Fetal distress early warning system based on fetal heart rate variability and deep learning model

Abstract

The invention discloses a fetal distress early warning system based on fetal heart rate variability and a deep learning model, and belongs to the technical field of fetal health monitoring. According to the system, a fetal heart rate acquisition module is used for acquiring fetal heart beat activity signals and generating time sequence data, and a signal preprocessing module is used for generating fetal heart rate variability signals through operations such as abnormal heart beat interval elimination and time axis resampling in sequence; the segmentation and feature extraction module is combined with multi-scale wavelet transform and a hidden Markov model to realize dynamic adaptive segmentation of signals and extract a high-dimensional time sequence feature matrix, and the deep learning recognition module completes feature learning and pattern recognition through a hybrid model and a multi-head self-attention mechanism. The risk judgment and early warning module generates a dynamic risk level and a multi-mode early warning prompt based on an integrated learning strategy; the system solves the problems that traditional fetal distress monitoring is high in subjectivity, low in early recognition rate, lagged in early warning and the like, the accuracy and the real-time performance of distress recognition are remarkably improved, and the risk of bad pregnancy outcome is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of fetal health monitoring technology, and more specifically, to a fetal distress early warning system based on fetal heart rate variability and a deep learning model. Background Technology

[0002] Fetal distress is a significant problem threatening the life and health of the fetus during the mid-to-late stages of pregnancy and delivery. Its core issue is the abnormal physiological state of the fetus related to hypoxia in the uterus. Timely monitoring and intervention are crucial to reducing adverse outcomes. Currently, the mainstream monitoring method in clinical practice is fetal heart rate monitoring, which collects fetal heart rate signals and maternal uterine contraction information to generate a continuous monitoring curve. Medical staff then interpret the curve characteristics based on their clinical experience. Some auxiliary solutions employ traditional statistical methods or basic algorithms to perform simple signal processing and extract a small number of features to assist in assessing the fetal condition. Simultaneously, basic filtering techniques are used to initially reduce the impact of environmental noise on the signal.

[0003] Existing monitoring technologies still face numerous practical problems in application. The manual interpretation of monitoring results is significantly influenced by the experience and differing judgment standards of medical staff, leading to inconsistent results and difficulty in accurately capturing early, subtle abnormal signals. Signal processing procedures are relatively simple, lacking effective interference separation mechanisms; physiological interferences such as maternal ECG and fetal movement are easily confused with fetal heart rate signals, resulting in inconsistent signal quality. Feature extraction dimensions are relatively singular, focusing primarily on surface-visible signal features and failing to fully uncover deeper information reflecting the fetal physiological state. Recognition models are insufficient in capturing the temporal changes in signals, and the decision-making process lacks sufficient transparency, making it difficult for medical staff to intuitively understand the judgment criteria. Early warnings are mostly static results, lacking dynamic analysis of risk trends and failing to quickly respond to real-time monitoring needs in clinical settings. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a fetal distress early warning system based on fetal heart rate variability and a deep learning model. The system addresses the problems of strong subjectivity, poor signal quality, incomplete feature extraction, insufficient recognition model performance, and delayed early warning in the prior art as described in the background section through the following solutions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a fetal distress early warning system based on fetal heart rate variability and a deep learning model, comprising: Fetal heart rate acquisition module: used to acquire fetal heartbeat signals and generate fetal heart rate time series data; Signal preprocessing module: used to preprocess the fetal heart rate time series data, including abnormal heartbeat interval removal, time axis resampling, data smoothing and continuity correction, to generate fetal heart rate variability signal; The segmentation and feature extraction module analyzes the instantaneous energy distribution of the signal through multi-scale wavelet transform and combines it with a hidden Markov model to decode the state of the fetal autonomic nervous activity pattern; by fusing energy changes and model state boundaries, it achieves dynamic adaptive segmentation of the signal; finally, it extracts high-dimensional feature vectors containing time domain, frequency domain, nonlinearity and entropy features from each segment and constructs a time-series feature matrix. Deep learning recognition module: Based on the temporal feature matrix, multi-scale feature learning is performed through a hybrid model of cascaded convolutional neural network and bidirectional long short-term memory network; the model introduces a multi-head self-attention mechanism to realize dynamic weighting of key feature channels and time steps, and integrates an interpretability module to output pattern recognition results and confidence scores related to fetal distress; Risk assessment and early warning module: This module combines pattern recognition results with confidence analysis, employs an ensemble learning strategy to perform multi-indicator fusion decision-making, and generates real-time dynamic risk levels and multimodal early warning alerts. The technical effects and advantages of this invention are as follows: 1. This invention employs an ultrasonic Doppler acquisition device paired with a broadband piezoelectric ceramic transducer, adapting to the fetal position characteristics at different gestational weeks. It combines standardized procedures such as pre-acquisition position optimization, coupling agent noise reduction, and probe fixation. Furthermore, it achieves interference prediction through surface electrocardiogram acquisition and fetal movement detection. With a triple screening mechanism of "interference correlation determination + statistical characteristics + physiological constraints," as well as a composite smoothing strategy of moving average and Gaussian filtering and dynamic continuity correction, it not only effectively separates interference components such as maternal electrocardiogram, fetal movement, and environmental noise, but also accurately completes missing or abnormal data segments. This ensures that the generated fetal heart rate variability signal has both high purity and continuity, and can truly reflect the state of fetal autonomic nervous activity, providing reliable data support for subsequent accurate analysis. 2. This invention performs multi-band decomposition of fetal heart rate variability signals through multi-scale wavelet transform, combines it with a hidden Markov model to dynamically decode fetal autonomic nervous activity patterns, integrates energy changes and state boundaries to achieve dynamic adaptive segmentation of signals, comprehensively extracts time-domain, frequency-domain, nonlinear and entropy features to construct a high-dimensional temporal feature matrix, and then uses a CNN-Bi-LSTM hybrid model to work together. It not only uses CNN to capture local key features, but also uses Bi-LSTM to mine temporal dependencies. Combined with a multi-head self-attention mechanism to dynamically strengthen the weight of core features, it effectively avoids the limitations of a single feature dimension or model architecture, greatly improves the accuracy of fetal distress pattern recognition, and especially enhances the ability to detect early weak abnormal signals. 3. This invention, by integrating gradient-weighted activation mapping and attention visualization interpretability modules, generates feature heatmaps and attention distribution maps, intuitively presenting the core basis and key time periods for model decision-making. This solves the "black box" problem of traditional deep learning models, helping medical staff quickly understand the judgment logic and verify the reliability of the results. At the same time, by integrating the decision results of multiple sub-models through ensemble learning strategies and combining time window analysis to dynamically capture risk change trends, it generates graded risk levels. Then, through multimodal early warning forms such as visual color coding, auditory graded alarms, and text natural language reports, coupled with real-time push functions on mobile devices, it achieves comprehensive monitoring, dynamic prediction, and rapid response to fetal distress risks, buying valuable time for timely clinical intervention and effectively protecting maternal and infant health. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the overall structure of the present invention; Figure 2 This is a schematic diagram of the process for obtaining fetal heart rate variability signals according to the present invention; Figure 3 This is a schematic diagram illustrating the construction process of the temporal feature matrix of the present invention; Figure 4 This is a schematic diagram illustrating the process of obtaining risk levels and early warning prompts for the present invention. Detailed Implementation

[0007] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0008] As attached Figures 1 to 4 The fetal distress early warning system shown, based on fetal heart rate variability and a deep learning model, includes: Fetal heart rate acquisition module: used to acquire fetal heartbeat signals and generate fetal heart rate time series data; The specific methods for acquiring the fetal heartbeat activity signal include: Equipment Selection and Composition: An ultrasonic Doppler fetal heart rate acquisition device is used as the core acquisition terminal. The device consists of an ultrasonic probe, a signal processing unit, and an acquisition control module. The ultrasonic probe uses a wideband piezoelectric ceramic transducer with an operating frequency adapted to the fetal position characteristics at different gestational weeks, ensuring effective reception of the fetal heartbeat reflection ultrasound signal in the mid-to-late stages of pregnancy. The signal processing unit integrates a dedicated preamplifier and bandpass filter for preliminary processing of the acquired reflection signal. The acquisition control module is responsible for coordinating core acquisition-related operations such as probe driving and signal acquisition timing control.

[0009] Pre-acquisition preparation procedure: The pregnant woman should lie supine or on her left side to avoid the uterus compressing the abdominal aorta and affecting the stability of maternal blood flow, which would interfere with signal acquisition; apply the coupling agent attached to the ultrasound probe to the contact point between the probe and the maternal abdominal skin to reduce the difference in acoustic impedance and thus reduce signal attenuation; the operator should locate the projection position of the fetal precordial region on the maternal abdomen according to the fetal ultrasound image, fix the ultrasound probe in this position and secure it with an elastic band to prevent the probe from shifting due to slight changes in the pregnant woman's position or fetal movement during the acquisition process, and maintain the stability of the acquisition position.

[0010] Fetal cardiac activity signal acquisition: The ultrasound probe emits a continuous wave ultrasound signal, which is directed towards the fetal heart region. When the ultrasound signal encounters moving tissues such as fetal heart valves and ventricular walls, it is reflected. The reflected signal undergoes a frequency shift due to the Doppler effect, and the amount of the shift is directly related to the fetal heart rate. This reflected signal is the core acquisition signal containing fetal cardiac activity information. The signal processing unit performs preliminary processing on the reflected echo signal to synchronize the acquisition. It filters out environmental noise and clutter signals reflected from maternal tissues through a bandpass filter, and then amplifies the effective cardiac activity signal to an amplitude range suitable for subsequent acquisition and processing through a programmable gain amplifier.

[0011] Interference suppression acquisition: The maternal electrocardiogram signal is acquired through an additional surface electrocardiogram acquisition electrode and used as an interference reference signal for interference separation of subsequent acquisition signals; the fetal movement signal is acquired synchronously through a fetal movement detection sensor, and the time period of fetal movement is marked in real time, providing an identification basis at the acquisition level for distinguishing between fetal movement interference and normal heartbeat signals.

[0012] Fetal heart rate characteristic peak acquisition and timestamp recording: Real-time peak detection is performed on the acquired signal after preliminary filtering and amplification. The characteristic peak of the fetal heart rate signal is identified by an adaptive threshold method. The acquisition timestamp corresponding to each characteristic peak is recorded synchronously to form the original acquisition data pair of "characteristic peak - acquisition timestamp".

[0013] Fetal heart rate time series data generation: Based on the collected raw data pairs of "characteristic peak - collection timestamp", the time interval between two adjacent characteristic peaks is calculated sequentially according to the time stamp order. This time interval is converted into instantaneous fetal heart rate values. Then, all instantaneous fetal heart rate values ​​are associated and arranged according to the corresponding timestamp order to form continuous fetal heart rate time series data.

[0014] Signal preprocessing module: used to preprocess the fetal heart rate time series data, including abnormal heartbeat interval removal, time axis resampling, data smoothing and continuity correction, to generate fetal heart rate variability signal; The preprocessing of the fetal heart rate time series data specifically includes: Abnormal cardiac interval removal: A triple screening mechanism of "interference correlation determination + statistical characteristics + physiological constraints" is adopted. First, interference correlation determination is performed, using maternal electrocardiogram signals acquired during the acquisition phase as a reference. Independent component analysis (ICA) is used to separate maternal electrocardiogram interference components from the signal, and cardiac intervals contaminated by interference are removed. At the same time, the fetal movement time period marked by the fetal movement detection sensor is combined. If the abnormal cardiac interval occurs within this time period, it is directly judged as a false abnormality caused by fetal movement interference. Then, the statistical characteristics of all cardiac intervals in the fetal heart rate time series are calculated, including the mean and standard deviation. The statistical screening threshold is set as the mean plus or minus three times the standard deviation. The standard deviation is used to mark heart intervals outside the range as statistical outliers. Simultaneously, a physiological constraint range is set based on fetal physiological characteristics, referencing the range of normal fetal heart intervals in clinical practice. Heart intervals outside this range are marked as physiological outliers. For heart intervals marked as both statistical and physiological outliers, the mean of one normal heart interval before and after the interval is used for replacement. For heart intervals that only meet a single outlier label, linear interpolation is prioritized to complete the range based on the changing trends of two normal heart intervals before and after the interval. If there are fewer than two normal heart intervals on one side, the mean of the existing normal heart intervals is combined with the complete data from the other side for interpolation.

[0015] Time axis resampling: A fixed-frequency resampling strategy is used to standardize the fetal heart rate time series; the resampling frequency is set to a fixed value to adapt to subsequent multi-scale wavelet transform and feature extraction, and the corresponding fetal heart rate value is calculated at uniformly distributed time nodes by linear interpolation, so that the resampled fetal heart rate time series has a uniform time interval; the change trend of the original data is strictly followed during the resampling process.

[0016] Data smoothing: A composite smoothing algorithm combining moving average and Gaussian filtering is employed. First, a fixed-length moving average window is set, with the window size determined based on the time scale of fetal heart rate changes. The mean of fetal heart rate values ​​within the window is calculated as the smoothing value for the current moment, initially filtering out random noise. Subsequently, Gaussian filtering is applied to perform a second smoothing on the data processed by the moving average. By setting an appropriate standard deviation for the Gaussian kernel function, the filtering process preserves the main trend of fetal heart rate changes while further attenuating high-frequency interference components. The parameters for both smoothing operations are optimized and determined based on the statistical characteristics of clinical fetal heart rate data.

[0017] Continuity Correction: A dynamic detection and adaptive completion mechanism is established. A sliding window is used to traverse the fetal heart rate time series, calculating the rate of change of adjacent data points within the window. When the rate of change exceeds a preset threshold, it is determined to be an abnormal data continuity. For short continuous abnormal segments, cubic spline interpolation is used to complete the data based on the trend of three normal data points before and after the abnormal segment, ensuring that the completed data maintains the same variation pattern as the original data. For longer continuous abnormal segments, combining the short-term stability characteristics of fetal heart rate variability, a combination of piecewise linear interpolation and trend prediction is used for completion. First, the overall trend of the completion interval is determined by the variation trend of adjacent normal data segments, and then the specific values ​​within the interval are filled using linear interpolation. After correction, the data series is traversed again to verify continuity, ensuring the data is continuous and smooth.

[0018] After a full-process preprocessing process including abnormal cardiac interval removal, time axis resampling, data smoothing, and continuity correction, a fetal heart rate variability signal that can accurately reflect the state of fetal autonomic nervous activity is finally generated.

[0019] The segmentation and feature extraction module: Based on the fetal heart rate variability signal, the instantaneous energy distribution of the signal is analyzed through multi-scale wavelet transform, and the state decoding of the fetal autonomic nervous activity pattern is performed by combining the hidden Markov model; by fusing energy changes and model state boundaries, dynamic adaptive segmentation of the signal is achieved; finally, high-dimensional feature vectors containing time domain, frequency domain, nonlinearity and entropy features are extracted from each segment, and a time-series feature matrix is ​​constructed. It should be specifically noted that, based on the aforementioned fetal heart rate variability signal, the instantaneous energy distribution of the signal is analyzed through multi-scale wavelet transform, and the state decoding of the fetal autonomic nervous activity pattern is performed in conjunction with a hidden Markov model. The specific process is as follows: Multi-scale wavelet transform analysis of instantaneous energy distribution: An adaptive wavelet basis was used to decompose the fetal heart rate variability signal at multiple scales. The wavelet basis type selected was the db4 wavelet, suitable for biomedical signal analysis. The decomposition scale was optimized and determined based on the frequency range of the fetal heart rate variability signal. Through multi-scale decomposition, the original signal was divided into detail components and approximate components at different frequency bands. The detail components correspond to the high-frequency fluctuations of the signal, and the approximate components correspond to the low-frequency trend components. The instantaneous energy of the detail components and approximate components at each scale was calculated separately. The sliding window method was used to calculate the signal energy value within each window. The window length was set based on the time-varying characteristics of fetal heart rate variability. The instantaneous energy was calculated using the following formula: In the formula, E is the instantaneous energy of the signal within the window; N is the length of the sliding window; w k Let be the k-th wavelet coefficient within the window; k is the index of the wavelet coefficient within the window.

[0020] Hidden Markov Models for State Decoding of Fetal Autonomic Nervous Activity Patterns: A Hidden Markov Model (HMM) adapted to recognize fetal autonomic nervous activity patterns is constructed. The model is defined as λ=(A,B,π), where A is the state transition probability matrix, B is the observation probability matrix, and π is the initial state probability vector. The hidden states of the model are defined as different patterns of fetal autonomic nervous activity, including three categories: sympathetic excitation state (S), parasympathetic excitation state (P), and sympathetic-parasympathetic balance state (B), i.e., the state set Ω={S,P,B}. The model training data consisted of clinically labeled fetal heart rate variability signals from normal and distressed fetuses. The model parameters were iteratively optimized using the Baum-Welch algorithm, with the parameters updated using the following formula during the iteration process: In the formula, The probability of transitioning from state i to state j after optimization; T is the signal sequence length, which refers to the total number of time points contained in the fetal heart rate variability signal sequence input to the Hidden Markov Model. Let be the probability that a state is in state j at time t and that it transitions to state j at time t+1. Let t be the probability of being in state i at time t; t is the time node index of the signal sequence; i and j are the indices of the hidden states; In the formula, This represents the probability of being in state i at the initial moment after optimization. The probability of being in state i at the first time point; The fetal heart rate variability signal after multi-scale wavelet transform processing is used as the model input. The Viterbi algorithm is used to decode the signal sequence with the optimal state path to obtain the state of the fetal autonomic nervous activity pattern at each time point, thus completing the dynamic tracking and state labeling of the fetal autonomic nervous activity pattern.

[0021] It should be further explained that dynamic adaptive segmentation of the signal is achieved by fusing energy changes with model state boundaries. The specific process is as follows: Dynamic adaptive segmentation is achieved by integrating energy change and model state boundaries: Based on the instantaneous energy distribution curve obtained by multi-scale wavelet transform and the state sequence decoded by the hidden Markov model, a dual boundary judgment criterion is established to achieve dynamic adaptive segmentation of the signal; firstly, energy abrupt change points in the instantaneous energy distribution curve are extracted, and an energy change rate threshold ΔE is set. th The rate of change of energy ΔE is calculated using the following formula: In the formula, ΔE is the rate of energy change between two adjacent windows; E t E represents the energy value of the window at time t. t+1 The energy value at time t+1; When ΔE>ΔE th At that time, the potential segmentation boundary is marked; at the same time, the state transition points in the state sequence of the Hidden Markov Model are extracted, and the time nodes when the state changes are marked as the core segmentation boundary; the potential segmentation boundary and the core segmentation boundary are fused and screened to remove false boundaries caused by noise and retain the effective boundaries that truly reflect the switching of fetal autonomic neural activity patterns and significant changes in signal energy; the signal segments are divided according to the effective boundaries, and the length of each segment is controlled within a reasonable range to match the feature extraction, ensuring that each segment contains the complete cycle of fetal autonomic neural activity patterns.

[0022] It should be further explained that high-dimensional feature vectors containing time-domain, frequency-domain, nonlinear, and entropy features are extracted from each segment, and a time-series feature matrix is ​​constructed. The specific process is as follows: Temporal characteristics: Indicators reflecting the temporal statistical characteristics and fluctuation patterns of the signal are extracted, including mean, standard deviation (SDNN), root mean square of adjacent heart rate interval differences (RMSSD), and percentage of adjacent heart rate interval differences greater than 50ms (PNN50). The mean is the arithmetic mean of all fetal heart rate variability signal values ​​within the segment, the standard deviation is the dispersion of all fetal heart rate variability signal values ​​within the segment relative to the mean, the root mean square of the sum of the squares of the differences between adjacent fetal heart rate variability signal values ​​within the segment is the square root of the mean, and the percentage of adjacent heart rate interval differences greater than 50ms is the proportion of the number of adjacent fetal heart rate variability signal values ​​with an absolute value greater than 50ms within the segment to the total number of adjacent differences.

[0023] Frequency domain features: Fourier transform (FFT) is performed on each segmented signal to convert the time-domain signal into a frequency-domain signal and obtain the power spectral density. Energy characteristic indexes corresponding to frequency bands are extracted, including the energy in the low-frequency band (0.04 - 0.15 Hz), the energy in the high-frequency band (0.15 - 0.4 Hz), and the low-to-high frequency energy ratio (LF / HF). The energy of each frequency band is obtained by integrating the power spectral density within the corresponding frequency band.

[0024] Nonlinear features: Nonlinear dynamic analysis methods are used to extract features, including the correlation dimension D2, which is obtained by fitting the correlation integral C(r) of the signal sequence. The correlation integral formula is: where C(r) is the correlation integral of the signal sequence; L is the length of the segmented signal, which refers to the total number of data points contained in the fetal heart rate variability signal within a single segment; r is the given radius, which is the distance threshold set for the correlation integral calculation; H(·) is the step function, where H(·) = 1 when the input is greater than 0, otherwise 0; x i is the value of the i-th fetal heart rate variability signal within the segment; x j is the value of the j-th fetal heart rate variability signal within the segment; i and j are the indexes of the signal data points within the segment, and i ≠ j; the correlation dimension D2 is the slope after the linear fitting of lnC(r) and lnr.

[0025] Entropy features: Sample entropy (SampEn) is extracted, which is obtained through steps such as constructing a vector sequence and calculating the proportion of vector similarities. The core calculation formula is: where m < L. In the formula, SampEn is the sample entropy, which is used to quantify the uncertainty and disorder of the signal sequence; L is the length of the segmented signal; m is the embedding dimension; are the proportions of similarities with the i-th vector as the reference when the embedding dimensions are m and m + 1 respectively; r is the similarity tolerance.

[0026] Construction of the time series feature matrix: The high-dimensional feature vectors corresponding to each segment are arranged in order according to the time sequence of the signal segments to construct the time series feature matrix , where X is the time series feature matrix; represents a real matrix with K rows and D columns; K is the number of signal segments; D is the total dimension of the extracted features; the rows of the matrix correspond to each signal segment, and the columns correspond to various types of extracted feature indexes. The matrix element X k,d is the d-th eigenvalue of the k-th segment, where X k,d is the element in the k-th row and d-th column of the time series feature matrix. k is the segment index, and d is the feature index; at the same time, a time dimension identifier is added to the matrix to record the time interval of the signal segment corresponding to each feature vector, ensuring the time series correlation of the feature matrix.

[0027] Deep learning recognition module: Based on the temporal feature matrix, multi-scale feature learning is performed through a hybrid model of cascaded convolutional neural network and bidirectional long short-term memory network; the model introduces a multi-head self-attention mechanism to realize dynamic weighting of key feature channels and time steps, and integrates an interpretability module to output pattern recognition results and confidence scores related to fetal distress; The hybrid model is constructed as follows: Hybrid Model Construction: A hybrid architecture consisting of a cascaded convolutional neural network (CNN) and a bidirectional long short-term memory network (Bi-LSTM) is employed to perform multi-scale feature learning on the temporal feature matrix; temporal feature matrix As the direct input to the model, K is the number of signal segments (i.e., the total number of segments obtained after the fetal heart rate variability signal is dynamically and adaptively segmented), and D is the total feature dimension (i.e., the total number of time-domain, frequency-domain, nonlinear, and entropy features extracted from each segment). The CNN part uses multiple one-dimensional convolutional layers to extract local spatiotemporal patterns. The kernel size is set according to the short-term correlation of the fetal heart rate variability signal, usually 3 or 5. The convolutional layers are followed by max pooling layers to reduce dimensionality and enhance robustness. The local features extracted by the CNN are then input into the Bi-LSTM layer. The Bi-LSTM consists of long short-term memory units in both forward and backward directions, which can simultaneously capture the temporal dependencies of the sequence. Finally, the hidden states in the two directions are concatenated to form a complete contextual representation. The entire hybrid model uses an end-to-end training method to optimize parameters, and the cross-entropy loss function is selected. In the formula, Loss is the total loss value of the model, used to guide parameter optimization; i is the index of the fetal state category (the value corresponds to three categories: normal, mild distress, and severe distress); p i The true label distribution of the sample (for a single sample, typically the corresponding class is 1, and other classes are 0); q i This represents the probability distribution predicted by the model for the current sample; the loss function trains the model by minimizing the difference between the predicted distribution and the true distribution.

[0028] The multi-head attention mechanism is introduced in the following way: Multi-head self-attention mechanism introduction: To enable the model to dynamically focus on the feature channels and time steps most relevant to fetal distress, a multi-head self-attention mechanism is integrated after the Bi-LSTM layer; the attention score is calculated through query-key matching and normalized by Softmax. In the formula, α t,d e represents the attention weight for the d-th feature channel at the t-th time step (corresponding to the t-th signal segment) (its value ranges from 0 to 1, and the sum of all channel weights is 1); t,dd' is the unnormalized attention score, determined by the matching degree between the query vector and the key vector; d' is the traversal index of the feature channels, used for normalizing the denominator calculation; the multi-head design divides the attention into 8 independent heads, each focusing on a different feature subspace, and finally concatenates the outputs of all heads; through this mechanism, the model can automatically increase the weight of key indicators (such as high-frequency energy changes or abnormal fluctuations in nonlinear entropy features) while reducing the influence of noise or irrelevant features.

[0029] The integration method of the interpretability module is as follows: Interpretability Module Integration: To enhance the transparency and clinical credibility of model decision-making, an interpretability module combining Gradient Weighted Class Activation Mapping (Grad-CAM) and attention visualization is integrated. By calculating the gradient sensitivity of the predicted category to the intermediate feature map, a global average gradient weight is generated, and a heatmap is produced accordingly to highlight the feature regions and time periods that contribute the most to the final judgment. The heatmap is directly overlaid on the time-series feature matrix, making it easier for clinicians to intuitively understand the model's focus. At the same time, the distribution of multi-head attention weights is visualized to form an attention distribution map. The model finally outputs the fetal distress identification result (normal / mild distress / severe distress), confidence score (range 0-1, the closer to 1, the more certain the judgment), as well as the heatmap and attention distribution map.

[0030] Risk assessment and early warning module: It combines pattern recognition results with confidence analysis, uses an integrated learning strategy to make multi-indicator fusion decisions, and generates real-time dynamic risk levels and multimodal early warning prompts.

[0031] An ensemble learning strategy is used to perform multi-indicator fusion decision-making, generating real-time dynamic risk levels and multimodal early warning prompts. The specific process is as follows: Multi-indicator fusion decision-making: An ensemble learning strategy is employed to fuse the recognition results (normal / mild distress / severe distress) and confidence scores output by the deep learning recognition module, constructing a multi-model voting mechanism. First, the recognition results are converted into numerical scores, e.g., 0 for normal, 1 for mild distress, and 2 for severe distress. Then, multiple sub-models are integrated, including a rule-based threshold classifier, a random forest classifier, and a gradient boosting machine (GBM). Each sub-model independently processes input features, including the recognition category score, confidence score, and auxiliary indicators (such as baseline variability of fetal heart rate). The fusion process uses a weighted voting method, and the final fusion score is calculated using the following formula. In the formula, S is the final fusion risk score, used for subsequent risk level classification; u is the sub-model index; U is the total number of sub-models (typically 3-5, optimized based on clinical validation); w uThe weights of the u-th sub-model (ranging from 0 to 1, with the sum of all weights being 1, determined through cross-validation based on historical data performance; for example, random forests have higher weights to emphasize non-linear relationships); s u The prediction score of the u-th sub-model for the current input (range 0-2, reflecting the severity of distress).

[0032] Risk level generation: Real-time risk levels are dynamically generated based on the fusion score, employing a threshold-based grading strategy; a low-risk threshold Th is set. l =0.5, high-risk threshold Th h =1.8 (the threshold is statistically optimized based on clinical guidelines and retrospective data, and can be dynamically adjusted); when S <Th l The time was determined to be low risk (normal state). l ≤S <Th h At the time of medium risk (mild distress), S≥Th h This indicates extremely high risk (severe distress); it also incorporates time window analysis to calculate the most recent T. step Time step (T) step Based on the monitoring cycle setting, such as T step The dynamic risk trend is calculated using the average score and rate of change of (=5) according to the following formula: In the formula, ΔF is the rate of change of risk at the current time step, used to detect sudden distress; F t F is the fusion score at the current time step. t−1 The fusion score is the score of the previous time step; F0=0.5 is the baseline score. When ΔF>0.3, a trend alarm is triggered, which improves the overall risk assessment even if the current level is not high.

[0033] Multimodal early warning prompts: Multimodal outputs are generated based on risk level and trends, including visual, auditory, and textual prompts. Visual prompts display a risk dashboard through a graphical interface, using color coding (green for low risk, yellow for medium risk, and red for high risk), and overlaying a heatmap to highlight abnormal segments. Auditory prompts employ tiered alarm sounds: no sound for low risk, intermittent beeping for medium risk, and continuous alarms for high risk, with volume and frequency adjusted according to confidence level. Textual prompts generate natural language reports, such as "Mild distress detected, confidence level 0.85, immediate re-examination recommended." Multimodal fusion is managed through priority queues to ensure all modalities are activated synchronously during high-risk situations. It also integrates a mobile push interface to send warning information to healthcare personnel's devices in real time, supporting remote intervention.

[0034] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fetal distress early warning system based on fetal heart rate variability and a deep learning model, characterized in that, include: Fetal heart rate acquisition module: used to acquire fetal heartbeat signals and generate fetal heart rate time series data; Signal preprocessing module: used to preprocess the fetal heart rate time series data, including abnormal heartbeat interval removal, time axis resampling, data smoothing and continuity correction, to generate fetal heart rate variability signal; The segmentation and feature extraction module analyzes the instantaneous energy distribution of the signal through multi-scale wavelet transform and combines it with a hidden Markov model to decode the state of the fetal autonomic nervous activity pattern; by fusing energy changes and model state boundaries, it achieves dynamic adaptive segmentation of the signal; finally, it extracts high-dimensional feature vectors containing time domain, frequency domain, nonlinearity and entropy features from each segment and constructs a time-series feature matrix. Deep learning recognition module: Based on the temporal feature matrix, multi-scale feature learning is performed through a hybrid model of cascaded convolutional neural network and bidirectional long short-term memory network; The model introduces a multi-head self-attention mechanism to achieve dynamic weighting of key feature channels and time steps, and integrates an interpretable module to output pattern recognition results and confidence scores related to fetal distress. Risk assessment and early warning module: It combines pattern recognition results with confidence analysis, uses an integrated learning strategy to make multi-indicator fusion decisions, and generates real-time dynamic risk levels and multimodal early warning prompts.

2. The fetal distress early warning system based on fetal heart rate variability and a deep learning model according to claim 1, characterized in that: The fetal heart rate time series data was obtained in the following way: The ultrasound probe emits continuous wave ultrasound signals and receives signals reflected by the fetal heart's moving tissues. The signal processing unit filters and amplifies the reflected signals. The maternal electrocardiogram (ECG) signals are acquired through additional surface ECG acquisition electrodes, and fetal movement signals are acquired through a fetal movement detection sensor. The processed signals are subjected to real-time peak detection using an adaptive threshold method, characteristic peaks are identified, and corresponding timestamps are recorded to form raw data pairs. The time interval between adjacent characteristic peaks is calculated according to the timestamp order and converted into instantaneous fetal heart rate values. All instantaneous fetal heart rate values ​​are arranged in timetamp order to generate fetal heart rate time series data.

3. The fetal distress early warning system based on fetal heart rate variability and a deep learning model according to claim 1, characterized in that: The preprocessing method is as follows: Abnormal cardiac interval elimination: A triple screening mechanism combining interference correlation judgment, statistical characteristics and physiological constraints is adopted. Maternal electrocardiogram signals are used to separate interference components, and fetal movement time periods are used to mark false abnormalities. Statistical thresholds are set based on the mean and standard deviation of cardiac intervals, and physiological constraint ranges are set based on the normal cardiac intervals of clinical fetuses. Abnormal cardiac intervals are filled by mean replacement or linear interpolation. Time axis resampling: A fixed frequency strategy is adopted, and the fetal heart rate value is calculated at uniform time nodes by linear interpolation to form a sequence with uniform time intervals; Data smoothing: A composite algorithm combining moving average and Gaussian filtering is used. First, random noise is filtered through a moving average window, and then high-frequency interference is attenuated through Gaussian filtering. Continuity correction: A dynamic detection and adaptive completion mechanism is established, which detects the rate of change of data through a sliding window and completes abnormal segments.

4. The fetal distress early warning system based on fetal heart rate variability and a deep learning model according to claim 1, characterized in that: The state decoding process is as follows: A hidden Markov model was constructed, with hidden states including sympathetic excitation, parasympathetic excitation, and sympathetic-parasympathetic balance. The model training data consisted of clinically labeled fetal heart rate variability signals from normal and distressed fetuses. The model parameters were iteratively optimized using the Baum-Welch algorithm. The fetal heart rate variability signals, processed by multi-scale wavelet transform, were input into the model, and the Viterbi algorithm was used to decode the signal sequence using the optimal state path to obtain the fetal autonomic nervous activity pattern state at each time point.

5. The fetal distress early warning system based on fetal heart rate variability and a deep learning model according to claim 1, characterized in that: The implementation process of the dynamic adaptive segmentation is as follows: Based on the instantaneous energy distribution curve obtained by multi-scale wavelet transform and the state sequence decoded by the Hidden Markov Model, a dual boundary judgment criterion is established; energy abrupt change points in the instantaneous energy distribution curve are extracted, and an energy change rate threshold is set. When the energy change rate of adjacent windows exceeds the threshold, it is marked as a potential segmentation boundary; state transition points in the state sequence of the Hidden Markov Model are extracted and marked as core segmentation boundaries; the potential segmentation boundaries and core segmentation boundaries are fused and screened to remove false boundaries and retain valid boundaries.

6. The fetal distress early warning system based on fetal heart rate variability and a deep learning model according to claim 1, characterized in that: The temporal feature matrix is ​​constructed as follows: Extract time-domain features, frequency-domain features, nonlinear features, and entropy features from each signal segment; The time-domain features include mean, standard deviation, root mean square of the difference between adjacent heartbeats, and percentage of the difference between adjacent heartbeats greater than fifty milliseconds; The frequency domain characteristics include low-frequency energy, high-frequency energy, and the low-to-high frequency energy ratio; The nonlinear feature is the correlation dimension; The entropy feature is the sample entropy; Arrange the high-dimensional feature vectors corresponding to each segment in order of the time sequence of the signal segments to construct a time-series feature matrix.

7. The fetal distress early warning system based on fetal heart rate variability and a deep learning model according to claim 1, characterized in that: The process of multi-scale feature learning is as follows: The hybrid model consists of a convolutional neural network and a bidirectional long short-term memory network cascaded together, with the temporal feature matrix serving as the model input. The local features extracted by the convolutional neural network are input into the bidirectional long short-term memory network, which is composed of long short-term memory units in both forward and backward directions. This network captures the temporal dependencies between sequences and concatenates the hidden states in both directions to form a complete contextual representation. The model uses an end-to-end training method to optimize parameters, and the loss function is cross-entropy loss.

8. The fetal distress early warning system based on fetal heart rate variability and a deep learning model according to claim 1, characterized in that: The pattern recognition results are obtained in the following way: A multi-head self-attention mechanism is integrated after the bidirectional long short-term memory network. This mechanism consists of eight independent heads, each focusing on a different feature subspace. Attention scores are calculated through query-key matching and normalized to achieve dynamic weighting of key feature channels and time steps. An interpretable module combining gradient-weighted class activation mapping and attention visualization is integrated to generate heatmaps and attention distribution maps. After processing by the above mechanism, the model outputs pattern recognition results related to fetal distress.

9. The fetal distress early warning system based on fetal heart rate variability and a deep learning model according to claim 1, characterized in that, include: The multi-indicator fusion decision-making process is as follows: The system integrates rule-based threshold classifier, random forest classifier, and gradient boosting machine sub-models. Each sub-model independently processes the identification category score, confidence value, and auxiliary indicators. A weighted voting method is used to calculate the final fusion score. Based on the fusion score, a threshold is set to divide the risk into three levels: low, medium, and high. By incorporating time window analysis, the average score and rate of change of time steps are calculated to analyze risk trends; Finally, multimodal early warning prompts are generated based on risk level and trend.

Images