System and method for monitoring and managing abnormal conditions and health information during pregnancy

Abstract

This invention discloses a management system and method for monitoring and managing abnormal states and health information during pregnancy and childbirth, belonging to the field of medical monitoring technology. The system includes: collecting monitoring data and fusing it with hospital time-series medical records using a timestamp synchronization algorithm to form a user health information profile; inputting an adaptive pregnancy and childbirth risk model and outputting abnormality judgment results and health indices with clinical interpretation labels; implementing a four-level tiered early warning system (from low to critical) based on the results; optimizing intervention plans through digital health twin simulation; generating monitoring parameter adjustment instructions and physical intervention equipment plans; and combining real-time data with prediction results, employing an extended Kalman filter dynamic calibration model and strategy. This invention achieves accurate identification, tiered early warning, and personalized intervention for pregnancy and childbirth risks, improving the intelligence, reliability, and timeliness of monitoring and management, and effectively protecting maternal and infant health.

Description

Technical Field

[0001] This invention belongs to the field of medical monitoring technology, specifically a management system and method for monitoring and managing abnormal conditions and health information during pregnancy and childbirth. Background Technology

[0002] Pregnancy and childbirth are critical stages for maternal and infant health. Real-time monitoring and early warning of abnormal physiological parameters such as uterine contractions, fetal heart rate variability, dynamic blood glucose, and postpartum hemorrhage are essential for reducing maternal and infant health risks. However, existing methods for monitoring pregnancy and childbirth have many shortcomings: First, they often rely on single devices to collect data, resulting in incomplete data dimensions and temporal asynchrony issues among multiple data sources, affecting the comprehensiveness and accuracy of monitoring. Second, data transmission lacks secure encryption mechanisms, making it prone to privacy leaks or data tampering, posing data security risks. Third, traditional risk assessment models are mostly static, making it difficult to dynamically adapt to individual differences and real-time changes in maternal and infant physiological states, and the interpretability of abnormality judgments is insufficient, making it impossible to accurately pinpoint the root cause of risks. Fourth, intervention plans are mostly based on clinical experience, lacking simulation optimization of the expected effects of different plans, and lacking dynamic calibration mechanisms, making it difficult to adjust monitoring parameters and intervention strategies based on real-time physiological data, resulting in poor timeliness and accuracy of risk management. Therefore, there is an urgent need for a monitoring and management method that integrates multimodal data collection, secure fusion, intelligent risk assessment, personalized intervention, and dynamic calibration to improve the intelligence and reliability of pregnancy and childbirth health protection. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a monitoring and management system and method for abnormal states and health information during pregnancy and childbirth. The system collects monitoring data and integrates it with hospital time-series medical records using a timestamp synchronization algorithm to create a user health profile. After inputting into an adaptive pregnancy and childbirth risk model, it outputs abnormality assessment results and health indices with clinical interpretation labels. Based on the results, it implements a four-level tiered early warning system from low to critical levels, optimizes intervention plans through digital health twin simulation, and generates monitoring parameter adjustment instructions and physical intervention equipment plans. Combining real-time data with predicted results, it employs an extended Kalman filter dynamic calibration model and strategy. This invention achieves accurate identification, tiered early warning, and personalized intervention for pregnancy and childbirth risks, improving the intelligence, reliability, and timeliness of monitoring and management, and effectively safeguarding maternal and infant health.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] Methods for managing and monitoring abnormal conditions and health information during pregnancy and childbirth include:

[0006] S1: Synchronously collect user monitoring data through a multimodal wearable monitoring device group;

[0007] S2: The monitoring data is fused with time-series medical record data in the hospital information system using a timestamp synchronization algorithm to form a user health information profile. This profile is then input into an adaptive pregnancy and childbirth risk model for analysis, outputting a personal health risk assessment result. The personal health risk assessment result includes a comprehensive health index score and anomaly judgment results with clinical interpretation labels. The adaptive pregnancy and childbirth risk model is dynamically updated based on the fusion features. The clinical interpretation labels are used to indicate to clinicians the main physiological parameters and their temporal context that cause changes or abnormalities in the health index score.

[0008] S3: Based on individual health risk assessment results and clinical interpretation labels, hierarchical early warning is generated, and early warning instructions are generated. At the same time, the early warning instructions and user health information profiles are input into a pre-built digital health twin and the expected effects of different intervention programs are simulated to generate a comprehensive management strategy that includes instructions for adjusting monitoring parameters and intervention equipment parameter plans.

[0009] S4: Send the monitoring parameter adjustment command to the wearable monitoring device group, preload the intervention device parameter plan into the physical intervention device, compare the newly collected real-time monitoring data with the prediction results of the digital health twin, and recalculate the health index based on the new data. If the actual value of the health index deviates from the predicted value, dynamically calibrate the status parameters and comprehensive management strategy of the digital health twin.

[0010] Specifically, the timestamp synchronization algorithm in S2 is as follows:

[0011] The timestamps of each data channel are parsed from the monitoring data;

[0012] Extract time-series medical record data with system timestamps from the parent medical record database of the hospital information system; the time-series medical record data includes prenatal check-up records, medication records, laboratory test results, and imaging reports;

[0013] The timestamps of the monitoring data and the system timestamps of the time-series medical records are aligned using a dynamic time warping algorithm to obtain the timestamp-aligned monitoring values ​​and medical record text descriptions. The timestamp-aligned monitoring values ​​and medical record text descriptions are then mapped to feature vectors of a unified dimension through an embedding layer and concatenated along the feature dimension to form a user health information profile.

[0014] Specifically, the adaptive pregnancy and childbirth risk model includes a feature encoder, a spatiotemporal attention network, a health index calculation module, and a risk classifier connected in sequence:

[0015] The feature encoder is composed of a cascaded one-dimensional convolutional neural network and a long short-term memory network. The input of the one-dimensional convolutional neural network receives the user's health information profile and extracts local temporal features through its convolutional and pooling layers. Its output is connected to the input of the long short-term memory network to capture long-term dependencies and output the encoded feature sequence.

[0016] The input of the spatiotemporal attention network is connected to the output of the feature encoder, and it contains a temporal attention module and a feature attention module. The temporal attention module is used to calculate the importance weights of different time steps in the encoded feature sequence. The feature attention module is used to calculate the importance weights of different physiological feature dimensions at the same time step. After weighted fusion, a context-enhanced feature representation is generated.

[0017] The input of the health index calculation module is connected to the output of the spatiotemporal attention network. It consists of a fully connected layer and a sigmoid activation function, and is used to map the context-enhanced feature representation into a scalar value between 0 and 1, which serves as the health index score.

[0018] The input of the risk classifier is connected to the output of the health index calculation module. It consists of a fully connected layer and a Softmax function. It is used to receive feature representations associated with the health index score and output anomaly judgment results with clinical interpretation labels. The clinical interpretation labels are used to indicate the main feature dimensions that cause the anomaly.

[0019] Specifically, the hierarchical early warning in S3 is as follows:

[0020] Establish a mapping relationship between health index score ranges and abnormal judgment results and preset risk level thresholds; the risk levels include low risk, medium risk, high risk and critical risk;

[0021] Based on the health index score and the clinical interpretation label, the current risk level and main abnormal characteristics are determined;

[0022] If the risk level is determined to be low, a daily follow-up reminder instruction will be generated; if the risk level is determined to be medium, an enhanced monitoring and outpatient follow-up warning instruction will be generated; if the risk level is determined to be high, an inpatient observation and special examination warning instruction will be generated; if the risk level is determined to be critical, an immediate medical intervention warning instruction will be generated.

[0023] The warning instructions and their corresponding risk levels, health index scores, anomaly judgment results, clinical interpretation labels, and user health information profiles are all input into the digital health twin.

[0024] Specifically, the construction and simulation process of the digital health twin is as follows:

[0025] Based on the maternal medical records and historical monitoring data, a maternal-fetal physiological system mechanism model described by a set of coupled differential equations is constructed. Upon receiving a clinical warning instruction, the model simulates and executes various intervention programs that conform to clinical guidelines in the digital space, using the current patient health information profile as the initial state, and predicts the evolution trajectory of health indices under each intervention program. The intervention programs include adjusting the maternal position, adjusting the medication regimen, adjusting the fluid resuscitation rate, and preparing for surgical intervention.

[0026] By comparing the evolution trajectory and clinical safety thresholds of each intervention program, the optimal intervention program that enables the health index to recover to the safe range as quickly as possible or maintain the most stable state is selected, and a comprehensive management strategy is generated that includes the monitoring parameter adjustment instructions and intervention equipment parameter contingency plans corresponding to the optimal intervention program.

[0027] Specifically, the process of generating the monitoring parameter adjustment instruction is as follows:

[0028] The evolution trajectory corresponding to the optimal intervention plan is analyzed to identify the monitoring target values ​​that need to be adjusted to maintain physiological parameters within a safe range;

[0029] The monitored target value is converted into configuration parameters for specific sensors in the multimodal wearable monitoring device group, and the configuration parameters are encapsulated into monitoring parameter adjustment instructions; the configuration parameters include sampling frequency, alarm threshold, and data upload interval.

[0030] Specifically, the process of generating and preloading the intervention device parameter plan is as follows:

[0031] The optimal intervention plan is analyzed to identify the required physical intervention equipment and its operating parameters; the physical intervention equipment includes an infusion pump, a uterine contraction inhibitor delivery device, a fetal monitor, and a postpartum uterine compression device;

[0032] Based on the prediction results of the maternal-fetal physiological system mechanism model, the activation sequence, dosage, flow rate or pressure parameters of the physical intervention equipment are calculated to form a structured parameter plan;

[0033] The parameter plan is preloaded into the corresponding physical intervention device via the medical device communication protocol, so that the physical intervention device is in standby mode.

[0034] Specifically, the dynamic calibration process in S4 is as follows:

[0035] After implementing the comprehensive management strategy, real-time monitoring data is collected through a multimodal wearable monitoring device group;

[0036] The real-time monitoring data is input into the adaptive pregnancy and childbirth risk model to recalculate the actual health index score. The actual health index score is then compared with the health index score predicted by the digital health twin at the same time to calculate the deviation value of the health index.

[0037] If the deviation value of the health index continuously exceeds the preset dynamic calibration threshold and reaches the set time, the calibration process is triggered; the calibration process includes: using real-time monitoring data as the observation value, and using the extended Kalman filter algorithm to update the state parameters of the maternal-fetal physiological system mechanism model in the digital health twin in reverse.

[0038] Based on the updated status parameters, the maternal-fetal physiological system mechanism model was rerun to correct the evolution trajectory of health indices and key physiological parameters, and the monitoring parameter adjustment instructions and intervention equipment parameter contingency plans in the comprehensive management strategy were adjusted.

[0039] Specifically, the generation of the clinical interpretability labels also employs a model interpretability method based on sapride and interpretation, specifically:

[0040] For the anomaly determination results output by the risk classifier, select the top N feature dimensions with the highest contribution and their corresponding key time steps;

[0041] Calculate the Shapley value of each feature in the user health information profile at the key time step; the Shapley value is used to quantify the contribution of the user health information profile to the anomaly detection result;

[0042] The user health information profiles are sorted from highest to lowest according to the Shapley value, and the top N features and their corresponding time steps are marked as the core content of the clinical interpretation label.

[0043] The system for monitoring and managing abnormal conditions and health information during pregnancy and childbirth includes a transmission and fusion module, a risk assessment module, a hierarchical early warning module, a digital health twin module, a dynamic calibration module, and a device linkage module.

[0044] The transmission and fusion module uses a timestamp synchronization algorithm to align monitoring data with hospital time-series medical record data, and fuses them to form a complete user health information profile.

[0045] The risk assessment module is used to analyze the health information profile and output the health index score and the abnormal judgment result with clinical interpretation label;

[0046] The hierarchical early warning module is used to establish a four-level risk mapping relationship of low, medium, high and critical, generate corresponding early warning instructions, and synchronize them to the digital health twin.

[0047] The digital health twin module is used to simulate the expected effects of various intervention programs, screen the optimal program, and generate instructions for adjusting monitoring parameters and intervention equipment parameter plans, providing personalized comprehensive management strategies.

[0048] The dynamic calibration module is used to compare real-time monitoring data with twin prediction results. When the deviation exceeds the threshold, the model parameters are calibrated by the extended Kalman filter algorithm to adjust the comprehensive management strategy.

[0049] The device linkage module is used to receive intervention device parameter plans and preload them into the physical intervention device, so that it is in a standby state.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] 1. Comprehensive and secure data acquisition and transmission lay a solid foundation for prenatal and postnatal monitoring. A multimodal wearable monitoring device suite simultaneously collects key physiological data from both the fetus and mother, including uterine contraction gradient, fetal heart rate variability, and dynamic blood glucose levels. Kalman filtering and wavelet transform algorithms optimize data processing, ensuring data accuracy. Data transmission utilizes a blockchain-encrypted channel, employing mechanisms such as asymmetric encryption, distributed notarization, and hash verification to effectively prevent privacy leaks and data tampering, ensuring data integrity and authenticity.

[0052] 2. Significantly improved accuracy and interpretability of risk identification, facilitating precise identification of the root causes of abnormalities. The adaptive perinatal risk model integrates a one-dimensional convolutional neural network, a long short-term memory network, and a spatiotemporal attention network, enabling it to extract both local temporal features and capture long-term dependencies, thus accurately identifying perinatal abnormalities. Combined with Shapleyga and interpretation methods, it generates clinical interpretive labels, clarifying the core feature dimensions and key time steps of abnormalities. Simultaneously, the model dynamically iterates and updates through an online gradient descent algorithm, continuously adapting to maternal and infant physiological differences, making risk assessment more targeted.

[0053] 3. Optimize tiered early warning and personalized intervention plans to achieve scientific and efficient risk management. Establish a four-level risk early warning mechanism (low, medium, high, and critical) to generate differentiated early warning instructions for different risk levels, such as daily follow-up, enhanced monitoring, hospitalization observation, and immediate medical intervention, to avoid insufficient or excessive intervention. The digital health twin, based on the maternal-fetal physiological system mechanism model and combined with a deep reinforcement learning optimization module, simulates multiple intervention plans, selects the optimal strategy, and generates instructions for adjusting monitoring parameters and contingency plans for activating physical intervention equipment, making intervention measures more tailored to individual needs and improving the scientific nature and timeliness of risk management.

[0054] 4. A dynamic calibration mechanism ensures real-time adaptation of management strategies and continuously improves the reliability of monitoring and management. By comparing real-time collected physiological data with the prediction results of the digital health twin, if the health index deviation continuously exceeds the threshold and reaches the set time, the extended Kalman filter algorithm is used to update the state parameters of the mechanism model in reverse, correcting the evolution trajectory of the health index and key physiological parameters, and simultaneously adjusting monitoring parameters and intervention plans. This ensures that the management strategy always adapts to the dynamic changes in the physiological state of mothers and infants, effectively avoiding management failures caused by physiological fluctuations. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the method for monitoring and managing abnormal conditions and health information during pregnancy and childbirth according to the present invention.

[0056] Figure 2 This is a flowchart illustrating the principle of the method for monitoring and managing abnormal conditions and health information during pregnancy and childbirth according to the present invention.

[0057] Figure 3 This is an architecture diagram of the management system for monitoring and managing abnormal conditions and health information during pregnancy and childbirth, as described in this invention. Detailed Implementation

[0058] Example 1:

[0059] Please see Figure 1 and Figure 2 The present invention provides an embodiment of a method for monitoring and managing abnormal conditions and health information during pregnancy and childbirth, comprising the following steps:

[0060] S1: Synchronously collect user monitoring data through a multimodal wearable monitoring device group;

[0061] Furthermore, the monitoring data includes uterine contraction gradient changes, fetal heart rate variability, fetal movement frequency and duration, maternal heart rate variability, dynamic blood glucose levels, and dynamic curves of postpartum hemorrhage.

[0062] S2: The monitoring data is fused with time-series medical record data in the hospital information system using a timestamp synchronization algorithm to form a user health information profile. This profile is then input into an adaptive pregnancy and childbirth risk model for analysis, outputting a personal health risk assessment result. The personal health risk assessment result includes a comprehensive health index score and anomaly judgment results with clinical interpretation labels. The adaptive pregnancy and childbirth risk model is dynamically updated based on the fusion features. The clinical interpretation labels are used to indicate to clinicians the main physiological parameters and their temporal context that cause changes or abnormalities in the health index score.

[0063] S3: Based on individual health risk assessment results and clinical interpretation labels, hierarchical early warning is generated, and early warning instructions are generated. At the same time, the early warning instructions and user health information profiles are input into a pre-built digital health twin and the expected effects of different intervention programs are simulated to generate a comprehensive management strategy that includes instructions for adjusting monitoring parameters and intervention equipment parameter plans.

[0064] S4: Send the monitoring parameter adjustment command to the wearable monitoring device group, preload the intervention device parameter plan into the physical intervention device, compare the newly collected real-time monitoring data with the prediction results of the digital health twin, and recalculate the health index based on the new data. If the actual value of the health index deviates from the predicted value, dynamically calibrate the status parameters and comprehensive management strategy of the digital health twin.

[0065] Regarding S1 above:

[0066] Through a multimodal wearable monitoring device array, comprehensive, synchronous, and high-precision collection of key physiological parameters of the fetus and mother is achieved, providing a high-quality data foundation for risk assessment and intervention decisions. The monitoring data covers six core indicators: changes in uterine contraction force gradient, fetal heart rate variability, frequency and duration of fetal movements, maternal heart rate variability, dynamic blood glucose levels, and dynamic curves of postpartum hemorrhage. These indicators correspond to key health statuses at different physiological stages of pregnancy and can comprehensively reflect the mother's pregnancy tolerance and the fetus's intrauterine safety.

[0067] The hardware configuration of the multimodal wearable monitoring device group needs to meet the requirements of accuracy in clinical monitoring and comfort of wearing. Specifically, it includes five core components: a distributed uterine contraction pressure sensor array, a fetal electrocardiogram and Doppler ultrasound probe, a maternal electrocardiogram patch, a continuous glucose monitor, and a postpartum hemorrhage optical monitoring module. Each component adopts a lightweight design and is worn in a position that conforms to the physiological structure of the human body. For example, the uterine contraction pressure sensor array is attached to the mother's abdomen, the fetal electrocardiogram probe is integrated into the fetal heart rate monitoring belt, and the continuous glucose monitor uses a minimally invasive sensor implanted on the outer side of the upper arm. The overall device has a battery life of no less than 24 hours and supports Bluetooth 5.0 and LoRa dual-mode communication to ensure the stability of data transmission and low power consumption.

[0068] Furthermore, the gradient change values ​​of uterine contraction force were collected:

[0069] This is achieved through a distributed uterine contraction pressure sensor array in a multimodal wearable monitoring device group. The array consists of 16 high-sensitivity piezoresistive sensors with a 2cm spacing, arranged in a matrix on an elastic monitoring band. When worn, it covers the bottom to the lower part of the mother's uterus, ensuring that pressure changes in different parts of the uterus can be captured during contractions.

[0070] The sampling frequency is set to 50 Hz. The raw pressure data collected by the sensor contains noise signals such as respiratory interference and body position changes, which need to be smoothed by Kalman filtering. In the specific implementation, the pressure difference sequence of adjacent sensor nodes within the same time window is first calculated. In this embodiment, it is set to 100 milliseconds. This sequence can reflect the transmission gradient of uterine contraction pressure on the uterine surface. Then, a Kalman filter model is constructed. The initialization process noise variance is 0.01, the measurement noise variance is 0.05, and the initial error covariance matrix is ​​an identity matrix. The pressure difference sequence is denoised through an update iteration process. Finally, the smoothed uterine contraction force gradient change value is output. Kalman filtering is a conventional method that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method.

[0071] For example, when a pregnant woman at 38 weeks of gestation wore a distributed uterine contraction pressure sensor array, the raw pressure data collected showed obvious peak fluctuations during contractions. The raw pressure difference between adjacent sensors was between 0.3 and 1.5 kPa / cm. After Kalman filtering, noise components were effectively suppressed, and the fluctuation range of the uterine contraction force gradient change value was stabilized between 0.5 and 1.2 kPa / cm. It was possible to clearly identify a regular uterine contraction pattern with a duration of about 40 seconds and an interval of 5 minutes. Here, kPa / cm is the unit of pressure gradient, used to characterize the uterine contraction force gradient change value. For example, 1.5 kPa / cm means that the pressure difference between two adjacent uterine contraction pressure sensors during contractions, after calculation, shows an average pressure change of 1.5 kPa per centimeter.

[0072] Optionally, the sampling frequency can be dynamically adjusted according to the pregnancy and childbirth stages, set to 20 Hz in the early and middle stages of pregnancy, and increased to 50 Hz in the late stages of pregnancy and during labor; the Kalman filter can be replaced by the particle filter algorithm, which is suitable for signal processing in non-Gaussian noise environments. When there is strong motion interference in the monitoring environment, the smoothing effect of the particle filter is better; the number of nodes in the sensor array can be adaptively adjusted according to the size of the mother's abdomen. For pregnant women with a larger body size, it can be expanded to 20 sensors to ensure the integrity of pressure coverage.

[0073] Furthermore, fetal heart rate variability, fetal movement frequency, and duration are collected: Fetal heart rate signals and fetal movement ultrasound echo signals are simultaneously collected using the fetal electrocardiogram and Doppler ultrasound probes in the multimodal wearable monitoring device group. Fetal heart rate variability features are extracted from the fetal heart rate signals using wavelet transform and template matching algorithms, and fetal movement events are identified and fetal movement frequency and duration are statistically analyzed from the fetal movement ultrasound echo signals. Wavelet transform and template matching algorithms are conventional methods that can be understood and implemented by those skilled in the art, and this application is not limited to specific partitioning methods.

[0074] Furthermore, wavelet transform and template matching algorithms were used to extract fetal heart rate variability features from the fetal heart rate electrocardiogram (ECG) signal. This included: firstly, using the db4 wavelet to decompose the ECG signal into five levels, obtaining detail coefficients and approximation coefficients at different scales; then, reconstructing the detail coefficients at levels 3-5 to separate the QRS complex from the ECG signal; subsequently, constructing a normal fetal heart rate QRS template, which was trained based on ECG data from 100,000 healthy fetuses; using a template matching algorithm to identify consecutive QRS peaks, calculating the time interval between adjacent peaks to obtain the instantaneous fetal heart rate sequence; and finally, extracting fetal heart rate variability features by calculating parameters such as the standard deviation, coefficient of variation, and long-term variability of the sequence.

[0075] Furthermore, the QRS complex is a characteristic waveform in an electrocardiogram that reflects the ventricular depolarization process. It is the core basis for extracting fetal health indicators such as fetal heart rate variability. The Q wave is a downward initial waveform, the R wave is a positive high-amplitude wave, and the S wave is a downward waveform. Together, the three constitute the complete electrical signal trajectory of ventricular depolarization.

[0076] Furthermore, the identification of fetal movement events is based on the amplitude changes of Doppler ultrasound echo signals. When the 2 MHz ultrasound waves emitted by the ultrasound probe encounter fetal limb movements, the amplitude of the echo signal will show a significant jump. An amplitude threshold of 8 millivolts is set. When the amplitude of the ultrasound echo signal exceeds this threshold and the duration is greater than or equal to 10 seconds, it is determined as a valid fetal movement event. The fetal movement frequency is obtained by counting the number of fetal movements per unit time, usually 1 hour, and the start and end times of each fetal movement are recorded to calculate the fetal movement duration.

[0077] For example, a pregnant woman at 32 weeks of gestation with a history of gestational diabetes, after wearing a fetal electrocardiogram and Doppler ultrasound probe, successfully identified a clear QRS complex after wavelet transform decomposition of the acquired fetal heart rate signal. The instantaneous fetal heart rate sequence fluctuated between 135 and 150 beats per minute, and the calculated standard deviation of fetal heart rate variability was 12 beats per minute, with a coefficient of variation of 8%, which meets the standard for a normal fetus. The ultrasound echo signal showed 10 fluctuations with amplitudes exceeding 8 millivolts between 10:00 and 11:00, each lasting 20 to 40 seconds. The fetal movement frequency was calculated to be 10 times per hour, with a cumulative fetal movement duration of 320 seconds, both within the normal reference range for mid-pregnancy. If the fetal movement frequency is greater than or equal to 3 times per hour, the cumulative duration is greater than or equal to 120 seconds.

[0078] Optionally, wavelet transform can be replaced by short-time Fourier transform. When the fetal heart rate signal is less disturbed, short-time Fourier transform can extract time-frequency features more quickly. The template matching algorithm can be optimized using dynamic time warping algorithm, which is suitable for scenarios where there are individual differences in the morphology of fetal heart rate signals. The fetal movement recognition threshold can be dynamically adjusted according to gestational age, set to 5 mV in early pregnancy and increased to 8-10 mV in mid-to-late pregnancy to avoid misjudgment due to fetal maturity. Among these, short-time Fourier transform is a conventional method that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0079] Furthermore, maternal heart rate variability and dynamic blood glucose value acquisition: Through the maternal ECG patch and continuous glucose monitor in the multimodal wearable monitoring device group, maternal ECG signals and tissue fluid glucose concentration signals are acquired simultaneously, and maternal heart rate variability features are extracted from the maternal ECG signals using a time-frequency analysis method. The tissue fluid glucose concentration signal is converted into dynamic blood glucose values ​​through a dynamic calibration algorithm. The time-frequency analysis method is a conventional means that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0080] Furthermore, the maternal ECG patch employs a dry electrode design, attached to the fourth intercostal space on the left side of the mother's chest, to collect maternal ECG signals. The sampling frequency is set to 250 Hz, with a bandwidth filtering range of 0.05~100 Hz, effectively removing respiratory interference and motion artifacts. Short-time Fourier transform (SFT) is used in time-frequency analysis to extract maternal heart rate variability features. The specific steps are as follows: the collected maternal ECG signal is segmented into 2-second segments, with each segment overlapping by 1 second. After windowing each segment using a Hanning window, a Fourier transform is performed to obtain the power spectral density of each segment. Subsequently, low-frequency and high-frequency segments are divided, such as low-frequency segment: 0.04~0.15 Hz, high-frequency segment: 0.15~0.4 Hz. The ratio of low-frequency power to high-frequency power, LF / HF, is calculated. This ratio reflects the balance between the sympathetic and parasympathetic nervous systems and serves as a core characteristic parameter of maternal heart rate variability.

[0081] Furthermore, the continuous glucose monitor employs a minimally invasive glucose sensor. A dedicated implanter inserts the sensor probe into the subcutaneous tissue of the mother's upper arm. The probe is 5 mm long and 0.3 mm in diameter, causing minimal pain during implantation and not affecting daily activities. The sensor indirectly reflects blood glucose levels by detecting glucose concentration in tissue fluid, sampling every 5 minutes. The collected tissue fluid glucose concentration signal is converted into a dynamic blood glucose value using a dynamic calibration algorithm. The dynamic calibration algorithm constructs a calibration model based on the least squares method, using three daily finger-prick blood glucose values ​​as a reference standard to linearly correct the tissue fluid glucose concentration. The calibration formula is dynamically adjusted through real-time updated calibration coefficients to ensure conversion accuracy across different time periods and blood glucose concentration ranges. The final output dynamic blood glucose value error does not exceed ±10%.

[0082] For example, after a pregnant woman with gestational diabetes at 28 weeks of gestation wore a maternal ECG patch and a continuous glucose monitor, the acquired maternal ECG signal, after short-time Fourier transform processing, yielded an LF / HF ratio of 1.8, which is within the normal reference range (e.g., 1.0-2.5), indicating stable maternal autonomic nerve function. The tissue fluid glucose concentration collected by the continuous glucose monitor was 9.2 mmol / L one hour after breakfast. After conversion using a dynamic calibration algorithm, the dynamic blood glucose value was 9.8 mmol / L, with an error of 2% compared to the corresponding finger-prick blood glucose value of 10.0 mmol / L, meeting the clinical monitoring accuracy requirements. Two hours after lunch, the tissue fluid glucose concentration was 7.5 mmol / L, and the converted dynamic blood glucose value was 7.8 mmol / L, which is within the target range for gestational diabetes blood glucose control. If the value is less than or equal to 8.5 mmol / L two hours after a meal, it indicates effective dietary control. Here, mmol / L represents millimoles per liter.

[0083] Optionally, the time-frequency analysis method can be replaced with wavelet packet transform, which can more finely divide the frequency range and is suitable for nonlinear feature extraction of maternal heart rate variability signals; the dynamic calibration algorithm can adopt a support vector machine regression model. When the mother is on insulin therapy or has changes in dietary structure, the support vector machine regression has stronger adaptability and higher calibration accuracy; the sampling frequency of the continuous glucose monitor can be adjusted according to the blood glucose control situation. When blood glucose fluctuates greatly, it can be increased to once every 2 minutes, and kept once every 5 minutes during the stable period.

[0084] Furthermore, the dynamic curve of postpartum hemorrhage is collected: the postpartum hemorrhage optical monitoring module in the multimodal wearable monitoring device group collects the spectral data of wound bleeding, and calculates and generates the dynamic curve of postpartum hemorrhage in real time based on Lambert-Beer's law. The Lambert-Beer's law is a conventional method that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0085] Furthermore, the near-infrared light emitted by the optical monitoring module has a wavelength range of 400–800 nanometers. This wavelength range can be effectively absorbed by hemoglobin in the blood, and the penetration depth is moderate, so it will not cause radiation damage to the mother. When collecting spectral data of bleeding from the wound, the sampling frequency is set to 1 Hz, and the light absorption intensity at different wavelengths is recorded in real time. Based on Beer-Lambert's law, a bleeding volume calculation model is constructed. This model establishes the correspondence between light absorption intensity and bleeding volume through calibration experiments. Specifically, bleeding scenarios with different bleeding volumes are first simulated in vitro, and the corresponding spectral absorption data are measured. A linear regression equation between absorption intensity and bleeding volume is fitted as the basis for real-time calculation. In clinical applications, the instantaneous bleeding volume is calculated by substituting the real-time collected spectral data into the regression equation, and then a dynamic curve of postpartum hemorrhage volume is plotted according to the time series to intuitively reflect the changing trend of blood volume.

[0086] For example, after a vaginal delivery, the mother wears an optical monitoring module for postpartum hemorrhage. The spectral data collected within 1 hour postpartum is calculated using Beer-Lambert's law to obtain instantaneous blood loss of 20 ml, 35 ml, 45 ml, 55 ml, and 60 ml, respectively. The dynamic curve shows a slow upward trend, with a cumulative blood loss of 60 ml, which is within the normal range for postpartum hemorrhage, such as less than or equal to 500 ml. At 2 hours postpartum, the instantaneous blood loss increases to 80 ml, the slope of the curve increases, and the cumulative blood loss is 140 ml. The monitoring system promptly marks this trend and alerts medical staff. At 3 hours postpartum, the instantaneous blood loss stabilizes at 70 ml, the cumulative blood loss is 210 ml, and the curve tends to flatten, confirming no risk of active bleeding.

[0087] Optionally, the wavelength range of the optical monitoring module can be extended to 900 nanometers to enhance the spectral recognition capability in high bleeding scenarios; the bleeding calculation model can incorporate machine learning algorithms and combine clinical parameters such as maternal weight, delivery method, and coagulation function to optimize the accuracy of bleeding estimation; the sampling frequency is set to 2 Hz within 2 hours postpartum and then reduced to 1 Hz after 2 hours to balance monitoring sensitivity and energy consumption control.

[0088] Regarding S2 above:

[0089] The timestamp synchronization algorithm in S2 is as follows:

[0090] S2.1: Parse the timestamps of each data channel from the monitoring data;

[0091] S2.2: Extract time-series medical record data with system timestamps from the parent medical record database of the hospital information system; the time-series medical record data includes prenatal check-up records, medication records, laboratory test results, and imaging reports;

[0092] Furthermore, prenatal checkup records include blood pressure, weight, fundal height, abdominal circumference, and fetal heart rate monitoring results from each prenatal checkup; medication records include drug name, dosage, administration time, and frequency; laboratory test results include the test values ​​and times of indicators such as complete blood count, urinalysis, liver and kidney function, blood glucose, and coagulation function; imaging reports include the conclusions, examination times, and key measurement data of examinations such as B-ultrasound and color Doppler ultrasound, such as fetal biparietal diameter, amniotic fluid index, and placental grading. The system timestamps for the time-series medical record data are automatically generated by the hospital information system to ensure the accuracy of time recording.

[0093] For example, in the chronological medical record data of a pregnant woman at 35 weeks of gestation, the prenatal examination record shows that the blood pressure was 135 / 85 mmHg at 32 weeks of gestation and 140 / 90 mmHg at 34 weeks of gestation; the medication record shows that labetalol 100mg was taken orally twice a day starting at 34 weeks of gestation; the laboratory test results show that the fasting blood glucose was 5.8 mmol / L and the hemoglobin was 110 g / L at 34 weeks of gestation; the imaging report shows that the B-ultrasound at 34 weeks of gestation showed that the fetal biparietal diameter was 8.8 cm, the amniotic fluid index was 12 cm, and the placenta was grade II. All these data have corresponding system timestamps, forming a time series correspondence with the timestamps of the monitoring data.

[0094] S2.3: The timestamps of the monitoring data and the system timestamps of the time-series medical record data are aligned using a dynamic time warping algorithm to obtain the timestamp-aligned monitoring values ​​and medical record text descriptions. The timestamp-aligned monitoring values ​​and medical record text descriptions are mapped to feature vectors of a unified dimension through an embedding layer, and then concatenated along the feature dimension to form a user health information profile. The dynamic time warping algorithm is a conventional method that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0095] Furthermore, the process of generating monitoring values ​​and medical record text descriptions includes: first, constructing a distance matrix between the monitoring data time series and the medical record data time series, using Euclidean distance as the distance metric; then, finding the matching path with the minimum sum of distances through a dynamic time warping algorithm, which is the optimal alignment relationship between the two time series; finally, based on the optimal path, unifying the timestamps of the monitoring data and the system timestamps of the medical record data into an aligned standard time axis, resulting in timestamp-aligned monitoring values ​​and medical record text descriptions.

[0096] Furthermore, the aligned data needs to undergo feature mapping and concatenation to form a user health information profile. First, the timestamp-aligned monitoring values, such as continuous data like uterine contraction gradient changes and dynamic blood glucose levels, are standardized and converted into normalized values ​​between 0 and 1. For medical record text descriptions, word embedding technology is used to convert the text into vector representations, where medical terminology is mapped using a pre-trained medical word vector model to ensure semantic accuracy. Then, the normalized monitoring value vectors and text embedding vectors are mapped to a unified-dimensional feature vector through an embedding layer, set to 256 dimensions. The feature vectors are then concatenated to form a 256-dimensional user health information profile. This profile includes both the dynamic features of real-time physiological monitoring data and the static features of historical medical record data, comprehensively reflecting the user's health status.

[0097] For example, in the pregnant woman's monitoring data, the dynamic blood glucose level at 10:00 AM on week 35 of pregnancy was 6.5 mmol / L, which was normalized to 0.6, and the uterine contraction gradient change value was 0.5 kPa / cm, which was normalized to 0.4. In the corresponding medical record data, the medication record at week 34 of pregnancy, "oral labetalol 100mg, twice daily", was converted into a 256-dimensional vector. The ultrasound report, "amniotic fluid index 12cm", was also converted into a 256-dimensional vector. After the embedding layer mapping, both the monitoring value vector and the text vector were converted into 256 dimensions. After splicing, a 256-dimensional user health information profile was formed. This profile includes the current dynamic status of blood glucose and uterine contractions, as well as the previous blood pressure control and fetal development.

[0098] Furthermore, the data transmission between the multimodal wearable monitoring device group and the hospital information system adopts a blockchain-based encrypted data channel, specifically:

[0099] (1) At the data sending end, an asymmetric encryption algorithm is used to encrypt the monitoring data stream and its timestamp, and generate the corresponding data hash value. The asymmetric encryption algorithm is a conventional means that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method.

[0100] (2) Upload the encrypted data packet and data hash value to the distributed nodes in the permissioned blockchain network for storage;

[0101] (3) At the data receiving end, the hospital information system downloads data packets from the blockchain network, decrypts them using a private key, and verifies the data hash value to ensure data integrity and authenticity of the source;

[0102] (4) Record every data access and fusion operation in the blockchain network to form an immutable audit log.

[0103] The adaptive pregnancy and childbirth risk model comprises a feature encoder, a spatiotemporal attention network, a health index calculation module, and a risk classifier connected in sequence.

[0104] A1: The feature encoder is composed of a cascaded one-dimensional convolutional neural network and a long short-term memory network; the input of the one-dimensional convolutional neural network receives the user's health information profile and extracts local temporal features through its convolutional and pooling layers; its output is connected to the input of the long short-term memory network to capture long-term dependencies and output the encoded feature sequence.

[0105] Furthermore, the one-dimensional convolutional neural network comprises three convolutional layers and two pooling layers. The first convolutional layer uses 64 3×1 convolutional kernels with a stride of 1 and the activation function ReLU, which can extract local detailed features of the data. The first pooling layer uses max pooling with a kernel size of 2×1 and a stride of 2 to reduce feature dimensionality and retain key features. The second convolutional layer uses 128 3×1 convolutional kernels with a stride of 1 and the activation function ReLU, which further extracts deep local features. The third convolutional layer uses 256 3×1 convolutional kernels with a stride of 1 and the activation function ReLU, which enhances feature representation capabilities. The second pooling layer also uses max pooling with a kernel size of 2×1 and a stride of 2, outputting a local temporal feature vector. ReLU is a conventional technique that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0106] Furthermore, the Long Short-Term Memory (LSTM) network comprises two hidden layers, each with 128 hidden units. The activation function is tanh, and the activation functions for the forget gate, input gate, and output gate are sigmoid. The local temporal feature vector output from the one-dimensional convolutional neural network is input into the LSM network. Through the gating mechanism, important information is selectively retained while redundant information is forgotten, capturing the long-term dependencies between features at different time steps. The final output is an encoded feature sequence with a length of 100. The sigmoid and tanh activation functions are conventional methods that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0107] For example, the uterine contraction gradient change value in the user's health information profile shows a low-high-low change pattern at different time steps. A one-dimensional convolutional neural network extracts local features of this pattern through convolutional layers, reduces the dimensionality through pooling layers, and then inputs them into a long short-term memory network. The long short-term memory network forgets the early irrelevant low pressure value features through the forget gate, retains the key features of the mid-term high pressure value through the input gate, and outputs the encoded feature sequence containing the trend of uterine contraction changes through the output gate. This sequence can clearly reflect the intensity change and duration pattern of uterine contractions.

[0108] A2: The input of the spatiotemporal attention network is connected to the output of the feature encoder, and it contains a temporal attention module and a feature attention module. The temporal attention module is used to calculate the importance weights of different time steps in the encoded feature sequence. The feature attention module is used to calculate the importance weights of different physiological feature dimensions at the same time step. After weighted fusion, a context-enhanced feature representation is generated.

[0109] Furthermore, the time attention module calculates the importance weights of different time steps in the encoded feature sequence, specifically through the Softmax function: first, a linear transformation is performed on the feature vector of each time step of the encoded feature sequence to obtain a time attention score; then, the score is normalized using the Softmax function to obtain the attention weight for each time step, with weight values ​​between 0 and 1, and the sum of the weights is 1. The time step weights can highlight key time nodes that have a significant impact on health risks, such as periods of frequent uterine contractions or moments of sharp fluctuations in blood sugar. The Softmax function is a conventional method that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0110] Furthermore, the feature attention module calculates the importance weights of different physiological feature dimensions at the same time step, also using the Softmax function: a linear transformation is performed on the feature vector at each time step to obtain the attention score for each feature dimension; after Softmax normalization, the feature dimension weights are obtained, with weight values ​​between 0 and 1, and the sum of the weights is 1. Feature dimension weights can highlight indicators that contribute significantly to health risks, such as the uterine contraction gradient change value in late pregnancy and the dynamic blood glucose value in patients with gestational diabetes.

[0111] A3: The input of the health index calculation module is connected to the output of the spatiotemporal attention network. It consists of a fully connected layer and a sigmoid activation function. It maps the context-enhanced feature representation into a scalar value between 0 and 1, which serves as the health index score. The fully connected layer contains two layers. The first fully connected layer has an output dimension of 64 and uses ReLU activation function. The second fully connected layer has an output dimension of 1 and uses sigmoid activation function. The sigmoid function maps the input value to the range of 0 to 1, making it easier to intuitively reflect the degree of health risk: the closer the health index score is to 0, the better the health status and the lower the risk; the closer it is to 1, the worse the health status and the higher the risk.

[0112] Furthermore, in the calculation of the health index, the model is trained using a cross-entropy loss function to ensure that the health index output by the model accurately reflects the user's actual health risks. The training data uses monitoring data and clinical outcome data from 100,000 pregnant and postpartum women. Clinical outcome data includes normal pregnancy outcomes, premature birth, pregnancy complications, postpartum hemorrhage, etc. Clinical outcomes are assigned label values ​​between 0 and 1 according to their severity. The model parameters are optimized using a backpropagation algorithm to minimize the error between the health index score and the clinical outcome label values.

[0113] For example, a pregnant woman in her second trimester with good health has a health index score of 0.2 after the context-enhanced feature representation is processed by a fully connected layer, indicating low risk; a pregnant woman in her third trimester with gestational hypertension and poor blood pressure control has a health index score of 0.75 after the context-enhanced feature representation is processed, indicating high risk; and a postpartum woman with continuously increasing postpartum hemorrhage has a health index score of 0.9, indicating critical risk.

[0114] A4: The input of the risk classifier is connected to the output of the health index calculation module. It consists of a fully connected layer and a Softmax function. It is used to receive feature representations associated with the health index score and output anomaly judgment results with clinical interpretation labels. The fully connected layer contains two layers. The output dimension of the first fully connected layer is 32, and the activation function is ReLU. The output dimension of the second fully connected layer is the preset number of risk categories. In this embodiment, the risk categories include seven categories: normal, abnormal uterine contractions, abnormal fetal heart rate, abnormal maternal heart rate, abnormal blood sugar, abnormal postpartum hemorrhage, and multiple indicator combined abnormalities. The activation function is Softmax. The Softmax function can normalize the scores of each risk category and output the probability of each category. The category with the highest probability is the anomaly judgment result.

[0115] Furthermore, the generation of clinical explanatory labels is based on the output of the risk classifier, combined with the weight information and feature contribution of the spatiotemporal attention network, to clearly indicate the main physiological parameters and their temporal context that lead to the abnormality. Specifically, firstly, the feature dimension corresponding to the category with the highest output probability of the risk classifier is extracted, and then the time step with the highest temporal attention weight and the physiological parameter with the highest feature attention weight are combined to generate clinical explanatory labels containing gestational age, abnormal indicators, abnormal time, and abnormal manifestations. These labels can provide clinicians with intuitive and specific explanations of the causes of abnormalities and assist in the formulation of intervention plans.

[0116] For example, for a pregnant woman at 39 weeks of gestation, the probability of the risk classifier outputting the abnormal uterine contraction category is 0.85, and the probabilities of the remaining categories are all below 0.1. Combining the weights of the spatiotemporal attention network, the weight of the time step 16:00~17:00 is the highest, such as 0.75, and the weight of the uterine contraction force gradient change value in the feature dimension is the highest, such as 0.65. The final generated clinical interpretation label is "At 39 weeks of gestation, the uterine contraction force gradient change value increased from 0.6kPa / cm to 1.3kPa / cm between 16:00 and 17:00, the uterine contraction intensity increased, and the frequency reached 5 minutes / time, indicating the start of labor."

[0117] Furthermore, the personal health risk assessment result is composed of the health index score and the anomaly determination result, wherein the health index score provides continuous risk quantification, and the anomaly determination result provides discrete risk classification.

[0118] The generation of the clinical interpretability labels also employs a model interpretability method based on sapride and interpretation, specifically:

[0119] B1: For the anomaly determination results output by the risk classifier, select the top N feature dimensions with the highest contribution and their corresponding key time steps. In this embodiment, N=3.

[0120] B2: Calculate the Shapley value of each feature in the user health information profile at the key time step; the Shapley value is used to quantify the contribution of the user health information profile to the anomaly detection result;

[0121] In the specific calculation of the Shapley value, each feature is regarded as a player in the game, and the combination of all features constitutes the game set. By traversing all possible feature subsets, the contribution of the feature to the anomaly judgment result after being added to the subset is calculated. Finally, the average value of the contribution changes of all subsets is taken as the Shapley value of the feature. The larger the Shapley value, the higher the contribution of the feature to the anomaly judgment result, and the more likely it is to be the main factor leading to the anomaly.

[0122] B3: Sort user health information profiles from high to low according to sapride values, and mark the top N features and their corresponding time steps as the core content of clinical interpretation labels.

[0123] For example, in a pregnant woman with gestational diabetes at 36 weeks of gestation, the risk classifier outputs abnormal blood glucose. The top three features with the highest attention weight are dynamic blood glucose level, maternal heart rate variability, and fetal movement frequency. The top three time steps with the highest time attention weight are 1 hour after breakfast, 1 hour after lunch, and 1 hour after dinner. After calculating the sapride value, the dynamic blood glucose level, the time step 1 hour after breakfast, and the time step 1 hour after lunch ranked in the top three, with sapride values ​​of 0.4, 0.3, and 0.2, respectively. The optimized clinical interpretation label is "36 weeks of gestation, abnormal blood glucose: dynamic blood glucose level reached 9.5 mmol / L 1 hour after breakfast and 8.8 mmol / L 1 hour after lunch, both higher than the control target of 7.8 mmol / L. The maternal heart rate variability LF / HF ratio is 1.9, indicating that poor blood glucose control may affect cardiovascular regulatory function."

[0124] Regarding S3 above:

[0125] The hierarchical early warning system in S3 specifically refers to:

[0126] S3.1: Establish a mapping relationship between health index score ranges and abnormal judgment results and preset risk level thresholds; the risk levels include low risk, medium risk, high risk and critical risk;

[0127] Further, the risk levels are categorized as follows: Low risk: Health index score 0-0.3, no abnormalities or only minor, temporary abnormalities, such as slightly lower frequency of fetal movements followed by a return to normal; Medium risk: Health index score 0.3-0.6, clear abnormalities, but the abnormal indicators do not meet the critical threshold, such as continuous glucose levels consistently above 7.8 mmol / L but below 11.1 mmol / L, or irregular uterine contraction frequency but not meeting the diagnostic criteria for preterm labor; High risk: Health index score 0.6-0.8, clear abnormalities, and the abnormal indicators are close to the critical threshold, such as cumulative postpartum hemorrhage of 300 ml or fetal heart rate variability standard deviation below 5 beats / min; Critical risk: Health index score 0.8-1.0, abnormalities indicating severe abnormalities that may endanger the lives of the mother or fetus, such as cumulative postpartum hemorrhage of 500 ml or a fetal heart rate consistently below 110 beats / min or above 160 beats / min.

[0128] Furthermore, the risk level threshold can be fine-tuned based on the clinical level of medical institutions and the epidemiological characteristics of pregnancy and childbirth diseases in the region. For example, in areas with a high incidence of premature birth, the health index threshold for medium risk can be lowered to 0.25 to improve the sensitivity of early warning; in areas with strained medical resources, the high-risk threshold can be appropriately increased to avoid waste of medical resources caused by excessive early warning.

[0129] S3.2: Determine the current risk level and main abnormal characteristics based on the health index score and the clinical interpretation label, including: firstly, determine the preliminary risk level based on the health index score, then, in combination with the abnormal judgment result and the clinical interpretation label, correct the risk level to ensure that the risk level can truly reflect the health status, and at the same time, extract the core abnormal indicators, abnormal time and abnormal manifestations from the clinical interpretation label to clarify the main abnormal characteristics.

[0130] For example, a pregnant woman at 32 weeks of gestation has a health index score of 0.55, and is initially classified as medium risk. The abnormality assessment result is "abnormal blood glucose", and the clinical interpretation label is "at 32 weeks of gestation, the dynamic blood glucose value reached 8.6 mmol / L 2 hours after lunch and did not decrease for 3 hours, and the maternal heart rate variability LF / HF ratio was 2.3". The overall risk level is classified as medium risk, and the main abnormal feature is "poor blood glucose control in gestational diabetes mellitus, dynamic blood glucose value of 8.6 mmol / L 2 hours after lunch for 3 hours, accompanied by mild autonomic dysfunction".

[0131] S3.3: If the risk level is determined to be low, a daily follow-up reminder instruction will be generated, such as a routine prenatal check-up once a week, daily self-monitoring of fetal movement, wearing a blood glucose monitor for no less than 22 hours / day, and contacting the prenatal doctor in time if abnormal conditions such as decreased fetal movement, abdominal pain, or vaginal bleeding occur.

[0132] If the risk level is determined to be medium, an early warning instruction for enhanced monitoring and outpatient follow-up will be generated. For example, outpatient follow-up will be conducted every 3 days, the frequency of fetal heart monitoring will be increased to twice a day, such as at 10 am and 4 pm, the interval of dynamic blood glucose monitoring will be shortened to 2 hours / time, and the diet will be adjusted, such as reducing the intake of high sugar and high fat foods and increasing dietary fiber. If the blood glucose level is consistently higher than 8.5 mmol / L or regular uterine contractions occur, such as 5 minutes / time lasting for 30 minutes, go to the hospital immediately for treatment.

[0133] If the risk level is determined to be high, an early warning instruction for hospitalization and special examinations will be generated. For example, the patient will be hospitalized immediately and transferred to a high-risk obstetric ward for 24-hour continuous fetal heart rate monitoring and uterine contraction monitoring. Blood pressure and blood sugar will be measured every hour. Special examinations such as ultrasound, blood routine, coagulation function, liver and kidney function will be completed. If fetal distress occurs or the mother's condition worsens, the emergency intervention process will be initiated.

[0134] If a critical risk is identified, an immediate medical intervention warning instruction will be generated. For example, the emergency green channel will be activated immediately, and obstetricians, anesthesiologists, and neonatologists will be notified to be on site. Surgical intervention or resuscitation equipment, such as blood transfusion devices and ventilators, will be prepared. Vital signs and fetal safety will be continuously monitored. If critical situations such as maternal shock or fetal cardiac arrest occur, the resuscitation plan will be implemented immediately.

[0135] S3.4: Input the warning instruction and its corresponding risk level, health index score, abnormal judgment result, clinical interpretation label and user health information profile into the digital health twin.

[0136] Furthermore, the dynamic iterative update process of the adaptive pregnancy and childbirth risk model is as follows: when a new user health information profile and its corresponding clinical outcome label are input into the model, the online gradient descent algorithm is used to minimize the difference between the model prediction result and the clinical outcome label, and the weight parameters of the one-dimensional convolutional neural network and long short-term memory network in the feature encoder, the weight parameters of the attention module in the spatiotemporal attention network, and the weight parameters of the fully connected layer in the risk classifier are updated in real time. The online gradient descent algorithm is a conventional method that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0137] Furthermore, based on the abnormality assessment results and clinical interpretation labels, a tiered and graded early warning system is implemented, including:

[0138] (1) Establish a mapping relationship between anomaly judgment results and risk level thresholds in advance; the risk levels include low risk, medium risk, high risk and critical risk;

[0139] (2) Based on the current abnormal judgment result and clinical interpretation label, determine the corresponding risk level; if it is low risk, generate a daily follow-up reminder instruction; if it is medium risk, generate an enhanced monitoring and outpatient re-examination warning instruction; if it is high risk, generate an inpatient observation and special examination warning instruction; if it is critical risk, generate an immediate medical intervention warning instruction.

[0140] The construction and simulation process of the digital health twin is as follows:

[0141] C1: Based on the maternal medical record data and historical monitoring data, a maternal-fetal physiological system mechanism model described by a set of coupled differential equations is constructed. Upon receiving a clinical warning instruction, using the current patient health information profile as the initial state, multiple intervention programs conforming to clinical guidelines are simulated and executed in the digital space, and the evolution trajectory of health indices under each intervention program is predicted. The intervention programs include adjusting the maternal position, adjusting the medication regimen, adjusting the fluid resuscitation rate, and preparing for surgical intervention.

[0142] Furthermore, the maternal-fetal physiological system mechanism model can quantify the interaction between various physiological indicators and simulate the dynamic changes in physiological state. The maternal-fetal physiological system mechanism model covers four major subsystems: maternal cardiovascular system, metabolic system, reproductive system, and fetal circulatory system and nervous system. Each subsystem is coupled with each other through key physiological parameters, such as blood pressure, blood sugar, uterine contraction force, and fetal heart rate, to form a complete physiological system network.

[0143] Specifically, the cardiovascular system sub-model describes the relationship between blood pressure, heart rate, and cardiac output; the metabolic system sub-model describes the dynamic processes of glucose absorption, utilization, and insulin secretion; the reproductive system sub-model describes the correlation between uterine contraction intensity, frequency, and cervical maturity; the fetal circulatory system sub-model describes the relationship between fetal heart rate and blood oxygen saturation; and the nervous system sub-model describes the regulatory mechanisms of heart rate variability and autonomic nervous function. The coupled differential equations of each sub-model are established based on physiological and physical principles and clinical experimental data. The model parameters are obtained by fitting 100,000 cases of perinatal physiological data to ensure the accuracy and clinical applicability of the models.

[0144] Furthermore, upon receiving a clinical warning instruction, the system uses the current patient health profile as the initial state and simulates the execution of various intervention protocols in accordance with clinical guidelines in the digital space. These intervention protocols fall into four main categories: adjusting maternal position, adjusting medication regimens, adjusting fluid resuscitation rates, and preparing for surgical intervention. Each category includes multiple specific implementation methods. For example, adjusting position includes left lateral decubitus, right lateral decubitus, supine, and semi-recumbent positions; adjusting medication regimens includes increasing / decreasing drug dosage, changing drug types, and adjusting medication frequency; adjusting fluid resuscitation rates includes increasing / decreasing fluid volume and changing the type of fluid; and preparing for surgical intervention includes cesarean section and uterine artery embolization.

[0145] Furthermore, during the simulation, the simulation time step was set to 1 minute, and the simulation duration was determined based on the expected onset time of the intervention plan. For example, the simulation duration for postural adjustment was 30 minutes, for medication administration it was 2 hours, and for surgical intervention it was 4 hours. By solving the coupled differential equations, the evolution trajectory of health indices under each intervention plan was predicted. This trajectory can intuitively reflect the changing trend of health status after the implementation of the intervention plan. For example, the evolution trajectory of the blood glucose control plan can show the decrease in blood glucose levels over time, and the evolution trajectory of the uterine contraction suppression plan can show the decreasing trend of the uterine contraction force gradient.

[0146] For example, a pregnant woman at 34 weeks of gestation with threatened preterm labor has a current health index of 0.7, indicating high risk. The main abnormal characteristic is "regular uterine contractions, 5 minutes / time, lasting 30 minutes, with a uterine contraction force gradient change of 1.0 kPa / cm". A digital health twin simulated three intervention plans: Plan 1, left lateral decubitus position + oral ritodrine 10mg, simulation duration 2 hours, predicting the health index would decrease from 0.7 to 0.5; Plan 2, intravenous infusion of magnesium sulfate 2g / h, simulation duration 2 hours, predicting the health index would decrease from 0.7 to 0.4; Plan 3, bed rest + fetal heart rate monitoring, simulation duration 2 hours, predicting the health index would remain at 0.65. The evolution trajectory of the health index under the three plans showed that Plan 2 had the most significant intervention effect, with the fastest decrease in health index and the lowest final value.

[0147] C2: Compare the evolution trajectory of each intervention plan with the clinical safety threshold, select the optimal intervention plan that can restore the health index to the safe range as quickly as possible or maintain the most stable state, and generate a comprehensive management strategy that includes the monitoring parameter adjustment instructions and intervention equipment parameter plans corresponding to the optimal intervention plan. In this embodiment, the clinical safety threshold is set to a health index of less than or equal to 0.3.

[0148] Furthermore, the screening criteria include three core indicators: the time it takes for the health index to drop to the safe range, the stability of the health index, and the safety of the intervention program. The specific screening process is as follows: First, calculate the time it takes for the health index to drop to the safe range under each intervention program; the shorter the time, the higher the score. Second, calculate the fluctuation range of the health index within the safe range; the smaller the fluctuation range, the higher the score. Finally, assess the safety of the intervention program; programs with no risk of adverse reactions receive higher scores.

[0149] The specific process for generating the monitoring parameter adjustment command is as follows:

[0150] D1: Analyze the evolution trajectory corresponding to the optimal intervention plan and identify the monitoring target values ​​that need to be adjusted to maintain physiological parameters within a safe range;

[0151] Furthermore, the specific steps of D1 include:

[0152] (1) The system obtains the evolution trajectory corresponding to the optimal intervention plan generated by the digital health twin simulation, compares this evolution trajectory with the predefined individualized clinical safety threshold range point by point, identifies any trajectory segment in the evolution trajectory that deviates from the safety threshold range, and marks the starting point of these segments as potential risk points.

[0153] (2) For each identified potential risk point, the system uses the potential risk point as a benchmark to perform reverse derivation. Specifically, the system calculates a stricter mid-term target range of parameters that needs to be maintained between the current time point and the risk point in order to prevent the parameter value from really deviating from the safe threshold range at later time points, based on the rate and direction of change of the evolution trajectory at the risk point.

[0154] (3) The system then verifies the clinical operability of the intermediate target range. This verification process is carried out by querying the historical simulation database of the digital health twin to determine whether the target can be actually achieved by adjusting the monitoring strategy or initiating physical intervention. At the same time, the system analyzes the mutual influence relationship between the intermediate target ranges of different physiological parameters, establishes their temporal correlation network, and clarifies the auxiliary target value that parameter B needs to be achieved in advance in order to achieve the target of parameter A and its time constraints.

[0155] (4) The system prioritizes all operationally validated intermediate target ranges according to their clinical urgency and timing requirements, and integrates and encapsulates them into a structured hierarchical monitoring target instruction set as the monitoring target value output; the hierarchical monitoring target instruction set clearly specifies the specific physiological parameters that need to be adjusted for the monitoring strategy, their target values ​​or ranges, the final time limit for achieving the target, and the superior risk prevention and control purpose served by the target.

[0156] Taking blood glucose intervention in patients with gestational diabetes mellitus as an example, the optimal intervention plan is "dietary adjustment + subcutaneous insulin injection." The evolution trajectory shows that postprandial 2-hour blood glucose needs to be controlled between 4.4 and 6.7 mmol / L, and the peak postprandial blood glucose at 1 hour should not exceed 7.8 mmol / L. Based on this, the identified monitoring target values ​​are: continuous glucose monitoring (CGM) values ​​≤ 7.8 mmol / L at 1 hour postprandial, ≤ 6.7 mmol / L at 2 hours postprandial, and fasting blood glucose 3.3–5.6 mmol / L.

[0157] For high-risk preterm labor women with abnormal uterine contractions, the optimal intervention is "left lateral decubitus position + ritodrine intravenous infusion." The evolution trajectory shows that the uterine contraction force gradient change value needs to be reduced to 0.3~0.5 kPa / cm, and the contraction frequency is less than or equal to 10 times / hour. The corresponding monitoring target values ​​are set as follows: uterine contraction force gradient change value less than or equal to 0.5 kPa / cm, contraction frequency less than or equal to 10 times / hour, and fetal heart rate variability standard deviation greater than or equal to 8 beats / min.

[0158] D2: Convert the monitored target value into configuration parameters for specific sensors in the multimodal wearable monitoring device group, and encapsulate the configuration parameters into monitoring parameter adjustment instructions; the configuration parameters include sampling frequency, alarm threshold, and data upload interval.

[0159] Furthermore, the specific steps of D2 include:

[0160] (1) The system receives the monitoring target value, queries a predefined parameter mapping rule library, which defines the quantitative conversion relationship between various physiological parameter target values ​​and sensor configuration parameters. Based on the content of the monitoring target value, the system matches and calculates the sensor type that needs to be adjusted and its corresponding initial configuration parameter value through the parameter mapping rule library, and generates an initial parameter list. For example, if the target value is to increase the monitoring sensitivity of fetal heart rate variability by 20%, the rule library indicates that the sampling rate of the fetal electrocardiogram sensor needs to be increased from 125 Hz to 250 Hz.

[0161] (2) The initial parameter list is compared and verified with the known hardware performance parameters of each sensor in the multimodal wearable monitoring device group and the current system resource status. The verification includes confirming whether the parameter values ​​are within the range supported by the sensor hardware and whether the adjusted overall system power consumption and data transmission volume are within the available resource threshold of the edge computing gateway. If the verification fails, the system returns (1) and recalculates the parameter values ​​according to the alternative mapping rules in the rule base until a final parameter list that has passed the verification is generated.

[0162] (3) Each configuration parameter in the final parameter list that has passed the verification is converted into a structured instruction code according to the specific medical device communication protocol standard followed by the target sensor. The encapsulation process includes filling in the sensor's unique device identifier, the corresponding parameter setting operation code and the specific parameter value. Finally, all the instruction codes for a single sensor are aggregated and a unified instruction sequence number and timestamp are added to encapsulate a complete and scalable monitoring parameter adjustment instruction data package.

[0163] For example, the sampling frequency for uterine contraction monitoring has been increased from the usual 50 Hz to 100 Hz to ensure that instantaneous changes in the intensity of uterine contractions are captured; the sampling frequency for dynamic blood glucose monitoring has been adjusted from once every 5 minutes to once every 2 minutes to track the blood glucose decline trend after insulin injection in real time; the sampling frequency for postpartum hemorrhage monitoring remains at 1 Hz, but the data upload interval has been shortened from 30 seconds to 10 seconds to ensure timely feedback on changes in bleeding volume.

[0164] For example, the high alarm threshold for dynamic blood glucose is set at 7.8 mmol / L, and the low alarm threshold is set at 3.3 mmol / L; the high alarm threshold for uterine contraction gradient change value is set at 0.5 kPa / cm, and the high alarm threshold for uterine contraction frequency is set at 10 times / hour; the alarm threshold range for fetal heart rate is set at 110~160 beats / min, and an alarm is triggered if the heart rate is below or above this range.

[0165] For example, the data upload interval for uterine contractions and fetal heart rate of high-risk preterm pregnant women was shortened from the usual 1 minute to 10 seconds; the data upload interval for blood glucose intervention was 2 minutes in the early stage, and then restored to 5 minutes after blood glucose stabilized; the data upload interval for bleeding volume of postpartum hemorrhage patients was fixed at 10 seconds until the bleeding volume dropped to a safe range and the cumulative amount was less than or equal to 100 ml.

[0166] The process of generating and preloading the intervention device parameter plan is as follows:

[0167] E1: Analyze the optimal intervention plan and identify the required physical intervention equipment and its operating parameters; the physical intervention equipment includes an infusion pump, a uterine contraction inhibitor delivery device, a fetal monitor, and a postpartum uterine compression device;

[0168] For example, taking emergency intervention for postpartum hemorrhage as an example, the optimal intervention plan is "intravenous infusion of crystalloid solution + uterine contraction inhibitor + postpartum uterine compression". The corresponding physical intervention equipment includes an infusion pump, a uterine contraction inhibitor delivery device, and a postpartum uterine compression device. The core operating parameters include the infusion rate of the infusion pump, the dose and infusion speed of oxytocin, and the pressure and frequency of the uterine compression device.

[0169] E2: Based on the prediction results of the maternal-fetal physiological system mechanism model, calculate the activation sequence, dosage, flow rate or pressure parameters of the physical intervention equipment to form a structured parameter plan;

[0170] E3: The parameter plan is preloaded to the corresponding physical intervention device through the medical device communication protocol, so that the physical intervention device is in standby mode.

[0171] For example, the pre-loading process of the parameter plan for the oxytocin administration device for postpartum hemorrhage patients is as follows: The parameter plan includes "Drug name: oxytocin; Initial dose: 20U; Solvent: 500 ml of normal saline; Initial infusion rate: 10 ml / min; Adjustment gradient: if there is no improvement after 30 minutes, increase to 15 ml / min; Maximum infusion rate: 20 ml / min". This is transmitted to the uterine contraction inhibitor administration device via the HL7FHIR protocol. After receiving the data, the device automatically configures parameters such as drug concentration and infusion rate. The indicator light shows the standby status, and the infusion can be started after confirmation by medical staff.

[0172] The fetal monitor's parameter pre-loading includes: "Monitoring frequency: once every 10 seconds; fetal heart rate alarm threshold: 110~160 beats / min; uterine contraction alarm threshold: greater than or equal to 5 times / hour; abnormal judgment criteria: disappearance of fetal heart rate variability, late deceleration". After the device is pre-loaded, it will automatically adjust the monitoring mode and monitor fetal heart rate and uterine contraction data in real time. Once an abnormality occurs, an alarm will be triggered immediately.

[0173] Furthermore, the digital health twin also integrates a deep reinforcement learning-based policy optimization module for optimizing intervention plans, specifically:

[0174] (1) Use the aforementioned mechanism model as an environment simulator;

[0175] (2) Define the state space as the user's health information profile, the action space as a predefined combination of intervention measures, and the reward function as the degree of conformity between the evolution trajectory and the clinical safety threshold;

[0176] (3) A proximal policy optimization algorithm is used to train an agent; the agent learns the strategy of choosing the optimal action in a given state by interacting with an environment simulator. The proximal policy optimization algorithm is a conventional means that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0177] (4) When the digital health twin receives the warning instruction, it calls the intelligent agent to generate one or more candidate intervention schemes for use when simulating the expected effect.

[0178] Regarding S4 above:

[0179] The dynamic calibration process in S4 is as follows:

[0180] S4.1: After implementing the comprehensive management strategy, real-time monitoring data is collected through the multimodal wearable monitoring device group;

[0181] S4.2: Input the real-time monitoring data into the adaptive pregnancy and childbirth risk model, recalculate the actual health index score, and compare the actual health index score with the health index score predicted by the digital health twin at the same time to calculate the deviation value of the health index.

[0182] Furthermore, the magnitude of the deviation value reflects the difference between the prediction results of the digital health twin and the actual health status. The smaller the deviation value, the more accurate the prediction; the larger the deviation value, the more accurate the prediction, and the more necessary the dynamic calibration.

[0183] S4.3: If the deviation value of the health index continuously exceeds the preset dynamic calibration threshold and reaches the set time, the calibration process is triggered; the calibration process includes: using real-time monitoring data as the observation value, and using the extended Kalman filter algorithm to update the state parameters of the maternal-fetal physiological system mechanism model in the digital health twin in reverse.

[0184] Furthermore, the dynamic calibration threshold and setting time can be dynamically adjusted according to the patient's risk level: the calibration threshold for high-risk and critical-risk patients is set at 0.08, and the setting time is 15 minutes to ensure rapid response to prediction deviations; the calibration threshold for low-risk and medium-risk patients is set at 0.12, and the setting time is 45 minutes to avoid intervention fluctuations caused by over-calibration.

[0185] S4.4: Based on the updated status parameters, rerun the maternal-fetal physiological system mechanism model, correct the evolution trajectory of health index and key physiological parameters, and adjust the monitoring parameter adjustment instructions and intervention equipment parameter contingency plans in the comprehensive management strategy.

[0186] Furthermore, the specific implementation steps of the extended Kalman filter algorithm are as follows:

[0187] (1) Define the state parameters of the mechanism model in the digital health twin at time k as a state vector, and define the collected real-time monitoring data as an observation vector;

[0188] (2) Discretize the differential equations of the mechanism model and establish the state transition equations;

[0189] (3) Establish the observation equation based on the sensor measurement principle;

[0190] (4) In each data fusion cycle, perform the prediction step: use the state transition equation to predict the state vector and error covariance matrix at time k+1;

[0191] (5) Perform update steps: Calculate the Kalman gain, and update the state vector and error covariance matrix using the residual between the observation vector and the predicted observation and the Kalman gain;

[0192] (6) Inject the updated state vector into the mechanism model of the digital health twin to complete the dynamic calibration of the state parameters.

[0193] Example 2:

[0194] Please see Figure 3 Another embodiment of the present invention provides a monitoring and management system for abnormal conditions and health information during pregnancy and childbirth, comprising:

[0195] Transmission and fusion module, risk assessment module, hierarchical early warning module, digital health twin module, dynamic calibration module, and device linkage module;

[0196] The transmission and fusion module 10 uses a timestamp synchronization algorithm to align monitoring data with hospital time-series medical record data, and merges them to form a complete user health information profile;

[0197] Risk assessment module 20 is used to analyze health information profiles and output health index scores and abnormal judgment results with clinical interpretation labels;

[0198] The hierarchical early warning module 30 is used to establish a four-level risk mapping relationship of low, medium, high and critical, generate corresponding early warning instructions, and synchronize relevant information to the digital health twin.

[0199] The digital health twin module 40 is used to simulate the expected effects of various intervention programs, screen the optimal program, and generate instructions for adjusting monitoring parameters and intervention equipment parameter plans, providing personalized comprehensive management strategies.

[0200] The dynamic calibration module 50 is used to compare real-time monitoring data with twin prediction results. When the deviation exceeds the threshold, the model parameters are calibrated by the extended Kalman filter algorithm to adjust the comprehensive management strategy.

[0201] The device linkage module 60 is used to receive intervention device parameter plans and preload them into devices such as infusion pumps and fetal monitors, so that they are in a standby state and can quickly respond to intervention needs.

[0202] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A method for monitoring and managing abnormal states and health information during pregnancy, characterized by, include: S1: Synchronously collect user monitoring data through a multimodal wearable monitoring device group; S2: The monitoring data is integrated with the time-series medical record data in the hospital information system through a timestamp synchronization algorithm to form a user health information profile, which is then input into the adaptive pregnancy and childbirth risk model for analysis, and the personal health risk assessment result is output. The personal health risk assessment results include a comprehensive health index score and abnormality judgment results with clinical interpretation labels; The adaptive pregnancy and childbirth risk model is dynamically updated based on the fusion features. The clinical interpretation labels are used to indicate to clinicians the main physiological parameters and their temporal context that cause changes or abnormalities in health index scores. S3: Based on individual health risk assessment results and clinical interpretation labels, hierarchical early warning is generated, and early warning instructions are generated. At the same time, the early warning instructions and user health information profiles are input into a pre-built digital health twin and the expected effects of different intervention programs are simulated to generate a comprehensive management strategy that includes instructions for adjusting monitoring parameters and intervention equipment parameter plans. S4: Send the monitoring parameter adjustment command to the wearable monitoring device group, preload the intervention device parameter plan into the physical intervention device, compare the newly collected real-time monitoring data with the prediction results of the digital health twin, and recalculate the health index based on the new data. If the actual value of the health index deviates from the predicted value, dynamically calibrate the status parameters and comprehensive management strategy of the digital health twin.

2. The method of claim 1, wherein the abnormal state of pregnancy is one of a plurality of abnormal states of pregnancy, and the health information is one of a plurality of health information. The timestamp synchronization algorithm in S2 is specifically as follows: The timestamps of each data channel are parsed from the monitoring data; Extract time-series medical record data with system timestamps from the parent medical record database of the hospital information system; the time-series medical record data includes prenatal check-up records, medication records, laboratory test results, and imaging reports; The timestamps of the monitoring data and the system timestamps of the time-series medical records are aligned using a dynamic time warping algorithm to obtain the timestamp-aligned monitoring values ​​and medical record text descriptions. The timestamp-aligned monitoring values ​​and medical record text descriptions are then mapped to feature vectors of a unified dimension through an embedding layer and concatenated along the feature dimension to form a user health information profile.

3. The method of claim 1, wherein the abnormal state of pregnancy is one of a plurality of abnormal states of pregnancy, and the health information is one of a plurality of health information. The adaptive pregnancy and childbirth risk model comprises a feature encoder, a spatiotemporal attention network, a health index calculation module, and a risk classifier connected in sequence. The feature encoder is composed of a cascaded one-dimensional convolutional neural network and a long short-term memory network; the input of the one-dimensional convolutional neural network receives the user's health information profile and extracts local temporal features through its convolutional and pooling layers. Its output is connected to the input of the long short-term memory network to capture long-term dependencies and output the encoded feature sequence; The input of the spatiotemporal attention network is connected to the output of the feature encoder, and it contains a temporal attention module and a feature attention module. The temporal attention module is used to calculate the importance weights of different time steps in the encoded feature sequence. The feature attention module is used to calculate the importance weights of different physiological feature dimensions at the same time step. Weighted fusion generates context-enhanced feature representations; The input of the health index calculation module is connected to the output of the spatiotemporal attention network. It consists of a fully connected layer and a sigmoid activation function, and is used to map the context-enhanced feature representation into a scalar value between 0 and 1, which serves as the health index score. The input of the risk classifier is connected to the output of the health index calculation module. It consists of a fully connected layer and a Softmax function. It is used to receive feature representations associated with the health index score and output anomaly judgment results with clinical interpretation labels. The clinical interpretation labels are used to indicate the main feature dimensions that cause the anomaly.

4. The monitoring management method for abnormal state during pregnancy according to claim 1, wherein, The hierarchical early warning system in S3 specifically refers to: Establish a mapping relationship between health index score ranges and abnormal judgment results and preset risk level thresholds; the risk levels include low risk, medium risk, high risk and critical risk; Based on the health index score and the clinical interpretation label, the current risk level and main abnormal characteristics are determined; If the risk level is determined to be low, a daily follow-up reminder instruction will be generated. If the risk level is determined to be medium, an early warning instruction for enhanced monitoring and outpatient follow-up will be generated. If the risk level is determined to be high, an early warning instruction for inpatient observation and special examinations will be generated; if the risk level is determined to be critical, an early warning instruction for immediate medical intervention will be generated. The warning instructions and their corresponding risk levels, health index scores, anomaly judgment results, clinical interpretation labels, and user health information profiles are all input into the digital health twin.

5. The monitoring management method for abnormal state during pregnancy according to claim 1, wherein, The construction and simulation process of the digital health twin is as follows: Based on the maternal medical records and historical monitoring data, a maternal-fetal physiological system mechanism model described by a set of coupled differential equations is constructed. Upon receiving a clinical warning instruction, the model simulates and executes various intervention programs that conform to clinical guidelines in the digital space, using the current patient health information profile as the initial state, and predicts the evolution trajectory of health indices under each intervention program. The intervention programs include adjusting the maternal position, adjusting the medication regimen, adjusting the fluid resuscitation rate, and preparing for surgical intervention. By comparing the evolution trajectory and clinical safety thresholds of each intervention program, the optimal intervention program that enables the health index to recover to the safe range as quickly as possible or maintain the most stable state is selected, and a comprehensive management strategy is generated that includes the monitoring parameter adjustment instructions and intervention equipment parameter contingency plans corresponding to the optimal intervention program.

6. The method for monitoring and managing abnormal conditions and health information during pregnancy and childbirth as described in claim 5, characterized in that, The specific process for generating the monitoring parameter adjustment instruction is as follows: The evolution trajectory corresponding to the optimal intervention plan is analyzed to identify the monitoring target values ​​that need to be adjusted to maintain physiological parameters within a safe range; The monitored target value is converted into configuration parameters for specific sensors in the multimodal wearable monitoring device group, and the configuration parameters are encapsulated into monitoring parameter adjustment instructions; the configuration parameters include sampling frequency, alarm threshold, and data upload interval.

7. The method for monitoring and managing abnormal conditions and health information during pregnancy and childbirth as described in claim 6, characterized in that, The process of generating and preloading the intervention device parameter plan is as follows: The optimal intervention plan is analyzed to identify the required physical intervention equipment and its operating parameters; the physical intervention equipment includes an infusion pump, a uterine contraction inhibitor delivery device, a fetal monitor, and a postpartum uterine compression device; Based on the prediction results of the maternal-fetal physiological system mechanism model, the activation sequence, dosage, flow rate or pressure parameters of the physical intervention equipment are calculated to form a structured parameter plan; The parameter plan is preloaded into the corresponding physical intervention device via the medical device communication protocol, so that the physical intervention device is in standby mode.

8. The method for monitoring and managing abnormal conditions and health information during pregnancy and childbirth as described in claim 1, characterized in that, The dynamic calibration process in S4 is as follows: After implementing the comprehensive management strategy, real-time monitoring data is collected through a multimodal wearable monitoring device group; The real-time monitoring data is input into the adaptive pregnancy and childbirth risk model to recalculate the actual health index score. The actual health index score is then compared with the health index score predicted by the digital health twin at the same time to calculate the deviation value of the health index. If the deviation value of the health index exceeds the preset dynamic calibration threshold continuously for a set period of time, the calibration process will be triggered. The calibration process includes: using real-time monitoring data as observations, and employing an extended Kalman filter algorithm to back-update the state parameters of the maternal-fetal physiological system mechanism model in the digital health twin; Based on the updated status parameters, the maternal-fetal physiological system mechanism model was rerun to correct the evolution trajectory of health indices and key physiological parameters, and the monitoring parameter adjustment instructions and intervention equipment parameter contingency plans in the comprehensive management strategy were adjusted.

9. The method for monitoring and managing abnormal conditions and health information during pregnancy and childbirth as described in claim 3, characterized in that, The generation of the clinical interpretability labels also employs a model interpretability method based on sapride and interpretation, specifically: For the anomaly determination results output by the risk classifier, select the top N feature dimensions with the highest contribution and their corresponding key time steps; Calculate the Shapley value of each feature in the user health information profile at the key time step; the Shapley value is used to quantify the contribution of the user health information profile to the anomaly detection result; The user health information profiles are sorted from highest to lowest according to the Shapley value, and the top N features and their corresponding time steps are marked as the core content of the clinical interpretation label.

10. A monitoring and management system for abnormal conditions and health information during pregnancy and childbirth, used to implement the monitoring and management method for abnormal conditions and health information during pregnancy and childbirth as described in any one of claims 1-9, characterized in that, include: Transmission and fusion module, risk assessment module, hierarchical early warning module, digital health twin module, dynamic calibration module, and device linkage module; The transmission and fusion module uses a timestamp synchronization algorithm to align monitoring data with hospital time-series medical record data, and fuses them to form a complete user health information profile. The risk assessment module is used to analyze the health information profile and output the health index score and the abnormal judgment result with clinical interpretation label; The hierarchical early warning module is used to establish a four-level risk mapping relationship of low, medium, high and critical, generate corresponding early warning instructions, and synchronize them to the digital health twin. The digital health twin module is used to simulate the expected effects of various intervention programs, screen the optimal program, and generate instructions for adjusting monitoring parameters and intervention equipment parameter plans, providing personalized comprehensive management strategies. The dynamic calibration module is used to compare real-time monitoring data with twin prediction results. When the deviation exceeds the threshold, the model parameters are calibrated by the extended Kalman filter algorithm to adjust the comprehensive management strategy. The device linkage module is used to receive intervention device parameter plans and preload them into the physical intervention device, so that it is in a standby state.

Images