Fetal heart rate estimation and event detection method and system based on multi-modal quality factor driving

Abstract

The application relates to the technical field of wearable fetal heart detection, in particular to a fetal heart rate estimation and event detection method and system based on a multi-modal quality factor driving, which comprises the following steps: collecting multi-modal data in multiple channels and aligning; selecting a maternal R peak sequence as a time anchor point in electrocardiogram data to construct a maternal heart sound template and a maternal electrocardiogram template and calculate heart sound residuals and electrocardiogram residuals; performing blind source separation to obtain a plurality of different classified sources and performing optimization sorting; performing time-frequency transformation on a target source and decomposing a plurality of bases, and applying rhythm constraints to construct a plurality of candidate FHR paths and perform cutting; detecting a fetal heart R peak sequence, comparing the fetal heart R peak sequence with the maternal R peak sequence, estimating a heart rate curve and realizing maternal-fetal separation; constructing a quality factor, classifying a segment set corresponding to each time window based on the quality factor, and outputting a fetal heart rate result. The application can improve fetal heart rate monitoring and analysis effect.

Description

Technical Field

[0001] This application relates to the technical field of wearable fetal heart rate detection, and in particular to a method and system for fetal heart rate estimation and event detection based on multimodal quality factor driven. Background Technology

[0002] Fetal heart rate monitoring is a diagnostic tool used during pregnancy to assess the health of the fetus. Currently, common clinical methods for fetal heart rate monitoring include:

[0003] Ultrasonic Doppler fetal heart rate monitoring (CTG):

[0004] Fetal cardiac mechanical activity and fetal heart rate are estimated by emitting a sound beam and receiving the reflected signal using a Doppler ultrasound probe. This method has a well-established monitoring procedure, but it suffers from drawbacks such as subjective interpretation, a high false-positive rate, and the inability to be used continuously at home for extended periods.

[0005] Non-invasive fetal electrocardiography (NI-fECG):

[0006] Electrodes are attached to the surface of the pregnant woman's abdomen to collect mixed maternal and fetal electrocardiograms, which are then separated into fetal components using a specific algorithm. This method does not use ultrasound energy, which is beneficial for long-term monitoring. However, in the mid- and early-pregnancy stages, the fetal electrocardiogram amplitude is extremely low and easily masked by noise.

[0007] Fetal heart rate (fPCG) monitoring:

[0008] Fetal heart rate mechanoacoustic signals are acquired using piezoelectric sensors or microphones. The advantages are that it is completely passive, radiation-free, and can provide information related to mechanical activity (S1 / S2, systole / diastole, etc.). However, a single fPCG is sensitive to changes in fetal position and environmental noise.

[0009] With the development of sensor technology and wearable electronics, some products have begun to combine surface electrophysiology with fetal heart sounds, using simple signal processing algorithms (such as bandpass filtering and threshold detection) to estimate fetal heart rate. However, significant shortcomings remain in the following aspects:

[0010] Multi-channel, multi-modal information is often not utilized by the system, resulting in "piling up channels but not integrating them".

[0011] Signal quality assessment (QF) only focuses on a single dimension such as SNR or contact impedance, without forming a closed loop that runs through acquisition, processing, and output.

[0012] The algorithm relies heavily on empirical thresholds and rules, and is not robust enough to difficult scenarios such as mid-pregnancy (early months), high BMI, and placenta previa.

[0013] Without an explainable event detection and reporting mechanism, doctors find it difficult to trace why data from a particular period is "credible / unreliable".

[0014] To address the aforementioned shortcomings, existing technologies and some products have attempted to employ the following methods: 1. Matrix component suppression based on template subtraction; 2. Source decomposition of multi-channel PCG / ECG based on blind source separation (BSS) using second-order statistics; 3. Joint modeling of the spectrum and time series using NMF (non-negative matrix factorization) or HMM (hidden Markov model).

[0015] These methods are effective to some extent, but they have certain problems in engineering implementation and expansion in lower months:

[0016] Signal quality model not displayed:

[0017] Most algorithms assume that all input segments can be used for FHR estimation, lacking explicit evaluation of body movement, contact, and environmental disturbances, resulting in results that "appear accurate but are actually unreliable" on poorly performing segments.

[0018] Inadequate system for handling maternal-fetal confusion:

[0019] When the fetal heart rate is close to the maternal heart rate, simply relying on frequency band or amplitude separation often fails, and the MHR is easily reported as FHR, especially within a short resting window.

[0020] Prior relationships between multimodalities are not fully utilized:

[0021] Electroacoustics essentially describes the electrical and mechanical activities of the same cardiac cycle, with a clear phase coupling relationship between the two. Existing algorithms typically only perform simple time alignment without building a strong prior of "phase consistency" for quality assessment and path pruning.

[0022] Insufficient edge-cloud collaboration and version auditing:

[0023] In real-world regulatory environments, algorithms need to support online updates, canary releases, offline recalculation, and version rollback. Existing solutions often treat algorithms as black boxes, which is detrimental to compliance management.

[0024] Furthermore, existing technologies generally neglect the collaborative mechanism between the physical structure of multimodal sensors and algorithms. For example, the spatial distribution of array sensors has a significant impact on the propagation paths of near-field fetal heart rate sources and far-field noise sources, but existing algorithms do not utilize such measurable spatial attenuation patterns for signal separation. At the same time, IMU and skin temperature data are often treated as independent quantitative indicators rather than used to construct an engineering closed loop of "contact state-signal availability," making it difficult for algorithms to operate stably in real dynamic wearing scenarios.

[0025] In summary, existing methods have failed to solve engineering challenges such as signal coupling, complex noise, and large fluctuations in availability in wearable multimodal fetal heart monitoring. There is an urgent need to propose a complete technical solution that combines hardware structure characteristics, physical propagation models, and multimodal quality factors. Summary of the Invention

[0026] To improve the fetal heart rate monitoring and analysis results, this application provides a method and system for fetal heart rate estimation and event detection based on multimodal quality factor driven technology.

[0027] Firstly, this application provides a fetal heart rate estimation and event detection method driven by multimodal quality factors, employing the following technical solution:

[0028] A method for fetal heart rate estimation and event detection based on multimodal quality factor driving includes the following steps:

[0029] Multimodal data from multiple channels are acquired and aligned. The multimodal data includes heart sound data from the heart sound channel, electrocardiogram data from the electrocardiogram channel, IMU data, and skin temperature.

[0030] In the electrocardiogram data, the maternal R-peak sequence is selected as the time anchor point to extract a set of segments within a preset time window. Based on the set of segments, maternal heart sound template and maternal electrocardiogram template are constructed respectively. Template cancellation is performed to calculate the heart sound residual and electrocardiogram residual.

[0031] By combining the coherence and geometric distribution between channels, blind source separation is performed on the heart sound residuals of the multi-channel to obtain several sources of different classifications, and target sources are selected for optimal ranking.

[0032] The target source is subjected to time-frequency transformation and decomposed into several bases including fetal heart rate, maternal heart rate and noise, and rhythm constraints are applied to construct multiple candidate FHR paths and perform pruning.

[0033] The fetal heart rate R-peak sequence is detected in the electrocardiogram residual, compared with the maternal R-peak sequence, and the heart rate curve is estimated to achieve maternal-fetal separation. After maternal-fetal separation, the maternal heart rate curve and the fetal heart rate curve are obtained.

[0034] Quality features under each time window are obtained to construct a quality factor. The set of segments corresponding to each time window is classified based on the quality factor. Event detection is performed based on the classification results, the maternal heart rate curve and the fetal heart rate curve to output the fetal heart rate result.

[0035] In some embodiments, the maternal R-peak sequence is selected from the electrocardiogram data as a time anchor to extract a set of segments within a preset time window. Based on the set of segments, a maternal heart sound template and a maternal electrocardiogram template are constructed, including the following steps:

[0036] Select one or more of the ECG channels with the highest spatial signal-to-noise ratio, detect the position of the maternal R peak, and combine it with the heartbeat index to obtain the maternal R peak sequence;

[0037] Obvious anomalies were removed by RR interval statistics and HRV analysis, and the parent R peak sequence was interpolated and corrected.

[0038] Using each of the parent R-peak sequences as the time anchor point, time windows of the same fixed duration are extracted from the electrocardiogram channel and the heart sound channel to form electrocardiogram segments and heart sound segments, respectively.

[0039] The ECG segments and heart sound segments are normalized and aligned, and the maternal ECG template and maternal heart sound template are constructed by weighted average or principal component analysis.

[0040] In some embodiments, template cancellation is performed to calculate the heart sound residuals and ECG residuals, including the following steps:

[0041] Based on the heartbeat index, the amplitude ratio of the heart sound data in each of the segment sets to the maternal heart sound template is calculated to calculate the heart sound channel gain; the amplitude ratio of the electrocardiogram data in each of the segment sets to the maternal electrocardiogram template is calculated to calculate the electrocardiogram channel gain.

[0042] In each of the aforementioned ECG channels, the maternal ECG template is mapped according to the heartbeat index and the ECG channel gain to calculate the ECG residual; in each of the aforementioned heart sound channels, the maternal ECG template is mapped according to the heartbeat index and the heart sound channel gain to calculate the heart sound residual.

[0043] In some embodiments, by combining inter-channel coherence and geometric distribution, blind source separation is performed on the multi-channel heart sound residuals to obtain several sources of different classifications, including the following steps:

[0044] Several of the aforementioned heart sound residuals are stacked by channel to form a multi-channel matrix;

[0045] In the multi-channel matrix, the covariance matrix and the delay covariance matrix between channels are calculated respectively, and a number of independent sources are decomposed using a second-order blind source separation algorithm.

[0046] Extract the IMU data to construct a body motion reference signal, calculate the correlation coefficient between each source and the body motion reference signal, and define the source with the correlation coefficient greater than a preset value as a noise source and perform downweighting or masking.

[0047] Spectral analysis is performed on several of the sources to obtain energy frequency band characteristics, rhythmic characteristics, and phase relationship characteristics to classify the sources, wherein the classification includes fetal heart sources, maternal heart residual sources, and noise sources.

[0048] In some embodiments, selecting target sources for priority ranking includes the following steps:

[0049] Select the fetal heart rate source as the target source;

[0050] The peak sharpness, temporal envelope stability, electro-acoustic phase coupling degree and spatial attenuation consistency of the target source are calculated, and preset weights are configured to calculate the index scores by weighting.

[0051] The scores of the indicators are sorted, and the target sources with the highest scores are selected as candidate sources to complete the optimal sorting.

[0052] In some embodiments, the target source is subjected to time-frequency transformation and decomposed into several bases including fetal heart rate, maternal heart rate, and noise, and rhythm constraints are applied to construct and prune multiple candidate FHR paths, including the following steps:

[0053] The target source is subjected to time-frequency transformation to obtain a power spectrum matrix, and the power spectrum matrix is ​​decomposed into several non-negative bases;

[0054] The non-negative bases are respectively designated as fetal heart base, maternal heart base and noise base, and the harmonic ranges corresponding to different bases are specified to perform base constraint.

[0055] At each time frame, several strongest peak positions are extracted from the power spectrum of the fetal heart base activation to be converted into candidate heart rate trajectories. For each candidate heart rate trajectory, local smoothness, maternal heart frequency interval, and derived heart rate difference are calculated as peak features.

[0056] Several hidden states are generated based on fetal rhythm information, and the state transition probabilities between each hidden state are set.

[0057] Calculate the observation probability of being in the corresponding spectral peak feature at each time frame, select the cause state with the highest cumulative probability as the final state based on the state transition probability and the observation probability, and perform reverse backtracking based on the final state to select the optimal FHR time series from several candidate heart rate trajectories.

[0058] In some embodiments, the fetal heart rate R-peak sequence is detected in the electrocardiogram residual, compared with the maternal R-peak sequence, and the heart rate curve is estimated to achieve maternal-fetal separation. After maternal-fetal separation, the maternal heart rate curve and the fetal heart rate curve are obtained, including the following steps:

[0059] The electrocardiogram residual is decomposed by wavelet to obtain several scales and the detail coefficients are enhanced. Based on the enhanced scales, the signal is reconstructed and the R-peak is detected to obtain the fetal heart R-peak sequence.

[0060] Remove fetal heart rate R-peak sequences that highly overlap with the maternal R-peak sequence, and verify based on the RR interval corresponding to the fetal heart rate R-peak sequence;

[0061] The mean of the RR intervals is calculated using a sliding window to estimate the derived heart rate curve;

[0062] The ECG residuals are projected onto the phase space to estimate the spatial phase heart rate curve;

[0063] A fused fetal heart rate curve is generated based on the FHR time series, the derived heart rate curve, and the spatial phase heart rate curve. Maternal-fetal analysis and consistency determination are then performed based on the fused fetal heart rate curve.

[0064] In some embodiments, quality features under each time window are obtained to construct a quality factor, and the set of segments corresponding to each time window is classified based on the quality factor, including the following steps:

[0065] Acquire several quality features and input them into a preset lightweight quality model to output the quality factor, wherein the quality factor includes a channel-level quality factor and a global-level quality factor;

[0066] The quality factor is compared with preset upper and lower quality limits;

[0067] If the quality value is greater than the upper limit, it is classified as a high-quality segment; if it is between the upper and lower limits, it is classified as a medium-quality segment; if it is less than the lower limit, it is classified as a low-quality segment.

[0068] In some embodiments, event detection is performed based on the classification results, the maternal heart rate curve, and the fetal heart rate curve to output a fetal heart rate result, including the following steps:

[0069] A graph model is constructed based on the maternal heart rate curve, the fetal heart rate curve, the quality factor, and the body motion reference signal as nodes, and the edges in the graph model represent the relationships between the nodes.

[0070] Consistency analysis is performed based on the graph model to calculate the RMC score, and low consistency scenarios are identified and marked based on the RMC score.

[0071] Select the high-quality segment and the medium-quality segment, perform event detection based on the fetal heart rate curve, and generate an interpretable report corresponding to the event based on the quality factor associated with the event and the RMC score.

[0072] Secondly, this application provides a fetal heart rate estimation and event monitoring system driven by multimodal quality factors, which adopts the following technical solution:

[0073] A fetal heart rate estimation and event detection system driven by multimodal quality factors includes a fetal heart rate monitoring sensor consisting of a central control area and several star-shaped fetal heart rate monitoring units, wherein each star-shaped fetal heart rate monitoring unit corresponds to the acquisition of multimodal data from one channel;

[0074] The central control area is equipped with a control unit, which communicates with the cloud service unit and implements the above method through end-to-cloud collaboration.

[0075] The technical solutions provided by the embodiments of this application have the following technical effects:

[0076] (1) Significantly improves the signal-to-noise ratio of fetal heart components and increases the detectability rate of pregnant women in early stages and with high BMI;

[0077] (2) Effectively reduces the probability of maternal-fetal confusion in complex scenarios such as similar maternal and fetal heart rates and large body movements;

[0078] (3) Real-time FHR output is achieved in a low-power manner in the end-side device, and the reliability of clinical reports is improved by recalculating with high-precision algorithms in the cloud.

[0079] (4) Through the consistency between quality factors and relational models, interpretable event detection and unusable fragment prompts are achieved. Attached Figure Description

[0080] Figure 1 This is a schematic diagram illustrating the steps of a fetal heart rate estimation and event detection method based on multimodal quality factor driven by this embodiment.

[0081] Figure 2 This is a schematic diagram illustrating the implementation logic of a fetal heart rate estimation and event detection method driven by multimodal quality factors, provided in an embodiment of this application.

[0082] Figure 3 This is a schematic diagram of the fetal heart rate monitoring sensor in the embodiments of this application. Detailed Implementation

[0083] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. In some cases, to avoid obscuring various aspects of this application due to unnecessary description, well-known methods, processes, systems, components, and / or circuits already described at a higher level will not be elaborated upon. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0084] It should be noted that the descriptions of these embodiments are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0085] It should be understood that the embodiments described herein may be implemented in hardware, software, firmware, middleware, microcode, or any combination thereof. For hardware implementations, the processor may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, other electronic units designed to perform the functions described herein, or combinations thereof.

[0086] In the description of this application, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0087] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples.

[0088] like Figure 1 and Figure 2 As shown in the figure, this application discloses a method for fetal heart rate estimation and event detection based on multimodal quality factor driven by the following steps:

[0089] S100 acquires and aligns multimodal data from multiple channels.

[0090] First, it should be noted that the fetal heart rate monitoring sensor in this application includes a central control area and several radial extension arms. Each extension arm is equipped with a fetal heart rate monitoring unit, and the central control area is equipped with a control unit. All end-side processing in the subsequent method is performed in the control unit.

[0091] The fetal heart rate monitoring unit is equipped with a co-position acoustic-electric integrated sensor, which includes an acoustic sensing module for collecting fetal heart sounds and a fetal heart rate electrode module for collecting electrocardiogram signals. Temperature sensors and IMU sensors are also integrated into the central control area or the fetal heart rate monitoring unit.

[0092] In this application, "multi-channel" refers to multiple fetal heart rate monitoring units in the fetal heart rate monitoring sensor, with each fetal heart rate monitoring unit collecting data corresponding to one channel of data.

[0093] Multimodal data includes heart sound data in the heart sound channel, electrocardiogram data in the electrocardiogram channel, IMU data, and skin temperature.

[0094] Under multimodal input including multi-channel fPCG (fetal heart rate), fECG (electrocardiogram), IMU, and skin temperature, stable estimation of fetal heart rate (FHR) and maternal heart rate (MHR) is performed, and the quality confidence level at each time step is given. Specifically:

[0095] First, the aforementioned multimodal data is collected using a fetal heart rate monitoring unit;

[0096] fPCG: At least one heart sound channel is acquired from each acoustic sensor module in the star array, denoted as fPCG. , where i = 1, 2, ..., N.

[0097] fECG: ECG data is acquired via a fetal heart electrode module. , i=1,2,...,N.

[0098] IMU data: Collects triaxial acceleration and triaxial angular velocity, denoted as IMU(t).

[0099] Skin temperature: The local skin temperature T (t) is collected.

[0100] After acquiring multimodal data, the PCG, ECG, IMU, and T data are synchronized and aligned using a hardware shared clock and software timestamps. Hardware synchronization means that all sensing units share the same crystal oscillator, and a hardware clock stamp is embedded during data acquisition from each sensor to ensure that the acquisition time deviation of the multi-channel data is not too large. Software calibration is used because the IMU and skin temperature sensors have different sampling rates, requiring linear interpolation to map the low sampling rate data to the same 500Hz time base.

[0101] Meanwhile, in order to preserve the effective signal and suppress targeted noise, it is also necessary to filter fPCG and fECG separately.

[0102] For the fPCG channel, a 20-200Hz bandpass filter is applied using a fourth-order Butterworth filter. The low cutoff frequency of 20Hz is used to suppress breathing noise (0.1-10Hz) and ergonomic noise (10-20Hz), while the high cutoff frequency of 200Hz is used to suppress ambient high-frequency noise. DC component removal (using a moving average method with a 100ms window strength) can be performed before filtering to avoid signal distortion caused by baseline drift.

[0103] For the fECG channel, a 6th-order Chebyshev filter is used for bandpass filtering from 0.5 to 100 Hz. The low cutoff frequency of 0.5 Hz suppresses baseline drift (such as slow potential changes caused by electrode polarization and respiration), while the high cutoff frequency of 100 Hz suppresses electromyographic noise (20-100 Hz) and high-frequency interference. Simultaneously, for power frequency interference (50 Hz / 60 Hz) and its harmonics (100 Hz, 150 Hz), an adaptive notch filter is used. By monitoring the amplitude and phase of the power frequency interference in real time, the notch filter parameters are dynamically adjusted to avoid attenuation of the fetal heart rate signal by a fixed notch filter.

[0104] All channel data were uniformly resampled to 500Hz, and the format of the resampled data was standardized as follows:

[0105] Data type: 16-bit signed integer (range -32768~32767), quantization precision ≤1μV (ECG), ≤0.01Pa (fPCG);

[0106] Data frame structure: Each frame contains a timestamp (8 bytes), N fPCG data streams (N×2 bytes), N ECG data streams (N×2 bytes), 1 IMU data stream (6×2 bytes), 1 skin temperature data stream (2 bytes), and a checksum (2 bytes). The frame length is dynamically adjusted according to N to ensure the integrity of data transmission and storage.

[0107] S200: Select the maternal R-peak sequence in the electrocardiogram data as the time anchor point to extract a set of segments within a preset time window. Based on the set of segments, construct the maternal heart sound template and the maternal electrocardiogram template respectively, and perform template cancellation to calculate the heart sound residual and electrocardiogram residual.

[0108] This step uses the maternal R-peak as an anchor for template modeling, establishing personalized, time-varying ECG and heart sound templates respectively, and then performing template cancellation in the PCG and ECG channels to obtain residual signals rich in fetal components.

[0109] It is important to emphasize that the template modeling process in this step not only provides a mathematically meaningful baseline state for subsequent algorithms, but also designs the signal propagation path based on wearable sensors. The amplitude, delay, and directionality of maternal electrocardiograms and maternal heart sounds in each monitoring sub-unit exhibit a measurable spatial pattern. By learning the template gain, phase shift, and attenuation coefficient of different channels, strong interference components dominated by the mother in the array can be effectively suppressed, providing the necessary prerequisites for the extraction of fetal heart sounds.

[0110] Specifically, the accurate maternal heartbeat timeline is obtained by first detecting the R-peak of the maternal electrocardiogram and correcting abnormal points through HRV (heart rate variability) analysis. Based on the R-peak, electrocardiogram and heart sound segments are extracted, and individualized time-varying templates are constructed through normalization, phase alignment, PCA, and other methods to avoid the problem of poor adaptability of fixed templates. The generated template is used to cancel the maternal principal components in the original signal to obtain a residual signal rich in fetal heart components, thereby filtering out strong noise (maternal interference) in the signal and making the weak fetal heart signal stand out.

[0111] S300 combines inter-channel coherence and geometric distribution to perform blind source separation on multi-channel heart sound residuals to obtain several sources of different classifications, and selects target sources for optimal ranking.

[0112] The BSS model inputting multi-channel heart sound residuals into the spatial domain, combined with the coherence and geometric distribution between channels, is decomposed into "near-field fetal heart source", "maternal / environmental far-field source" and "noise source", and different output sources are selected optimally.

[0113] First, time-domain and frequency-domain joint BSS (such as SOBI and JADE variants) is performed using multi-channel heart sound residuals to decompose independent components. Then, IMU data is introduced to assist in directly labeling components highly correlated with body motion as noise sources, reducing their weight or eliminating them. Using array spatial priors, components that conform to the propagation characteristics of fetal heart sources are identified, and then high-confidence candidate fetal heart sources are selected through some predetermined indicators.

[0114] S400 performs time-frequency transformation on the target source and decomposes it into several bases containing fetal heart rate, maternal heart rate and noise, and applies rhythm constraints to construct multiple candidate FHR paths and perform pruning.

[0115] The candidate fetal heart channel output by blind source separation is subjected to time-frequency transformation. The power spectrum is decomposed into three or more bases: fetal heart, maternal heart and noise using structured nonnegative matrix factorization (sNMF). The specified range of different bases is defined and cross-channel consistency constraints are added to ensure that the fetal heart spectrum peak is stable in multiple channels.

[0116] Candidate spectral peak trajectories are extracted from the fetal heart rate baseline, and rhythmic constraints are applied in time to construct and prune multi-hypothesis FHR paths. The globally optimal trajectory is then selected to avoid misjudgment of spectral peak detection at a single moment, thus achieving stable tracking in the time dimension.

[0117] The S500 detects the fetal heart rate R-peak sequence in the electrocardiogram residual, compares it with the maternal R-peak sequence, and estimates the heart rate curve to achieve maternal-fetal separation. After maternal-fetal separation, the maternal heart rate curve and the fetal heart rate curve are obtained.

[0118] The wavelet / filter-based R-peak detection and phase space-based period detection algorithms were applied in parallel to the ECG residual channel to obtain the heart rate estimation curve. After obtaining the high-confidence MHR (maternal heart rate) and candidate FHR (fetal heart rate) curves, the consistency with the heart rate estimation curve was evaluated to achieve maternal-fetal separation and remove maternal-fetal confusion segments.

[0119] S600 acquires quality features for each time window to construct quality factors, classifies the set of segments corresponding to each time window based on the quality factors, and performs event detection based on the classification results, maternal heart rate curve, and fetal heart rate curve to output fetal heart rate results.

[0120] A quality factor comprising multiple quality feature sub-items is constructed and summarized into a scalar QF in the 0-1 interval using a lightweight classification / regression model, which drives subsequent weighting and segment dropping decisions. In this embodiment, quality features include SNR proxy, spectral peak sharpness, rhythm smoothness, electro-acoustic phase consistency, inter-channel consistency, contact and body motion scores, etc.

[0121] By integrating maternal heart rate curves, fetal heart rate curves, and quality factor time axes, a consistency (RMC) index is constructed. This index is used to identify different fetal heart rate monitoring scenarios, such as "maternal-fetal confusion, prolonged overlap, and abnormal events," and to classify and distinguish events such as late deceleration, reduced variability, transient acceleration, and prolonged Brady / Tachy syndrome.

[0122] Furthermore, some other embodiments also include an end-to-cloud collaboration and version auditing mechanism, which runs a lightweight real-time algorithm on the worn fetal heart monitoring sensor, performs high-precision recalculation and model updates in the cloud, and maintains a standard replay set and a versioned algorithm library to enable canary releases, rollbacks, and traceable report generation.

[0123] Specifically, the end-side operation includes data preprocessing, maternal template modeling, rapid estimation of quality factor (QF), simple fetal heart rate (FHR) estimation and event coarse screening, and outputs real-time fetal heart rate values ​​and immediate quality prompts to meet the real-time requirements for home use.

[0124] The cloud platform runs a full chain of algorithms based on the raw or compressed multimodal data, including complex NMF / HMM, BSS optimization, and fine-grained event classification.

[0125] Meanwhile, the automatic recalculation in the cloud involves recalculating all data uploaded from the mobile device. The recalculation result is compared with the result on the device. If the difference is greater than the preset bpm value and exceeds a certain period of time, or if the event detection results are inconsistent, the cloud result will be used to update the report. For suspected abnormal events marked on the device, the cloud will prioritize recalculation and push a report update reminder after the recalculation is completed. Manual recalculation is also supported, allowing users to adjust algorithm parameters for custom recalculation.

[0126] Support for audit and traceability mechanisms:

[0127] Data traceability: Each monitoring report includes "data traceability information", including the original data storage address, acquisition device number, algorithm version number, number of recalculations and time;

[0128] Algorithm traceability: Regulatory agencies can query algorithm version update records, performance evaluation reports, and canary release data to verify the compliance of the algorithm;

[0129] Dispute resolution: In the event of a clinical dispute (such as an algorithm misjudgment), the corresponding version of the algorithm can be called offline for recalculation using the version number and original data address in the report, thus restoring the original processing procedure and providing technical evidence for dispute resolution.

[0130] Log recording: All algorithm operation logs (such as parameter adjustments, recalculation operations, and version updates) are recorded in the cloud to meet the regulatory requirements for medical software.

[0131] The above scheme provides stable estimation of fetal and maternal heart rates under multi-channel, multi-modal input, and gives the quality confidence level at each moment. It effectively suppresses maternal components and environmental noise by using maternal R-peak template creation, blind source separation, and NMF / HMM joint spectral-temporal tracking, enhancing the detection capability of low SNR (signal-to-noise ratio) fetal heart signals. A multi-dimensional quality factor based on "electro-acoustic phase coupling consistency" is constructed, forming an automatic decision-making mechanism of "weighting reduction - segment loss - medical advice." It supports mid-pregnancy and even earlier gestational weeks, providing quantifiable boundary displays for "detectability rate - coverage rate - consistency error" under limited inclusion and exclusion criteria. It supports edge-cloud collaborative computing and version auditing, meeting the requirements of medical software lifecycle and risk management.

[0132] Compared with existing fetal heart rate estimation techniques based on single modality or empirical thresholds, this invention achieves the following quantifiable technical effects by establishing a continuous technical chain of "acquisition-template modeling-source separation-spectral-temporal joint tracking-quality factor-driven-event detection" on a wearable star array structure:

[0133] (4) Significantly improves the signal-to-noise ratio of fetal heart components and increases the detectability rate in early-term and high-BMI pregnant women;

[0134] (5) Effectively reduces the probability of maternal-fetal confusion in complex scenarios such as similar maternal and fetal heart rates and large body movements;

[0135] (6) Real-time FHR output is achieved in a low-power manner in the end-side device, and the reliability of clinical reports is improved by recalculating with high-precision algorithms in the cloud;

[0136] (7) Through the consistency between quality factors and relational models, interpretable event detection and unusable fragment prompts are achieved.

[0137] In other embodiments, the maternal R-peak sequence is selected from the electrocardiogram data as a time anchor point to extract a set of segments within a preset time window. Based on the set of segments, a maternal heart sound template and a maternal electrocardiogram template are constructed, including the following steps:

[0138] S210: Select one or more ECG channels with the highest spatial signal-to-noise ratio, detect the position of the maternal R peak, and obtain the maternal R peak sequence by combining the heartbeat index.

[0139] First, the signal-to-noise ratio (SNR) of each ECG channel is calculated. The SNR is defined as the ratio between the QRS complex energy and the noise energy. Then, one or more ECG channels with the highest SNR are selected, and the R-peak position is detected using a combination of improved Pan-Tompkins, wavelet transform, and energy operators to obtain the parent R-peak sequence. , where k is the heart rate index.

[0140] The algorithms described above all employ existing methods and only require achieving the desired technical effect for this step.

[0141] S211, through RR interval statistics and HRV analysis to remove obvious abnormalities, interpolation correction was performed on the parent R peak sequence.

[0142] Calculate the RR interval between adjacent R peaks. If a certain RR interval exceeds the range of "mean of the previous 5 RR intervals ± 30%", it is determined to be an abnormal interval. Abnormal R peaks are supplemented or eliminated by linear interpolation.

[0143] The corrected maternal R-peak sequence is smoothed (e.g., by a moving average window of 3 heartbeats) to obtain the final maternal R-peak sequence. This ensures that the positional error of the R peak is less than the preset value.

[0144] S212 uses each parent R-peak sequence as a time anchor point, and extracts time windows of the same fixed duration in the ECG channel and the heart sound channel to form ECG segments and heart sound segments.

[0145] S213, normalize and align the ECG and heart sound segments, and construct the maternal ECG and maternal heart sound templates through weighted average or principal component analysis.

[0146] With each parent R peak sequence As an anchor point, a window of fixed duration is captured in the ECG channel. τ_pre, The duration window is then used as a set of cardiac segment data. In some embodiments, the window length is τ_pre=100ms (before the R peak) and τ_post=300ms (after the R peak) for ECG segments (total length 400ms, covering the QRS complex and the first half of the T wave). This window length is determined based on the maternal cardiac cycle in the mid-to-late stages of pregnancy (600-1000ms) to ensure that the core morphology of a single cardiac beat is fully contained.

[0147] Simultaneously, each segment needs to be preprocessed. First, the amplitude of each segment is normalized to eliminate the influence of amplitude differences in different heartbeat cycles. Then, QRS alignment is performed, using the R peak position as a reference, and all segments are aligned on the time axis to ensure the consistency of QRS complex positions. Finally, baseline correction is performed to remove baseline drift in the segments and ensure the accuracy of the template shape.

[0148] We used weighted average and principal component analysis (PCA) to construct the maternal ECG template T_ECG(t,k). Specifically, we selected the ECG segments corresponding to the first 20 corrected R peaks and calculated a weighted average (the weights are related to the SNR of each segment, and the larger the SNR, the greater the weight) to obtain the initial template. We then formed a sample matrix from all the heartbeat segments and performed PCA dimensionality reduction, retaining the first few principal components and reconstructing the template to obtain a low-noise template shape, thus avoiding template distortion caused by individual differences.

[0149] Meanwhile, the ECG template is updated slowly over time, continuously updating the latest weighted average template of the five heartbeat segments at a low update rate. This ensures that the template can adapt to the slow changes in the mother's ECG morphology (such as minor changes caused by changes in body position or emotional fluctuations), while avoiding template drift caused by updating too quickly.

[0150] After the maternal electrocardiogram template is generated, the maternal R-peak sequence is used. Using this as an anchor point, the same time window as that in the ECG channel is extracted from the corresponding PCG heart sound channel to ensure time synchronization between the heart sound segment and the ECG segment.

[0151] After acquiring the heart sound segments, power normalization and phase alignment are required to obtain the parent heart sound template T_PCG(t,k). Power normalization involves calculating the power spectral density for each heart sound segment and normalizing the maximum value of the power spectral density to 1, eliminating the influence of differences in heart sound amplitude across different cardiac cycles. Phase alignment is based on the peak position of the S1 sound (determined by detecting the peak amplitude of the heart sound segment, with the time delay between this peak and the R peak within the range of 50-80ms), and all segments are phase aligned to ensure that the positions of S1 and S2 are consistent.

[0152] The core components of the maternal heart sound template include S1 (atrioventricular valve closure), S2 (aortic and pulmonary valve closure), and additional noise (such as valvular murmurs). The maternal heart sound template also employs a slow update mechanism to adapt to minor changes in maternal heart sound morphology.

[0153] In other embodiments, template cancellation is performed to calculate the heart sound residuals and electrocardiogram residuals, including the following steps:

[0154] S220, calculate the amplitude ratio of the central sound data of each segment set to the maternal heart sound template based on the heartbeat index to calculate the heart sound channel gain, and calculate the amplitude ratio of the electrocardiogram data of each segment set to the maternal electrocardiogram template to calculate the electrocardiogram channel gain.

[0155] Gain G(k) represents the amplitude ratio between the original ECG signal or original PCG signal and the template in the k-th heartbeat cycle. It is solved by least squares fitting. In order to avoid abnormal gain estimation, the gain range is set between 0.5 and 2.0. If it exceeds this range, the gain value of the previous heartbeat cycle is used.

[0156] S221, In each ECG channel, the maternal ECG template is mapped according to the heartbeat index and the ECG channel gain to calculate the ECG residual; in each heart sound channel, the maternal ECG template is mapped according to the heartbeat index and the heart sound channel gain to calculate the heart sound residual.

[0157] After calculating the gain, the ECG residual and heart sound residual are obtained using the following formulas:

[0158] ;

[0159] .

[0160] Characterized as electrocardiogram residuals, Characterized by heart sound residuals, For ECG channel gain, This is for the heart sound channel gain.

[0161] The obtained ECG residuals and heart sound residuals can be further processed by adaptive noise suppression to remove residual noise that was not canceled by the template, thus preserving the fetal heart component.

[0162] The residual signal is characterized by the significant suppression of the main components of maternal electrocardiogram and maternal heart sounds, while the fetal heart signal is completely preserved due to its large difference in morphology and phase from the maternal template, laying the foundation for subsequent fetal heart source extraction. At the same time, a small portion of unmodeled noise is also preserved in the residual signal, which will be separated and suppressed in the subsequent blind source separation.

[0163] In other embodiments, by combining inter-channel coherence and geometric distribution, blind source separation is performed on the multi-channel heart sound residuals to obtain several sources of different classifications, including the following steps:

[0164] S310 stacks several heart sound residuals by channel to form a multi-channel matrix.

[0165] Stack the N heart sound residuals by channel to form a matrix. .

[0166] S311: Calculate the covariance matrix and delay covariance matrix between channels in the multi-channel matrix, and decompose several independent sources using a second-order blind source separation algorithm.

[0167] First, calculate the covariance matrix and delay covariance matrix between channels within a sliding window.

[0168] The covariance matrix represents the covariance of the signals of the i-th channel and the j-th channel at the same time, reflecting the degree of synchronous linear correlation between the two channels. For example, if the fetal heart signal appears synchronously in multiple channels, the covariance will be high, while if the noise source is randomly distributed in multiple channels, the covariance will be low. This provides a spatial correlation basis for subsequent source separation.

[0169] The time delay covariance matrix represents the covariance between the current time of the i-th channel and the signal of the j-th channel after a delay of τ steps, reflecting the temporal coherence of the two channel signals.

[0170] It captures the temporal rhythm characteristics of the signal. The fetal heart signal is a periodic signal, and it still maintains a strong correlation (high time delay covariance) after a delay of τ steps, while noise is a random signal, and its correlation drops sharply after a delay (time delay covariance is close to 0). By using the time delay covariance matrix corresponding to multiple τ steps, it is possible to further distinguish between "rhythmic fetal heart sources" and "non-rhythmic noise sources".

[0171] The covariance matrix is ​​decomposed into eigenvalues ​​and eigenvectors. The multi-channel signal is transformed into a whitened signal by whitening transformation to eliminate the correlation between channels. All time delay covariance matrices are jointly diagonalized to find the separation matrix. The independent component matrix is ​​obtained by combining the whitened signal. Each row of the matrix corresponds to an independent source S_j(t).

[0172] S312, extract IMU data to construct a body motion reference signal, calculate the correlation coefficient between each source and the body motion reference signal, and define sources with correlation coefficients greater than preset values ​​as noise sources and reduce their weight or mask them.

[0173] Furthermore, classical correlation analysis (CAA) is introduced, treating components highly correlated with the IMU signal as sources of motion noise, and pre-masking or downweighting them before subsequent source type identification.

[0174] Extract the triaxial acceleration RMS values ​​from the IMU data to construct the body motion reference signal M(t). Calculate the correlation coefficient between each independent source S_j(t) and M(t). The threshold for the correlation coefficient is generally set at 0.6. If the calculated... If the threshold is exceeded, the component is determined to be a body motion noise source, and it will be downweighted or masked in subsequent processing to avoid interference from body motion noise in fetal heart rate extraction.

[0175] S313, Perform spectral analysis on several sources to obtain energy frequency band characteristics, rhythmic characteristics, and phase relationship characteristics to classify the sources.

[0176] Preliminary spectral analysis was performed on several obtained source S_j(t) signals, based on their main energy ranges, rhythmicity, and resemblance to the parent R-peak sequence. The phase relationship between the sources is used to classify each source into candidate "fetal heart source", "maternal heart remnant source" and "noise source".

[0177] Energy frequency band characteristics: The main energy of fetal heart sounds is concentrated in 1-4Hz, the energy of residual maternal heart sounds is concentrated in 0.5-3Hz, and the energy distribution of noise sources has no obvious pattern. Candidate fetal heart sources can be preliminarily determined by calculating the energy proportion of the power spectrum in the 1-4Hz frequency band.

[0178] Rhythmic characteristics: The fetal heart signal has a stable periodicity. Calculate the autocorrelation function of the components. If the peak value of the autocorrelation function at period T (T=250-1000ms) is greater than the preset value, it indicates that the rhythmicity is good and consistent with the characteristics of the fetal heart source.

[0179] Phase relationship characteristics: parent R-peak sequence There is a fixed phase relationship with the residual source of maternal heart sounds (the S1 and R peaks are delayed by 50-80ms), and the calculated components are related to... If the proportion of phase differences within the range of 50-80ms is greater than the preset value, it is determined to be a residual source of maternal heartbeat; if the phase difference has no fixed pattern and conforms to the fetal heartbeat rhythm, it is determined to be a candidate source of fetal heartbeat.

[0180] In other embodiments, selecting target sources for priority ranking includes the following steps:

[0181] S320, select fetal heart rate source as target source.

[0182] S321 calculates the spectral peak sharpness, temporal envelope stability, electro-acoustic phase coupling, and spatial attenuation consistency of the target source, and configures preset weights to calculate the index scores.

[0183] S322: Sort the indicator scores and select the target sources with the highest indicator scores as candidate sources to complete the optimal sorting.

[0184] Components classified as candidate fetal heart sources were sorted according to indicators such as peak sharpness, temporal envelope stability, electro-acoustic phase coupling, and spatial attenuation consistency, and several were selected for subsequent spectral-temporal tracking.

[0185] Peak sharpness: The power spectrum of the fetal heart source has obvious sharp peaks in the 1-4Hz frequency band. Peak sharpness is defined as "the ratio of the peak value to the mean value within 10Hz on both sides of the peak value". The higher the ratio, the sharper the peak. It corresponds to a weight of 0.3.

[0186] Temporal envelope stability: The fetal heart rate source is subjected to Hilbert transformation to extract the envelope, and the coefficient of variation (standard deviation / mean) of the envelope is calculated. If the coefficient of variation is less than 0.2, it is characterized as envelope stability, and a weight of 0.3 is matched accordingly.

[0187] Electro-acoustic phase coupling: Calculate the phase difference between the potential ECG signal in the fetal heart source and the ECG residual channel. If the proportion of the phase difference within a fixed range (such as the R peak of fECG and the S1 delay of fPCG 20-40ms) exceeds the preset value, the coupling is high and a weight of 0.3 is matched.

[0188] Spatial attenuation consistency: The channel covariance and time delay covariance used in blind source separation are strongly correlated with the geometric layout of the star array. By utilizing the known distances and directions between each monitoring unit and the central control area in the fetal heart rate monitoring sensor, the phase difference and energy attenuation mode of the near-field fetal heart rate source in different channels can be determined. Specifically, based on the positional relationship between each monitoring unit and the central control area, the amplitude attenuation coefficient of the fetal heart rate source in each channel is calculated. If the goodness of fit of this coefficient to the "inverse square law of distance" is greater than a preset value, it conforms to the near-field source characteristics, and the corresponding weight is 0.2.

[0189] Among them, the spatial prior at the hardware level, spatial decay consistency, significantly enhances the identifiability of blind source separation (BSS) and is one of the key technical points that distinguish this application from traditional pure algorithm separation methods.

[0190] Based on the weighted score (maximum score is 1), the top 2-3 candidate fetal heart sources are selected for subsequent spectrum-time joint tracking, which ensures signal redundancy and avoids excessive sources that would increase computational complexity.

[0191] In other embodiments, the target source is subjected to time-frequency transformation and decomposed into several bases including fetal heart rate, maternal heart rate and noise, and rhythm constraints are applied to construct multiple candidate FHR paths and perform pruning.

[0192] First, it should be noted that the spectral decomposition in this application is not the decomposition of an abstract mathematical matrix, but rather a combination of the physiological frequency band priors of the fetal and maternal cardiac cycles, the differences in energy distribution between array channels, and the physical intensity characteristics of the residual signal. By limiting the base frequency range and introducing engineering conditions such as cross-channel consistency constraints, the decomposition results can correspond to specific fetal heart mechanical vibration modes, thus achieving a clear effect.

[0193] Specifically, it includes the following steps:

[0194] S410 performs a time-frequency transformation on the target source to obtain the power spectrum matrix, and decomposes the power spectrum matrix into several non-negative bases.

[0195] To achieve time-frequency representation, short-time Fourier transform or continuous wavelet transform is used to perform time-frequency transformation on the fetal heart source S_j(t) to obtain the power spectrum matrix V(f,n). Here, f is the frequency index and n is the time frame index.

[0196] The choice between short-time Fourier transform and continuous wavelet transform can be adjusted based on the actual stage of pregnancy. For example, continuous wavelet transform can be used for scenarios with relatively low signal-noise ratios, such as the first trimester.

[0197] S420 specifies several non-negative bases as fetal heart base, maternal heart base and noise base respectively, and specifies the harmonic range corresponding to different bases to perform base constraint.

[0198] The power spectral matrix V(f,n) is decomposed using structured natural mathematical function (NMF). The core of structured NMF is to give the decomposition results explicit physiological meaning through prior constraints. Specifically:

[0199] .

[0200] Where R is the number of basis points. Its value can be optimized based on clinical data, such as determining it to be R=5 (2 fetal heart basis points, 1 maternal heart basis point, and 2 noise basis points), which ensures the flexibility of decomposition and avoids overfitting. Let f be the spectral shape of the r-th basis (f∈0-250Hz). Let f be the activation intensity corresponding to the r-th basis (n∈1-N_f).

[0201] At the same time, it is necessary to ensure that the final basis is non-negative, so additional constraints are added: >0, >0.

[0202] Perform basal constraint design, and place the fetal heart base The frequency range is limited to 1-4Hz and its harmonics (2-8Hz), and the frequency range corresponding to the base is constrained by a Gaussian function; the maternal base is... Limited to the range of 0.5-3Hz and its harmonics (1-6Hz), the frequency range corresponding to the basis is also constrained by a Gaussian function; while for the noisy basis... There are no strict frequency constraints, but the sparsity of its activation intensity can be constrained by group sparse regularization (L1 regularization) to avoid the noise substrate from overfitting the effective signal.

[0203] Furthermore, by introducing group sparsity regularization and cross-channel consistency constraints into the loss function, the fetal heart base shares shape but has different activation intensities among multiple candidate source channels, thereby improving the stability of the fetal heart spectrum peaks.

[0204] Specifically, the loss function is as follows:

[0205] .

[0206] Among them, the first item The first term represents the reconstruction error (Frobenius norm); the second term represents the reference constraint of the fetal heart base (…). (A standard fetal heart rate spectrum template based on clinical data). =0.1; the third term is the l1 regularization of activation strength (group sparsity constraint). =0.05; the fourth term is the cross-channel consistency constraint (ensuring that the fetal heart base morphology of different candidate fetal heart source channels is similar). =0.03.

[0207] S430: At each time frame, extract the positions of several strongest spectral peaks from the power spectrum of the fetal heart base activation to convert them into candidate heart rate trajectories. Calculate the local smoothness, maternal heart rate interval, and derived heart rate difference as spectral peak features for each candidate heart rate trajectory.

[0208] At each time frame n, the power spectrum corresponding to the fetal heart base activation (i.e. · ), and extract the positions of the three strongest spectral peaks. (c=1, 2, 3), where the peak extraction uses a peak detection algorithm:

[0209] First, perform first-order difference on the power spectrum to find the point where the difference changes from positive to negative (candidate peak point); second, eliminate candidate points whose peak amplitude is less than twice the average power spectrum value of the frame; finally, sort them in descending order of peak amplitude and select the top three as candidate spectral peaks.

[0210] Position of the spectral peak Convert to candidate instantaneous heart rate trajectory The conversion formula is:

[0211] (Unit: bpm).

[0212] make sure If the fetal heart rate is within the range of physiological heart rate, the candidate trajectory will be directly eliminated if it is outside the range.

[0213] For each candidate trajectory, features such as local smoothness, maternal heart rate interval, and derived heart rate difference are calculated for subsequent HMM path selection. Specifically:

[0214] Local smoothness: Calculate the average absolute value of the difference between adjacent heart rates. The smaller the local smoothness, the smoother the trajectory, which conforms to the pattern of fetal heart rate changes.

[0215] Interval with maternal heart rate: The average difference between the candidate trajectory and the maternal heart rate (MHR) is calculated. The larger the value, the higher the distinction between the candidate trajectory and the maternal heart rate, thus reducing the risk of maternal-fetal confusion.

[0216] Difference between derived heart rate and potential fECG heart rate: Extract potential fECG heart rate from ECG residuals and calculate the average difference between the potential and candidate trajectories. The smaller the final value, the better the acoustic-electrical consistency and the higher the reliability.

[0217] S440 generates several hidden states based on fetal rhythm information and sets the state transition probability between each hidden state.

[0218] The fetal heart rate range is discretized into n states (each state corresponds to a bpm interval of length m, such as state 1: 80-82 bpm, state 2: 83-85 bpm, etc.). Each state also includes the heart rate change trend (stable, slow rise, slow fall, rapid rise, rapid fall). The final number of hidden states is n*5, which ensures both coverage of the heart rate range and captures the dynamics of heart rate changes.

[0219] A transition probability matrix is ​​established based on fetal physiological rhythms. The transition probability is highest between stable states, followed by states with a slow rise / slow fall, and extremely low between states with a rapid rise / rapid fall, consistent with the physiological characteristic that fetal heart rate does not fluctuate drastically. The state transition probabilities can be updated and optimized via the cloud based on large-scale clinical data. The state transition probabilities dynamically change with the fetal physiological cycle and gestational age, and are represented as the probability of transitioning from state x in the previous moment to state y in the current moment.

[0220] S450: Calculate the observation probability of being at the corresponding spectral peak feature in each time frame. Based on the state transition probability and the observation probability, select the hidden state with the highest cumulative probability as the final state. Based on the final state, perform reverse backtracking to select the optimal FHR time series from several candidate heart rate trajectories.

[0221] The observation probability is represented as the observation of candidate spectral peak features when the nth frame is in state s. The probability can also be understood as the probability that the collected signal features are true fetal heart signals if the current heart rate is in state s. Among these, candidate spectral peak features... It includes comprehensive factors such as the power of candidate spectral peaks, the difference in maternal core frequency, and the local score under the quality factor calculated subsequently.

[0222] Specifically, a Gaussian mixture model is trained for each hidden state to fit the distribution of observed features in that state. The observation probability is specifically the feature probability in the Gaussian mixture model. The probability density value at a given location indicates that the more closely the feature matches the distribution of that state, the higher the probability of observation.

[0223] The Viterbi algorithm is used to find the globally optimal FHR time path. Specifically,

[0224] Calculate the initial probability of all hidden states in the first frame (based on the prior probability of the fetal heart rate distribution, such as the initial probability of the 120-160 bpm state is higher), calculate the maximum cumulative probability of each hidden state s for each time frame n, find the hidden state with the largest cumulative probability in the N_f frame as the final state, backtrack from the final state to obtain the optimal hidden state sequence, and map it to the FHR time series HR_{NMF-HMM}(t).

[0225] The maximum cumulative probability is represented as the complete confidence of an FHR path. Its value is equal to the product of the state transition probabilities and observation probabilities of all hidden states under the path. By recursively calculating frame by frame, the "total confidence" of each possible trajectory is recorded to avoid selecting the wrong global trajectory due to the high probability of a single frame.

[0226] To further optimize the path smoothness, median filtering was applied to the backtracked FHR sequence to remove unreasonable jumps and ensure that the FHR time series conforms to the fetal physiological rhythm.

[0227] In other embodiments, the fetal heart rate R-peak sequence is detected in the electrocardiogram residual, compared with the maternal R-peak sequence, and the heart rate curve is estimated to achieve maternal-fetal separation. After maternal-fetal separation, the maternal heart rate curve and the fetal heart rate curve are obtained, including the following steps:

[0228] S510 performs wavelet decomposition on the ECG residuals to obtain several scales and enhances the detail coefficients. Based on the enhanced scales, the signal is reconstructed and the R-peak is detected to obtain the fetal heart rate R-peak sequence.

[0229] By using multi-scale wavelet decomposition, the ECG residuals were decomposed into 4-scale db6 wavelets. The energy of the R-peak of the ECG signal was mainly concentrated in the detail coefficients of scale 2 (corresponding to frequencies of 12.5-25Hz) and scale 3 (corresponding to frequencies of 6.25-12.5Hz), while the residual parent ECG components were mainly concentrated in scale 4 (corresponding to frequencies of 3.125-6.25Hz).

[0230] Morphological filtering is performed on the detail coefficients of scales 2 and 3, and thresholding is used to enhance the R-peak signal and suppress noise. In this embodiment, a hard threshold is used.

[0231] Signal reconstruction was performed using only the enhanced scale 2 and 3 detail coefficients and scale 4 approximation coefficients to obtain a reconstructed signal that highlights the ECG R peak.

[0232] Detecting the R-peak in the reconstructed signal to obtain candidate fetal heart rate R-peak sequences. .

[0233] S520, remove fetal heart rate R-peak sequences that highly overlap with the maternal R-peak sequence, and verify based on the RR interval corresponding to the fetal heart rate R-peak sequence.

[0234] Calculation of fetal heart rate R-wave sequence With the parent R peak sequence The time difference, if the time difference is within the parent QRS group interval [ τ_pre, If the proportion within the range exceeds the preset value, it is considered that there is a high degree of overlap between the two, and it is judged as a residual R peak of the parent body and is removed.

[0235] Calculate the RR interval of each R peak in the fetal heart rate R peak sequence. If the proportion of the RR interval in the interval corresponding to the fetal physiological heart rate exceeds the preset value, then the R peak is retained.

[0236] S530 uses a sliding window to calculate the mean of the RR interval to estimate the derived heart rate curve.

[0237] For the fetal heart rate R-peak sequence after completing the two screening steps of S510, the average value of the RR interval is calculated using a sliding window to obtain the estimated derived heart rate curve HR_{ECG}(t).

[0238] S540 projects the ECG residuals onto the phase space to estimate the spatial phase heart rate curve.

[0239] Phase spatial period detection methods are insensitive to signal amplitude and can serve as a supplement to the wavelet transform and morphological methods mentioned above.

[0240] First, the one-dimensional ECG residual projection is reconstructed into a d-dimensional phase space vector. The Euclidean distance between each vector in the phase space is calculated. If there exists a set of vectors whose Euclidean distance is less than a preset distance and whose time intervals are basically consistent, it is considered to satisfy the closed-loop cycle, and the corresponding cycle is output. The corresponding cycle is converted into heart rate, and heart rate values ​​within the fetal heart rate range are selected to form the spatial phase heart rate curve HR_m(t).

[0241] The S550 generates a fused fetal heart rate curve based on the FHR time series, derived heart rate curve, and spatial phase heart rate curve, and realizes maternal-fetal analysis and consistency determination based on the fused fetal heart rate curve.

[0242] The FHR time series, derived heart rate curve, and spatial phase heart rate curve are integrated and a weighted average method is used. The weights of different curves are adjusted based on the quality scores of each curve. The quality scores are determined by indicators such as the coefficient of variation of the curve and its consistency with other curves. The higher the consistency and the smaller the coefficient of variation, the greater the corresponding weight.

[0243] At the same time, it is also necessary to compare the differences between different curves to determine whether there is confusion in recognizing the maternal and fetal heart rates. Confusion includes the following types:

[0244] Frequency confusion: Calculate the average difference between the fused fetal heart rate curve and the maternal heart rate curve. If the average difference is less than 10 bpm and the duration exceeds a certain period of time, it is judged as frequency confusion.

[0245] Spatial characteristic confusion: If the energy distribution of the fetal heart rate source in multiple channels does not conform to the near-field source law, and the fused fetal heart rate curve is close to the maternal heart rate curve, it is judged as spatial characteristic confusion.

[0246] Electro-acoustic consistency confusion: If the phase coupling between the heart sound signal and the electrocardiogram signal corresponding to the fused fetal heart rate curve is less than the preset value, and the fused fetal heart rate curve is close to the maternal heart rate curve, then it is determined to be electro-acoustic consistency confusion.

[0247] When any of the above-mentioned confusion scenarios are detected, a relational consistency analysis is performed to calculate the RMC score (the specific method will be explained later). When the RMC score is below half the value, the interval is marked as a "high-risk confusion segment", which will be marked in the generated report later.

[0248] In other embodiments, quality features for each time window are obtained to construct a quality factor, and the set of segments corresponding to each time window is classified based on the quality factor, including the following steps:

[0249] S610: Acquire several quality features and input them into a preset lightweight quality model to output quality factors, where the quality factors include channel-level quality factors and global-level quality factors.

[0250] The quality characteristics cover the entire signal acquisition, processing, and modeling chain, and the following characteristics are calculated for each time window:

[0251] Spectral domain features include:

[0252] Fetal heart rate peak SNR: Since directly calculating the signal-to-noise ratio of the fetal heart rate signal requires knowledge of the pure fetal heart rate signal, the ratio of the fetal heart rate peak amplitude to the noise amplitude is used as a proxy for SNR in the embodiments of this application.

[0253] Half-width of the main peak: Calculate the half-width of the fetal heart rate main peak. When the half-width is less than the preset value, it indicates that the spectral peak is sharp and the signal purity is high.

[0254] Harmonic energy ratio: The ratio of the total energy of the fetal heart main peak harmonics (2nd harmonic, 3rd harmonic) to the main peak energy is calculated. If the ratio is greater than the preset value, it indicates that the harmonic structure of the fetal heart signal is complete and has high reliability.

[0255] Temporal characteristics include:

[0256] Consistency of fetal heart rate morphology: Calculate the average correlation coefficient of 5 consecutive fetal heart rate cycles. If it is greater than the preset value, it indicates that the morphology is consistent.

[0257] Envelope smoothness: Extract the envelope from the heart sound signal, calculate the ratio of the standard deviation of the envelope to the mean to obtain the coefficient of variation. A low coefficient of variation indicates a smooth envelope.

[0258] Electro-acoustic phase coupling characteristics:

[0259] Stability of the relationship between the parent R peak and S1 / S2 in PCG: Calculate the time delay between the parent R peak and S1 and S2, and count the proportion of delays within the normal range (S1: 50-80ms, S2: 200-250ms). If the proportion exceeds the threshold, it is characterized as high stability.

[0260] Alignment degree between fetal heart rate and heart sound cycle: Calculate the difference between the theoretical cycle corresponding to the fetal heart rate and the actual cycle of the heart sound signal. If the difference is less than the percentage corresponding to the preset value, it indicates a high degree of alignment.

[0261] Multi-channel consistency characteristics:

[0262] FHR estimate standard deviation: Calculate the standard deviation of the FHR estimate in N channels. If the standard deviation is less than the preset value, it indicates that the consistency of the multiple channels is good.

[0263] Median deviation: Calculate the mean absolute deviation between the estimated FHR values ​​of each channel and the median. If the mean is less than the preset value, it indicates good consistency.

[0264] Contact and body movement characteristics:

[0265] IMU motion intensity: Calculate the RMS value of the IMU's three-axis acceleration and determine the magnitude of motion based on the RMS value;

[0266] Skin temperature change rate: Calculate the moving average change rate of skin temperature and determine whether the contact is stable based on the change rate;

[0267] Contact impedance estimation: The quality of contact is determined by measuring the AC impedance of the electrodes used to acquire ECG signals.

[0268] In addition, quality features can also include internal diagnostic features of the model, composed of indicators such as NMF residual energy, HMM path posterior probability, and BSS separation degree. The appropriate feature can be selected based on the actual scenario.

[0269] In this embodiment of the application, a lightweight quality model, such as logistic regression, random forest, or small neural network, is used to generate quality factors. This model is obtained through supervised learning on a large-scale dataset labeled by domain experts. The dataset includes multiple types of data under different gestational weeks, different BMIs, and different scenarios. At the same time, each sample is labeled with a quality level (1-5) by multiple senior obstetricians. Then, label mapping is performed to map different quality levels to continuous labels in the 0-1 range for subsequent model training.

[0270] The aforementioned quality characteristics are input into a lightweight quality model to output corresponding quality factors, including independent channel-level quality factors for each channel. It reflects the signal quality of a single channel and can be used for channel selection; it also includes a global-level quality factor for the overall output. It integrates all channels through a weighted average. The results reflect the overall monitoring quality.

[0271] S611, compare the quality factor with the preset upper and lower quality limits.

[0272] S612, if the quality value is greater than the upper limit, it is classified as a high-quality segment; if it is between the upper and lower limits, it is classified as a medium-quality segment; if it is less than the lower limit, it is classified as a low-quality segment.

[0273] Set two thresholds , These correspond to the upper and lower quality limits, respectively. The segments corresponding to each time window are classified by comparing the quality factor QF with the two thresholds.

[0274] When QF≥ This is defined as a high-quality fragment that can be directly used for subsequent heart sound reports and event analysis.

[0275] when >QF≥ This is defined as a medium-quality segment, used for smoothing and interpolation, but will be indicated as having a low confidence level in subsequent reports;

[0276] when >QF is defined as a low-quality segment, which is not used to output heart sound reports or is only used as a "segment unavailable" message.

[0277] Furthermore, the aforementioned thresholds can be dynamically adjusted based on specific scenarios, such as:

[0278] Mid-pregnancy (20-28 weeks): Reduced to 0.65, The value was lowered to 0.35 to accommodate weaker fetal heart rate signals.

[0279] Pregnant women with high BMI (BMI>30): Reduced to 0.65, Maintain a value of 0.4 to balance the detectability rate and the reliability of the results.

[0280] The quality factor in this application is not an abstract scoring of signal segments, but rather a model that combines multiple features directly derived from physical acquisition, such as sensor status, IMU body kinetic energy, fetal heart rate coupling, electro-acoustic phase consistency, and multi-channel spatial differences. Changes in the quality factor are highly consistent with objective physical phenomena such as actual loosening of the device, violent body movements, and array offset. This automatically drives the algorithm to take measures such as weighting, segment dropping, or unavailability alerts, thereby achieving closed-loop optimization between the acquisition and processing ends.

[0281] In other embodiments, event detection is performed based on classification results, maternal heart rate curves, and fetal heart rate curves to output fetal heart rate results, including the following steps:

[0282] S620 constructs a graph model based on maternal heart rate curve, fetal heart rate curve, quality factor, and body motion reference signal as nodes. The edges in the graph model represent the relationships between nodes.

[0283] A graph model is constructed, in which nodes include FHR (fetal heart rate curve), MHR (maternal heart rate curve), QF (quality factor), and IMU (body motion reference signal). Different edges correspond to the relationships between different nodes. The edge (FHR, MHR) represents the frequency difference between the two, the edge (FHR, QF) represents the positive correlation between FHR confidence and QF, the edge (FHR, IMU) represents the negative correlation between FHR confidence and body motion intensity, and the edge (MHR, QF) represents the positive correlation between MHR confidence and QF.

[0284] S621 performs consistency analysis based on a graph model to calculate the RMC score, and identifies and marks low consistency scenarios based on the RMC score.

[0285] The RMC score is the weighted sum of all edge weights in the graph model, reflecting the consistency of relationships between nodes. When the RMC score is lower than a preset value, the following low consistency scenarios are identified and corresponding measures are taken:

[0286] Scenario 1: If the FHR and MHR overlap for a long period and the QF is low, it is marked as "high risk of maternal-fetal confusion" and highlighted in the report. It is recommended that the doctor confirm the diagnosis with ultrasound examination.

[0287] Scenario 2: FHR / MHR exceeds the physiologically reasonable range: Mark as "abnormal heart rate pending verification". If it continues for more than 10 seconds, the app will push a reminder and suggest that the user seek medical attention in time.

[0288] Scenario 3: The FHR curve shows a sharp jump in the low QF segment: mark it as "signal interference causing the jump", remove the jump point from the FHR curve, replace it with the difference value, and indicate the cause of the interference.

[0289] S622 selects high-quality and medium-quality segments, performs event detection based on fetal heart rate curves, and generates an interpretable report corresponding to the event based on the quality factor and RMC score associated with the event.

[0290] Event detection is performed only on high-quality and medium-quality segments, based on morphological features of the FHR curve, persistent events, etc.

[0291] Fetal heart rate deceleration events:

[0292] Early deceleration, late deceleration, and variable deceleration are classified according to their onset time, duration, and morphological characteristics.

[0293] Fetal heart rate acceleration and changes in baseline variability:

[0294] Short-term acceleration and long-term reduced variability correspond to fetal responsiveness and potential distress signals.

[0295] Persistent bradycardia / tachycardia:

[0296] FHR is below or above a given threshold for an extended period, and artifacts are eliminated by combining QF and MHR.

[0297] Rhythm abnormalities:

[0298] Features include irregular RR intervals, excessive or low HRV.

[0299] Each event is associated with a corresponding QF range and RMC score to form an interpretable report that combines numerical values ​​with credibility.

[0300] Finally, several scenario examples will be used to further illustrate the technical content of this application:

[0301] Scenario 1: Baseline algorithm chain for in-hospital testing in the third trimester:

[0302] Scenario and Data:

[0303] Subjects: 50 pregnant women aged 36–40 weeks of gestation;

[0304] Monitoring duration: 30–60 minutes per person;

[0305] Concurrent measurement equipment: the system and method of this invention + standard CTG.

[0306] Algorithm parameter configuration:

[0307] The sliding window lasts 8 seconds with a 50% overlap.

[0308] NMF base number R=5, fetal heart rate base=2, maternal heart rate base=1, noise base=2;

[0309] The number of hidden states in the HMM is 40, corresponding to a discrete interval of FHR 80–200 bpm;

[0310] =0.7, =0.4.

[0311] Results shown:

[0312] Within high-quality segments, compared to CTG, FHR has a mean absolute error (MAE) of approximately 2.5 bpm and a coverage of ≥92%.

[0313] When segments with QF < 0.4 were removed, the overall error decreased by about 30%, and the proportion of doctors who subjectively rated the image quality as "consistent with the prompts" exceeded 90%.

[0314] Scenario 2: In-hospital verification and boundary demonstration during mid-pregnancy (28–32 weeks):

[0315] Scenario and Data:

[0316] Subjects: 40 pregnant women aged 28–32 weeks of gestation;

[0317] The area is characterized by weak signals and high noise levels, making it a "low-month window" that traditional devices struggle to cover.

[0318] Algorithm enhancements:

[0319] Increase the IMU gating weight and decrease the PCG channel weight when the body movement is intense;

[0320] Extend the frequency prior of the NMF fetal heart rate base to give it flexibility in a higher frequency (faster heart rate) range;

[0321] In HMM, an "undetectable" hidden state is added, which outputs "unavailable" instead of a false alarm when the observation does not support any reasonable path.

[0322] Results shown:

[0323] Within the resting window (low body movement, QF ≥ 0.7), the detectability rate compared to CTG reached approximately 80%.

[0324] The system clearly marks unusable segments and prompts "excessive body movement / weak signal", effectively reducing the risk of misuse.

[0325] Scenario 3: Low-power edge deployment and buffered cloud migration:

[0326] End-side simplified algorithm:

[0327] The complex NMF / HMM was removed, and only constrained peak tracking and simple HMM were retained;

[0328] The QF model uses a linear classifier and only employs SNR, IMU, and contact features.

[0329] Cloud-based recomputation mechanism:

[0330] Perform a full recalculation of segments marked as "near critical events" on the end side;

[0331] Doctors can trigger a cloud-based recalculation and report update for a specific time period with a single click at their workstation.

[0332] Balancing power consumption and performance:

[0333] The measured CPU utilization of the edge-side algorithm is controlled below 20%, and the overall working time can be increased from 8 hours to more than 12 hours.

[0334] The consistency rate between the determination of key clinical indicators (such as whether late deceleration exists within 10 minutes) and the results of the cloud-based complete algorithm exceeds 95%.

[0335] Scenario 4: Maternal-fetal separation in multiple births (twins):

[0336] Scene:

[0337] Ten pregnant women carrying twins underwent simultaneous testing in late pregnancy.

[0338] Fetuses A and B are located in different spatial positions on the array.

[0339] Algorithm changes:

[0340] In the BSS stage, the number of sources is increased, and the goal is to decompose the source into 2 fetal heart near-field sources + 1 maternal source + several noise sources.

[0341] In the NMF / HMM stage, separate FHR paths are established for different sources, allowing for the existence of two stable but different FHR sequences;

[0342] The two FHR sequences were assigned identities using array spatial weights and medical priors (such as fetal position regions determined by ultrasound).

[0343] Effect illustration:

[0344] For more than 70% of the monitoring time, FHR curves can be stably output for two fetuses separately;

[0345] During the remaining time periods, QF decreases and is marked as an unavailable window with "fetal position overlap / signal indivisible".

[0346] Scenario 5: Self-explanatory report for outpatient home monitoring:

[0347] User scenarios:

[0348] Pregnant women should wear the device for 30–60 minutes daily at home for one week.

[0349] The data is uploaded to the cloud via a mobile app and reviewed by doctors regularly.

[0350] Report content:

[0351] FHR curve + MHR curve + QF time axis;

[0352] Label the segments as "high confidence monitoring segment", "low confidence monitoring segment", and "not worn / fallen off segment";

[0353] Automatically summarize "total duration of high-confidence monitoring in this session", "number and time of suspected abnormal events", and "whether it is recommended to seek medical attention in advance".

[0354] Doctor's feedback assessment:

[0355] Doctors rated the overall readability, credibility, and interpretability of the report (out of 1–5) on average ≥4.3;

[0356] Subjective rating for “consistency between algorithm suggestions and clinical judgment” ≥ 4 / 5.

[0357] Scenario 6: Algorithm Version Management and Regulatory Audit Applications

[0358] Repository and Replay Set:

[0359] Multiple algorithm versions (such as v1.0, v1.1, v2.0, etc.) are maintained in the cloud, and each version is evaluated on the "gold standard replay set" before being upgraded;

[0360] The replay set contains multimodal data for different gestational weeks, different BMIs, and different placental locations.

[0361] Gray-scale release process:

[0362] The new version will be rolled out in a limited number of off-site users, while retaining the parallel output of the original version.

[0363] If it significantly outperforms the old version in key metrics (FHR consistency, QF relevance, event detection rate, etc.), it will be gradually rolled out.

[0364] Audit and traceability:

[0365] In the event of a dispute or security incident, the original data at that time can be recalculated and analyzed based on the algorithm version number recorded in the report;

[0366] Regulatory agencies can examine algorithm update logs, performance evaluation reports, and QF / event detection logs to verify that the system complies with software lifecycle and risk management requirements.

[0367] like Figure 3 As shown in the embodiments of this application, a fetal heart rate estimation and event detection system based on multimodal quality factor is also disclosed, including a fetal heart rate monitoring sensor composed of a central control area and several star-shaped fetal heart rate monitoring units, wherein each star-shaped fetal heart rate monitoring unit corresponds to the acquisition of multimodal data from one channel;

[0368] The central control area is equipped with a control unit that communicates with the cloud service unit and implements the above method through end-to-cloud collaboration.

[0369] The implementation principle is as follows:

[0370] Under multi-channel, multi-modal input, the system stably estimates fetal and maternal heart rates and provides quality confidence at each moment. It utilizes maternal R-peak template creation, blind source separation, and NMF / HMM joint spectral-temporal tracking to effectively suppress maternal components and environmental noise, enhancing the detection capability for low SNR (signal-to-noise ratio) fetal heart signals. A multi-dimensional quality factor based on "electro-acoustic phase coupling consistency" is constructed, forming an automatic decision-making mechanism of "weighting reduction - segment loss - medical advice." It supports mid-pregnancy and even earlier gestational weeks, providing quantifiable boundary displays for "detectability rate - coverage rate - consistency error" under defined inclusion and exclusion criteria. It supports edge-cloud collaborative computing and version auditing, meeting the requirements of medical software lifecycle and risk management.

[0371] It should be understood that although the steps in the flowcharts in the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise expressly stated herein, there is no strict order in which these steps are performed, and they may be performed in other orders.

[0372] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for fetal heart rate estimation and event detection based on multimodal quality factor driven by a method characterized in that, Includes the following steps: Multimodal data from multiple channels are acquired and aligned. The multimodal data includes heart sound data from the heart sound channel, electrocardiogram data from the electrocardiogram channel, IMU data, and skin temperature. In the electrocardiogram data, the maternal R-peak sequence is selected as the time anchor point to extract a set of segments within a preset time window. Based on the set of segments, maternal heart sound template and maternal electrocardiogram template are constructed respectively. Template cancellation is performed to calculate the heart sound residual and electrocardiogram residual. By combining the coherence and geometric distribution between channels, blind source separation is performed on the heart sound residuals of the multi-channel to obtain several sources of different classifications, and target sources are selected for optimal ranking. The target source is subjected to time-frequency transformation and decomposed into several bases including fetal heart rate, maternal heart rate and noise. Rhythm constraints are applied to construct multiple candidate FHR paths and pruning is performed to obtain the optimal FHR time series among the multiple candidate FHR paths. The fetal heart rate R-wave sequence is detected in the electrocardiogram residuals, compared with the maternal R-wave sequence, and the heart rate curve is estimated to achieve maternal-fetal separation. After maternal-fetal separation, maternal and fetal heart rate curves are obtained. Specifically... The electrocardiogram residual is decomposed by wavelet to obtain several scales and the detail coefficients are enhanced. Based on the enhanced scales, the signal is reconstructed and the R-peak is detected to obtain the fetal heart R-peak sequence. Remove fetal heart rate R-peak sequences that highly overlap with the maternal R-peak sequence, and verify based on the RR interval corresponding to the fetal heart rate R-peak sequence; The mean of the RR intervals is calculated using a sliding window to estimate the derived heart rate curve; The ECG residuals are projected onto the phase space to estimate the spatial phase heart rate curve; A fetal heart rate curve is generated based on the FHR time series, the derived heart rate curve, and the spatial phase heart rate curve. Maternal-fetal analysis and consistency determination are then performed based on the fetal heart rate curve. Quality features under each time window are obtained to construct a quality factor. The set of segments corresponding to each time window is classified based on the quality factor. Event detection is performed based on the classification results, the maternal heart rate curve and the fetal heart rate curve to output the fetal heart rate result.

2. The method for fetal heart rate estimation and event detection based on multimodal quality factor driving according to claim 1, characterized in that, In the electrocardiogram (ECG) data, the maternal R-peak sequence is selected as a time anchor point to extract a set of segments within a preset time window. Based on the set of segments, a maternal heart sound template and a maternal ECG template are constructed, including the following steps: Select one or more of the ECG channels with the highest spatial signal-to-noise ratio, detect the position of the maternal R peak, and combine it with the heartbeat index to obtain the maternal R peak sequence; Obvious anomalies were removed by RR interval statistics and HRV analysis, and the parent R peak sequence was interpolated and corrected. Using each of the parent R-peak sequences as the time anchor point, time windows of the same fixed duration are extracted from the electrocardiogram channel and the heart sound channel to form electrocardiogram segments and heart sound segments, respectively. The ECG segments and heart sound segments are normalized and aligned, and the maternal ECG template and maternal heart sound template are constructed by weighted average or principal component analysis.

3. The method for fetal heart rate estimation and event detection based on multimodal quality factor driving according to claim 2, characterized in that, Template cancellation is performed to calculate the heart sound residuals and ECG residuals, including the following steps: Based on the heartbeat index, the amplitude ratio of the heart sound data in each of the segment sets to the maternal heart sound template is calculated to calculate the heart sound channel gain; the amplitude ratio of the electrocardiogram data in each of the segment sets to the maternal electrocardiogram template is calculated to calculate the electrocardiogram channel gain. In each of the aforementioned ECG channels, the maternal ECG template is mapped according to the heartbeat index and the ECG channel gain to calculate the ECG residual; in each of the aforementioned heart sound channels, the maternal ECG template is mapped according to the heartbeat index and the heart sound channel gain to calculate the heart sound residual.

4. The method for fetal heart rate estimation and event detection based on multimodal quality factor driving according to claim 3, characterized in that, Combining inter-channel coherence and geometric distribution, blind source separation is performed on the multi-channel heart sound residuals to obtain several sources of different classifications, including the following steps: Several of the aforementioned heart sound residuals are stacked by channel to form a multi-channel matrix; In the multi-channel matrix, the covariance matrix and the delay covariance matrix between channels are calculated respectively, and a number of independent sources are decomposed using a second-order blind source separation algorithm. Extract the IMU data to construct a body motion reference signal, calculate the correlation coefficient between each source and the body motion reference signal, and define the source with the correlation coefficient greater than a preset value as a noise source and perform downweighting or masking. Spectral analysis is performed on several of the sources to obtain energy frequency band characteristics, rhythmic characteristics, and phase relationship characteristics to classify the sources, wherein the classification includes fetal heart sources, maternal heart residual sources, and noise sources.

5. The method for fetal heart rate estimation and event detection based on multimodal quality factor driving according to claim 4, characterized in that, The selection of target sources for optimal ranking includes the following steps: Select the fetal heart rate source as the target source; The peak sharpness, temporal envelope stability, electro-acoustic phase coupling degree and spatial attenuation consistency of the target source are calculated, and preset weights are configured to calculate the index scores by weighting. The scores of the indicators are sorted, and the target sources with the highest scores are selected as candidate sources to complete the optimal sorting.

6. The method for fetal heart rate estimation and event detection based on multimodal quality factor driving according to claim 5, characterized in that, The target source is subjected to time-frequency transformation and decomposed into several bases including fetal heart rate, maternal heart rate, and noise. Rhythm constraints are applied to construct multiple candidate FHR paths and then prune them, including the following steps: The target source is subjected to time-frequency transformation to obtain a power spectrum matrix, and the power spectrum matrix is ​​decomposed into several non-negative bases; The non-negative bases are respectively designated as fetal heart base, maternal heart base and noise base, and the harmonic ranges corresponding to different bases are specified to perform base constraint. At each time frame, several strongest peak positions are extracted from the power spectrum of the fetal heart base activation to be converted into candidate heart rate trajectories. For each candidate heart rate trajectory, local smoothness, maternal heart frequency interval, and derived heart rate difference are calculated as peak features. Several hidden states are generated based on fetal rhythm information, and the state transition probabilities between each hidden state are set. Calculate the observation probability of being in the corresponding spectral peak feature at each time frame, select the hidden state with the largest cumulative probability as the final state based on the state transition probability and the observation probability, and perform reverse backtracking based on the final state to select the optimal FHR time series from several candidate heart rate trajectories.

7. The method for fetal heart rate estimation and event detection based on multimodal quality factor driving according to claim 4, characterized in that, The process involves obtaining quality features for each time window to construct a quality factor, and classifying the set of segments corresponding to each time window based on the quality factor, including the following steps: Acquire several quality features and input them into a preset lightweight quality model to output the quality factor, wherein the quality factor includes a channel-level quality factor and a global-level quality factor; The quality factor is compared with preset upper and lower quality limits; If the quality value is greater than the upper limit, it is classified as a high-quality segment; if it is between the upper and lower limits, it is classified as a medium-quality segment; if it is less than the lower limit, it is classified as a low-quality segment.

8. The method for fetal heart rate estimation and event detection based on multimodal quality factor driving according to claim 7, characterized in that, Based on the classification results, the maternal heart rate curve, and the fetal heart rate curve, event detection is performed to output the fetal heart rate result, including the following steps: A graph model is constructed based on the maternal heart rate curve, the fetal heart rate curve, the quality factor, and the body motion reference signal as nodes, and the edges in the graph model represent the relationships between the nodes. Consistency analysis is performed based on the graph model to calculate the RMC score, and low consistency scenarios are identified and marked based on the RMC score. Select the high-quality segment and the medium-quality segment, perform event detection based on the fetal heart rate curve, and generate an interpretable report corresponding to the event based on the quality factor associated with the event and the RMC score.

9. A fetal heart rate estimation and event detection system driven by multimodal quality factors, characterized in that, It includes a fetal heart monitoring sensor consisting of a central control area and several star-shaped fetal heart monitoring units, wherein each star-shaped fetal heart monitoring unit corresponds to the acquisition of multimodal data from one channel; The central control area is equipped with a control unit that communicates with the cloud service unit and implements the method described in any one of claims 1-8 through end-to-cloud collaboration.

Images