A fetal distress early warning and identification system based on fetal heart monitoring

Abstract

The application relates to the technical field of fetal monitoring, and discloses a fetal distress early warning and identification system based on fetal heart monitoring, which comprises a data acquisition module, a data preprocessing module, a warning calculation module, a warning identification module and a man-machine interaction module; through multi-modal maternal-fetal synchronous sensing, individualized baseline calibration and TGAT warning technology with fusion timing attention, different fetal distress detection dimensions can be covered, not only traditional indexes such as fetal heart rate and uterine contraction are paid attention to, but also elements such as maternal blood oxygen, umbilical blood flow and uterine tension are included, and the timing pathological characteristics of maternal-fetal coupling can be deeply captured, so that the warning result can more accurately reflect the fetal hypoxia state, and a scientific basis can be provided for early intervention in obstetrics.

Description

Technical Field

[0001] This invention relates to the field of fetal monitoring technology, and more specifically discloses an early warning and identification system for fetal distress based on fetal heart rate monitoring. Background Technology

[0002] With the development of medicine and the advancement of medical sensor technology, electronic fetal heart rate monitoring has become a core means for obstetrics to assess the fetal health status in utero. By continuously monitoring changes in fetal heart rate and its correspondence with uterine contractions and fetal movements, it indirectly reflects whether the fetus is hypoxic or at risk of acidosis in utero.

[0003] The existing patent document with authorization announcement number CN120859434A discloses "A Precise Early Warning System for Fetal Distress Based on Sensor Fusion," which includes: a detection module implanted in the user's body to acquire key physiological information such as electromyographic fluctuations, vital signs, fetal movement frequency, and fetal position parameters in the pregnant woman's abdominal region in real time. A central control module, as the core processing unit, integrates multi-source data and uses an algorithm model to calculate a scoring index representing the state of fetal distress. The early warning module determines the risk level based on this score and promptly sends the early warning information to the medical reception system or user terminal via a built-in communication module, realizing dynamic monitoring and real-time early warning of fetal distress.

[0004] The patent document with authorization announcement number CN116458864A discloses "an intelligent fetal health early warning system," which includes a fetal intelligent detection terminal, a fetal detection data processing terminal, an abnormality early warning terminal, and a risk assessment terminal. The fetal intelligent detection terminal is used to perform intelligent detection on the fetus and generate fetal data information. The fetal detection data processing terminal is used to process the detection data based on the fetal data information and generate detection data processing information. The abnormality early warning terminal is used to perform early warning analysis based on the detection data processing information and generate abnormality early warning information. The risk assessment terminal is used to perform fetal risk assessment based on the fetal data information, detection data processing information, and abnormality early warning information and generate corresponding fetal risk assessment information.

[0005] While existing technologies have achieved the detection and correction of fetal heart rate, fetal movement, and intrauterine pressure data through multi-terminal layered design, and can dynamically adjust detection parameters based on the individual circumstances of pregnant women, thus improving the basic accuracy of fetal health early warning, and employ in-vivo implantable multi-sensor fusion technology to achieve the collection and coupled modeling of multimodal data such as maternal electromyography fluctuations and fetal position, and improve the timeliness of risk identification and early warning response through quantitative scoring and graded early warning mechanisms, thus constructing a closed-loop control from preliminary risk identification to early warning response, existing technologies have not achieved deep temporal coupling feature extraction and analysis of multiple maternal and fetal indicators, lack an individualized dynamic calibration mechanism for fetal heart rate baseline based on individual fetal characteristics and real-time maternal physiological status, have a single early warning dimension, and have not formed a clinically applicable quantitative identification and scoring system, and the synchronization of data collection, iterative optimization of models, and clinical adaptability are insufficient, making it difficult to achieve ultra-early accurate early warning and scientific quantitative identification of fetal distress, prone to early warning lag, and the identification results are easily influenced by human experience, failing to provide accurate and unified decision-making basis for clinical intervention. Summary of the Invention

[0006] The main technical problem solved by this invention is to provide an early warning and identification system for fetal distress based on fetal heart rate monitoring, which can solve the problems mentioned in the background art.

[0007] To address the aforementioned technical problems, according to one aspect of the present invention, more specifically, a fetal distress early warning and identification system based on fetal heart rate monitoring, comprising: a data acquisition module, a data preprocessing module, an early warning calculation module, an early warning identification module, and a human-computer interaction module; the data acquisition module synchronously acquires multi-dimensional physiological data of the fetus and mother; the data preprocessing module performs denoising, temporal alignment, and standardization processing on the raw maternal and fetal data, extracting clinically interpretable basic maternal and fetal characteristics; the early warning calculation module completes fetal heart rate baseline calibration, maternal and fetal indicator feature extraction, and calculation of fetal distress risk probability based on the basic maternal and fetal characteristics; the early warning identification module combines the fetal distress risk probability and risk development trend to achieve graded early warning, and completes the scientific identification and classification of fetal distress through multi-indicator quantitative scoring; the human-computer interaction module provides visualized monitoring data, multimodal early warning prompts, and clinical intervention suggestions at the medical staff end and the patient end, respectively.

[0008] Furthermore, the data acquisition module includes: a fetal data acquisition module and a maternal data acquisition module; Fetal data acquisition module: Acquires raw fetal heart rate signals, quantifies fetal movement signals, and collects fetal umbilical artery blood flow velocity and blood flow-related index data; Maternal data acquisition module: Collects maternal blood oxygen saturation and peripheral pulse rate, uterine tension and uterine contraction pressure, maternal systolic blood pressure, diastolic blood pressure and mean arterial pressure data.

[0009] Furthermore, the data preprocessing module includes: a fetal heart rate signal denoising module, a time sequence alignment module, and a standardization and feature mapping module; Fetal heart rate signal denoising module: Receives data output from the data acquisition module, performs noise removal processing on the raw fetal heart rate signal, and retains the subtle features of fetal heart rate variability; Timing alignment module: Receives data output from the fetal heart rate signal denoising module, performs interpolation and resampling on maternal and fetal data collected at different sampling rates, and completes missing data; Standardization and Feature Mapping Module: Receives data output from the temporal alignment module, performs standardization transformation on it, and extracts clinically interpretable multidimensional maternal-fetal basic features.

[0010] Furthermore, the early warning calculation module includes: a baseline calibration module, a feature extraction module, and a risk early warning calculation module; Baseline calibration module: Receives data output from the data preprocessing module, generates individualized fetal heart rate baseline thresholds based on individual fetal characteristics and maternal physiological state, and dynamically adjusts them; Feature extraction module: Receives data output from the baseline calibration module, performs time-series analysis on it, extracts the time-series coupling features between multiple maternal and fetal indicators, and outputs them to the risk warning calculation module; Risk warning calculation module: Based on the features output by the feature extraction module, a dynamic graph network is constructed to calculate and output the probability of fetal distress risk, and at the same time output the risk contribution index.

[0011] Furthermore, the early warning identification module includes: a graded early warning module and an identification scoring module; Tiered early warning module: Receives the fetal distress risk probability and risk contribution index output by the early warning calculation module, performs tiered early warning, and outputs the early warning level, risk contribution index and corresponding clinical intervention suggestions; The identification and scoring module receives data from the graded early warning module, constructs a quantitative scoring system based on core maternal and fetal indicators to complete the fetal distress score, and combines the scoring results with risk contribution indicators to classify the types of fetal distress and rank abnormal indicators.

[0012] Furthermore, the human-computer interaction module includes: a doctor-side module and a patient-side module; Doctor-side module: Displays real-time curves of multimodal maternal-fetal monitoring data, individualized fetal heart rate baseline dynamic curves, and early warning identification results, supporting data playback, feature annotation, and manual review; Patient-side module: Receives data output from the doctor-side module, displays simplified maternal and fetal monitoring data and basic risk warnings, supports emergency alarms for abnormal situations, and provides a one-click function to contact medical staff.

[0013] Furthermore, it also includes edge-cloud collaboration modules: local inference module, hospital data management module, and data optimization module; Local inference module: Equipped with a lightweight data preprocessing algorithm and an early warning and identification inference model, it enables real-time local fetal distress early warning and identification calculation; Hospital Data Management Module: Receives data output from the local inference module, completes data storage and management, and performs local parameter fine-tuning on the local inference model; Data optimization module: Receives data output from the hospital data management module, continuously optimizes the core model, performs lightweight processing on the optimized model, and distributes the lightweight model data to the local inference module and the hospital data management module respectively.

[0014] The beneficial effects of this invention's fetal distress early warning and identification system based on fetal heart rate monitoring are as follows: Through multimodal maternal-fetal synchronous sensing, individualized baseline calibration, and TGAT early warning technology incorporating temporal attention, it can cover different dimensions of fetal distress detection. It not only focuses on traditional indicators such as fetal heart rate and uterine contractions but also incorporates factors such as maternal blood oxygenation, umbilical blood flow, and uterine tension, and deeply captures the temporal pathological characteristics of maternal-fetal coupling, making the early warning results more accurately reflect the fetal hypoxic state and providing a scientific basis for early obstetric intervention. Furthermore, through an edge-cloud collaborative architecture and multi-system interface technology, it achieves fetal heart rate monitoring... The effective integration of the nursing system with the hospital's HIS / LIS system breaks down data barriers, enabling real-time sharing and interaction of maternal and fetal monitoring data among the systems. This improves the efficiency and scientific rigor of obstetric clinical decision-making and enhances the accuracy and adaptability of early warning models. Simultaneously, through multimodal early warning prompts, FDS quantitative scoring, and real-time risk trend visualization technology, it provides richer and more intuitive assessment methods, clearly displaying the dynamic changes in fetal distress risk. Furthermore, the decision support system based on real-time data and the TGAT model can quickly provide scientific intervention suggestions in emergency situations, significantly improving the decision-making level and emergency response capabilities of obstetric management. Attached Figure Description

[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0016] Figure 1 This is a schematic diagram of the system module architecture. Detailed Implementation

[0017] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0018] According to one aspect of the invention, such as Figure 1As shown, a fetal distress early warning and identification system based on fetal heart rate monitoring is provided, including: a data acquisition module that simultaneously collects multi-dimensional physiological data of the fetus and the mother. This module includes: Fetal data acquisition module: Acquires raw fetal heart rate signals, quantifies fetal movement signals, and collects fetal umbilical artery blood flow velocity and blood flow-related index data; Specifically, a sensing device with dual 2MHz ultrasonic transducers is used to extract the raw fetal heart rate signal. This device can track fetal heart sounds and suppress environmental interference such as uterine contractions and maternal vascular murmurs through a signal anti-interference algorithm, achieving accurate acquisition of the raw fetal heart rate signal. Simultaneously, a MEMS triaxial accelerometer is attached to the corresponding position of the fetal limbs on the mother's abdomen to capture the physical characteristics of fetal movement, such as acceleration, duration, and frequency. A quantization conversion algorithm transforms these physical characteristics into standardized quantized fetal movement signals, as shown in the following formula: In the formula, This refers to the acceleration during fetal movement. , , The X, Y, and Z axis acceleration values ​​were collected by a MEMS triaxial accelerometer. Finally, a 4MHz pulse Doppler ultrasound sensor was used to achieve non-invasive detection of fetal umbilical artery blood flow velocity. At the same time, the detected blood flow velocity data was analyzed in real time through a built-in algorithm to calculate blood flow-related indices such as umbilical artery pulsatility index and resistance index. All fetal physiological data were collected using a uniform sampling rate of 100Hz.

[0019] Maternal data acquisition module: Collects maternal blood oxygen saturation and peripheral pulse rate, uterine tension and uterine contraction pressure, maternal systolic blood pressure, diastolic blood pressure and mean arterial pressure data; Specifically, a dual-wavelength photoelectric sensor using red / infrared light is employed in a clip-on design to collect maternal blood oxygen saturation and peripheral pulse rate. This device is attached to the mother's fingertip for signal acquisition, with a sampling rate of 100Hz. The photoelectric signal analysis algorithm directly outputs digitized maternal blood oxygen saturation and peripheral pulse rate data. The specific algorithm is shown below: In the formula, The ratio of red light to infrared light absorption. This refers to the AC component of the red light photoelectric signal. This refers to the DC component of the red light photoelectric signal. The AC component of the infrared photoelectric signal. The system uses the DC component of infrared photoelectric signals. Simultaneously, a pressure-sensitive thin-film sensor is attached to the fundus of the uterus. This integrated sensor design allows for the simultaneous capture of dynamic changes in uterine contraction pressure (TOCO) and basal uterine tension at the same acquisition point, with a sampling rate of 100Hz. A pressure signal calibration algorithm reconstructs the true data on uterine myometrial perfusion and contractions. Then, a non-invasive blood pressure sensor based on the oscillation principle collects maternal systolic and diastolic blood pressure, with a sampling rate of 1Hz adapted to the dynamic changes in maternal blood pressure. The device's built-in pressure sensing and data analysis algorithms calculate and output the maternal mean arterial pressure based on the collected systolic and diastolic pressure data. All collected raw physiological data is transmitted locally in real-time via Bluetooth Low Energy 5.0, and 5G slicing transmission technology enables high-speed, low-latency uploading of collected data to the cloud platform, suitable for clinical applications of mobile monitoring during labor.

[0020] The data preprocessing module performs denoising, temporal alignment, and standardization on the raw maternal-fetal data to extract clinically interpretable basic maternal-fetal characteristics. This module includes: Fetal heart rate signal denoising module: Receives data output from the data acquisition module, performs noise removal processing on the raw fetal heart rate signal, and retains the subtle features of fetal heart rate variability; First, the wavelet basis function that is precisely matched with the noise characteristics of the dual ultrasound transducer sensing device is retrieved. Then, the original fetal heart rate signal is subjected to multi-scale wavelet decomposition, which decomposes the original signal into high-frequency noise components and low-frequency effective signal components. The high-frequency components correspond to various interference signals such as uterine contraction murmurs, maternal vascular murmurs, and environmental electromagnetic noise, while the low-frequency components are the effective signals that carry the core characteristics of fetal heart rate variation. Then, based on the factory noise calibration parameters of the dual ultrasound transducer sensing device and the noise samples collected clinically, a dynamic adaptive threshold is set to perform threshold quantization screening on the decomposed high-frequency noise components, accurately removing invalid noise points that exceed the threshold, while completely preserving the small effective feature signals representing the fine and coarse variations of fetal heart rate in the high-frequency components, thus avoiding the loss of early distress characteristics. Finally, the high-frequency components after thresholding and the unprocessed low-frequency effective components are subjected to inverse wavelet transform to reconstruct a denoised pure fetal heart rate signal. This signal can completely restore the various dynamic variation characteristics of the fetal heart rate, providing a high-precision signal basis for subsequent time series analysis.

[0021] Timing alignment module: Receives data output from the fetal heart rate signal denoising module, performs interpolation and resampling on maternal and fetal data collected at different sampling rates, and completes missing data; Specifically, the process begins by using a unified timestamp based on a crystal oscillator clock to precisely anchor all maternal and fetal data in the time dimension, assigning a unique time axis coordinate to each data set and establishing a unified time reference system for all maternal and fetal data. Then, linear interpolation is used to synchronously resample maternal and fetal physiological data at different sampling rates. Data such as fetal heart rate, fetal movement, maternal blood oxygen saturation, and uterine tension from a 100Hz high sampling rate, along with maternal blood pressure data from a 1Hz sampling rate, are uniformly resampled to a standard 10Hz sampling rate. This significantly reduces the computational load of subsequent algorithms, improves processing efficiency, and preserves the temporal dynamic characteristics of the data to the greatest extent possible. To address the issue of temporary data loss due to fetal movement obstruction or temporary signal interruptions from sensing devices, an autoregressive completion algorithm based on adjacent temporal features is employed. Using feature data from several (e.g., five) consecutive effective time nodes before and after the missing data point as training samples, a local autoregressive prediction model is constructed. The model calculates and completes the missing data points in real time, ensuring the continuity, integrity, and feature authenticity of all temporal data and avoiding interference from data distortion with subsequent analysis results.

[0022] Standardization and Feature Mapping Module: Receives data output from the temporal alignment module, performs standardization transformation on it, and extracts clinically interpretable multidimensional maternal-fetal baseline features; First, the Z-score normalization method was used to fully normalize all resampled maternal-fetal time-series data to eliminate differences in dimensions and orders of magnitude among different physiological indicators. The specific calculation formula is as follows: In the formula, This is the original data. The mean, The standard deviation is used to uniformly map all maternal and fetal physiological data to the standard numerical range of [-1, 1], which meets the input data format requirements of the subsequent algorithm model. Finally, clinically interpretable basic features were accurately extracted from the normalized maternal-fetal time-series data. Core basic features were extracted, including fetal heart rate variability, duration of fetal heart rate acceleration / deceleration, umbilical blood flow pulsatility index, and umbilical blood flow resistance index. Maternal features included blood oxygen saturation variability, mean uterine tension, uterine tension fluctuation coefficient, blood pressure fluctuation coefficient, and peak uterine contraction pressure. All extracted basic features were highly compatible with the pathophysiological characteristics of fetal distress, providing a standardized, clinical, and systematic input feature set for the subsequent early warning calculation module.

[0023] The early warning calculation module, based on the aforementioned maternal and fetal baseline characteristics, completes fetal heart rate baseline calibration, extracts maternal and fetal indicator features, and calculates the probability of fetal distress risk. This module includes: Baseline calibration module: Receives data output from the data preprocessing module, generates individualized fetal heart rate baseline thresholds based on individual fetal characteristics and maternal physiological state, and dynamically adjusts them; First, individual fetal characteristics such as gestational age, ultrasound-estimated weight, and fetal developmental grade, along with real-time maternal physiological characteristics such as baseline maternal blood oxygen saturation, mean maternal arterial pressure, and baseline uterine tension, are used as model input variables. This set of input variables is derived from standardized maternal-fetal baseline characteristics output by the data preprocessing module. Then, the Gradient Boosting Regression Tree (GBRT) algorithm is used to build the baseline calculation model, as shown in the following formula: In the formula, This represents the overall predicted value after the m-th iteration. This represents the predicted value of the model after the (m-1)th iteration. For learning rate, The regression tree-based learner obtained in the m-th training round is used. Then, the clinical dataset labeled with maternal and fetal features and the corresponding actual values ​​of the fetal heart rate baseline is input into the model for training. At the same time, the rules and constraints formulated by the obstetric clinical guidelines are incorporated into the model training process to avoid the model output results deviating from the actual clinical diagnosis and treatment standards. After training, the model can output dynamic thresholds of the upper and lower limits of the fetal heart rate baseline that are adapted to the individual. Finally, the output baseline threshold is dynamically adjusted based on real-time changes in the mother's physiological state. When changes in the mother's physiological state such as decreased blood oxygen or abnormal blood pressure are detected, the baseline threshold is adaptively corrected by adjusting the step size at 1 time / minute, thereby achieving real-time dynamic calibration of the fetal heart rate baseline and adapting to the dynamic changes in the mother and fetus's state during labor.

[0024] Feature extraction module: Receives data output from the baseline calibration module, performs time-series analysis on it, extracts the time-series coupling features between multiple maternal and fetal indicators, and outputs them to the risk warning calculation module; The maternal-fetal time-series data output by the baseline calibration module after individualized baseline correction is integrated with the preprocessed clinically interpretable maternal-fetal basic characteristics into a time-series feature sequence, which serves as the basic input data for feature extraction. This data sequence completely preserves the time dimension change information of maternal-fetal indicators. Temporal convolutional networks (TCNs) perform deep analysis of single-index feature sequences by stacking dilated convolutional layers, capturing long-term temporal correlation features of single indicators such as fetal heart rate variability and maternal blood oxygenation changes. For example, they can extract the trend of fetal heart rate variability within a 5-minute time window. At the same time, they solve the gradient vanishing problem of deep networks through residual connections, ensuring the effectiveness of long-term feature extraction. In addition, a cross-fetal attention mechanism is introduced to quantify the correlation between all maternal and fetal indicators. Strong coupling correlations between maternal and fetal indicators are identified through attention weight allocation, such as calculating the coupling coefficient between the decrease in maternal blood oxygen saturation and the disappearance of fetal heart rate fine variability. Finally, the long-term features of single indicators and the coupling features of maternal and fetal indicators are fused to extract the time-series coupling features of maternal and fetal multi-indicators with clinical and pathological significance, which are used as the core input for risk warning calculation.

[0025] Risk warning calculation module: Based on the features output by the feature extraction module, a dynamic graph network is constructed to calculate and output the probability of fetal distress risk, and at the same time output the risk contribution index; Specifically, the maternal-fetal temporal coupling features output by the feature extraction module are used as nodes in a graph network. Based on the pathological association rules of fetal distress as defined by obstetric clinicians, corresponding pathological association weights are assigned to each node and constructed as edges. In this way, a maternal-fetal dynamic graph network adapted to the pathological mechanism of fetal distress is built. This network can accurately characterize the pathological association between maternal and fetal indicators. Next, a temporal attention mechanism was incorporated into the dynamic graph network, assigning differentiated attention weights to graph networks with different time windows of 1 minute, 3 minutes, and 5 minutes. The graph network with a short time window of 1 minute was given higher weights to focus on capturing subtle abnormal features in the early stages of fetal distress. Then, a clinically labeled dataset containing early, middle, and late fetal distress samples was input into the graph neural network (TGAT) with temporal attention fusion for model training. After training, the model can perform real-time calculations on the coupled features of the input and output the fetal distress risk probability of 0-100%. At the same time, through feature importance analysis, the core risk contribution indicators that lead to fetal distress risk were identified and output to clarify the specific risk triggers.

[0026] The early warning and identification module combines the probability and trend of fetal distress risk to achieve graded early warning, and uses multi-indicator quantitative scoring to scientifically identify and classify fetal distress. This module includes: Tiered early warning module: Receives the fetal distress risk probability and risk contribution index output by the early warning calculation module, performs tiered early warning, and outputs the early warning level, risk contribution index and corresponding clinical intervention suggestions; First, the 0-100% fetal distress risk probability value and corresponding risk contribution index are received from the early warning calculation module. At the same time, the multi-time window risk probability time series data output by the TGAT model in the risk early warning calculation module are retrieved. By calculating the slope and magnitude of the change in risk probability within the continuous time window, the risk development trend is quantitatively determined to be one of four categories: stable, declining, rising, or rapidly rising, thus providing basic data support for dual-dimensional early warning judgment. Then, based on the preset classification rules, the risk probability value range is precisely matched with the risk development trend to define four warning levels: no risk, low risk, medium risk, and high risk. Among them, no risk is a risk probability of <10% and a stable / decreasing trend, low risk is 10%-30% and a stable trend, medium risk is 30%-60% and an increasing trend, and high risk is >60% and a rapidly increasing trend, thus completing the automatic determination of the warning level. Finally, the determined warning levels, core risk contribution indicators, and intervention recommendations adapted to obstetric clinical guidelines are linked and integrated. Warning prompts are output through a multimodal approach of sound, light, and text. Different warning levels are matched with different sound and light prompt frequencies and text annotation styles. At the same time, the core risk causes and corresponding clinical intervention measures are clearly marked in the prompt information, providing medical staff with direct intervention decision-making basis.

[0027] The identification and scoring module receives data from the graded early warning module, constructs a quantitative scoring system based on core maternal and fetal indicators to complete the fetal distress score, and combines the scoring results with risk contribution indicators to classify the types of fetal distress and rank abnormal indicators. First, clinically interpretable core maternal-fetal indicators are extracted from the data output by the graded early warning module. These indicators include individualized fetal heart rate baseline deviation, fetal heart rate fine variability disappearance time, umbilical blood flow PI value, maternal blood oxygen saturation, maternal mean arterial pressure, uterine tension, fetal movement frequency, and acceleration / deceleration frequency. At the same time, the pathological coupling weights of each indicator calculated by the TGAT model in the core algorithm layer are retrieved, and dynamic weight values ​​are assigned to the core indicators. The weight values ​​are adjusted in real time according to the degree of pathological correlation between the indicators and fetal distress, avoiding scoring bias caused by fixed weights. Next, a fetal distress quantitative scoring system (FDS score) with a score of 0-10 was constructed. Each core indicator was assigned a quantitative value of 0-10 according to its clinical abnormality. The quantitative values ​​of each indicator were then weighted and summed with their corresponding dynamic weights to calculate the comprehensive FDS score. A score ≥6 was set as the threshold for fetal distress determination. The higher the comprehensive score, the more severe the fetal distress. This completed the quantitative identification of fetal distress. Finally, by combining the FDS comprehensive score results, the risk contribution indicators of the graded early warning module, and the degree of abnormal deviation of each core indicator, the fetal distress type is accurately classified. It is mainly divided into acute fetal distress-maternal hypoxia type, acute fetal distress-excessive uterine contractions type, chronic fetal distress-placental insufficiency type, and chronic fetal distress-umbilical cord abnormality type. At the same time, the core indicators are sorted from high to low abnormal deviation and from large to small pathological coupling weight, and the abnormal indicator ranking results are output to achieve accurate classification of fetal distress and clear definition of abnormal causes.

[0028] The human-computer interaction module provides visualized monitoring data, multimodal early warning prompts, and clinical intervention suggestions to both medical staff and patients. This module includes: Doctor-side module: Displays real-time curves of multimodal maternal-fetal monitoring data, individualized fetal heart rate baseline dynamic curves, and early warning identification results, supporting data playback, feature annotation, and manual review; Specifically, a split-screen, partitioned display logic is adopted to show real-time dynamic curves of multimodal maternal-fetal monitoring data, such as fetal heart rate, uterine contraction pressure, maternal blood oxygen saturation, and umbilical blood flow PI / RI values. Simultaneously, individualized dynamic threshold curves for the upper and lower limits of the fetal heart rate baseline are overlaid on the fetal heart rate curve area. All curves support free scaling and timeline dragging switching at multiple time scales (1 minute, 5 minutes, 10 minutes), enabling refined viewing of monitoring data. The core display area of ​​the interface is refreshed in real-time with fetal distress risk probability values ​​and continuous time-series trend graphs, FDS quantitative score values ​​and score change curves, and simultaneously displays color-coded warning levels (no risk green, low risk yellow, medium risk orange, high risk red), core risk contribution indicators, and matching clinical intervention suggestions. All warning assessment results are accompanied by annotations of the data calculation basis. Medical staff can select specific... Data playback can be triggered by either a time interval or by inputting key time points. During playback, the system simultaneously displays the warning identification results and characteristic data for the corresponding time period, and supports pause, fast forward, and slow motion operations. Medical staff can manually click and mark abnormal characteristic points on the monitoring curve interface and add custom diagnostic and treatment notes. The system supports manual correction and review of the warning level, FDS score, distress type, and other identification results output by the system. The corrected results will generate review records and be stored in association with the original data. A standardized HL7 / CDA data interaction interface has been developed to achieve docking with the hospital's HIS / LIS system. All obstetric diagnosis and treatment data, including monitoring data, warning identification results, and manual review records, can be synchronized to the hospital's electronic medical record system to complete the electronic storage, classification, and retrieval of data, adapting to the standardized workflow of obstetric clinical practice.

[0029] Patient-side module: Receives data output from the doctor-side module, displays simplified maternal and fetal monitoring data and basic risk warnings, supports emergency alarms for abnormal situations, and provides a one-click function to contact medical staff; Specifically, the system performs clinical information filtering and simplification on the multimodal monitoring data output by doctors, displaying only simple, easy-to-read, non-professional monitoring data such as real-time fetal heart rate values ​​and simplified trend curves, time-based statistics on fetal movement frequency, and baseline maternal blood oxygen saturation. All professional diagnostic indicators and quantitative scoring data are removed to avoid causing patient anxiety due to complex information. Basic risk warnings are displayed in a minimalist format combining text and cartoon icons. In no-risk / low-risk states, only a green safety icon and the text "Monitoring Normal" are displayed, without outputting any risk-related information to ensure the patient's psychological stability. When the system determines a medium-risk / high-risk warning level, the patient's end will immediately... Triggering a local audible and visual alarm triggers a large emergency notification pop-up on the interface, displaying the prominent text "Please contact medical staff immediately." The alarm signal is then pushed in real-time via an edge-cloud collaborative architecture to the corresponding doctor's backend, the obstetrics nurse station's work terminal, and the responsible medical staff's handheld monitor, achieving multi-terminal synchronous reach of the alarm signal. Simultaneously, when the patient clicks the "One-click Call Medical Staff" button on the interface, a voice call request is initiated to the obstetrics nursing station, and a text message containing the patient's bed number, current monitoring data, and alarm type is pushed to the responsible medical staff's work terminal, enabling rapid two-way linkage between the patient and medical staff and ensuring the efficiency of emergency treatment response during labor.

[0030] The edge-cloud collaboration module enables real-time data processing, iterative model optimization, and individualized adaptation. This module includes: Local inference module: Equipped with a lightweight data preprocessing algorithm and an early warning and identification inference model, it enables real-time local fetal distress early warning and identification calculation; Specifically, lightweight processing chips are embedded in local hardware carriers such as bedside monitoring terminals and handheld monitors to optimize data preprocessing algorithms such as denoising, linear interpolation, and Z-score standardization. Redundant calculation steps are eliminated, and algorithm iteration logic is simplified to adapt to local embedded computing environments and meet the high-speed data processing requirements under low computing power. At the same time, the GBRT baseline calibration model, TCN feature extraction model, and TGAT risk warning model, which have undergone knowledge distillation and channel pruning in the cloud, are integrated and packaged and implanted into the local inference module to build an integrated local early warning and identification inference model library. All model inference calculations are completed locally without relying on real-time computing power support from the cloud. Finally, a local circular caching mechanism is configured to temporarily encrypt and cache the preprocessed feature data, model input and output data, and early warning identification results during the inference process. The data is uploaded to the hospital data management module at a preset interval of 5 minutes. It also supports the function of resuming transmission after network failure. After the network is restored, the cached data is automatically re-uploaded to ensure the continuous and stable operation of local fetal distress monitoring and early warning identification work when the network is abnormal.

[0031] Hospital Data Management Module: Receives data output from the local inference module, completes data storage and management, and performs local parameter fine-tuning on the local inference model; Specifically, a hospital-level dedicated maternal-fetal monitoring database was built. Data indexes were constructed based on patient unique IDs, bed numbers, and data collection times. A distributed storage architecture was used to classify and store all data uploaded by the local inference module, including raw data, preprocessed features, and early warning identification results. Data encryption and access control mechanisms were configured to ensure the security and compliance of clinical data. Uploaded data was cleaned, deduplicated, and outlier removed. Combined with manual annotation by medical staff, effective clinical data was selected to build a hospital-specific model training and fine-tuning dataset. Based on this dataset, a lightweight local inference model was fine-tuned using mini-batch gradient descent. Model weights were optimized to suit the clinical case characteristics and treatment habits of the hospital's obstetrics department. After validity verification, the fine-tuned model was pushed to all local inference modules within the hospital to update the model, achieving individualized adaptation of the model within the hospital.

[0032] Data optimization module: Receives data output from the hospital data management module, continuously optimizes the core model, performs lightweight processing on the optimized model, and distributes the lightweight model data to the local inference module and the hospital data management module respectively. Specifically, the system receives clinical data uploaded from the data management modules of various hospitals, removes patient privacy information, and constructs a large-scale cross-hospital maternal-fetal monitoring dataset. Based on this dataset, it retrains and optimizes models such as the GBRT baseline calibration model, TCN feature extraction model, and TGAT core early warning model. By increasing the training sample size and sample diversity, it improves the generalization ability and early warning identification accuracy of the core model. The optimized model is then transformed into a lightweight inference model adapted to the local terminal's computing power and storage requirements using a combination of knowledge distillation, channel pruning, and floating-point quantization techniques. This minimizes the computational load and storage space of the model while ensuring controllable loss of model accuracy. The lightweight optimized model is then distributed to the hospital data management modules of each hospital through an encrypted cloud-dedicated push channel, along with model version update logs and deployment instructions. Each hospital then distributes the optimized model to its local inference module, achieving end-to-end model iterative updates. A cloud-based model version management system is also built to support model backtracking and rollback operations, ensuring system stability.

[0033] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention are also within the protection scope of the present invention.

Claims

1. A fetal distress early warning and identification system based on fetal heart rate monitoring, characterized in that, include: Data acquisition module, data preprocessing module, early warning calculation module, early warning identification module, human-computer interaction module; The data acquisition module synchronously collects multi-dimensional physiological data of the fetus and the mother. The data preprocessing module performs noise reduction, temporal alignment, and standardization on the raw maternal and fetal data, and extracts clinically interpretable basic maternal and fetal characteristics. The early warning calculation module completes fetal heart rate baseline calibration, maternal and fetal indicator feature extraction, and calculation of fetal distress risk probability based on the basic maternal and fetal characteristics. The early warning identification module combines the fetal distress risk probability and risk development trend to achieve graded early warning. It completes the scientific identification and classification of fetal distress through multi-indicator quantitative scoring. The human-computer interaction module provides visualized monitoring data, multimodal early warning prompts, and clinical intervention suggestions on the medical staff end and the patient end, respectively.

2. The fetal distress early warning and identification system based on fetal heart rate monitoring according to claim 1, characterized in that: The data acquisition module includes: a fetal data acquisition module and a maternal data acquisition module; Fetal data acquisition module: Acquires raw fetal heart rate signals, quantifies fetal movement signals, and collects fetal umbilical artery blood flow velocity and blood flow-related index data; Maternal data acquisition module: Collects maternal blood oxygen saturation and peripheral pulse rate, uterine tension and uterine contraction pressure, maternal systolic blood pressure, diastolic blood pressure and mean arterial pressure data.

3. The fetal distress early warning and identification system based on fetal heart rate monitoring according to claim 1, characterized in that: The data preprocessing module includes: a fetal heart rate signal denoising module, a time alignment module, and a standardization and feature mapping module; Fetal heart rate signal denoising module: Receives data output from the data acquisition module, performs noise removal processing on the raw fetal heart rate signal, and retains the subtle features of fetal heart rate variability; Timing alignment module: Receives data output from the fetal heart rate signal denoising module, performs interpolation and resampling on maternal and fetal data collected at different sampling rates, and completes missing data; Standardization and Feature Mapping Module: Receives data output from the temporal alignment module, performs standardization transformation on it, and extracts clinically interpretable multidimensional maternal-fetal basic features.

4. The fetal distress early warning and identification system based on fetal heart rate monitoring according to claim 1, characterized in that: The early warning calculation module includes: a baseline calibration module, a feature extraction module, and a risk early warning calculation module; Baseline calibration module: Receives data output from the data preprocessing module, generates individualized fetal heart rate baseline thresholds based on individual fetal characteristics and maternal physiological state, and dynamically adjusts them; Feature extraction module: Receives data output from the baseline calibration module, performs time-series analysis on it, extracts the time-series coupling features between multiple maternal and fetal indicators, and outputs them to the risk warning calculation module; Risk warning calculation module: Based on the features output by the feature extraction module, a dynamic graph network is constructed to calculate and output the probability of fetal distress risk, and at the same time output the risk contribution index.

5. The fetal distress early warning and identification system based on fetal heart rate monitoring according to claim 1, characterized in that: The early warning identification module includes: a graded early warning module and an identification scoring module; Tiered early warning module: Receives the fetal distress risk probability and risk contribution index output by the early warning calculation module, performs tiered early warning, and outputs the early warning level, risk contribution index and corresponding clinical intervention suggestions; The identification and scoring module receives data from the graded early warning module, constructs a quantitative scoring system based on core maternal and fetal indicators to complete the fetal distress score, and combines the scoring results with risk contribution indicators to classify the types of fetal distress and rank abnormal indicators.

6. The fetal distress early warning and identification system based on fetal heart rate monitoring according to claim 1, characterized in that: The human-computer interaction module includes: a doctor-side module and a patient-side module; Doctor-side module: Displays real-time curves of multimodal maternal-fetal monitoring data, individualized fetal heart rate baseline dynamic curves, and early warning identification results, supporting data playback, feature annotation, and manual review; Patient-side module: Receives data output from the doctor-side module, displays simplified maternal and fetal monitoring data and basic risk warnings, supports emergency alarms for abnormal situations, and provides a one-click function to contact medical staff.

7. The fetal distress early warning and identification system based on fetal heart rate monitoring according to claim 1 further includes an edge-cloud collaborative module: a local inference module, a hospital data management module, and a data optimization module; Local inference module: Equipped with a lightweight data preprocessing algorithm and an early warning and identification inference model, it enables real-time local fetal distress early warning and identification calculation; Hospital Data Management Module: Receives data output from the local inference module, completes data storage and management, and performs local parameter fine-tuning on the local inference model; Data optimization module: Receives data output from the hospital data management module, continuously optimizes the core model, performs lightweight processing on the optimized model, and distributes the lightweight model data to the local inference module and the hospital data management module respectively.

Images