network outputs the forward hidden state and backward hidden state corresponding to each time point; a risk warning distribution part, which is used to perform time-frequency domain analysis on the health index vector at each time point, and extract the wavelet transform coefficient at each time point; calculate the maximum Lyapunov exponent of the health index vector at each time point; perform weighted feature fusion on the wavelet transform coefficient and the maximum Lyapunov exponent at each time point, and obtain the fused feature at each time point The forward hidden state of the point is fitted by the first Gaussian mixture model to obtain the corresponding forward hidden state distribution, the forward hidden state distribution is analyzed for anomalies, the anomaly score of the forward hidden state is calculated, and the forward hidden state with an anomaly score exceeding the set threshold is taken as the forward abnormal hidden state; the backward hidden state of all time points is fitted by the second Gaussian mixture model to obtain the corresponding backward hidden state distribution, the backward hidden state distribution is analyzed for anomalies, the anomaly score of the backward hidden state is calculated, and the backward hidden state with an anomaly score exceeding the set threshold is taken as the backward abnormal hidden state; the forward hidden state distribution is taken as the first prior knowledge, the anomaly score of the forward abnormal hidden state is taken as the first observation data, and the first risk probability is calculated by the Bayesian method; the backward hidden state distribution is taken as the second prior knowledge, the anomaly score of the backward abnormal hidden state is taken as the second observation data, and the second risk probability is calculated by the Bayesian method; and the mean of the first risk probability and the second risk probability is taken as the final risk probability, a risk judgment threshold is set, and if the final risk probability exceeds the risk judgment threshold, a risk warning is issued.

[0022] The risk assessment and early warning method and system for older pregnant women of the present invention have the following beneficial effects: the present invention greatly improves the feature extraction capability of complex changes in health status by combining wavelet transform with the maximum Lyapunov exponent. Wavelet transform can finely decompose the health index vector in the time-frequency domain, and capture short-term fluctuations and local anomalies in the health status of pregnant women. This is particularly critical for dynamic monitoring of health status, because the physiological indicators of older pregnant women often fluctuate in a short period of time, and these fluctuations may be early signals of potential risks. At the same time, the maximum Lyapunov exponent is used to quantify the chaotic characteristics and system instability of the health status, which can help identify the long-term risks hidden in the health status of pregnant women. This dual feature extraction method combines local changes and global dynamics, making the system more sensitive and accurate in capturing key changes in health status. The present invention introduces a bidirectional LSTM network, which greatly improves the analysis capability of time series data. The LSTM network is good at processing time series data with long-term dependence, and can effectively avoid the gradient disappearance problem in the traditional RNN network. The design of the bidirectional LSTM in the present invention not only considers the impact of historical changes in health status on the current situation (forward analysis), but also combines the prediction of current health risks with future potential trends (backward analysis). This bidirectional time-dependent processing method enables the system to more comprehensively evaluate changes in health status, especially when health status has complex dependencies, and can provide more accurate risk prediction results through bidirectional analysis. This design enables the system to not only detect historical health problems, but also warn of potential health risks in the future, further enhancing the accuracy of risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic diagram of the method flow of the risk assessment and early warning method for older pregnant women provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0025] Example 1: Reference Figure 1 , a risk assessment and early warning method for older pregnant women, the method comprising:

[0026] Step 1: Collect the physiological index data of the target parturient at each time point, standardize each physiological index data, and use it as an element to construct the health index vector at each time point; use the health index vector at each time point as an element to construct the health index time series of the target parturient;

[0027] Each physiological index data collected may have different dimensions and ranges. For example, the magnitude of blood pressure and heart rate is quite different. Directly comparing and analyzing these data of different scales will lead to deviation. In order to eliminate these differences, each physiological index data needs to be standardized. Through standardization, different indicators are adjusted to the same scale, so that both the fluctuation of heart rate and the change of blood pressure can be compared and analyzed on the same scale. The purpose of this standardization is to ensure that different physiological indicators have equal importance in constructing the health index, and will not have too much impact on the overall health assessment due to the large dimension of a certain indicator. After standardization, all physiological indicators of the parturient at a certain point in time will be integrated into a vector, which is called the health index vector. Each element of the health index vector represents a standardized physiological indicator data, so it can comprehensively reflect the overall health status of the parturient at a specific point in time. Then, as time goes by, the health index vectors of each time point are arranged in sequence to form a time series. This time series not only contains the health status of the parturient at a certain moment, but also reflects the trend of her health status over time.

[0028] Step 2: Use wavelet transform to perform time-frequency domain analysis on the health index vector at each time point, extract the wavelet transform coefficient at each time point; calculate the maximum Lyapunov exponent of the health index vector at each time point; perform weighted feature fusion on the wavelet transform coefficient and the maximum Lyapunov exponent at each time point to obtain the fused feature at each time point;

[0029] Wavelet transform is a commonly used signal processing method, especially suitable for analyzing non-stationary signals. The physiological indicators of older pregnant women are such complex data that fluctuate over time and have non-stationary characteristics. By performing wavelet transform on the health index vector, we can decompose the health status of pregnant women into details at different scales and capture the local changes of physiological indicators. This is particularly important for older pregnant women, because as they age, their physiological systems may undergo sudden and unpredictable changes at certain moments, and wavelet transform can amplify and reveal these subtle fluctuations. The Lyapunov exponent is an indicator used to measure the sensitivity of a system to initial conditions, which can reflect the chaotic characteristics of the system. In the health monitoring of older pregnant women, the calculation of the maximum Lyapunov exponent can help assess whether the health status of pregnant women is in an unstable or chaotic state. For older pregnant women, the uncertainty of health status may be greater, and the maximum Lyapunov exponent can quantify this uncertainty and identify potential crises in health status. By combining the analysis of wavelet transform and Lyapunov exponent, the system can simultaneously obtain the local detailed features of the health status and the global dynamic behavior features. This multi-level feature extraction greatly enhances the insight into the health status. After obtaining the wavelet transform coefficients and Lyapunov exponents at each time point, the two are not isolated, but are further combined through the weighted feature fusion method. The significance of this step is that it can integrate the advantages of different analysis methods to form a more accurate representation of the health status. The wavelet transform reveals the local change trend of maternal physiological indicators, while the Lyapunov exponent reveals the global stability or chaos of the system. Through weighted fusion, these features can more comprehensively reflect the health status of the mother at a certain point in time, especially those potential risk signals.

[0030] Step 3: Take the fused features at each time point as input and input them into a bidirectional LSTM network. The bidirectional LSTM network outputs the forward hidden state and backward hidden state corresponding to each time point.

[0031] For older pregnant women, the health status of the body during pregnancy may show complex fluctuations and instability, and the changes in health indicators are not always linear or predictable. The advantage of the bidirectional LSTM model is that it can remember long-term changes in health status while avoiding forgetting important short-term fluctuations. In the bidirectional LSTM, the forward network is used to extract features from the past health index time series, while the backward network infers possible impacts from future health status. This bidirectional processing mechanism ensures that the health status analysis at each time point is not only based on previous data, but also includes predictions of possible future trends. In specific applications, the health index time series is subjected to wavelet transformation and maximum Lyapunov exponent extraction to obtain fused features, which are passed as input to the bidirectional LSTM model. The memory unit of LSTM can effectively handle long-term dependencies in time series, and due to its gating mechanism, it can selectively retain important information to avoid information loss or excessive forgetting. Especially when processing the health data of older pregnant women, some important health signals may appear at an earlier time point, and the bidirectional LSTM can ensure that these key information will not be ignored in subsequent analysis. This is extremely important for risk warning for older pregnant women, because some potential risks may show signs as early as the early stages of pregnancy, rather than only manifesting themselves in the later stages. In addition, the forward hidden state and backward hidden state output by the bidirectional LSTM capture the dual dependencies of the mother's health status. The forward hidden state reflects the trend of health status changes from the past to the present moment, while the backward hidden state reveals the health prediction from the present moment to the future. Through this bidirectional analysis, the system can form a more accurate and comprehensive understanding of the health status at each time point, thereby improving the accuracy of risk assessment. Compared with the traditional unidirectional time series analysis model, this bidirectional LSTM design can more effectively capture the complex dependencies and potential risk patterns in the health status.

[0032] Step 4: Fit the forward hidden state of all time points through the first Gaussian mixture model to obtain the corresponding forward hidden state distribution, perform anomaly analysis on the forward hidden state distribution, calculate the anomaly score of the forward hidden state, and take the forward hidden state whose anomaly score exceeds the set threshold as the forward abnormal hidden state; fit the backward hidden state of all time points through the second Gaussian mixture model to obtain the corresponding backward hidden state distribution, perform anomaly analysis on the backward hidden state distribution, calculate the anomaly score of the backward hidden state, and take the backward hidden state whose anomaly score exceeds the set threshold as the backward abnormal hidden state;

[0033] The forward and backward hidden states represent the health status of the mother at different times. These hidden states contain multi-dimensional information about the health status, but due to the complexity of the time series, directly analyzing the changes in the hidden states may not be enough to reveal potential health risks. Therefore, the Gaussian mixture model is introduced to fit these hidden states to capture their potential probability distribution. The Gaussian mixture model approximates the true distribution of the data through multiple Gaussian distributions. This multimodal probability model can effectively handle complex and changeable health status data. For older mothers, the changes in health status may present multiple patterns, such as slight fluctuations, sudden abnormalities, etc., and the multimodal characteristics of the Gaussian mixture model are just adapted to this complex change. After fitting the Gaussian mixture model for the forward and backward hidden states, the system will obtain the corresponding hidden state distribution. These distributions reflect the normal range and fluctuation range of the health status. Once these distributions are established, the system can further perform abnormal analysis on the hidden states. The process of abnormal analysis is to calculate the degree of deviation between the hidden state and the fitted distribution at each time point, and then determine whether the health status at that moment deviates from the normal range. In the health monitoring of older pregnant women, some hidden risk signals may appear in slight deviations from the health status, and the Gaussian mixture model can effectively identify these subtle anomalies. When the system detects the anomaly score of the hidden state, it calculates the size of the anomaly score and marks the hidden state whose anomaly score exceeds the set threshold as abnormal. This means that if the health status at a certain point in time exceeds the normal fluctuation range, the system will identify the potential health risk at that time. The accuracy of this anomaly analysis depends on the fitting effect of the Gaussian mixture model on the data distribution. The more accurate the model, the better the effect of the anomaly analysis. This step greatly improves the system's sensitivity to potential risks and can capture slight changes in health status at an early stage, thereby issuing risk signals in advance. In addition, by performing anomaly analysis on the forward and backward hidden states separately, the system can comprehensively consider the historical and future changes in health status. This two-way anomaly analysis method ensures that the system not only relies on historical data to identify risks, but also can discover possible anomalies by predicting future trends. For older pregnant women, changes in health status are often sudden, and historical data alone may not be able to fully capture potential risks. By analyzing both forward and backward anomaly scores simultaneously, the system can more comprehensively assess the degree of abnormality in health status.

[0034] Step 5: Take the forward hidden state distribution as the first prior knowledge, the anomaly score of the forward abnormal hidden state as the first observation data, and use the Bayesian method to calculate the first risk probability; take the backward hidden state distribution as the second prior knowledge, the anomaly score of the backward abnormal hidden state as the second observation data, and use the Bayesian method to calculate the second risk probability; and take the average of the first risk probability and the second risk probability as the final risk probability, set a risk judgment threshold, and issue a risk warning if the final risk probability exceeds the risk judgment threshold.

[0035] The core of the Bayesian method is to obtain a more accurate posterior probability through the fusion of prior knowledge and observation data. In the present invention, the forward hidden state distribution is used as the first prior knowledge, the abnormal score of the forward abnormal hidden state is used as the first observation data, the backward hidden state distribution is used as the second prior knowledge, and the abnormal score of the backward abnormal hidden state is used as the second observation data. Through Bayesian reasoning, the system can calculate the forward and backward risk probabilities respectively. The importance of this step lies in that it can effectively combine the prior health distribution and the current abnormal situation to achieve dynamic quantification of risks. For older pregnant women, the health status of the body may be affected by a variety of complex factors, and the Bayesian method provides an intelligent way to organically combine historical experience and current health data, providing a robust basis for systematic risk assessment.

[0036] After the forward and backward risk probabilities are calculated respectively, the system will average the two to obtain the final comprehensive risk probability. The advantage of this processing method is that it comprehensively considers the two-way changes in the health status of the mother, including both retrospective risk assessment based on historical data and predictive risk assessment based on future health status. This two-way risk assessment method ensures that the system can monitor the health status of the mother in an all-round way. Whether the risk comes from a certain health fluctuation in the past or a potential abnormality that may occur in the future, the system can capture it in advance and make corresponding warnings. In order to ensure the effectiveness and timeliness of risk assessment, the system sets a risk judgment threshold. Once the comprehensive risk probability exceeds the set threshold, the system will immediately issue a risk warning signal to remind relevant medical staff or patients to conduct further examinations or interventions. This risk assessment and warning mechanism based on Bayesian reasoning greatly improves the accuracy and reliability of risk judgment. Especially in the health monitoring of older mothers, due to individual differences and the complexity of health status, a single monitoring method is often difficult to accurately identify potential risks. Through Bayesian reasoning, this method can dynamically adjust the risk assessment of health status, ensuring the timeliness and accuracy of warnings.

[0037] Embodiment 2: The physiological index data at least include: systolic blood pressure, diastolic blood pressure, heart rate, blood sugar, body temperature, blood oxygen saturation, respiratory rate, fetal heart rate, weight, fetal biparietal diameter, fetal femur length, fetal weight estimation, and amniotic fluid index.

[0038] Specifically, systolic and diastolic blood pressure are two main indicators of blood pressure, which can reflect the vascular resistance and cardiac workload of the mother. Hypertension is a common risk factor during pregnancy, especially in older mothers, and may cause serious complications such as pregnancy-induced hypertension syndrome and preeclampsia. Heart rate is a direct manifestation of heart function. By monitoring changes in the mother's heart rate, it can warn of possible problems in the cardiovascular system. Blood sugar is another key physiological indicator. High blood sugar may lead to gestational diabetes, which is more common in older mothers and may have an adverse effect on the health of mothers and babies. Body temperature and blood oxygen saturation are basic indicators that reflect the overall metabolic level and respiratory function of the mother. Fluctuations in body temperature may indicate infection or metabolic abnormalities, while a decrease in blood oxygen saturation may mean a problem with the respiratory system. Respiratory rate is also a direct reflection of the health of the mother's respiratory system. Abnormal respiratory rate may indicate an early sign of lung disease or heart problems. Monitoring of fetal heart rate is particularly important and can directly reflect the health of the fetus in the uterus. Abnormal fetal heart rate may be an early sign of fetal hypoxia, especially in older mothers, who are at a higher risk of fetal hypoxia due to placental dysfunction or other complications. By continuously monitoring the fetal heart rate, the system can identify fetal stress responses and potential health risks. In addition, weight is one of the important indicators in pregnancy management. Too fast or too slow weight gain may mean health problems. The biparietal diameter and femur length of the fetus are important parameters for assessing fetal development and can help predict fetal growth. The estimated fetal weight is a comprehensive result obtained from these measurements, which is used to determine whether the fetus is developmentally delayed or too large, which has an important impact on the management of labor and the choice of delivery method. The amniotic fluid index reflects the amount of amniotic fluid. Abnormal amniotic fluid volume may indicate fetal developmental abnormalities, placental insufficiency or other pregnancy complications.

[0039] Example 3: In step 1, set Indicates the target maternal age at time The Physiological index data, assuming that Different physiological index data; each physiological index data at each time point is standardized: ;in, It's time Time Normalized values ​​of physiological index data; It is The mean of physiological index data; It is The standard deviation of the physiological index data; each time point All standardized physiological index data Combined into a vector, the health index vector is ;in It's time The health index vector at each moment contains all Normalized physiological index data; By transforming the health index vector at each time point Arranged in chronological order, the target maternal health index time series is formed as follows: ; Indicates the total number of time points for data collection, It means that the mother has Time series of health index at each time point; It's time The standardized value of the first physiological index data; It's time The standardized value of the second physiological index data; It's time Time Normalized values ​​of physiological index data; is the health index vector at the first time point; is the health index vector at the second time point; It is A vector of health indices at each time point.

[0040] Specifically, due to the particularity of physiological characteristics, older parturients face a higher risk of pregnancy, and the dynamic changes of multiple key physiological indicators such as blood pressure, heart rate, blood sugar, etc. can effectively reflect their physical condition. However, directly using these raw data for analysis will make it difficult to accurately assess health risks due to the differences in the dimensions and ranges of different indicators. Therefore, the primary task of Example 3 is to standardize each physiological indicator so that they are compared at the same level. This step can eliminate the differences between the data and ensure that the model will not be biased due to the large numerical range of a certain indicator. After standardization, the physiological indicators of all time points are combined into a health index vector, which is a comprehensive representation of all physiological conditions of the parturient at a certain time point. By arranging these health index vectors in chronological order, the system can obtain a time series reflecting the changes in the health of the parturient. The construction of the time series enables the system to dynamically track the evolution of health status, rather than just relying on data at a specific time point. This is particularly important in the risk assessment of older parturients, because their health status may fluctuate significantly over time, and these fluctuations may be early signals of potential health risks. Time series can not only capture dramatic changes in maternal mortality over a short period of time, but also identify long-term trends, thereby providing richer information for risk prediction.

[0041] Embodiment 4: In step 2, the health index vector at each time point is analyzed in the time-frequency domain using wavelet transform, and the wavelet transform coefficients at each time point are extracted as follows: ; Calculate the maximum Lyapunov exponent of the health index vector at each time point as ; The wavelet transform coefficient and the maximum Lyapunov exponent at each time point are weighted and fused by the following formula to obtain the fused feature at each time point : ;in, ; is the wavelet weight, is the Lyapunov weight, and The value range of ; represents wavelet transform, is the scale of wavelet transform; is the displacement of wavelet transform.

[0042] Specifically, wavelet transform, as a time-frequency analysis tool, can decompose the maternal health index vector The wavelet transform can capture complex signals in the image, especially local changes in health status in a short period of time. The health status of older pregnant women often shows non-stationarity, that is, physiological indicators such as heart rate, blood pressure, fetal heart rate, etc. may fluctuate significantly in a short period of time, and these fluctuations are crucial for judging the health risks of pregnant women. and displacement parameters By carefully characterizing these changes, signal components of different frequencies can be captured at different times. For example, at a certain moment, the blood pressure or fetal heart rate of the parturient may undergo a short and sharp change, which may be a precursor to potential risks, and the wavelet transform can effectively extract this short-term fluctuation information, so that the system can respond in time. However, simple local feature analysis is not enough to fully evaluate the health status of the parturient, especially the health problems of older parturients are often caused by long-term systemic changes. Therefore, the calculation of the maximum Lyapunov exponent is introduced in Example 4. The Lyapunov exponent is an indicator of the sensitivity of the system to the initial conditions. Specifically, it can reflect the chaos of the system. For older parturients, the health status of the body may show instability due to age, decreased metabolic function, etc., and even tend to a chaotic state in some cases. By calculating the maximum Lyapunov exponent at each time point, the system can evaluate the global dynamic changes in the health status of the parturient. This dynamic change information can reveal the long-term trend of the health status and help identify those health instability factors that may cause serious complications. For example, long-term fluctuations in blood pressure or abnormal heart rate may indicate potential cardiovascular problems, and the Lyapunov index can provide a quantitative reference for this risk.

[0043] The combination of wavelet transform and maximum Lyapunov exponent does not use these two analysis methods in isolation, but combines their advantages organically through weighted feature fusion. This fusion not only enables the system to capture local changes in health status, but also comprehensively evaluate the global dynamic characteristics of health status. and The weight setting depends on the different analysis requirements of the health status. If the health status of a parturient fluctuates significantly in a short period of time, such as a sharp increase in fetal heart rate or unstable blood pressure, the local features extracted by the wavelet transform will play a more critical role. In other cases, if the health problems of the mother are more reflected in the long-term trend, such as the continuous increase in blood pressure during multiple prenatal examinations or abnormal fetal development, the global dynamic information reflected by the Lyapunov exponent is particularly important. The weight of the system is used to highlight the analysis of the dynamic characteristics of the system. Through this weighted fusion, the system can flexibly adjust the analysis strategy according to the health status of different mothers, ensuring that both short-term drastic changes and long-term trend changes can be accurately captured and analyzed. The generation of provides a more accurate and comprehensive basis for subsequent risk assessment, enabling the system to identify potential health problems in advance and issue early warnings. For example, if the fusion characteristics of a pregnant woman show that her health status has fluctuated sharply locally, and the system dynamics also show high instability, this will significantly increase the confidence of the risk assessment. The system will issue a warning of health risks at an early stage and recommend further medical intervention measures in a timely manner.

[0044] Example 5: In step 3, each time point The fusion feature vector Input into the bidirectional LSTM network; the bidirectional LSTM network contains a forward LSTM network unit and a backward LSTM network unit; the fusion feature vector After inputting into the forward LSTM network unit, the forward hidden state is output ;in, For time The forward hidden state of Represents the forward LSTM network unit; fused feature vector After inputting into the backward LSTM network unit, the backward hidden state is output ; For time The backward hidden state of Represents a backward LSTM network unit.

[0045] Specifically, the design of the LSTM network can solve the gradient vanishing problem encountered by traditional recurrent neural networks (RNNs) when dealing with long-term dependencies. By introducing memory units, the LSTM network can retain important historical information and forget irrelevant information when appropriate. Therefore, when the fusion feature vector When input to the forward LSTM network unit, the forward LSTM is fed into the network according to the past hidden state. and the current input Calculate the forward hidden state at the current time point This hidden state is actually the system's understanding of the health status of the mother at the current time point, combining the current health characteristics and previous historical information. Similarly, the backward LSTM network unit receives the same input , but its calculation is based on the information of subsequent time points. Backward LSTM refers to the features of future time points And hidden state , output the backward hidden state at the current time point This design allows the system to consider not only historical information but also potential future changes. This is of great significance in the health monitoring of older pregnant women, because health status is often affected not only by historical information but also by the potential impact of future health risks.

[0046] The bidirectional LSTM network is designed to capture both forward and backward time dependencies, and it can analyze time series data more comprehensively. The forward LSTM network provides historical background information on health status, while the backward LSTM network provides future prediction references for current health status. The advantage of this bidirectional analysis is that it can help the system better understand the overall picture of the mother's health status, especially when assessing health risks, it can refer to both past historical trends and potential changes in the future. For older mothers, changes in health status are often accompanied by complex time dependencies, which may change significantly before or after a certain point in time, and unidirectional time series models find it difficult to fully capture these changes. In practice, the forward hidden state and the backward hidden state Both represent the system's understanding of the health status of the mother at the current time point, but they start from different time directions, providing the system with a dual perspective of analysis. The forward hidden state reflects how the current health status is affected by the previous time point, while the backward hidden state predicts how the current health status may affect changes at future time points. This forward-looking and backward-looking design greatly improves the system's prediction accuracy and sensitivity, especially in complex health status monitoring, and can capture potential health risk signals in a timely manner. Through this bidirectional processing mechanism of the bidirectional LSTM network, the system can identify more subtle risk signals during the change of health status. For example, the health characteristics at a certain point in time Abnormal fluctuations may be shown in both forward and backward analysis, so the system can make a more accurate risk assessment based on the hidden states in these two directions. This approach not only improves the reliability of health status analysis, but also enhances the sensitivity to health risks, ensuring that early warning signals can be issued.

[0047] Embodiment 6: The process of obtaining the forward abnormal hidden state in step 4 specifically includes: forming the forward hidden states of all time points into a forward hidden state sequence , it is fitted by the first Gaussian mixture model, where the probability density function of the first Gaussian mixture model is It is expressed as: ;in, It is The mixing coefficient of the Gaussian distribution satisfies ; It is The probability density function of a Gaussian distribution with a mean of , the covariance matrix is ; and They are The mean and covariance of a Gaussian distribution; is the number of preset Gaussian distributions; calculate the Mahalanobis distance between each forward hidden state and its corresponding Gaussian distribution: ;in, is the forward hidden state With Gaussian distribution mean The Mahalanobis distance between is the inverse of the covariance matrix; each forward hidden state is assigned a forward anomaly score based on the size of the Mahalanobis distance , the specific formula is: ;in, is the forward hidden state The forward anomaly score of yes The maximum Mahalanobis distance from the center of all Gaussian distributions; The forward hidden state exceeding the set threshold is regarded as the forward abnormal hidden state.

[0048] Specifically, the Gaussian mixture model is a commonly used probability density estimation method that can effectively process complex data with multi-peak characteristics. In the present invention, the health status of older mothers is generated through a bidirectional LSTM network to generate a forward hidden state sequence. These hidden states represent the system's understanding of the health status of the mothers at different time points. However, the complexity of the health status determines that it cannot be accurately described by a single normal distribution. Therefore, the use of a Gaussian mixture model can approximate the actual distribution of the health status of the mother through multiple Gaussian distributions. Each hidden state The probability density of can be expressed as a weighted sum of multiple Gaussian distributions, which means that the system can capture different patterns or trends in health status. For example, the health status may show a pattern at certain moments, but over time, the state may change and show a new pattern. In this way, the Gaussian mixture model is able to decompose the complexity of the hidden state into several Gaussian distributions, each defined by its mean and covariance matrix.

[0049] In order to evaluate the abnormality of each forward hidden state, the present invention uses the statistical tool of Mahalanobis distance. Mahalanobis distance is a standardized method to measure the distance between a data point and the center of the distribution, which can take into account the covariance structure of the data. Unlike Euclidean distance, Mahalanobis distance not only considers the straight-line distance between the data point and the center, but also combines the covariance matrix of the data, so that in data with different variance directions, it can more accurately reflect the degree of deviation of the data point. In the present invention, Mahalanobis distance is used to measure the abnormality of each forward hidden state. With Gaussian distribution center The distance between the hidden state and the center of the Gaussian distribution corresponding to it can be judged by the size of this distance whether the hidden state deviates from the normal health state distribution. If the distance between a hidden state and the center of its corresponding Gaussian distribution is too large, it may mean that the health status of the parturient at that time point has changed abnormally. In order to quantify abnormality, the present invention further introduces the concept of forward anomaly score. By calculating the Mahalanobis distance between each hidden state and the center of all Gaussian distributions, the system can determine the maximum degree of abnormality of the hidden state. Specifically, the anomaly score is the maximum value of all Mahalanobis distances, which means that the system not only considers the deviation of a single Gaussian distribution, but also comprehensively evaluates the performance of the hidden state under multiple distributions. In this way, the system can more comprehensively identify abnormalities in the health state. For example, if a hidden state shows obvious deviation under multiple Gaussian distributions, it means that the state may be extremely abnormal and needs to attract the system's high attention. On the contrary, if the deviation is large only under a certain distribution, but is relatively normal in other distributions, the degree of abnormality of the hidden state may be relatively low.

[0050] This anomaly detection method based on Gaussian mixture model and Mahalanobis distance greatly improves the system's ability to identify health risks. In the health monitoring of older pregnant women, the changes in health status are often complex and difficult to predict, and a single time point or a single analysis method may not be able to capture all potential risks. Through the comprehensive analysis of multi-distribution fitting and Mahalanobis distance, the system can better cope with this complexity and promptly detect abnormal fluctuations in health status. Especially in the group of older pregnant women, some health problems may not appear in a short period of time, but gradually manifest themselves through long-term cumulative effects. For example, high blood pressure or abnormal blood sugar may only fluctuate slightly in the early stage, but over time, if these states continue to be abnormal, they may lead to serious complications. The combination of Gaussian mixture model and Mahalanobis distance enables the system to capture this potential risk in advance. After the system calculates the anomaly score of each hidden state, it can mark those states whose anomaly scores exceed the threshold as abnormal states according to the set threshold. The setting of this threshold is based on the assessment of maternal health risks. For example, when the health status of a pregnant woman begins to fluctuate unusually, the system will first capture this signal through the change of the anomaly score. If the anomaly score exceeds the preset threshold, the system will consider that the health status of the mother has deviated from the normal range and there may be potential health problems. For example, if the anomaly score of a hidden state exceeds the threshold, it may mean that the mother's key physiological indicators such as heart rate and blood pressure have changed significantly at that time, prompting the system to issue an early warning and recommend further medical intervention. In this way, the system can monitor and evaluate the health status of the mother in real time, capture potential health risks at an early stage, and provide timely early warnings. The core of this process is to fit the complex health status through multiple distributions using a Gaussian mixture model, and combine the Mahalanobis distance and anomaly score to quantify the anomaly, thereby providing the system with powerful anomaly detection capabilities. Especially in the special group of older mothers, health problems are often cumulative and hidden, and this method can effectively capture early signals and issue early warnings, thereby providing more comprehensive and intelligent support for maternal health management.

[0051] Embodiment 7: The process of obtaining the backward abnormal hidden state in step 4 specifically includes: forming the backward hidden states at all time points into a backward hidden state sequence , and fit it through the second Gaussian mixture model, where the probability density function of the second Gaussian mixture model is It is expressed as: ; among them, among them, It is The mixing coefficient of the Gaussian distribution satisfies ; It is The probability density function of a Gaussian distribution with a mean of , the covariance matrix is ; and They are The mean and covariance of a Gaussian distribution; is the number of preset Gaussian distributions; calculate the Mahalanobis distance between each backward hidden state and its corresponding Gaussian distribution: ;in, is the backward hidden state With Gaussian distribution mean The Mahalanobis distance between is the inverse matrix of the covariance matrix; each backward hidden state is assigned a backward anomaly score based on the size of the Mahalanobis distance , the specific formula is: ;in, is the backward hidden state The backward anomaly score of yes The maximum Mahalanobis distance from the center of all Gaussian distributions; The backward hidden state exceeding the set threshold is regarded as the backward abnormal hidden state.

[0052] Specifically, the process of determining the backward abnormal hidden state is similar to the process of determining the forward abnormal hidden state.

[0053] Embodiment 8: In step 5, the first risk probability is calculated by the following formula: ;in, is the first risk probability, which represents the anomaly score given the forward anomaly hidden state The mother is at risk The probability of is the first likelihood function, indicating that the mother is in a risk state When , the anomaly score of the forward anomaly hidden state The probability of is the prior probability, which indicates the prior knowledge that the mother is in a risky state; is the first marginal likelihood, representing the anomaly score of the forward hidden state The overall probability of the second risk is calculated by the following formula: ;in, is the second risk probability, representing the anomaly score given the backward anomaly hidden state The mother is at risk The probability of is the second likelihood function, indicating that the mother is in a risk state When , the anomaly score of the backward anomaly hidden state The probability of is the prior probability, which indicates the prior knowledge that the mother is in a risky state; is the second marginal likelihood, representing the abnormality score of the backward hidden state The overall probability ; The final risk probability is calculated using the following formula: .

[0054] Specifically, when calculating the first risk probability, the formula Represents the abnormality score of a given forward hidden state Here, the role of Bayesian theorem is to adjust the prediction of risk status by combining the observed value of the abnormal score with the existing prior knowledge. This method can update the judgment of the system through real-time data, so that the evaluation results take into account historical risk characteristics and adapt to new health information. In Bayesian reasoning, As a likelihood function, it means that the mother is in a risk state When , the anomaly score of the forward hidden state The likelihood function plays a crucial role in the present invention because it quantifies the correlation between the anomaly score and the risk state. If the anomaly score of the forward hidden state is high, it means that the system detects a deviation from the healthy state, and the likelihood function helps the system identify the possibility that the mother is at health risk by modeling the probability of these deviations. In other words, when the observation results of the forward anomaly score are significant, the system assigns more weight to the risk state through the Bayesian formula, thereby improving the prediction accuracy of the risk probability.

[0055] Another key factor is the prior probability , which reflects the initial assessment that the mother is at risk when there is no new observation data. In the health monitoring of older mothers, the prior probability may be determined based on previous data or medical knowledge, such as risk factors such as hypertension and diabetes in older mothers. This prior information provides a basic framework for risk assessment, especially when new data is insufficient or the abnormal score is not obvious, the system can still make reasonable judgments based on prior knowledge. In order to rationalize and standardize the evaluation results, the system uses marginal likelihood Normalize the calculation results. Marginal likelihood represents the overall probability of abnormal scores of the forward hidden state in all possible health states. Its introduction ensures that when the system calculates the risk probability, it will not produce an overly high or undervalued risk assessment due to the deviation of the observed data. The normalization process puts the results within a reasonable probability range, making the system more robust and stable.

[0056] Similar to the forward hidden state, the system also calculates the risk probability of the backward hidden state through Bayes’ theorem, that is, the second risk probability This process is the same as the forward risk probability calculation principle, but it is based on the anomaly score of the backward hidden state In the scenario of health monitoring for older mothers, the analysis of the backward hidden state provides predictive information about future health changes, while the forward hidden state reflects historical data. By calculating the risk probabilities from these two different perspectives, the system can more comprehensively capture the changing trends of maternal health conditions. In the framework of the Bayesian formula, the risk probability of the backward hidden state is also determined by the likelihood function, prior probability, and marginal likelihood. Likelihood function The probability of the backward anomaly score appearing when there is a health risk is quantified. This function provides the system with a basis for judging future health status. If the backward anomaly score deviates significantly from the normal range, the system will adjust the prediction of future health risks according to the likelihood function, thereby increasing the system's sensitivity to sudden changes in health status. Prior probability and marginal likelihood It also plays a key role in the calculation of the backward risk probability. Prior probability provides the basic knowledge for risk assessment, and marginal likelihood ensures the rationality and consistency of probability calculation. The advantage of this framework is that when the observed data and anomaly scores change, the system can flexibly adjust the risk assessment so that it depends on new observed data while maintaining continuous monitoring of the risk status. Finally, by averaging the forward and backward risk probabilities, the system obtains the final risk probability . The design of this averaging processing method reflects the core concept of the two-way time series analysis in the present invention. By combining the risk probabilities of the forward and backward hidden states, the system can comprehensively evaluate the health status of the mother. The forward risk probability reflects the impact of the mother's historical health data on the current risk, while the backward risk probability provides the system with predictive information on future health trends. Through this dual analysis, the system can not only detect current health risks, but also predict potential future risks, thereby providing more comprehensive early warnings in the health management of older mothers.

[0057] Embodiment 9: In step 5, the first likelihood function is expressed by the following formula:

[0058] ;in, is the number of forward anomaly hidden states; is the subscript index; is the Gaussian kernel function; the second likelihood function is expressed by the following formula: ;in, is the number of backward anomaly hidden states; the first marginal likelihood Use the following formula to calculate: ;in, is the prior probability that the mother is not at risk; is the distribution of the forward hidden state anomaly scores in the non-risk state; Indicates not being at risk; is the number of forward anomaly hidden states; the second marginal likelihood Use the following formula to calculate: ;in, is the distribution of the backward hidden state anomaly scores in the non-risk state.

[0059] Specifically, the first likelihood function The mother is at risk of health When , the anomaly score of the forward anomaly hidden state The probability of occurrence. This likelihood function uses the Gaussian kernel function The purpose of the estimation is to smooth the probability distribution between anomaly scores and risk states through the kernel function. The formula is expressed as a typical form in Gaussian kernel density estimation, which is based on averaging from a set of anomaly scores and weighting the differences between anomaly scores using a kernel function. Through the kernel function, the system can capture the distribution pattern of the hidden state of the anomaly, so as to more accurately calculate the probability of the occurrence of anomaly scores under a given risk state. Specifically, the kernel function is a smooth function, usually using the Gaussian kernel function, because it can smooth between data points and assign different weights to different anomaly scores. The weight here is determined by the distance between the anomaly scores. The smaller the distance, the greater the weight assigned to the anomaly score, and vice versa. Therefore, the formula uses kernel density estimation to make probability inferences on the entire anomaly score distribution, and derives the probability of a specific anomaly score when the mother is in a risk state.

[0060] Similarly, the second likelihood function The anomaly score of the backward anomaly hidden state is estimated in the same way as the forward likelihood function. The backward anomaly hidden state represents the system's prediction of future health trends, and the distribution of the backward anomaly score is also estimated using a Gaussian kernel function to capture the health risks that may occur to the mother in the future. In order to calculate the marginal likelihood , the formula introduces a weighted combination of risk states and non-risk states. The marginal likelihood represents the probability of the forward abnormal score appearing in all possible health states, so it not only considers the distribution of abnormal scores when the mother is in a risk state, but also comprehensively considers the distribution of abnormal scores in a non-risk state. In the formula, the marginal likelihood is calculated by the weighted sum of two parts: one is based on the probability that the mother is in a risk state , and the other is based on the probability that the mother is not at risk This enables the system to calculate risk probabilities not only by relying on observed anomaly scores, but also by incorporating prior probabilities of different health states, thereby providing a more comprehensive assessment of health risks. represents the distribution of anomaly scores in the forward hidden state when the mother is not in a health risk state. This distribution reflects the usual range and probability of occurrence of forward anomaly scores in normal situations without health problems. Combining the distribution of anomaly scores in risk and non-risk states, the system is able to provide a more accurate context for the calculation of health risk probabilities. This not only allows the system to capture potential risk signals behind anomaly scores, but also helps it avoid over-predicting risks due to a single abnormal data point, thereby improving the robustness of the assessment. Similarly, the marginal likelihood of the backward anomaly score It is calculated in a similar way. It also combines the probability estimates of when the mother is in a risk state and a non-risk state, and weights the distribution of abnormal scores in these two states. The backward abnormal score distribution can provide predictive information about future health status, especially in trend changes in health status, this distribution helps to capture potential risk signals. By combining the marginal likelihood of forward and backward anomaly scores, the system can simultaneously evaluate historical and future health states and generate a comprehensive risk probability.

[0061] Embodiment 10: A risk assessment and early warning system for older pregnant women, the system comprising: a data collection part, used to collect physiological indicator data of target pregnant women at each time point, and after data standardization of each physiological indicator data, as an element, construct a health index vector at each time point; the health index vector at each time point is used as an element to constitute the health index time series of the target pregnant woman; a feature extraction and fusion part, used to use wavelet transform to perform time-frequency domain analysis on the health index vector at each time point, and extract the wavelet transform coefficient at each time point; calculate the maximum Lyapunov exponent of the health index vector at each time point; perform weighted feature fusion on the wavelet transform coefficient and the maximum Lyapunov exponent at each time point to obtain the fused feature at each time point; a bidirectional LSTM network part, used to input the fused feature at each time point as input into a bidirectional LSTM network, and the bidirectional LSTM network outputs the forward hidden state and backward hidden state corresponding to each time point; risk warning distribution, used to analyze the forward hidden state of all time points The state is fitted by a Gaussian mixture model to obtain the corresponding forward hidden state distribution, the forward hidden state distribution is analyzed for anomalies, the anomaly score of the forward hidden state is calculated, and the forward hidden state with an anomaly score exceeding the set threshold is taken as the forward abnormal hidden state; the backward hidden state at all time points is fitted by a Gaussian mixture model to obtain the corresponding backward hidden state distribution, the backward hidden state distribution is analyzed for anomalies, the anomaly score of the backward hidden state is calculated, and the backward hidden state with an anomaly score exceeding the set threshold is taken as the backward abnormal hidden state; the forward hidden state distribution is taken as the first prior knowledge, the anomaly score of the forward abnormal hidden state is taken as the first observation data, and the first risk probability is calculated using the Bayesian method; the backward hidden state distribution is taken as the second prior knowledge, the anomaly score of the backward abnormal hidden state is taken as the second observation data, and the second risk probability is calculated using the Bayesian method; and the average of the first risk probability and the second risk probability is taken as the final risk probability, a risk judgment threshold is set, and if the final risk probability exceeds the risk judgment threshold, a risk warning is issued.

[0062] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A risk assessment and early warning method for older pregnant women, characterized in that: The method comprises: Step 1: Collect the physiological index data of the target parturient at each time point, standardize each physiological index data, and use it as an element to construct the health index vector at each time point; use the health index vector at each time point as an element to construct the health index time series of the target parturient; Step 2: Use wavelet transform to perform time-frequency domain analysis on the health index vector at each time point, extract the wavelet transform coefficient at each time point; calculate the maximum Lyapunov exponent of the health index vector at each time point; perform weighted feature fusion on the wavelet transform coefficient and the maximum Lyapunov exponent at each time point to obtain the fused feature at each time point; Step 3: The fused features at each time point are used as input into a bidirectional LSTM network, which outputs the forward hidden state and backward hidden state corresponding to each time point; Step 4: Fit the forward hidden state of all time points through the first Gaussian mixture model to obtain the corresponding forward hidden state distribution, perform anomaly analysis on the forward hidden state distribution, calculate the anomaly score of the forward hidden state, and take the forward hidden state whose anomaly score exceeds the set threshold as the forward abnormal hidden state; fit the backward hidden state of all time points through the second Gaussian mixture model to obtain the corresponding backward hidden state distribution, perform anomaly analysis on the backward hidden state distribution, calculate the anomaly score of the backward hidden state, and take the backward hidden state whose anomaly score exceeds the set threshold as the backward abnormal hidden state; Step 5: Take the forward hidden state distribution as the first prior knowledge, take the anomaly score of the forward abnormal hidden state as the first observation data, and use the Bayesian method to calculate the first risk probability; take the backward hidden state distribution as the second prior knowledge, take the anomaly score of the backward abnormal hidden state as the second observation data, and use the Bayesian method to calculate the second risk probability; and take the mean of the first risk probability and the second risk probability as the final risk probability, set a risk judgment threshold, and issue a risk warning if the final risk probability exceeds the risk judgment threshold; The process of obtaining the forward abnormal hidden state in step 4 specifically includes: forming a forward hidden state sequence from the forward hidden states of all time points ; Indicates the total number of time points for data collection; is time; it is fitted by the first Gaussian mixture model, where the probability density function of the first Gaussian mixture model is It is expressed as: ； in, It is The mixing coefficient of the Gaussian distribution satisfies ; It is The probability density function of a Gaussian distribution with a mean of , the covariance matrix is ; and They are The mean and covariance of a Gaussian distribution; is the number of preset Gaussian distributions; calculate the Mahalanobis distance between each forward hidden state and its corresponding Gaussian distribution: ； in, is the forward hidden state With Gaussian distribution mean The Mahalanobis distance between is the inverse of the covariance matrix; each forward hidden state is assigned a forward anomaly score based on the size of the Mahalanobis distance , the specific formula is: ； in, is the forward hidden state The forward anomaly score of yes The maximum Mahalanobis distance from the center of all Gaussian distributions; The forward hidden state that exceeds the set threshold is regarded as the forward abnormal hidden state; The process of obtaining the backward abnormal hidden state in step 4 specifically includes: forming a backward hidden state sequence of all time points into a backward hidden state sequence , and fit it through the second Gaussian mixture model, where the probability density function of the second Gaussian mixture model is It is expressed as: ； in, It is The mixing coefficient of the Gaussian distribution satisfies ; It is The probability density function of a Gaussian distribution with a mean of , the covariance matrix is ; and They are The mean and covariance of a Gaussian distribution; is the number of preset Gaussian distributions; calculate the Mahalanobis distance between each backward hidden state and its corresponding Gaussian distribution: ； in, is the backward hidden state With Gaussian distribution mean The Mahalanobis distance between is the inverse matrix of the covariance matrix; each backward hidden state is assigned a backward anomaly score based on the size of the Mahalanobis distance , the specific formula is: ; in, is the backward hidden state The backward anomaly score of yes The maximum Mahalanobis distance from the center of all Gaussian distributions; The backward hidden state exceeding the set threshold is regarded as the backward abnormal hidden state.

2. The risk assessment and early warning method for older pregnant women according to claim 1, characterized in that: The physiological index data at least include: systolic blood pressure, diastolic blood pressure, heart rate, blood sugar, body temperature, blood oxygen saturation, respiratory rate, fetal heart rate, weight, fetal biparietal diameter, fetal femur length, fetal weight estimation, and amniotic fluid index.

3. The risk assessment and early warning method for older pregnant women according to claim 2, characterized in that: In step 1, assume Indicates the target maternal age at time The Physiological index data, assuming that Different physiological index data; each physiological index data at each time point is standardized: ;in, It's time Time Normalized values ​​of physiological index data; It is The mean of physiological index data; It is The standard deviation of the physiological index data; each time point All standardized physiological index data Combined into a vector, the health index vector is ;in It's time The health index vector at each moment contains all Normalized physiological index data; By transforming the health index vector at each time point Arranged in chronological order, the target maternal health index time series is formed as follows: ; Indicates the total number of time points for data collection, It means that the mother has Time series of health index at each time point; It's time The standardized value of the first physiological index data; It's time The standardized value of the second physiological index data; It's time Time Normalized values ​​of physiological index data; is the health index vector at the first time point; is the health index vector at the second time point; It is A vector of health indices at each time point.

4. The risk assessment and early warning method for older pregnant women according to claim 3, characterized in that: In step 2, wavelet transform is used to perform time-frequency domain analysis on the health index vector at each time point, and the wavelet transform coefficients at each time point are extracted as follows: ; Calculate the maximum Lyapunov exponent of the health index vector at each time point as ; The wavelet transform coefficient and the maximum Lyapunov exponent at each time point are weighted and fused by the following formula to obtain the fused feature at each time point : ； in, ; is the wavelet weight, is the Lyapunov weight, and The value range of ; represents wavelet transform, is the scale of wavelet transform; is the displacement of wavelet transform.

5. The risk assessment and early warning method for older pregnant women according to claim 4, characterized in that: In step 3, each time point The fusion feature vector Input into the bidirectional LSTM network; the bidirectional LSTM network contains a forward LSTM network unit and a backward LSTM network unit; the fusion feature vector After inputting into the forward LSTM network unit, the forward hidden state is output ;in, For time The forward hidden state of Represents the forward LSTM network unit; fused feature vector After inputting into the backward LSTM network unit, the backward hidden state is output ; For time The backward hidden state of Represents a backward LSTM network unit.

6. The risk assessment and early warning method for older pregnant women according to claim 5, characterized in that: In step 5, the first risk probability is calculated by the following formula: ； in, is the first risk probability, which represents the anomaly score given the forward anomaly hidden state The mother is at risk probability; is the first likelihood function, indicating that the mother is in a risk state When , the anomaly score of the forward anomaly hidden state probability; is the prior probability, which indicates the prior knowledge that the mother is in a risky state; is the first marginal likelihood, representing the anomaly score of the forward hidden state The overall probability of the second risk is calculated by the following formula: ； in, is the second risk probability, representing the anomaly score given the backward anomaly hidden state The mother is at risk probability; is the second likelihood function, indicating that the mother is in a risk state When , the anomaly score of the backward anomaly hidden state probability; is the prior probability, which indicates the prior knowledge that the mother is in a risky state; is the second marginal likelihood, representing the abnormality score of the backward hidden state The overall probability of the final risk Calculated using the following formula: 。 7. The risk assessment and early warning method for older pregnant women according to claim 6, characterized in that: In step 5, the first likelihood function is expressed by the following formula: ； in, is the number of forward anomaly hidden states; is the subscript index; is the Gaussian kernel function; the second likelihood function is expressed by the following formula: ； in, is the number of backward anomaly hidden states; the first marginal likelihood Use the following formula to calculate: ;in, is the prior probability that the mother is not at risk; is the distribution of the forward hidden state anomaly scores in the non-risk state; Indicates not being at risk; is the number of forward anomaly hidden states; the second marginal likelihood Use the following formula to calculate: ； in, is the distribution of the backward hidden state anomaly scores in the non-risk state.

8. A risk assessment and early warning system for older pregnant women, used to implement the risk assessment and early warning method for older pregnant women as described in any one of claims 1 to 7, characterized in that: The system includes: a data collection part, which is used to collect physiological index data of the target parturient at each time point, and after data standardization of each physiological index data, the data is used as an element to construct a health index vector at each time point; the health index vector at each time point is used as an element to form a health index time series of the target parturient; a feature extraction and fusion part, which is used to use wavelet transform to perform time-frequency domain analysis on the health index vector at each time point, and extract the wavelet transform coefficient at each time point; calculate the maximum Lyapunov exponent of the health index vector at each time point; perform weighted feature fusion on the wavelet transform coefficient and the maximum Lyapunov exponent at each time point to obtain the fused feature at each time point; a bidirectional LSTM network part, which is used to input the fused feature at each time point as input into a bidirectional LSTM network, and the bidirectional LSTM network outputs the forward hidden state and backward hidden state corresponding to each time point; a risk warning distribution part, which is used to perform forward hidden state analysis on all time points through the first Gaussian mixture model Fitting is performed to obtain the corresponding forward hidden state distribution, anomaly analysis is performed on the forward hidden state distribution, anomaly scores of the forward hidden state are calculated, and the forward hidden state with anomaly scores exceeding the set threshold is taken as the forward abnormal hidden state; the backward hidden state of all time points is fitted through the second Gaussian mixture model to obtain the corresponding backward hidden state distribution, anomaly analysis is performed on the backward hidden state distribution, anomaly scores of the backward hidden state are calculated, and the backward hidden state with anomaly scores exceeding the set threshold is taken as the backward abnormal hidden state; the forward hidden state distribution is taken as the first prior knowledge, the anomaly score of the forward abnormal hidden state is taken as the first observation data, and the first risk probability is calculated using the Bayesian method; the backward hidden state distribution is taken as the second prior knowledge, the anomaly score of the backward abnormal hidden state is taken as the second observation data, and the second risk probability is calculated using the Bayesian method; and the mean of the first risk probability and the second risk probability is taken as the final risk probability, a risk judgment threshold is set, and if the final risk probability exceeds the risk judgment threshold, a risk warning is issued.

Images